diff --git a/build/conf/compilers/gnu_compiler.conf b/build/conf/compilers/gnu_compiler.conf
index d7b1b3a32e83..a33343866830 100644
--- a/build/conf/compilers/gnu_compiler.conf
+++ b/build/conf/compilers/gnu_compiler.conf
@@ -101,14 +101,6 @@ when ($OS_EMSCRIPTEN == "yes") {
     CFLAGS+=-D__EMSCRIPTEN__
 }
 
-when ($ARCH_WASM64 == "yes") {
-    CFLAGS+=-D_LIBCPP_HAS_MUSL_LIBC
-}
-
-when ($ARCH_WASM32 == "yes") {
-    CFLAGS+=-D_musl_
-}
-
 CFLAGS+=$_C_FLAGS $DEBUG_INFO_FLAGS $_C_FOPTIONS $C_WARNING_OPTS $GCC_PREPROCESSOR_OPTS $USER_CFLAGS $USER_CFLAGS_GLOBAL
 CXXFLAGS+=$CFLAGS $_STD_CXX $CXX_WARNING_OPTS $USER_CXXFLAGS $USER_CXXFLAGS_GLOBAL
 CONLYFLAGS+=$USER_CONLYFLAGS $USER_CONLYFLAGS_GLOBAL
diff --git a/build/conf/settings.conf b/build/conf/settings.conf
index 2094fce5c50d..f645b56620f5 100644
--- a/build/conf/settings.conf
+++ b/build/conf/settings.conf
@@ -19,6 +19,7 @@ YMAKE_USE_NEW_UIDS=yes
 PIC_NO_PIE=no
 FAIL_PY2=no
 MAIN_OUTPUT_AS_EXTRA=yes
+USE_REACHABILITY_TO_REPORT_CONF_ERRORS = yes
 
 when ($YMAKE_USE_OLD_UIDS == "yes") {
     YMAKE_USE_NEW_UIDS=no
diff --git a/build/config/tests/cpp_style/config18.clang-format b/build/config/tests/cpp_style/config18.clang-format
new file mode 100644
index 000000000000..b2d89e2e756e
--- /dev/null
+++ b/build/config/tests/cpp_style/config18.clang-format
@@ -0,0 +1,97 @@
+Language:        Cpp
+AccessModifierOffset: -4
+ConstructorInitializerIndentWidth: 4
+AlignEscapedNewlinesLeft: false
+AlignEscapedNewlines: Left
+AlignTrailingComments:
+    Kind: Always
+    OverEmptyLines: 0
+AllowAllParametersOfDeclarationOnNextLine: true
+AllowShortBlocksOnASingleLine: false
+AllowShortIfStatementsOnASingleLine: false
+AllowShortLoopsOnASingleLine: false
+AllowShortFunctionsOnASingleLine: None
+AlwaysBreakTemplateDeclarations: true
+AlwaysBreakBeforeMultilineStrings: false
+BreakBeforeBinaryOperators: false
+BreakBeforeTernaryOperators: true
+BreakConstructorInitializersBeforeComma: true
+BinPackParameters: true
+ColumnLimit:     0
+ConstructorInitializerAllOnOneLineOrOnePerLine: false
+DerivePointerAlignment: false
+ExperimentalAutoDetectBinPacking: false
+IndentCaseLabels: true
+IndentWrappedFunctionNames: false
+IndentFunctionDeclarationAfterType: false
+MaxEmptyLinesToKeep: 1
+KeepEmptyLinesAtTheStartOfBlocks: false
+NamespaceIndentation: All
+FixNamespaceComments: true
+ShortNamespaceLines: 0
+ObjCSpaceAfterProperty: false
+ObjCSpaceBeforeProtocolList: true
+PackConstructorInitializers: Never
+PenaltyBreakBeforeFirstCallParameter: 19
+PenaltyBreakComment: 300
+PenaltyBreakString: 1000
+PenaltyBreakFirstLessLess: 120
+PenaltyExcessCharacter: 1000000
+PenaltyReturnTypeOnItsOwnLine: 60
+PointerAlignment: Left
+SpacesBeforeTrailingComments: 1
+SpacesInSquareBrackets: false
+Cpp11BracedListStyle: true
+Standard:        c++20
+IndentWidth:     4
+TabWidth:        4
+UseTab:          Never
+BreakBeforeBraces: Attach
+InsertBraces: true
+
+# NB: BraceWrapping has no effect unless BreakBeforeBraces is set to Custom
+BraceWrapping:
+ AfterClass: false
+ AfterControlStatement: false
+ AfterEnum: false
+ AfterFunction: false
+ AfterNamespace: false
+ AfterObjCDeclaration: false
+ AfterStruct: false
+ AfterUnion: false
+ BeforeCatch: false
+ BeforeElse: false
+ IndentBraces: true
+
+SpacesInParentheses: false
+SpacesInAngles:  false
+SpaceInEmptyParentheses: false
+SpacesInCStyleCastParentheses: false
+SpacesInContainerLiterals: true
+SpaceBeforeAssignmentOperators: true
+ContinuationIndentWidth: 4
+CommentPragmas:  '^ IWYU pragma:'
+ForEachMacros: [
+    foreach,
+    Q_FOREACH,
+    BOOST_FOREACH,
+    # Is not a foreach, but correct formatting is achieved
+    Y_DEFER
+]
+IfMacros: [
+    with_lock  # Is not an if, but correct formatting is achieved
+]
+SpaceBeforeParens: ControlStatements
+DisableFormat:   false
+SortIncludes: false
+IndentPPDirectives: BeforeHash
+
+SpaceBeforeInheritanceColon: false
+NamespaceMacros: [Y_UNIT_TEST_SUITE]
+AttributeMacros: [Y_PRINTF_FORMAT, Y_NO_SANITIZE, Y_FORCE_INLINE, Y_NO_INLINE, Y_WARN_UNUSED_RESULT, Y_HIDDEN, Y_PUBLIC, Y_PURE_FUNCTION]
+IndentExternBlock: Indent
+TypenameMacros: [Y_THREAD, Y_STATIC_THREAD, Y_POD_THREAD, Y_POD_STATIC_THREAD]
+StatementAttributeLikeMacros: [ythrow]
+SpacesInLineCommentPrefix:
+ Minimum: 1
+ Maximum: 1
diff --git a/build/config/tests/cpp_style/ya.make b/build/config/tests/cpp_style/ya.make
index bb5fccd5ea45..afb406a4fc64 100644
--- a/build/config/tests/cpp_style/ya.make
+++ b/build/config/tests/cpp_style/ya.make
@@ -2,6 +2,7 @@ LIBRARY()
 
 RESOURCE_FILES(
     config.clang-format
+    config18.clang-format
 )
 
 END()
diff --git a/build/mapping.conf.json b/build/mapping.conf.json
index b40d2cd31557..d3b644ceeb33 100644
--- a/build/mapping.conf.json
+++ b/build/mapping.conf.json
@@ -482,6 +482,8 @@
         "7750425697": "https://devtools-registry.s3.yandex.net/7750425697",
         "7808687279": "https://devtools-registry.s3.yandex.net/7808687279",
         "7808709138": "https://devtools-registry.s3.yandex.net/7808709138",
+        "7827427680": "https://devtools-registry.s3.yandex.net/7827427680",
+        "7827481464": "https://devtools-registry.s3.yandex.net/7827481464",
         "5486731632": "https://devtools-registry.s3.yandex.net/5486731632",
         "5514350352": "https://devtools-registry.s3.yandex.net/5514350352",
         "5514360398": "https://devtools-registry.s3.yandex.net/5514360398",
@@ -1645,6 +1647,8 @@
         "7750425697": "devtools/ya/test/programs/test_tool/bin/test_tool for linux",
         "7808687279": "devtools/ya/test/programs/test_tool/bin/test_tool for linux",
         "7808709138": "devtools/ya/test/programs/test_tool/bin/test_tool for linux",
+        "7827427680": "devtools/ya/test/programs/test_tool/bin/test_tool for linux",
+        "7827481464": "devtools/ya/test/programs/test_tool/bin/test_tool for linux",
         "5486731632": "devtools/ya/test/programs/test_tool/bin3/test_tool3 for linux",
         "5514350352": "devtools/ya/test/programs/test_tool/bin3/test_tool3 for linux",
         "5514360398": "devtools/ya/test/programs/test_tool/bin3/test_tool3 for linux",
diff --git a/build/platform/test_tool/host.ya.make.inc b/build/platform/test_tool/host.ya.make.inc
index bcdce412a739..d283e213faf7 100644
--- a/build/platform/test_tool/host.ya.make.inc
+++ b/build/platform/test_tool/host.ya.make.inc
@@ -1,12 +1,12 @@
 IF (HOST_OS_DARWIN AND HOST_ARCH_X86_64)
-    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7808708280)
+    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7827478825)
 ELSEIF (HOST_OS_DARWIN AND HOST_ARCH_ARM64)
-    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7808707972)
+    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7827476769)
 ELSEIF (HOST_OS_LINUX AND HOST_ARCH_X86_64)
-    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7808709138)
+    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7827481464)
 ELSEIF (HOST_OS_LINUX AND HOST_ARCH_AARCH64)
-    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7808707633)
+    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7827475180)
 ELSEIF (HOST_OS_WINDOWS AND HOST_ARCH_X86_64)
-    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7808708809)
+    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7827480238)
 
 ENDIF()
diff --git a/build/platform/test_tool/host_os.ya.make.inc b/build/platform/test_tool/host_os.ya.make.inc
index 3813fb019155..20fd205d7d62 100644
--- a/build/platform/test_tool/host_os.ya.make.inc
+++ b/build/platform/test_tool/host_os.ya.make.inc
@@ -1,12 +1,12 @@
 IF (HOST_OS_DARWIN AND HOST_ARCH_X86_64)
-    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7808686430)
+    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7827425105)
 ELSEIF (HOST_OS_DARWIN AND HOST_ARCH_ARM64)
-    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7808685838)
+    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7827423650)
 ELSEIF (HOST_OS_LINUX AND HOST_ARCH_X86_64)
-    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7808687279)
+    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7827427680)
 ELSEIF (HOST_OS_LINUX AND HOST_ARCH_AARCH64)
-    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7808685532)
+    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7827422227)
 ELSEIF (HOST_OS_WINDOWS AND HOST_ARCH_X86_64)
-    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7808686875)
+    DECLARE_EXTERNAL_RESOURCE(TEST_TOOL_HOST sbr:7827426459)
 
 ENDIF()
diff --git a/build/scripts/make_java_srclists.py b/build/scripts/make_java_srclists.py
index 02a9da78e015..97c80fb19660 100644
--- a/build/scripts/make_java_srclists.py
+++ b/build/scripts/make_java_srclists.py
@@ -24,7 +24,6 @@ def main():
     parser = argparse.ArgumentParser()
     parser.add_argument('--moddir')
     parser.add_argument('--java')
-    parser.add_argument('--groovy')
     parser.add_argument('--kotlin')
     parser.add_argument('--coverage')
     parser.add_argument('--source-root')
@@ -32,7 +31,6 @@ def main():
 
     java = []
     kotlin = []
-    groovy = []
     coverage = []
 
     cur_resources_list_file = None
@@ -74,8 +72,6 @@ def main():
             kotlin.append(src)
             if args.coverage and args.source_root:
                 add_rel_src_to_coverage(coverage, src, args.source_root)
-        elif args.groovy and src.endswith(".groovy"):
-            groovy.append(src)
         else:
             if src == '--resources':
                 if cur_resources_list_file is not None:
@@ -118,9 +114,6 @@ def main():
     if args.kotlin:
         with open(args.kotlin, 'w') as f:
             writelines(f, kotlin)
-    if args.groovy:
-        with open(args.groovy, 'w') as f:
-            writelines(f, groovy)
     if args.coverage:
         jcov.write_coverage_sources(args.coverage, args.source_root, coverage)
 
diff --git a/build/sysincl/stl-to-libcxx.yml b/build/sysincl/stl-to-libcxx.yml
index 22a93f7ad6fe..dc24ce5aa54b 100644
--- a/build/sysincl/stl-to-libcxx.yml
+++ b/build/sysincl/stl-to-libcxx.yml
@@ -50,25 +50,34 @@
 
 # Though these are headers provided by libcxx, we do not want to allow them to be included.
 # We are using non-existent filename to generate error instead, as there is no specific syntax for this case.
-- includes:
-  - ext/hash_map:                       DO_NOT_INCLUDE_NON_STANDARD_EXT_HASH_MAP
-  - ext/hash_set:                       DO_NOT_INCLUDE_NON_STANDARD_EXT_HASH_SET
-  - experimental/any:                   DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_ANY
-  - experimental/chrono:                DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_CHRONO
-  - experimental/coroutine:             DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_COROUTINE
-  - experimental/filesystem:            DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_FILESYSTEM
-  - experimental/numeric:               DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_NUMERIC
-  - experimental/memory:                DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_MEMORY
-  - experimental/optional:              DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_OPTIONAL
-  - experimental/ratio:                 DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_RATIO
-  - experimental/source_location:       DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_SOURCE_LOCATION
-  - experimental/string_view:           DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_STRING_VIEW
-  - experimental/system_error:          DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_SYSTEM_ERROR
-  - experimental/tuple:                 DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_TUPLE
-
+# GENERATED BY YM2
+- source_filter: "^(?!(contrib/libs/cxxsupp/libcxx)).*"
+  includes:
+    - ext/__hash:                                            DO_NOT_INCLUDE_NON_STANDARD_EXT___HASH
+    - ext/hash_map:                                          DO_NOT_INCLUDE_NON_STANDARD_EXT_HASH_MAP
+    - ext/hash_set:                                          DO_NOT_INCLUDE_NON_STANDARD_EXT_HASH_SET
+    - experimental/__config:                                 DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL___CONFIG
+    - experimental/__memory:                                 DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL___MEMORY
+    - experimental/__simd/abi_tag.h:                         DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL___SIMD_ABI_TAG_H
+    - experimental/__simd/aligned_tag.h:                     DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL___SIMD_ALIGNED_TAG_H
+    - experimental/__simd/declaration.h:                     DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL___SIMD_DECLARATION_H
+    - experimental/__simd/internal_declaration.h:            DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL___SIMD_INTERNAL_DECLARATION_H
+    - experimental/__simd/reference.h:                       DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL___SIMD_REFERENCE_H
+    - experimental/__simd/scalar.h:                          DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL___SIMD_SCALAR_H
+    - experimental/__simd/simd.h:                            DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL___SIMD_SIMD_H
+    - experimental/__simd/simd_mask.h:                       DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL___SIMD_SIMD_MASK_H
+    - experimental/__simd/traits.h:                          DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL___SIMD_TRAITS_H
+    - experimental/__simd/utility.h:                         DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL___SIMD_UTILITY_H
+    - experimental/__simd/vec_ext.h:                         DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL___SIMD_VEC_EXT_H
+    - experimental/iterator:                                 DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_ITERATOR
+    - experimental/memory:                                   DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_MEMORY
+#TODO: ban this header too
+#    - experimental/propagate_const:                          DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_PROPAGATE_CONST
+    - experimental/simd:                                     DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_SIMD
+    - experimental/type_traits:                              DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_TYPE_TRAITS
+    - experimental/utility:                                  DO_NOT_INCLUDE_NON_STANDARD_EXPERIMENTAL_UTILITY
 
 # This includes all headers needed to resolve includes in c-headers from libcxx listed above
-# GENERATED BY YM2
 - includes:
   - algorithm:                                             contrib/libs/cxxsupp/libcxx/include/algorithm
   - any:                                                   contrib/libs/cxxsupp/libcxx/include/any
diff --git a/build/ymake.core.conf b/build/ymake.core.conf
index b3bd862595bd..0dce97a19222 100644
--- a/build/ymake.core.conf
+++ b/build/ymake.core.conf
@@ -112,6 +112,7 @@ _TOOL_PROTOC_GEN_GO_GRPC_V2=vendor/google.golang.org/grpc/cmd/protoc-gen-go-grpc
 _TOOL_RESCOMPILER=tools/rescompiler
 _TOOL_RESCOMPRESSOR=tools/rescompressor
 _TOOL_RORESCOMPILER=tools/rorescompiler
+TOOL_CPPYNDEXER=
 
 when ($DISABLE_SEPARATE_AUX_CPP != "yes") {
     SEPARATE_AUX_CPP=yes
@@ -257,8 +258,15 @@ otherwise {
 
 # tag:codenav
 when ($CODENAVIGATION && $NOCODENAVIGATION != "yes") {
+    when ($TOOL_CPPYNDEXER == "") {
+        _TOOL_CPPYNDEXER_REAL=$CPPYNDEXER_RESOURCE_GLOBAL/yndexer
+    }
+    otherwise {
+        _TOOL_CPPYNDEXER_REAL=$TOOL_CPPYNDEXER
+    }
+
     PY_PROGRAM_LINK_EXE=$LINK_EXE && ${hide;kv:"pyndex $TARGET"}
-    YNDEXER_ARGS=$YMAKE_PYTHON ${input:"build/scripts/yndexer.py"} $CPPYNDEXER_RESOURCE_GLOBAL/yndexer 1500 $(SOURCE_ROOT) $ARCADIA_BUILD_ROOT ${input:SRC}
+    YNDEXER_ARGS=$YMAKE_PYTHON ${input:"build/scripts/yndexer.py"} $_TOOL_CPPYNDEXER_REAL 1500 $(SOURCE_ROOT) $ARCADIA_BUILD_ROOT ${input:SRC}
     YNDEXER_OUTPUT=${noauto;output;suf=${OBJ_CROSS_SUF}${COMPILE_OUT_SUFFIX}.ydx.pb2:SRC}
 }
 otherwise {
@@ -764,6 +772,9 @@ module _BASE_UNIT: _BARE_UNIT {
     when ($OS_EMSCRIPTEN == "yes") {
         when ($NOLIBC != "yes") {
             PEERDIR+=contrib/restricted/emscripten/include
+            PEERDIR+=contrib/restricted/emscripten/system/lib/c
+            PEERDIR+=contrib/restricted/emscripten/system/lib/dlmalloc
+            PEERDIR+=contrib/restricted/emscripten/system/lib/standalonewasm
         }
     }
 
diff --git a/contrib/libs/cxxsupp/libcxx/.yandex_meta/scripts/sysincls.py b/contrib/libs/cxxsupp/libcxx/.yandex_meta/scripts/sysincls.py
index a3fb43e6ea56..d1a87cb532b9 100644
--- a/contrib/libs/cxxsupp/libcxx/.yandex_meta/scripts/sysincls.py
+++ b/contrib/libs/cxxsupp/libcxx/.yandex_meta/scripts/sysincls.py
@@ -14,6 +14,8 @@
 assert sysincl.count("# END OF GENERATION") == 1
 
 headers = []
+ext_headers = []
+expr_headers = []
 
 for header in include_dir.glob("**/*"):
     if not header.is_file():
@@ -23,29 +25,56 @@
 
     if header.parent == include_dir:
         if header.name.startswith("__") or not '.' in header.name:
-            headers.append(rel_path)
+            headers.append(str(rel_path))
     else:
         str_path = str(rel_path)
-        if str_path.startswith('experimental') or str_path.startswith('ext'):
-            continue
-        headers.append(rel_path)
+        if str_path.startswith('experimental'):
+            expr_headers.append(str_path)
+        elif str_path.startswith('ext'):
+            ext_headers.append(str_path)
+        else:
+            headers.append(str_path)
 
-headers = sorted(headers, key = lambda header: (str(header).count('/') != 0, str(header).count('__') != 0, header))
+
+headers = sorted(headers, key = lambda header: (header.count('/') != 0, header.count('__') != 0, header))
+ext_headers = sorted(ext_headers)
+expr_headers = sorted(expr_headers)
 
 manual, generated = sysincl.split(GENERATED_BEG)
 generated, manual_end = generated.split(GENERATED_END)
 
 generated = []
+banned_generated = []
+
+for header in ext_headers + expr_headers:
+    banned_generated.append(
+        "    - {name}:{spaces}DO_NOT_INCLUDE_NON_STANDARD_{banned_name}".format(
+            name=str(header),
+            spaces=' ' * max(1, 54 - len(header)),
+            banned_name=header.replace('/', '_').replace('.', '_').upper(),
+        )
+    )
+    if "experimental/propagate_const" == header:
+        banned_generated[-1] = "#TODO: ban this header too\n#" + banned_generated[-1]
 
 for header in headers:
     generated.append(
         "  - {name}:{spaces}contrib/libs/cxxsupp/libcxx/include/{name}".format(
-            name=str(header),
-            spaces=' ' * max(1, 54 - len(str(header)))
+            name=header,
+            spaces=' ' * max(1, 54 - len(header))
         )
     )
 
-generated = "- includes:\n" + '\n'.join(generated)
+generated = """- source_filter: "^(?!(contrib/libs/cxxsupp/libcxx)).*"
+  includes:
+""" \
+    + '\n'.join(banned_generated) \
+    + """
+
+# This includes all headers needed to resolve includes in c-headers from libcxx listed above
+- includes:
+""" \
+    + '\n'.join(generated)
 
 with open(sysicncl_path, 'w') as f:
     f.write(f"{manual}{GENERATED_BEG}{generated}\n{GENERATED_END}{manual_end}")
diff --git a/contrib/libs/cxxsupp/libcxxabi/.yandex_meta/__init__.py b/contrib/libs/cxxsupp/libcxxabi/.yandex_meta/__init__.py
index 1ba6085ed2d9..c77e96a5b3ea 100644
--- a/contrib/libs/cxxsupp/libcxxabi/.yandex_meta/__init__.py
+++ b/contrib/libs/cxxsupp/libcxxabi/.yandex_meta/__init__.py
@@ -1,5 +1,7 @@
 from devtools.yamaker import fileutil
-from devtools.yamaker.project import CMakeNinjaNixProject
+from devtools.yamaker import pathutil
+from devtools.yamaker.modules import Library
+from devtools.yamaker.project import NixSourceProject
 
 
 def post_install(self):
@@ -17,25 +19,32 @@ def post_install(self):
         f"VERSION({self.version})",
     )
 
-    with self.yamakes["."] as libcxxabi:
+    self.yamakes["."] = self.module(
+        Library,
+        NO_UTIL=True,
+        NO_RUNTIME=True,
+        NO_COMPILER_WARNINGS=True,
         # Files are distributed between libcxxabi and libcxx in a weird manner
         # but we can not peerdir the latter to avoid loops (see below)
         # FIXME: sort includes open moving glibcxx-shims into its own dir
-        libcxxabi.ADDINCL = [
+        SRCS=fileutil.files(self.dstdir, rel=True, test=pathutil.is_source),
+        ADDINCL=[
             f"{self.arcdir}/include",
             "contrib/libs/cxxsupp/libcxx/include",
             # libcxxabi includes libcxx's private "include/refstring.h" header from src subdirectory
             "contrib/libs/cxxsupp/libcxx/src",
-        ]
-
-        # We link libpthread.so automatically depending on the target platform
-        libcxxabi.CFLAGS.remove("-D_LIBCXXABI_LINK_PTHREAD_LIB")
-
-        libcxxabi.NO_UTIL = True
-        libcxxabi.NO_RUNTIME = True
+        ],
+        PEERDIR=[
+            "contrib/libs/libunwind",
+        ],
+        CFLAGS=[
+            "-D_LIBCPP_BUILDING_LIBRARY",
+            "-D_LIBCXXABI_BUILDING_LIBRARY",
+        ],
+    )
 
-        # As of 1.2.3, musl libc does not provide __cxa_thread_atexit_impl.
-        libcxxabi.CFLAGS.remove("-DHAVE___CXA_THREAD_ATEXIT_IMPL")
+    with self.yamakes["."] as libcxxabi:
+        # As of 1.2.3, musl libc does not provide __cxa_thread_atexit_impl
         libcxxabi.after(
             "SRCS",
             """
@@ -68,13 +77,22 @@ def post_install(self):
         libcxxabi.PEERDIR.add("library/cpp/sanitizer/include")
 
 
-llvm_libcxxabi = CMakeNinjaNixProject(
+libcxxabi = NixSourceProject(
     owners=["g:cpp-committee", "g:cpp-contrib"],
     arcdir="contrib/libs/cxxsupp/libcxxabi",
     nixattr="llvmPackages_16.libcxxabi",
     copy_sources=[
         "include/__cxxabi_config.h",
         "include/cxxabi.h",
+        "src/*.cpp",
+        "src/*.h",
+        "src/demangle/*.cpp",
+        "src/demangle/*.def",
+        "src/demangle/*.h",
+    ],
+    copy_sources_except=[
+        # fake exception implementation which just invokes std::terminate
+        "src/cxa_noexception.cpp",
     ],
     disable_includes=[
         "aix_state_tab_eh.inc",
diff --git a/contrib/libs/cxxsupp/libcxxabi/.yandex_meta/devtools.licenses.report b/contrib/libs/cxxsupp/libcxxabi/.yandex_meta/devtools.licenses.report
index c8b0bda349a9..8c88510b2369 100644
--- a/contrib/libs/cxxsupp/libcxxabi/.yandex_meta/devtools.licenses.report
+++ b/contrib/libs/cxxsupp/libcxxabi/.yandex_meta/devtools.licenses.report
@@ -83,8 +83,6 @@ BELONGS ya.make
         Links           : http://www.apache.org/licenses/, http://www.apache.org/licenses/LICENSE-2.0, https://spdx.org/licenses/Apache-2.0
     Files with this license:
         include/__cxxabi_config.h [3:4]
-        include/c++/v1/__cxxabi_config.h [3:4]
-        include/c++/v1/cxxabi.h [3:4]
         include/cxxabi.h [3:4]
         src/abort_message.cpp [3:4]
         src/abort_message.h [3:4]
@@ -122,8 +120,6 @@ BELONGS ya.make
         Links           : http://llvm.org/foundation/relicensing/LICENSE.txt, https://spdx.org/licenses/LLVM-exception
     Files with this license:
         include/__cxxabi_config.h [3:4]
-        include/c++/v1/__cxxabi_config.h [3:4]
-        include/c++/v1/cxxabi.h [3:4]
         include/cxxabi.h [3:4]
         src/abort_message.cpp [3:4]
         src/abort_message.h [3:4]
@@ -201,8 +197,6 @@ BELONGS ya.make
         Links           : http://www.apache.org/licenses/, http://www.apache.org/licenses/LICENSE-2.0, https://spdx.org/licenses/Apache-2.0
     Files with this license:
         include/__cxxabi_config.h [5:5]
-        include/c++/v1/__cxxabi_config.h [5:5]
-        include/c++/v1/cxxabi.h [5:5]
         include/cxxabi.h [5:5]
         src/abort_message.cpp [5:5]
         src/abort_message.h [5:5]
@@ -240,8 +234,6 @@ BELONGS ya.make
         Links           : http://llvm.org/foundation/relicensing/LICENSE.txt, https://spdx.org/licenses/LLVM-exception
     Files with this license:
         include/__cxxabi_config.h [5:5]
-        include/c++/v1/__cxxabi_config.h [5:5]
-        include/c++/v1/cxxabi.h [5:5]
         include/cxxabi.h [5:5]
         src/abort_message.cpp [5:5]
         src/abort_message.h [5:5]
diff --git a/contrib/libs/cxxsupp/libcxxabi/.yandex_meta/override.nix b/contrib/libs/cxxsupp/libcxxabi/.yandex_meta/override.nix
index 6466e5c5e8b9..e0d5f583e07c 100644
--- a/contrib/libs/cxxsupp/libcxxabi/.yandex_meta/override.nix
+++ b/contrib/libs/cxxsupp/libcxxabi/.yandex_meta/override.nix
@@ -8,19 +8,7 @@ pkgs: attrs: with pkgs; with attrs; rec {
     hash = "sha256-fspqSReX+VD+Nl/Cfq+tDcdPtnQPV1IRopNDfd5VtUs=";
   };
 
-  patches = [
-    # https://reviews.llvm.org/D132298, Allow building libcxxabi alone
-    (fetchpatch {
-      url = "https://github.com/llvm/llvm-project/commit/e6a0800532bb409f6d1c62f3698bdd6994a877dc.patch";
-      sha256 = "1xyjd56m4pfwq8p3xh6i8lhkk9kq15jaml7qbhxdf87z4jjkk63a";
-      stripLen = 1;
-    })
-  ];
-
-  NIX_CFLAGS_COMPILE = [
-    # See DTCC-589 for the details
-    "-fno-integrated-cc1"
-  ];
+  patches = [];
 
   sourceRoot = "source/libcxxabi";
 }
diff --git a/contrib/libs/cxxsupp/libcxxabi/include/c++/v1/__cxxabi_config.h b/contrib/libs/cxxsupp/libcxxabi/include/c++/v1/__cxxabi_config.h
deleted file mode 100644
index c97dd656e165..000000000000
--- a/contrib/libs/cxxsupp/libcxxabi/include/c++/v1/__cxxabi_config.h
+++ /dev/null
@@ -1,106 +0,0 @@
-//===----------------------------------------------------------------------===//
-//
-// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
-// See https://llvm.org/LICENSE.txt for license information.
-// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
-//
-//===----------------------------------------------------------------------===//
-
-#ifndef ____CXXABI_CONFIG_H
-#define ____CXXABI_CONFIG_H
-
-#if defined(__arm__) && !defined(__USING_SJLJ_EXCEPTIONS__) &&                 \
-    !defined(__ARM_DWARF_EH__) && !defined(__SEH__)
-#define _LIBCXXABI_ARM_EHABI
-#endif
-
-#if !defined(__has_attribute)
-#define __has_attribute(_attribute_) 0
-#endif
-
-#if defined(__clang__)
-#  define _LIBCXXABI_COMPILER_CLANG
-#  ifndef __apple_build_version__
-#    define _LIBCXXABI_CLANG_VER (__clang_major__ * 100 + __clang_minor__)
-#  endif
-#elif defined(__GNUC__)
-#  define _LIBCXXABI_COMPILER_GCC
-#elif defined(_MSC_VER)
-#  define _LIBCXXABI_COMPILER_MSVC
-#elif defined(__IBMCPP__)
-#  define _LIBCXXABI_COMPILER_IBM
-#endif
-
-#if defined(_WIN32)
- #if defined(_LIBCXXABI_DISABLE_VISIBILITY_ANNOTATIONS)
-  #define _LIBCXXABI_HIDDEN
-  #define _LIBCXXABI_DATA_VIS
-  #define _LIBCXXABI_FUNC_VIS
-  #define _LIBCXXABI_TYPE_VIS
- #elif defined(_LIBCXXABI_BUILDING_LIBRARY)
-  #define _LIBCXXABI_HIDDEN
-  #define _LIBCXXABI_DATA_VIS __declspec(dllexport)
-  #define _LIBCXXABI_FUNC_VIS __declspec(dllexport)
-  #define _LIBCXXABI_TYPE_VIS __declspec(dllexport)
- #else
-  #define _LIBCXXABI_HIDDEN
-  #define _LIBCXXABI_DATA_VIS __declspec(dllimport)
-  #define _LIBCXXABI_FUNC_VIS __declspec(dllimport)
-  #define _LIBCXXABI_TYPE_VIS __declspec(dllimport)
- #endif
-#else
- #if !defined(_LIBCXXABI_DISABLE_VISIBILITY_ANNOTATIONS)
-  #define _LIBCXXABI_HIDDEN __attribute__((__visibility__("hidden")))
-  #define _LIBCXXABI_DATA_VIS __attribute__((__visibility__("default")))
-  #define _LIBCXXABI_FUNC_VIS __attribute__((__visibility__("default")))
-  #if __has_attribute(__type_visibility__)
-   #define _LIBCXXABI_TYPE_VIS __attribute__((__type_visibility__("default")))
-  #else
-   #define _LIBCXXABI_TYPE_VIS __attribute__((__visibility__("default")))
-  #endif
- #else
-  #define _LIBCXXABI_HIDDEN
-  #define _LIBCXXABI_DATA_VIS
-  #define _LIBCXXABI_FUNC_VIS
-  #define _LIBCXXABI_TYPE_VIS
- #endif
-#endif
-
-#if defined(_LIBCXXABI_COMPILER_MSVC)
-#define _LIBCXXABI_WEAK
-#else
-#define _LIBCXXABI_WEAK __attribute__((__weak__))
-#endif
-
-#if defined(__clang__)
-#define _LIBCXXABI_COMPILER_CLANG
-#elif defined(__GNUC__)
-#define _LIBCXXABI_COMPILER_GCC
-#endif
-
-#if __has_attribute(__no_sanitize__) && defined(_LIBCXXABI_COMPILER_CLANG)
-#define _LIBCXXABI_NO_CFI __attribute__((__no_sanitize__("cfi")))
-#else
-#define _LIBCXXABI_NO_CFI
-#endif
-
-// wasm32 follows the arm32 ABI convention of using 32-bit guard.
-#if defined(__arm__) || defined(__wasm32__) || defined(__ARM64_ARCH_8_32__)
-#  define _LIBCXXABI_GUARD_ABI_ARM
-#endif
-
-#if defined(_LIBCXXABI_COMPILER_CLANG)
-#  if !__has_feature(cxx_exceptions)
-#    define _LIBCXXABI_NO_EXCEPTIONS
-#  endif
-#elif defined(_LIBCXXABI_COMPILER_GCC) && !defined(__EXCEPTIONS)
-#  define _LIBCXXABI_NO_EXCEPTIONS
-#endif
-
-#if defined(_WIN32)
-#define _LIBCXXABI_DTOR_FUNC __thiscall
-#else
-#define _LIBCXXABI_DTOR_FUNC
-#endif
-
-#endif // ____CXXABI_CONFIG_H
diff --git a/contrib/libs/cxxsupp/libcxxabi/include/c++/v1/cxxabi.h b/contrib/libs/cxxsupp/libcxxabi/include/c++/v1/cxxabi.h
deleted file mode 100644
index 85cb4b36b811..000000000000
--- a/contrib/libs/cxxsupp/libcxxabi/include/c++/v1/cxxabi.h
+++ /dev/null
@@ -1,177 +0,0 @@
-//===----------------------------------------------------------------------===//
-//
-// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
-// See https://llvm.org/LICENSE.txt for license information.
-// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
-//
-//===----------------------------------------------------------------------===//
-
-#ifndef __CXXABI_H
-#define __CXXABI_H
-
-/*
- * This header provides the interface to the C++ ABI as defined at:
- *       https://itanium-cxx-abi.github.io/cxx-abi/
- */
-
-#include <stddef.h>
-#include <stdint.h>
-
-#include <__cxxabi_config.h>
-
-#define _LIBCPPABI_VERSION 15000
-#define _LIBCXXABI_NORETURN  __attribute__((noreturn))
-#define _LIBCXXABI_ALWAYS_COLD __attribute__((cold))
-
-#ifdef __cplusplus
-
-namespace std {
-#if defined(_WIN32)
-class _LIBCXXABI_TYPE_VIS type_info; // forward declaration
-#else
-class type_info; // forward declaration
-#endif
-}
-
-
-// runtime routines use C calling conventions, but are in __cxxabiv1 namespace
-namespace __cxxabiv1 {
-extern "C"  {
-
-// 2.4.2 Allocating the Exception Object
-extern _LIBCXXABI_FUNC_VIS void *
-__cxa_allocate_exception(size_t thrown_size) throw();
-extern _LIBCXXABI_FUNC_VIS void
-__cxa_free_exception(void *thrown_exception) throw();
-
-// 2.4.3 Throwing the Exception Object
-extern _LIBCXXABI_FUNC_VIS _LIBCXXABI_NORETURN void
-__cxa_throw(void *thrown_exception, std::type_info *tinfo,
-            void (_LIBCXXABI_DTOR_FUNC *dest)(void *));
-
-// 2.5.3 Exception Handlers
-extern _LIBCXXABI_FUNC_VIS void *
-__cxa_get_exception_ptr(void *exceptionObject) throw();
-extern _LIBCXXABI_FUNC_VIS void *
-__cxa_begin_catch(void *exceptionObject) throw();
-extern _LIBCXXABI_FUNC_VIS void __cxa_end_catch();
-#if defined(_LIBCXXABI_ARM_EHABI)
-extern _LIBCXXABI_FUNC_VIS bool
-__cxa_begin_cleanup(void *exceptionObject) throw();
-extern _LIBCXXABI_FUNC_VIS void __cxa_end_cleanup();
-#endif
-extern _LIBCXXABI_FUNC_VIS std::type_info *__cxa_current_exception_type();
-
-// 2.5.4 Rethrowing Exceptions
-extern _LIBCXXABI_FUNC_VIS _LIBCXXABI_NORETURN void __cxa_rethrow();
-
-// 2.6 Auxiliary Runtime APIs
-extern _LIBCXXABI_FUNC_VIS _LIBCXXABI_NORETURN void __cxa_bad_cast(void);
-extern _LIBCXXABI_FUNC_VIS _LIBCXXABI_NORETURN void __cxa_bad_typeid(void);
-extern _LIBCXXABI_FUNC_VIS _LIBCXXABI_NORETURN void
-__cxa_throw_bad_array_new_length(void);
-
-// 3.2.6 Pure Virtual Function API
-extern _LIBCXXABI_FUNC_VIS _LIBCXXABI_NORETURN void __cxa_pure_virtual(void);
-
-// 3.2.7 Deleted Virtual Function API
-extern _LIBCXXABI_FUNC_VIS _LIBCXXABI_NORETURN void __cxa_deleted_virtual(void);
-
-// 3.3.2 One-time Construction API
-#if defined(_LIBCXXABI_GUARD_ABI_ARM)
-extern _LIBCXXABI_FUNC_VIS _LIBCXXABI_ALWAYS_COLD int __cxa_guard_acquire(uint32_t *);
-extern _LIBCXXABI_FUNC_VIS _LIBCXXABI_ALWAYS_COLD void __cxa_guard_release(uint32_t *);
-extern _LIBCXXABI_FUNC_VIS _LIBCXXABI_ALWAYS_COLD void __cxa_guard_abort(uint32_t *);
-#else
-extern _LIBCXXABI_FUNC_VIS _LIBCXXABI_ALWAYS_COLD int __cxa_guard_acquire(uint64_t *);
-extern _LIBCXXABI_FUNC_VIS _LIBCXXABI_ALWAYS_COLD void __cxa_guard_release(uint64_t *);
-extern _LIBCXXABI_FUNC_VIS _LIBCXXABI_ALWAYS_COLD void __cxa_guard_abort(uint64_t *);
-#endif
-
-// 3.3.3 Array Construction and Destruction API
-extern _LIBCXXABI_FUNC_VIS void *
-__cxa_vec_new(size_t element_count, size_t element_size, size_t padding_size,
-              void (*constructor)(void *), void (*destructor)(void *));
-
-extern _LIBCXXABI_FUNC_VIS void *
-__cxa_vec_new2(size_t element_count, size_t element_size, size_t padding_size,
-               void (*constructor)(void *), void (*destructor)(void *),
-               void *(*alloc)(size_t), void (*dealloc)(void *));
-
-extern _LIBCXXABI_FUNC_VIS void *
-__cxa_vec_new3(size_t element_count, size_t element_size, size_t padding_size,
-               void (*constructor)(void *), void (*destructor)(void *),
-               void *(*alloc)(size_t), void (*dealloc)(void *, size_t));
-
-extern _LIBCXXABI_FUNC_VIS void
-__cxa_vec_ctor(void *array_address, size_t element_count, size_t element_size,
-               void (*constructor)(void *), void (*destructor)(void *));
-
-extern _LIBCXXABI_FUNC_VIS void __cxa_vec_dtor(void *array_address,
-                                               size_t element_count,
-                                               size_t element_size,
-                                               void (*destructor)(void *));
-
-extern _LIBCXXABI_FUNC_VIS void __cxa_vec_cleanup(void *array_address,
-                                                  size_t element_count,
-                                                  size_t element_size,
-                                                  void (*destructor)(void *));
-
-extern _LIBCXXABI_FUNC_VIS void __cxa_vec_delete(void *array_address,
-                                                 size_t element_size,
-                                                 size_t padding_size,
-                                                 void (*destructor)(void *));
-
-extern _LIBCXXABI_FUNC_VIS void
-__cxa_vec_delete2(void *array_address, size_t element_size, size_t padding_size,
-                  void (*destructor)(void *), void (*dealloc)(void *));
-
-extern _LIBCXXABI_FUNC_VIS void
-__cxa_vec_delete3(void *__array_address, size_t element_size,
-                  size_t padding_size, void (*destructor)(void *),
-                  void (*dealloc)(void *, size_t));
-
-extern _LIBCXXABI_FUNC_VIS void
-__cxa_vec_cctor(void *dest_array, void *src_array, size_t element_count,
-                size_t element_size, void (*constructor)(void *, void *),
-                void (*destructor)(void *));
-
-// 3.3.5.3 Runtime API
-// These functions are part of the C++ ABI, but they are not defined in libc++abi:
-//    int __cxa_atexit(void (*)(void *), void *, void *);
-//    void __cxa_finalize(void *);
-
-// 3.4 Demangler API
-extern _LIBCXXABI_FUNC_VIS char *__cxa_demangle(const char *mangled_name,
-                                                char *output_buffer,
-                                                size_t *length, int *status);
-
-// Apple additions to support C++ 0x exception_ptr class
-// These are primitives to wrap a smart pointer around an exception object
-extern _LIBCXXABI_FUNC_VIS void *__cxa_current_primary_exception() throw();
-extern _LIBCXXABI_FUNC_VIS void
-__cxa_rethrow_primary_exception(void *primary_exception);
-extern _LIBCXXABI_FUNC_VIS void
-__cxa_increment_exception_refcount(void *primary_exception) throw();
-extern _LIBCXXABI_FUNC_VIS void
-__cxa_decrement_exception_refcount(void *primary_exception) throw();
-
-// Apple extension to support std::uncaught_exception()
-extern _LIBCXXABI_FUNC_VIS bool __cxa_uncaught_exception() throw();
-extern _LIBCXXABI_FUNC_VIS unsigned int __cxa_uncaught_exceptions() throw();
-
-#if defined(__linux__) || defined(__Fuchsia__)
-// Linux and Fuchsia TLS support. Not yet an official part of the Itanium ABI.
-// https://sourceware.org/glibc/wiki/Destructor%20support%20for%20thread_local%20variables
-extern _LIBCXXABI_FUNC_VIS int __cxa_thread_atexit(void (*)(void *), void *,
-                                                   void *) throw();
-#endif
-
-} // extern "C"
-} // namespace __cxxabiv1
-
-namespace abi = __cxxabiv1;
-
-#endif // __cplusplus
-
-#endif // __CXXABI_H
diff --git a/contrib/python/argcomplete/py3/.dist-info/METADATA b/contrib/python/argcomplete/py3/.dist-info/METADATA
index 5298d4a927a8..fe74af8ae585 100644
--- a/contrib/python/argcomplete/py3/.dist-info/METADATA
+++ b/contrib/python/argcomplete/py3/.dist-info/METADATA
@@ -1,17 +1,18 @@
-Metadata-Version: 2.1
+Metadata-Version: 2.4
 Name: argcomplete
-Version: 3.5.2
+Version: 3.5.3
 Summary: Bash tab completion for argparse
-Home-page: https://github.com/kislyuk/argcomplete
-Author: Andrey Kislyuk
-Author-email: kislyuk@gmail.com
-License: Apache Software License
+Project-URL: Homepage, https://github.com/kislyuk/argcomplete
 Project-URL: Documentation, https://kislyuk.github.io/argcomplete
 Project-URL: Source Code, https://github.com/kislyuk/argcomplete
 Project-URL: Issue Tracker, https://github.com/kislyuk/argcomplete/issues
 Project-URL: Change Log, https://github.com/kislyuk/argcomplete/blob/develop/Changes.rst
-Platform: MacOS X
-Platform: Posix
+Author: Andrey Kislyuk
+Author-email: kislyuk@gmail.com
+License: Apache Software License
+License-File: LICENSE.rst
+License-File: NOTICE
+Classifier: Development Status :: 5 - Production/Stable
 Classifier: Environment :: Console
 Classifier: Intended Audience :: Developers
 Classifier: License :: OSI Approved :: Apache Software License
@@ -27,21 +28,18 @@ Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Classifier: Programming Language :: Python :: Implementation :: CPython
 Classifier: Programming Language :: Python :: Implementation :: PyPy
-Classifier: Development Status :: 5 - Production/Stable
 Classifier: Topic :: Software Development
 Classifier: Topic :: Software Development :: Libraries :: Python Modules
 Classifier: Topic :: System :: Shells
 Classifier: Topic :: Terminals
 Requires-Python: >=3.8
-Description-Content-Type: text/x-rst
-License-File: LICENSE.rst
-License-File: NOTICE
 Provides-Extra: test
-Requires-Dist: coverage; extra == "test"
-Requires-Dist: pexpect; extra == "test"
-Requires-Dist: wheel; extra == "test"
-Requires-Dist: ruff; extra == "test"
-Requires-Dist: mypy; extra == "test"
+Requires-Dist: coverage; extra == 'test'
+Requires-Dist: mypy; extra == 'test'
+Requires-Dist: pexpect; extra == 'test'
+Requires-Dist: ruff; extra == 'test'
+Requires-Dist: wheel; extra == 'test'
+Description-Content-Type: text/x-rst
 
 argcomplete - Bash/zsh tab completion for argparse
 ==================================================
@@ -85,6 +83,9 @@ follows:
     args = parser.parse_args()
     ...
 
+If using ``pyproject.toml`` ``[project.scripts]`` entry points, the ``PYTHON_ARGCOMPLETE_OK`` marker should appear
+at the beginning of the file that contains the entry point.
+
 Register your Python application with your shell's completion framework by running ``register-python-argcomplete``::
 
     eval "$(register-python-argcomplete my-python-app)"
@@ -268,9 +269,9 @@ multiple Python versions on the same system, the version being used to run the s
  ``/etc/shells``, and run ``chsh`` to change your shell). You can check the version of the running copy of bash with
  ``echo $BASH_VERSION``.
 
-.. note:: If you use setuptools/distribute ``scripts`` or ``entry_points`` directives to package your module,
+.. note:: If you use ``project.scripts`` directives to provide command line entry points to your package,
  argcomplete will follow the wrapper scripts to their destination and look for ``PYTHON_ARGCOMPLETE_OK`` in the
- destination code.
+ first kilobyte of the file containing the destination code.
 
 If you choose not to use global completion, or ship a completion module that depends on argcomplete, you must register
 your script explicitly using ``eval "$(register-python-argcomplete my-python-app)"``. Standard completion module
diff --git a/contrib/python/argcomplete/py3/README.rst b/contrib/python/argcomplete/py3/README.rst
index c134a566e6b2..13f0fa7ae366 100644
--- a/contrib/python/argcomplete/py3/README.rst
+++ b/contrib/python/argcomplete/py3/README.rst
@@ -40,6 +40,9 @@ follows:
     args = parser.parse_args()
     ...
 
+If using ``pyproject.toml`` ``[project.scripts]`` entry points, the ``PYTHON_ARGCOMPLETE_OK`` marker should appear
+at the beginning of the file that contains the entry point.
+
 Register your Python application with your shell's completion framework by running ``register-python-argcomplete``::
 
     eval "$(register-python-argcomplete my-python-app)"
@@ -223,9 +226,9 @@ multiple Python versions on the same system, the version being used to run the s
  ``/etc/shells``, and run ``chsh`` to change your shell). You can check the version of the running copy of bash with
  ``echo $BASH_VERSION``.
 
-.. note:: If you use setuptools/distribute ``scripts`` or ``entry_points`` directives to package your module,
+.. note:: If you use ``project.scripts`` directives to provide command line entry points to your package,
  argcomplete will follow the wrapper scripts to their destination and look for ``PYTHON_ARGCOMPLETE_OK`` in the
- destination code.
+ first kilobyte of the file containing the destination code.
 
 If you choose not to use global completion, or ship a completion module that depends on argcomplete, you must register
 your script explicitly using ``eval "$(register-python-argcomplete my-python-app)"``. Standard completion module
diff --git a/contrib/python/argcomplete/py3/argcomplete/completers.py b/contrib/python/argcomplete/py3/argcomplete/completers.py
index b62c593d64f3..610349f65278 100644
--- a/contrib/python/argcomplete/py3/argcomplete/completers.py
+++ b/contrib/python/argcomplete/py3/argcomplete/completers.py
@@ -59,13 +59,17 @@ def __call__(self, prefix, **kwargs):
         completion = []
         if self.allowednames:
             if self.directories:
-                files = _call(["bash", "-c", "compgen -A directory -- '{p}'".format(p=prefix)])
+                # Using 'bind' in this and the following commands is a workaround to a bug in bash
+                # that was fixed in bash 5.3 but affects older versions. Environment variables are not treated
+                # correctly in older versions and calling bind makes them available. For details, see
+                # https://savannah.gnu.org/support/index.php?111125
+                files = _call(["bash", "-c", "bind; compgen -A directory -- '{p}'".format(p=prefix)], stderr=subprocess.DEVNULL)
                 completion += [f + "/" for f in files]
             for x in self.allowednames:
-                completion += _call(["bash", "-c", "compgen -A file -X '!*.{0}' -- '{p}'".format(x, p=prefix)])
+                completion += _call(["bash", "-c", "bind; compgen -A file -X '!*.{0}' -- '{p}'".format(x, p=prefix)], stderr=subprocess.DEVNULL)
         else:
-            completion += _call(["bash", "-c", "compgen -A file -- '{p}'".format(p=prefix)])
-            anticomp = _call(["bash", "-c", "compgen -A directory -- '{p}'".format(p=prefix)])
+            completion += _call(["bash", "-c", "bind; compgen -A file -- '{p}'".format(p=prefix)], stderr=subprocess.DEVNULL)
+            anticomp = _call(["bash", "-c", "bind; compgen -A directory -- '{p}'".format(p=prefix)], stderr=subprocess.DEVNULL)
             completion = list(set(completion) - set(anticomp))
 
             if self.directories:
diff --git a/contrib/python/argcomplete/py3/ya.make b/contrib/python/argcomplete/py3/ya.make
index e54ca26a823e..dfa97494a92b 100644
--- a/contrib/python/argcomplete/py3/ya.make
+++ b/contrib/python/argcomplete/py3/ya.make
@@ -2,7 +2,7 @@
 
 PY3_LIBRARY()
 
-VERSION(3.5.2)
+VERSION(3.5.3)
 
 LICENSE(Apache-2.0)
 
diff --git a/contrib/python/pyparsing/py3/.dist-info/METADATA b/contrib/python/pyparsing/py3/.dist-info/METADATA
index ff6f9b622795..6b5fbefef606 100644
--- a/contrib/python/pyparsing/py3/.dist-info/METADATA
+++ b/contrib/python/pyparsing/py3/.dist-info/METADATA
@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: pyparsing
-Version: 3.2.0
+Version: 3.2.1
 Summary: pyparsing module - Classes and methods to define and execute parsing grammars
 Author-email: Paul McGuire <ptmcg.gm+pyparsing@gmail.com>
 Requires-Python: >=3.9
diff --git a/contrib/python/pyparsing/py3/pyparsing/__init__.py b/contrib/python/pyparsing/py3/pyparsing/__init__.py
index 543ceb62bd53..726c76cb2449 100644
--- a/contrib/python/pyparsing/py3/pyparsing/__init__.py
+++ b/contrib/python/pyparsing/py3/pyparsing/__init__.py
@@ -120,8 +120,8 @@ def __repr__(self):
         return f"{__name__}.{type(self).__name__}({', '.join('{}={!r}'.format(*nv) for nv in zip(self._fields, self))})"
 
 
-__version_info__ = version_info(3, 2, 0, "final", 1)
-__version_time__ = "13 Oct 2024 09:46 UTC"
+__version_info__ = version_info(3, 2, 1, "final", 1)
+__version_time__ = "31 Dec 2024 20:41 UTC"
 __version__ = __version_info__.__version__
 __versionTime__ = __version_time__
 __author__ = "Paul McGuire <ptmcg.gm+pyparsing@gmail.com>"
diff --git a/contrib/python/pyparsing/py3/pyparsing/actions.py b/contrib/python/pyparsing/py3/pyparsing/actions.py
index 1d2dce99e199..f491aab986e9 100644
--- a/contrib/python/pyparsing/py3/pyparsing/actions.py
+++ b/contrib/python/pyparsing/py3/pyparsing/actions.py
@@ -1,21 +1,34 @@
 # actions.py
+from __future__ import annotations
+
+from typing import Union, Callable, Any
 
 from .exceptions import ParseException
 from .util import col, replaced_by_pep8
+from .results import ParseResults
+
+
+ParseAction = Union[
+    Callable[[], Any],
+    Callable[[ParseResults], Any],
+    Callable[[int, ParseResults], Any],
+    Callable[[str, int, ParseResults], Any],
+]
 
 
 class OnlyOnce:
     """
     Wrapper for parse actions, to ensure they are only called once.
+    Note: parse action signature must include all 3 arguments.
     """
 
-    def __init__(self, method_call):
+    def __init__(self, method_call: Callable[[str, int, ParseResults], Any]):
         from .core import _trim_arity
 
         self.callable = _trim_arity(method_call)
         self.called = False
 
-    def __call__(self, s, l, t):
+    def __call__(self, s: str, l: int, t: ParseResults) -> ParseResults:
         if not self.called:
             results = self.callable(s, l, t)
             self.called = True
@@ -30,20 +43,20 @@ def reset(self):
         self.called = False
 
 
-def match_only_at_col(n):
+def match_only_at_col(n: int) -> ParseAction:
     """
     Helper method for defining parse actions that require matching at
     a specific column in the input text.
     """
 
-    def verify_col(strg, locn, toks):
+    def verify_col(strg: str, locn: int, toks: ParseResults) -> None:
         if col(locn, strg) != n:
             raise ParseException(strg, locn, f"matched token not at column {n}")
 
     return verify_col
 
 
-def replace_with(repl_str):
+def replace_with(repl_str: str) -> ParseAction:
     """
     Helper method for common parse actions that simply return
     a literal value.  Especially useful when used with
@@ -60,7 +73,7 @@ def replace_with(repl_str):
     return lambda s, l, t: [repl_str]
 
 
-def remove_quotes(s, l, t):
+def remove_quotes(s: str, l: int, t: ParseResults) -> Any:
     """
     Helper parse action for removing quotation marks from parsed
     quoted strings.
@@ -77,7 +90,7 @@ def remove_quotes(s, l, t):
     return t[0][1:-1]
 
 
-def with_attribute(*args, **attr_dict):
+def with_attribute(*args: tuple[str, str], **attr_dict) -> ParseAction:
     """
     Helper to create a validating parse action to be used with start
     tags created with :class:`make_xml_tags` or
@@ -133,17 +146,17 @@ def with_attribute(*args, **attr_dict):
         1 4 0 1 0
         1,3 2,3 1,1
     """
+    attrs_list: list[tuple[str, str]] = []
     if args:
-        attrs = args[:]
+        attrs_list.extend(args)
     else:
-        attrs = attr_dict.items()
-    attrs = [(k, v) for k, v in attrs]
+        attrs_list.extend(attr_dict.items())
 
-    def pa(s, l, tokens):
-        for attrName, attrValue in attrs:
+    def pa(s: str, l: int, tokens: ParseResults) -> None:
+        for attrName, attrValue in attrs_list:
             if attrName not in tokens:
                 raise ParseException(s, l, "no matching attribute " + attrName)
-            if attrValue != with_attribute.ANY_VALUE and tokens[attrName] != attrValue:
+            if attrValue != with_attribute.ANY_VALUE and tokens[attrName] != attrValue:  # type: ignore [attr-defined]
                 raise ParseException(
                     s,
                     l,
@@ -156,7 +169,7 @@ def pa(s, l, tokens):
 with_attribute.ANY_VALUE = object()  # type: ignore [attr-defined]
 
 
-def with_class(classname, namespace=""):
+def with_class(classname: str, namespace: str = "") -> ParseAction:
     """
     Simplified version of :class:`with_attribute` when
     matching on a div class - made difficult because ``class`` is
diff --git a/contrib/python/pyparsing/py3/pyparsing/common.py b/contrib/python/pyparsing/py3/pyparsing/common.py
index 649aad009617..e46511086dad 100644
--- a/contrib/python/pyparsing/py3/pyparsing/common.py
+++ b/contrib/python/pyparsing/py3/pyparsing/common.py
@@ -210,14 +210,14 @@ class pyparsing_common:
     """any numeric expression, returns the corresponding Python type"""
 
     fnumber = (
-        Regex(r"[+-]?\d+\.?\d*([eE][+-]?\d+)?")
+        Regex(r"[+-]?\d+\.?\d*(?:[eE][+-]?\d+)?")
         .set_name("fnumber")
         .set_parse_action(convert_to_float)
     )
     """any int or real number, returned as float"""
 
     ieee_float = (
-        Regex(r"(?i)[+-]?((\d+\.?\d*(e[+-]?\d+)?)|nan|inf(inity)?)")
+        Regex(r"(?i:[+-]?(?:(?:\d+\.?\d*(?:e[+-]?\d+)?)|nan|inf(?:inity)?))")
         .set_name("ieee_float")
         .set_parse_action(convert_to_float)
     )
diff --git a/contrib/python/pyparsing/py3/pyparsing/core.py b/contrib/python/pyparsing/py3/pyparsing/core.py
index 4f43c3bf991c..b884e2d4a40b 100644
--- a/contrib/python/pyparsing/py3/pyparsing/core.py
+++ b/contrib/python/pyparsing/py3/pyparsing/core.py
@@ -215,12 +215,7 @@ def _should_enable_warnings(
 _generatorType = types.GeneratorType
 ParseImplReturnType = tuple[int, Any]
 PostParseReturnType = Union[ParseResults, Sequence[ParseResults]]
-ParseAction = Union[
-    Callable[[], Any],
-    Callable[[ParseResults], Any],
-    Callable[[int, ParseResults], Any],
-    Callable[[str, int, ParseResults], Any],
-]
+
 ParseCondition = Union[
     Callable[[], bool],
     Callable[[ParseResults], bool],
@@ -486,6 +481,7 @@ def __init__(self, savelist: bool = False):
         self.callPreparse = True
         self.callDuringTry = False
         self.suppress_warnings_: list[Diagnostics] = []
+        self.show_in_diagram = True
 
     def suppress_warning(self, warning_type: Diagnostics) -> ParserElement:
         """
@@ -5558,7 +5554,8 @@ def __or__(self, other) -> ParserElement:
             not in self.suppress_warnings_
         ):
             warnings.warn(
-                "using '<<' operator with '|' is probably an error, use '<<='",
+                "warn_on_match_first_with_lshift_operator:"
+                " using '<<' operator with '|' is probably an error, use '<<='",
                 stacklevel=2,
             )
         ret = super().__or__(other)
@@ -5572,7 +5569,8 @@ def __del__(self):
             and Diagnostics.warn_on_assignment_to_Forward not in self.suppress_warnings_
         ):
             warnings.warn_explicit(
-                "Forward defined here but no expression attached later using '<<=' or '<<'",
+                "warn_on_assignment_to_Forward:"
+                " Forward defined here but no expression attached later using '<<=' or '<<'",
                 UserWarning,
                 filename=self.caller_frame.filename,
                 lineno=self.caller_frame.lineno,
@@ -5600,7 +5598,8 @@ def parseImpl(self, instring, loc, do_actions=True) -> ParseImplReturnType:
             else:
                 stacklevel = 2
             warnings.warn(
-                "Forward expression was never assigned a value, will not parse any input",
+                "warn_on_parse_using_empty_Forward:"
+                " Forward expression was never assigned a value, will not parse any input",
                 stacklevel=stacklevel,
             )
         if not ParserElement._left_recursion_enabled:
@@ -6157,7 +6156,7 @@ def autoname_elements() -> None:
     Utility to simplify mass-naming of parser elements, for
     generating railroad diagram with named subdiagrams.
     """
-    calling_frame = sys._getframe().f_back
+    calling_frame = sys._getframe(1)
     if calling_frame is None:
         return
     calling_frame = typing.cast(types.FrameType, calling_frame)
diff --git a/contrib/python/pyparsing/py3/pyparsing/diagram/__init__.py b/contrib/python/pyparsing/py3/pyparsing/diagram/__init__.py
index 7926f2c35528..56526b741b87 100644
--- a/contrib/python/pyparsing/py3/pyparsing/diagram/__init__.py
+++ b/contrib/python/pyparsing/py3/pyparsing/diagram/__init__.py
@@ -1,6 +1,7 @@
 # mypy: ignore-errors
 from __future__ import annotations
 
+import itertools
 import railroad
 import pyparsing
 import dataclasses
@@ -40,7 +41,7 @@
 {{ body | safe }}
 {% for diagram in diagrams %}
     <div class="railroad-group">
-        <h1 class="railroad-heading">{{ diagram.title }}</h1>
+        <h1 class="railroad-heading" id="{{ diagram.bookmark }}">{{ diagram.title }}</h1>
         <div class="railroad-description">{{ diagram.text }}</div>
         <div class="railroad-svg">
             {{ diagram.svg }}
@@ -56,8 +57,35 @@
 template = Template(jinja2_template_source)
 
 
+_bookmark_lookup = {}
+_bookmark_ids = itertools.count(start=1)
+
+def _make_bookmark(s: str) -> str:
+    """
+    Converts a string into a valid HTML bookmark (ID or anchor name).
+    """
+    if s in _bookmark_lookup:
+        return _bookmark_lookup[s]
+
+    # Replace invalid characters with hyphens and ensure only valid characters
+    bookmark = re.sub(r'[^a-zA-Z0-9-]+', '-', s)
+
+    # Ensure it starts with a letter by adding 'z' if necessary
+    if not bookmark[:1].isalpha():
+        bookmark = f"z{bookmark}"
+
+    # Convert to lowercase and strip hyphens
+    bookmark = bookmark.lower().strip('-')
+
+    _bookmark_lookup[s] = bookmark = f"{bookmark}-{next(_bookmark_ids):04d}"
+
+    return bookmark
+
+
 def _collapse_verbose_regex(regex_str: str) -> str:
-    collapsed = pyparsing.Regex(r"#.*").suppress().transform_string(regex_str)
+    if "\n" not in regex_str:
+        return regex_str
+    collapsed = pyparsing.Regex(r"#.*$").suppress().transform_string(regex_str)
     collapsed = re.sub(r"\s*\n\s*", "", collapsed)
     return collapsed
 
@@ -72,6 +100,11 @@ class NamedDiagram:
     index: int
     diagram: railroad.DiagramItem = None
 
+    @property
+    def bookmark(self):
+        bookmark = _make_bookmark(self.name)
+        return bookmark
+
 
 T = TypeVar("T")
 
@@ -99,7 +132,7 @@ class AnnotatedItem(railroad.Group):
     """
 
     def __init__(self, label: str, item):
-        super().__init__(item=item, label=f"[{label}]")
+        super().__init__(item=item, label=f"[{label}]" if label else "")
 
 
 class EditablePartial(Generic[T]):
@@ -162,7 +195,11 @@ def railroad_to_html(diagrams: list[NamedDiagram], embed=False, **kwargs) -> str
         title = diagram.name
         if diagram.index == 0:
             title += " (root)"
-        data.append({"title": title, "text": "", "svg": io.getvalue()})
+        data.append(
+            {
+                "title": title, "text": "", "svg": io.getvalue(), "bookmark": diagram.bookmark
+            }
+        )
 
     return template.render(diagrams=data, embed=embed, **kwargs)
 
@@ -336,6 +373,12 @@ def __delitem__(self, key: int):
     def __contains__(self, key: int):
         return key in self._element_diagram_states
 
+    def get(self, key, default=None):
+        try:
+            return self[key]
+        except KeyError:
+            return default
+
     def generate_unnamed(self) -> int:
         """
         Generate a number used in the name of an otherwise unnamed diagram
@@ -360,7 +403,8 @@ def extract_into_diagram(self, el_id: int):
 
         # Replace the original definition of this element with a regular block
         if position.parent:
-            ret = EditablePartial.from_call(railroad.NonTerminal, text=position.name)
+            href = f"#{_make_bookmark(position.name)}"
+            ret = EditablePartial.from_call(railroad.NonTerminal, text=position.name, href=href)
             if "item" in position.parent.kwargs:
                 position.parent.kwargs["item"] = ret
             elif "items" in position.parent.kwargs:
@@ -447,7 +491,7 @@ def _visible_exprs(exprs: Iterable[pyparsing.ParserElement]):
     return [
         e
         for e in exprs
-        if not (e.customName or e.resultsName or isinstance(e, non_diagramming_exprs))
+        if not isinstance(e, non_diagramming_exprs)
     ]
 
 
@@ -461,6 +505,7 @@ def _to_diagram_element(
     name_hint: str = None,
     show_results_names: bool = False,
     show_groups: bool = False,
+    show_hidden: bool = False,
 ) -> typing.Optional[EditablePartial]:
     """
     Recursively converts a PyParsing Element to a railroad Element
@@ -472,8 +517,9 @@ def _to_diagram_element(
     do so
     :param name_hint: If provided, this will override the generated name
     :param show_results_names: bool flag indicating whether to add annotations for results names
-    :returns: The converted version of the input element, but as a Partial that hasn't yet been constructed
     :param show_groups: bool flag indicating whether to show groups using bounding box
+    :param show_hidden: bool flag indicating whether to show elements that are typically hidden
+    :returns: The converted version of the input element, but as a Partial that hasn't yet been constructed
     """
     exprs = element.recurse()
     name = name_hint or element.customName or type(element).__name__
@@ -489,7 +535,7 @@ def _to_diagram_element(
             element,
             (
                 # pyparsing.TokenConverter,
-                # pyparsing.Forward,
+                pyparsing.Forward,
                 pyparsing.Located,
             ),
         ):
@@ -513,25 +559,33 @@ def _to_diagram_element(
 
     # If the element isn't worth extracting, we always treat it as the first time we say it
     if _worth_extracting(element):
-        if el_id in lookup and lookup[el_id].name is not None:
+        looked_up = lookup.get(el_id)
+        if looked_up and looked_up.name is not None:
             # If we've seen this element exactly once before, we are only just now finding out that it's a duplicate,
             # so we have to extract it into a new diagram.
-            looked_up = lookup[el_id]
             looked_up.mark_for_extraction(el_id, lookup, name=name_hint)
-            ret = EditablePartial.from_call(railroad.NonTerminal, text=looked_up.name)
+            href = f"#{_make_bookmark(looked_up.name)}"
+            ret = EditablePartial.from_call(railroad.NonTerminal, text=looked_up.name, href=href)
             return ret
 
         elif el_id in lookup.diagrams:
             # If we have seen the element at least twice before, and have already extracted it into a subdiagram, we
             # just put in a marker element that refers to the sub-diagram
+            text = lookup.diagrams[el_id].kwargs["name"]
             ret = EditablePartial.from_call(
-                railroad.NonTerminal, text=lookup.diagrams[el_id].kwargs["name"]
+                railroad.NonTerminal, text=text, href=f"#{_make_bookmark(text)}"
             )
             return ret
 
     # Recursively convert child elements
     # Here we find the most relevant Railroad element for matching pyparsing Element
     # We use ``items=[]`` here to hold the place for where the child elements will go once created
+
+    # see if this element is normally hidden, and whether hidden elements are desired
+    # if not, just return None
+    if not element.show_in_diagram and not show_hidden:
+        return None
+
     if isinstance(element, pyparsing.And):
         # detect And's created with ``expr*N`` notation - for these use a OneOrMore with a repeat
         # (all will have the same name, and resultsName)
@@ -566,7 +620,9 @@ def _to_diagram_element(
         if show_groups:
             ret = EditablePartial.from_call(AnnotatedItem, label="", item="")
         else:
-            ret = EditablePartial.from_call(railroad.Sequence, items=[])
+            ret = EditablePartial.from_call(
+                railroad.Group, item=None, label=element_results_name
+            )
     elif isinstance(element, pyparsing.TokenConverter):
         label = type(element).__name__.lower()
         if label == "tokenconverter":
@@ -607,10 +663,6 @@ def _to_diagram_element(
                 *args,
             )
         ret = EditablePartial.from_call(railroad.ZeroOrMore, item="")
-    elif isinstance(element, pyparsing.Group):
-        ret = EditablePartial.from_call(
-            railroad.Group, item=None, label=element_results_name
-        )
     elif isinstance(element, pyparsing.Empty) and not element.customName:
         # Skip unnamed "Empty" elements
         ret = None
@@ -619,10 +671,8 @@ def _to_diagram_element(
     elif len(exprs) > 0 and not element_results_name:
         ret = EditablePartial.from_call(railroad.Group, item="", label=name)
     elif isinstance(element, pyparsing.Regex):
-        patt = _collapse_verbose_regex(element.pattern)
-        element.pattern = patt
-        element._defaultName = None
-        ret = EditablePartial.from_call(railroad.Terminal, element.defaultName)
+        collapsed_patt = _collapse_verbose_regex(element.pattern)
+        ret = EditablePartial.from_call(railroad.Terminal, collapsed_patt)
     elif len(exprs) > 0:
         ret = EditablePartial.from_call(railroad.Sequence, items=[])
     else:
@@ -685,8 +735,10 @@ def _to_diagram_element(
     if el_id in lookup and lookup[el_id].extract and lookup[el_id].complete:
         lookup.extract_into_diagram(el_id)
         if ret is not None:
+            text = lookup.diagrams[el_id].kwargs["name"]
+            href = f"#{_make_bookmark(text)}"
             ret = EditablePartial.from_call(
-                railroad.NonTerminal, text=lookup.diagrams[el_id].kwargs["name"]
+                railroad.NonTerminal, text=text, href=href
             )
 
     return ret
diff --git a/contrib/python/pyparsing/py3/pyparsing/helpers.py b/contrib/python/pyparsing/py3/pyparsing/helpers.py
index d2bd05f3d39a..f781e8713271 100644
--- a/contrib/python/pyparsing/py3/pyparsing/helpers.py
+++ b/contrib/python/pyparsing/py3/pyparsing/helpers.py
@@ -199,7 +199,8 @@ def one_of(
         and __diag__.warn_on_multiple_string_args_to_oneof
     ):
         warnings.warn(
-            "More than one string argument passed to one_of, pass"
+            "warn_on_multiple_string_args_to_oneof:"
+            " More than one string argument passed to one_of, pass"
             " choices as a list or space-delimited string",
             stacklevel=2,
         )
@@ -779,25 +780,27 @@ def parseImpl(self, instring, loc, doActions=True):
     _FB.__name__ = "FollowedBy>"
 
     ret = Forward()
+    ret.set_name(f"{base_expr.name}_expression")
     if isinstance(lpar, str):
         lpar = Suppress(lpar)
     if isinstance(rpar, str):
         rpar = Suppress(rpar)
 
+    nested_expr = (lpar + ret + rpar).set_name(f"nested_{base_expr.name}")
+
     # if lpar and rpar are not suppressed, wrap in group
     if not (isinstance(lpar, Suppress) and isinstance(rpar, Suppress)):
-        lastExpr = base_expr | Group(lpar + ret + rpar).set_name(
-            f"nested_{base_expr.name}"
-        )
+        lastExpr = base_expr | Group(nested_expr)
     else:
-        lastExpr = base_expr | (lpar + ret + rpar).set_name(f"nested_{base_expr.name}")
-    root_expr = lastExpr
+        lastExpr = base_expr | nested_expr
 
     arity: int
     rightLeftAssoc: opAssoc
     pa: typing.Optional[ParseAction]
     opExpr1: ParserElement
     opExpr2: ParserElement
+    matchExpr: ParserElement
+    match_lookahead: ParserElement
     for operDef in op_list:
         opExpr, arity, rightLeftAssoc, pa = (operDef + (None,))[:4]  # type: ignore[assignment]
         if isinstance(opExpr, str_type):
@@ -809,9 +812,9 @@ def parseImpl(self, instring, loc, doActions=True):
                     "if numterms=3, opExpr must be a tuple or list of two expressions"
                 )
             opExpr1, opExpr2 = opExpr
-            term_name = f"{opExpr1}{opExpr2} term"
+            term_name = f"{opExpr1}{opExpr2} operations"
         else:
-            term_name = f"{opExpr} term"
+            term_name = f"{opExpr} operations"
 
         if not 1 <= arity <= 3:
             raise ValueError("operator must be unary (1), binary (2), or ternary (3)")
@@ -821,48 +824,62 @@ def parseImpl(self, instring, loc, doActions=True):
 
         thisExpr: ParserElement = Forward().set_name(term_name)
         thisExpr = typing.cast(Forward, thisExpr)
+        match_lookahead = And([])
         if rightLeftAssoc is OpAssoc.LEFT:
             if arity == 1:
-                matchExpr = _FB(lastExpr + opExpr) + Group(lastExpr + opExpr[1, ...])
+                match_lookahead = _FB(lastExpr + opExpr)
+                matchExpr = Group(lastExpr + opExpr[1, ...])
             elif arity == 2:
                 if opExpr is not None:
-                    matchExpr = _FB(lastExpr + opExpr + lastExpr) + Group(
-                        lastExpr + (opExpr + lastExpr)[1, ...]
-                    )
+                    match_lookahead = _FB(lastExpr + opExpr + lastExpr)
+                    matchExpr = Group(lastExpr + (opExpr + lastExpr)[1, ...])
                 else:
-                    matchExpr = _FB(lastExpr + lastExpr) + Group(lastExpr[2, ...])
+                    match_lookahead = _FB(lastExpr + lastExpr)
+                    matchExpr = Group(lastExpr[2, ...])
             elif arity == 3:
-                matchExpr = _FB(
+                match_lookahead = _FB(
                     lastExpr + opExpr1 + lastExpr + opExpr2 + lastExpr
-                ) + Group(lastExpr + OneOrMore(opExpr1 + lastExpr + opExpr2 + lastExpr))
+                )
+                matchExpr = Group(
+                    lastExpr + (opExpr1 + lastExpr + opExpr2 + lastExpr)[1, ...]
+                )
         elif rightLeftAssoc is OpAssoc.RIGHT:
             if arity == 1:
                 # try to avoid LR with this extra test
                 if not isinstance(opExpr, Opt):
                     opExpr = Opt(opExpr)
-                matchExpr = _FB(opExpr.expr + thisExpr) + Group(opExpr + thisExpr)
+                match_lookahead = _FB(opExpr.expr + thisExpr)
+                matchExpr = Group(opExpr + thisExpr)
             elif arity == 2:
                 if opExpr is not None:
-                    matchExpr = _FB(lastExpr + opExpr + thisExpr) + Group(
-                        lastExpr + (opExpr + thisExpr)[1, ...]
-                    )
+                    match_lookahead = _FB(lastExpr + opExpr + thisExpr)
+                    matchExpr = Group(lastExpr + (opExpr + thisExpr)[1, ...])
                 else:
-                    matchExpr = _FB(lastExpr + thisExpr) + Group(
-                        lastExpr + thisExpr[1, ...]
-                    )
+                    match_lookahead = _FB(lastExpr + thisExpr)
+                    matchExpr = Group(lastExpr + thisExpr[1, ...])
             elif arity == 3:
-                matchExpr = _FB(
+                match_lookahead = _FB(
                     lastExpr + opExpr1 + thisExpr + opExpr2 + thisExpr
-                ) + Group(lastExpr + opExpr1 + thisExpr + opExpr2 + thisExpr)
+                )
+                matchExpr = Group(lastExpr + opExpr1 + thisExpr + opExpr2 + thisExpr)
+
+        # suppress lookahead expr from railroad diagrams
+        match_lookahead.show_in_diagram = False
+
+        # TODO - determine why this statement can't be included in the following
+        #  if pa block
+        matchExpr = match_lookahead + matchExpr
+
         if pa:
             if isinstance(pa, (tuple, list)):
                 matchExpr.set_parse_action(*pa)
             else:
                 matchExpr.set_parse_action(pa)
+
         thisExpr <<= (matchExpr | lastExpr).setName(term_name)
         lastExpr = thisExpr
+
     ret <<= lastExpr
-    root_expr.set_name("base_expr")
     return ret
 
 
@@ -1009,10 +1026,9 @@ def checkUnindent(s, l, t):
     return smExpr.set_name("indented block")
 
 
-# it's easy to get these comment structures wrong - they're very common, so may as well make them available
-c_style_comment = Combine(Regex(r"/\*(?:[^*]|\*(?!/))*") + "*/").set_name(
-    "C style comment"
-)
+# it's easy to get these comment structures wrong - they're very common,
+# so may as well make them available
+c_style_comment = Regex(r"/\*(?:[^*]|\*(?!/))*\*\/").set_name("C style comment")
 "Comment of the form ``/* ... */``"
 
 html_comment = Regex(r"<!--[\s\S]*?-->").set_name("HTML comment")
@@ -1022,8 +1038,8 @@ def checkUnindent(s, l, t):
 dbl_slash_comment = Regex(r"//(?:\\\n|[^\n])*").set_name("// comment")
 "Comment of the form ``// ... (to end of line)``"
 
-cpp_style_comment = Combine(
-    Regex(r"/\*(?:[^*]|\*(?!/))*") + "*/" | dbl_slash_comment
+cpp_style_comment = Regex(
+    r"(?:/\*(?:[^*]|\*(?!/))*\*\/)|(?://(?:\\\n|[^\n])*)"
 ).set_name("C++ style comment")
 "Comment of either form :class:`c_style_comment` or :class:`dbl_slash_comment`"
 
diff --git a/contrib/python/pyparsing/py3/pyparsing/results.py b/contrib/python/pyparsing/py3/pyparsing/results.py
index 245847832a8f..be834b7e6077 100644
--- a/contrib/python/pyparsing/py3/pyparsing/results.py
+++ b/contrib/python/pyparsing/py3/pyparsing/results.py
@@ -523,6 +523,7 @@ def as_list(self, *, flatten: bool = False) -> list:
             result_list = result.as_list()
             print(type(result_list), result_list) # -> <class 'list'> ['sldkj', 'lsdkj', 'sldkj']
         """
+
         def flattened(pr):
             to_visit = collections.deque([*self])
             while to_visit:
diff --git a/contrib/python/pyparsing/py3/pyparsing/util.py b/contrib/python/pyparsing/py3/pyparsing/util.py
index 1487019c272d..03a60d4fddc2 100644
--- a/contrib/python/pyparsing/py3/pyparsing/util.py
+++ b/contrib/python/pyparsing/py3/pyparsing/util.py
@@ -192,6 +192,7 @@ class _GroupConsecutive:
             (2, iter(['m']))
             (3, iter(['p', 'q', 'r', 's']))
     """
+
     def __init__(self):
         self.prev = 0
         self.counter = itertools.count()
diff --git a/contrib/python/pyparsing/py3/ya.make b/contrib/python/pyparsing/py3/ya.make
index 1b36194fab48..e229986ca672 100644
--- a/contrib/python/pyparsing/py3/ya.make
+++ b/contrib/python/pyparsing/py3/ya.make
@@ -4,7 +4,7 @@ PY3_LIBRARY()
 
 PROVIDES(pyparsing)
 
-VERSION(3.2.0)
+VERSION(3.2.1)
 
 LICENSE(MIT)
 
diff --git a/ya b/ya
index bcacf384295a..a850fec58084 100755
--- a/ya
+++ b/ya
@@ -39,33 +39,33 @@ REGISTRY_ENDPOINT = os.environ.get("YA_REGISTRY_ENDPOINT", "https://devtools-reg
 PLATFORM_MAP = {
     "data": {
         "darwin": {
-            "md5": "950e88ffc7c851dfbafafb89ba1b1b77",
+            "md5": "85c76fbd0d23319cce3ad98cd6926ad7",
             "urls": [
-                f"{REGISTRY_ENDPOINT}/7808696359"
+                f"{REGISTRY_ENDPOINT}/7827466176"
             ]
         },
         "darwin-arm64": {
-            "md5": "9d15442d9bb0a0aa6669cea947fe336b",
+            "md5": "b29d3df1d6e67afaa304bdc0fc811966",
             "urls": [
-                f"{REGISTRY_ENDPOINT}/7808696176"
+                f"{REGISTRY_ENDPOINT}/7827465506"
             ]
         },
         "linux-aarch64": {
-            "md5": "9250ef732fc3a971ff28a126126b4a76",
+            "md5": "855754159dd5301a07b58ed0c01dc334",
             "urls": [
-                f"{REGISTRY_ENDPOINT}/7808695917"
+                f"{REGISTRY_ENDPOINT}/7827464666"
             ]
         },
         "win32-clang-cl": {
-            "md5": "00674a6eda2cf970ab4c73937c808050",
+            "md5": "fba50c68c7a0a106dc7370113172c77c",
             "urls": [
-                f"{REGISTRY_ENDPOINT}/7808696543"
+                f"{REGISTRY_ENDPOINT}/7827466887"
             ]
         },
         "linux": {
-            "md5": "9a6d64ab467535c82de4351766427700",
+            "md5": "8292d28960e74996f913db75c3a880f6",
             "urls": [
-                f"{REGISTRY_ENDPOINT}/7808696801"
+                f"{REGISTRY_ENDPOINT}/7827467700"
             ]
         }
     }
diff --git a/ydb/ci/rightlib.txt b/ydb/ci/rightlib.txt
index ca6d071a3a5a..f575271e186e 100644
--- a/ydb/ci/rightlib.txt
+++ b/ydb/ci/rightlib.txt
@@ -1 +1 @@
-fc7eb64e6147971fb5da235644f47a1df89e0de5
+96a6bd3bf87c39c8257a86654f45eabf6546a778
diff --git a/yql/essentials/core/peephole_opt/yql_opt_peephole_physical.cpp b/yql/essentials/core/peephole_opt/yql_opt_peephole_physical.cpp
index 6131a31b1542..f7ba75c96c1f 100644
--- a/yql/essentials/core/peephole_opt/yql_opt_peephole_physical.cpp
+++ b/yql/essentials/core/peephole_opt/yql_opt_peephole_physical.cpp
@@ -558,11 +558,15 @@ TExprNode::TPtr ExpandEquiJoinImpl(const TExprNode& node, TExprContext& ctx) {
     if (joinKind == "Cross") {
         return ctx.Builder(node.Pos())
             .Callable("FlatMap")
-                .Add(0, std::move(list1))
+                .Callable(0, "Collect")
+                    .Add(0, std::move(list1))
+                .Seal()
                 .Lambda(1)
                     .Param("left")
                     .Callable("Map")
-                        .Add(0, std::move(list2))
+                        .Callable(0, "Collect")
+                            .Add(0, std::move(list2))
+                        .Seal()
                         .Lambda(1)
                             .Param("right")
                             .Callable("AsStruct")
diff --git a/yql/essentials/core/yql_expr_type_annotation.cpp b/yql/essentials/core/yql_expr_type_annotation.cpp
index 7da6bd439abc..56626afe87ee 100644
--- a/yql/essentials/core/yql_expr_type_annotation.cpp
+++ b/yql/essentials/core/yql_expr_type_annotation.cpp
@@ -393,6 +393,25 @@ IGraphTransformer::TStatus TryConvertToImpl(TExprContext& ctx, TExprNode::TPtr&
                 .Build();
 
             return IGraphTransformer::TStatus::Repeat;
+        } else if (IsDataTypeDecimal(from) && IsDataTypeDecimal(to)) {
+            auto* sourceDecimal = sourceType.Cast<TDataExprParamsType>();
+            auto* expectedDecimal = expectedType.Cast<TDataExprParamsType>();
+            ui8 p1 = FromString(sourceDecimal->GetParamOne()), s1 = FromString(sourceDecimal->GetParamTwo());
+            ui8 p2 = FromString(expectedDecimal->GetParamOne()), s2 = FromString(expectedDecimal->GetParamTwo());
+            if (s1 > s2) {
+                TString message = TStringBuilder() << "Implicit decimal cast would lose precision";
+                auto issue = TIssue(node->Pos(ctx), message);
+                ctx.AddError(issue);
+                return IGraphTransformer::TStatus::Error;
+            }
+            if (p1 - s1 > p2 - s2) {
+                TString message = TStringBuilder() << "Implicit decimal cast would narrow the range";
+                auto issue = TIssue(node->Pos(ctx), message);
+                ctx.AddError(issue);
+                return IGraphTransformer::TStatus::Error;
+            }
+            allow = true;
+            useCast = true;
         }
 
         if (!allow || !isSafe) {
diff --git a/yql/essentials/minikql/aligned_page_pool.cpp b/yql/essentials/minikql/aligned_page_pool.cpp
index ec9199bc272c..3e69f83c9945 100644
--- a/yql/essentials/minikql/aligned_page_pool.cpp
+++ b/yql/essentials/minikql/aligned_page_pool.cpp
@@ -1,7 +1,7 @@
 #include "aligned_page_pool.h"
-#include "util/string/builder.h"
 
 #include <util/generic/yexception.h>
+#include <util/stream/file.h>
 #include <util/string/cast.h>
 #include <util/system/align.h>
 #include <util/system/compiler.h>
@@ -26,6 +26,30 @@ static_assert(MaxMidSize == 64 * 1024 * 1024, "Upper memory block 64 Mb");
 
 namespace {
 
+size_t GetMemoryMapsCount() {
+    size_t lineCount = 0;
+    TString line;
+#if defined(_unix_)
+    TFileInput file("/proc/self/maps");
+    while (file.ReadLine(line)) ++lineCount;
+#endif
+    return lineCount;
+}
+
+ui64 GetMaxMemoryMaps() {
+    ui64 maxMapCount = 0;
+#if defined(_unix_)
+    maxMapCount = FromString<ui64>(TFileInput("/proc/sys/vm/max_map_count").ReadAll());
+#endif
+    return maxMapCount;
+}
+
+TString GetMemoryMapsString() {
+    TStringStream ss;
+    ss << " (maps: " << GetMemoryMapsCount() << " vs " << GetMaxMemoryMaps() << ")";
+    return ss.Str();
+}
+
 template<typename T, bool SysAlign>
 class TGlobalPools;
 
@@ -146,9 +170,15 @@ class TGlobalPools {
 
     void DoMunmap(void* addr, size_t size) {
         if (Y_UNLIKELY(0 != T::Munmap(addr, size))) {
+            TStringStream mmaps;
+            const auto lastError = LastSystemError();
+            if (lastError == ENOMEM) {
+                mmaps << GetMemoryMapsString();
+            }
+
             ythrow yexception() << "Munmap(0x"
                 << IntToString<16>(reinterpret_cast<uintptr_t>(addr))
-                << ", " << size << ") failed: " << LastSystemErrorText();
+                << ", " << size << ") failed: " << LastSystemErrorText(lastError) << mmaps.Str();
         }
 
         i64 prev = TotalMmappedBytes.fetch_sub(size);
@@ -493,7 +523,14 @@ void* TAlignedPagePoolImpl<T>::Alloc(size_t size) {
         auto allocSize = size + ALLOC_AHEAD_PAGES * POOL_PAGE_SIZE;
         void* mem = globalPool.DoMmap(allocSize);
         if (Y_UNLIKELY(MAP_FAILED == mem)) {
-            ythrow yexception() << "Mmap failed to allocate " << (size + POOL_PAGE_SIZE) << " bytes: " << LastSystemErrorText();
+            TStringStream mmaps;
+            const auto lastError = LastSystemError();
+            if (lastError == ENOMEM) {
+                mmaps << GetMemoryMapsString();
+            }
+
+            ythrow yexception() << "Mmap failed to allocate " << (size + POOL_PAGE_SIZE) << " bytes: "
+                << LastSystemErrorText(lastError) << mmaps.Str();
         }
 
         res = AlignUp(mem, POOL_PAGE_SIZE);
@@ -641,7 +678,13 @@ void* GetAlignedPage(ui64 size) {
     auto allocSize = Max<ui64>(MaxMidSize, size);
     void* mem = pool.DoMmap(allocSize);
     if (Y_UNLIKELY(MAP_FAILED == mem)) {
-        ythrow yexception() << "Mmap failed to allocate " << allocSize << " bytes: " << LastSystemErrorText();
+        TStringStream mmaps;
+        const auto lastError = LastSystemError();
+        if (lastError == ENOMEM) {
+            mmaps << GetMemoryMapsString();
+        }
+
+        ythrow yexception() << "Mmap failed to allocate " << allocSize << " bytes: " << LastSystemErrorText(lastError) << mmaps.Str();
     }
 
     if (size < MaxMidSize) {
diff --git a/yql/essentials/minikql/comp_nodes/mkql_append.cpp b/yql/essentials/minikql/comp_nodes/mkql_append.cpp
index 41436b098555..82894c6d108e 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_append.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_append.cpp
@@ -53,44 +53,20 @@ class TAppendWrapper : public TMutableCodegeneratorNode<TAppendWrapper<IsVoid>>
             BranchInst::Create(work, done, check, block);
             block = work;
 
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                const auto funType = FunctionType::get(left->getType(), {factory->getType(), left->getType(), right->getType()}, false);
-                const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-                const auto res = CallInst::Create(funType, funcPtr, {factory, left, right}, "res", block);
-                result->addIncoming(res, block);
-            } else {
-                const auto retPtr = new AllocaInst(left->getType(), 0U, "ret_ptr", block);
-                const auto itemPtr = new AllocaInst(right->getType(), 0U, "item_ptr", block);
-                new StoreInst(left, retPtr, block);
-                new StoreInst(right, itemPtr, block);
-                const auto funType = FunctionType::get(Type::getVoidTy(context), {factory->getType(), retPtr->getType(), retPtr->getType(), itemPtr->getType()}, false);
-                const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-                CallInst::Create(funType, funcPtr, {factory, retPtr, retPtr, itemPtr}, "", block);
-                const auto res = new LoadInst(left->getType(), retPtr, "res", block);
-                result->addIncoming(res, block);
-            }
+            const auto funType = FunctionType::get(left->getType(), {factory->getType(), left->getType(), right->getType()}, false);
+            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+            const auto res = CallInst::Create(funType, funcPtr, {factory, left, right}, "res", block);
+            result->addIncoming(res, block);
 
             BranchInst::Create(done, block);
 
             block = done;
             return result;
         } else {
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                const auto funType = FunctionType::get(left->getType(), {factory->getType(), left->getType(), right->getType()}, false);
-                const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-                const auto res = CallInst::Create(funType, funcPtr, {factory, left, right}, "res", block);
-                return res;
-            } else {
-                const auto retPtr = new AllocaInst(left->getType(), 0U, "ret_ptr", block);
-                const auto itemPtr = new AllocaInst(right->getType(), 0U, "item_ptr", block);
-                new StoreInst(left, retPtr, block);
-                new StoreInst(right, itemPtr, block);
-                const auto funType = FunctionType::get(Type::getVoidTy(context), {factory->getType(), retPtr->getType(), retPtr->getType(), itemPtr->getType()}, false);
-                const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-                CallInst::Create(funType, funcPtr, {factory, retPtr, retPtr, itemPtr}, "", block);
-                const auto res = new LoadInst(left->getType(), retPtr, "res", block);
-                return res;
-            }
+            const auto funType = FunctionType::get(left->getType(), {factory->getType(), left->getType(), right->getType()}, false);
+            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+            const auto res = CallInst::Create(funType, funcPtr, {factory, left, right}, "res", block);
+            return res;
         }
     }
 #endif
diff --git a/yql/essentials/minikql/comp_nodes/mkql_block_compress.cpp b/yql/essentials/minikql/comp_nodes/mkql_block_compress.cpp
index 3db1e34a4c8d..74e0dca0d1d4 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_block_compress.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_block_compress.cpp
@@ -51,15 +51,12 @@ using TBaseComputation = TStatefulWideFlowCodegeneratorNode<TCompressWithScalarB
         auto& context = ctx.Codegen.GetContext();
 
         const auto valueType = Type::getInt128Ty(context);
-        const auto ptrValueType = PointerType::getUnqual(valueType);
         const auto statusType = Type::getInt32Ty(context);
         const auto bitmapType = Type::getInt8Ty(context);
 
         const auto name = "GetBitmapScalarValue";
         ctx.Codegen.AddGlobalMapping(name, reinterpret_cast<const void*>(&GetBitmapScalarValue));
-        const auto getBitmapType = NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget() ?
-            FunctionType::get(bitmapType, { valueType }, false):
-            FunctionType::get(bitmapType, { ptrValueType }, false);
+        const auto getBitmapType = FunctionType::get(bitmapType, { valueType }, false);
         const auto getBitmap = ctx.Codegen.GetModule().getOrInsertFunction(name, getBitmapType);
 
         const auto work = BasicBlock::Create(context, "work", ctx.Func);
@@ -83,7 +80,7 @@ using TBaseComputation = TStatefulWideFlowCodegeneratorNode<TCompressWithScalarB
         block = test;
 
         const auto bitmapValue = getres.second[BitmapIndex_](ctx, block);
-        const auto bitmap = CallInst::Create(getBitmap, { WrapArgumentForWindows(bitmapValue, ctx, block) }, "bitmap", block);
+        const auto bitmap = CallInst::Create(getBitmap, { bitmapValue }, "bitmap", block);
 
         ValueCleanup(EValueRepresentation::Any, bitmapValue, ctx, block);
 
@@ -156,7 +153,6 @@ using TBaseComputation = TStatelessWideFlowCodegeneratorNode<TCompressScalars>;
         auto& context = ctx.Codegen.GetContext();
 
         const auto valueType = Type::getInt128Ty(context);
-        const auto ptrValueType = PointerType::getUnqual(valueType);
         const auto statusType = Type::getInt32Ty(context);
         const auto sizeType = Type::getInt64Ty(context);
 
@@ -167,9 +163,7 @@ using TBaseComputation = TStatelessWideFlowCodegeneratorNode<TCompressScalars>;
 
         const auto name = "GetBitmapPopCountCount";
         ctx.Codegen.AddGlobalMapping(name, reinterpret_cast<const void*>(&GetBitmapPopCountCount));
-        const auto getPopCountType = NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget() ?
-            FunctionType::get(sizeType, { valueType }, false):
-            FunctionType::get(sizeType, { ptrValueType }, false);
+        const auto getPopCountType = FunctionType::get(sizeType, { valueType }, false);
         const auto getPopCount = ctx.Codegen.GetModule().getOrInsertFunction(name, getPopCountType);
 
         const auto loop = BasicBlock::Create(context, "loop", ctx.Func);
@@ -191,7 +185,7 @@ using TBaseComputation = TStatelessWideFlowCodegeneratorNode<TCompressScalars>;
         block = work;
 
         const auto bitmapValue = getres.second[BitmapIndex_](ctx, block);
-        const auto pops = CallInst::Create(getPopCount, { WrapArgumentForWindows(bitmapValue, ctx, block) }, "pops", block);
+        const auto pops = CallInst::Create(getPopCount, { bitmapValue }, "pops", block);
 
         ValueCleanup(EValueRepresentation::Any, bitmapValue, ctx, block);
 
@@ -413,7 +407,7 @@ using TBaseComputation = TStatefulWideFlowCodegeneratorNode<TCompressBlocks>;
         block = good;
 
         const auto bitmap = getres.second[BitmapIndex_](ctx, block);
-        const auto bitmapArg = WrapArgumentForWindows(bitmap, ctx, block);
+        const auto bitmapArg = bitmap;
 
         const auto stepType = Type::getInt8Ty(context);
         const auto checkFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TState::Check));
diff --git a/yql/essentials/minikql/comp_nodes/mkql_block_skiptake.cpp b/yql/essentials/minikql/comp_nodes/mkql_block_skiptake.cpp
index f5671b2635f6..2496b75ca60c 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_block_skiptake.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_block_skiptake.cpp
@@ -51,7 +51,6 @@ using TBaseComputation = TStatefulWideFlowCodegeneratorNode<TWideSkipBlocksWrapp
 
         const auto indexType = Type::getInt64Ty(context);
         const auto valueType = Type::getInt128Ty(context);
-        const auto ptrValueType = PointerType::getUnqual(valueType);
 
         const auto atTop = &ctx.Func->getEntryBlock().back();
 
@@ -64,9 +63,7 @@ using TBaseComputation = TStatefulWideFlowCodegeneratorNode<TWideSkipBlocksWrapp
 
         const auto name = "GetBlockCount";
         ctx.Codegen.AddGlobalMapping(name, reinterpret_cast<const void*>(&GetBlockCount));
-        const auto getCountType = NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget() ?
-            FunctionType::get(indexType, { valueType }, false):
-            FunctionType::get(indexType, { ptrValueType }, false);
+        const auto getCountType = FunctionType::get(indexType, { valueType }, false);
         const auto getCount = ctx.Codegen.GetModule().getOrInsertFunction(name, getCountType);
 
         const auto init = BasicBlock::Create(context, "init", ctx.Func);
@@ -119,7 +116,7 @@ using TBaseComputation = TStatefulWideFlowCodegeneratorNode<TWideSkipBlocksWrapp
         block = test;
 
         const auto countValue = getres.second.back()(ctx, block);
-        const auto height = CallInst::Create(getCount, { WrapArgumentForWindows(countValue, ctx, block) }, "height", block);
+        const auto height = CallInst::Create(getCount, { countValue }, "height", block);
 
         ValueCleanup(EValueRepresentation::Any, countValue, ctx, block);
 
@@ -272,7 +269,6 @@ using TBaseComputation = TStatefulWideFlowCodegeneratorNode<TWideTakeBlocksWrapp
 
         const auto indexType = Type::getInt64Ty(context);
         const auto valueType = Type::getInt128Ty(context);
-        const auto ptrValueType = PointerType::getUnqual(valueType);
 
         const auto atTop = &ctx.Func->getEntryBlock().back();
 
@@ -285,9 +281,7 @@ using TBaseComputation = TStatefulWideFlowCodegeneratorNode<TWideTakeBlocksWrapp
 
         const auto name = "GetBlockCount";
         ctx.Codegen.AddGlobalMapping(name, reinterpret_cast<const void*>(&GetBlockCount));
-        const auto getCountType = NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget() ?
-            FunctionType::get(indexType, { valueType }, false):
-            FunctionType::get(indexType, { ptrValueType }, false);
+        const auto getCountType = FunctionType::get(indexType, { valueType }, false);
         const auto getCount = ctx.Codegen.GetModule().getOrInsertFunction(name, getCountType);
 
         const auto init = BasicBlock::Create(context, "init", ctx.Func);
@@ -332,7 +326,7 @@ using TBaseComputation = TStatefulWideFlowCodegeneratorNode<TWideTakeBlocksWrapp
         block = good;
 
         const auto countValue = getres.second.back()(ctx, block);
-        const auto height = CallInst::Create(getCount, { WrapArgumentForWindows(countValue, ctx, block) }, "height", block);
+        const auto height = CallInst::Create(getCount, { countValue }, "height", block);
 
         ValueCleanup(EValueRepresentation::Any, countValue, ctx, block);
 
diff --git a/yql/essentials/minikql/comp_nodes/mkql_blocks.cpp b/yql/essentials/minikql/comp_nodes/mkql_blocks.cpp
index f269264fa23f..5947cb7a161b 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_blocks.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_blocks.cpp
@@ -557,7 +557,6 @@ using TBaseComputation = TStatefulWideFlowCodegeneratorNode<TWideFromBlocksWrapp
 
         const auto width = Types_.size();
         const auto valueType = Type::getInt128Ty(context);
-        const auto ptrValueType = PointerType::getUnqual(valueType);
         const auto statusType = Type::getInt32Ty(context);
         const auto indexType = Type::getInt64Ty(context);
         const auto arrayType = ArrayType::get(valueType, width);
@@ -575,9 +574,7 @@ using TBaseComputation = TStatefulWideFlowCodegeneratorNode<TWideFromBlocksWrapp
 
         const auto name = "GetBlockCount";
         ctx.Codegen.AddGlobalMapping(name, reinterpret_cast<const void*>(&GetBlockCount));
-        const auto getCountType = NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget() ?
-            FunctionType::get(indexType, { valueType }, false):
-            FunctionType::get(indexType, { ptrValueType }, false);
+        const auto getCountType = FunctionType::get(indexType, { valueType }, false);
         const auto getCount = ctx.Codegen.GetModule().getOrInsertFunction(name, getCountType);
 
         const auto make = BasicBlock::Create(context, "make", ctx.Func);
@@ -633,7 +630,7 @@ using TBaseComputation = TStatefulWideFlowCodegeneratorNode<TWideFromBlocksWrapp
         block = good;
 
         const auto countValue = getres.second.back()(ctx, block);
-        const auto height = CallInst::Create(getCount, { WrapArgumentForWindows(countValue, ctx, block) }, "height", block);
+        const auto height = CallInst::Create(getCount, { countValue }, "height", block);
 
         ValueCleanup(EValueRepresentation::Any, countValue, ctx, block);
 
@@ -867,18 +864,9 @@ using TBaseComputation = TMutableCodegeneratorNode<TAsScalarWrapper>;
         const auto self = CastInst::Create(Instruction::IntToPtr, ConstantInt::get(Type::getInt64Ty(context), uintptr_t(this)), ptrType, "self", block);
         const auto asScalarFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TAsScalarWrapper::AsScalar));
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto asScalarType = FunctionType::get(Type::getInt128Ty(context), {self->getType(), value->getType(), ctx.Ctx->getType()}, false);
-            const auto asScalarFuncPtr = CastInst::Create(Instruction::IntToPtr, asScalarFunc, PointerType::getUnqual(asScalarType), "function", block);
-            return CallInst::Create(asScalarType, asScalarFuncPtr, {self, value, ctx.Ctx}, "scalar", block);
-        } else {
-            const auto valuePtr = new AllocaInst(value->getType(), 0U, "value", block);
-            new StoreInst(value, valuePtr, block);
-            const auto asScalarType = FunctionType::get(Type::getVoidTy(context), {self->getType(), valuePtr->getType(), valuePtr->getType(), ctx.Ctx->getType()}, false);
-            const auto asScalarFuncPtr = CastInst::Create(Instruction::IntToPtr, asScalarFunc, PointerType::getUnqual(asScalarType), "function", block);
-            CallInst::Create(asScalarType, asScalarFuncPtr, {self, valuePtr, valuePtr, ctx.Ctx}, "", block);
-            return new LoadInst(value->getType(), valuePtr, "result", block);
-        }
+        const auto asScalarType = FunctionType::get(Type::getInt128Ty(context), {self->getType(), value->getType(), ctx.Ctx->getType()}, false);
+        const auto asScalarFuncPtr = CastInst::Create(Instruction::IntToPtr, asScalarFunc, PointerType::getUnqual(asScalarType), "function", block);
+        return CallInst::Create(asScalarType, asScalarFuncPtr, {self, value, ctx.Ctx}, "scalar", block);
     }
 #endif
 private:
@@ -932,20 +920,9 @@ using TBaseComputation = TMutableCodegeneratorNode<TReplicateScalarWrapper>;
         const auto self = CastInst::Create(Instruction::IntToPtr, ConstantInt::get(Type::getInt64Ty(context), uintptr_t(this)), ptrType, "self", block);
         const auto replicateFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TReplicateScalarWrapper::Replicate));
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto replicateType = FunctionType::get(Type::getInt128Ty(context), {self->getType(), value->getType(), count->getType(), ctx.Ctx->getType()}, false);
-            const auto replicateFuncPtr = CastInst::Create(Instruction::IntToPtr, replicateFunc, PointerType::getUnqual(replicateType), "function", block);
-            return CallInst::Create(replicateType, replicateFuncPtr, {self, value, count, ctx.Ctx}, "replicate", block);
-        } else {
-            const auto valuePtr = new AllocaInst(value->getType(), 0U, "value", block);
-            const auto countPtr = new AllocaInst(count->getType(), 0U, "count", block);
-            new StoreInst(value, valuePtr, block);
-            new StoreInst(count, countPtr, block);
-            const auto replicateType = FunctionType::get(Type::getVoidTy(context), {self->getType(), valuePtr->getType(), valuePtr->getType(), countPtr->getType(), ctx.Ctx->getType()}, false);
-            const auto replicateFuncPtr = CastInst::Create(Instruction::IntToPtr, replicateFunc, PointerType::getUnqual(replicateType), "function", block);
-            CallInst::Create(replicateType, replicateFuncPtr, {self, valuePtr, valuePtr, countPtr, ctx.Ctx}, "", block);
-            return new LoadInst(value->getType(), valuePtr, "result", block);
-        }
+        const auto replicateType = FunctionType::get(Type::getInt128Ty(context), {self->getType(), value->getType(), count->getType(), ctx.Ctx->getType()}, false);
+        const auto replicateFuncPtr = CastInst::Create(Instruction::IntToPtr, replicateFunc, PointerType::getUnqual(replicateType), "function", block);
+        return CallInst::Create(replicateType, replicateFuncPtr, {self, value, count, ctx.Ctx}, "replicate", block);
     }
 #endif
 private:
diff --git a/yql/essentials/minikql/comp_nodes/mkql_callable.cpp b/yql/essentials/minikql/comp_nodes/mkql_callable.cpp
index 4f49d4e4fef5..76f703137f0a 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_callable.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_callable.cpp
@@ -104,24 +104,16 @@ class TCallableWrapper : public TCustomValueCodegeneratorNode<TCallableWrapper>
         const auto argsType = ArrayType::get(valueType, ArgNodes.size());
         const auto contextType = GetCompContextType(context);
 
-        const auto funcType = codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows ?
-            FunctionType::get(valueType, {PointerType::getUnqual(contextType), PointerType::getUnqual(argsType)}, false):
-            FunctionType::get(Type::getVoidTy(context), {PointerType::getUnqual(valueType), PointerType::getUnqual(contextType), PointerType::getUnqual(argsType)}, false);
+        const auto funcType =
+            FunctionType::get(valueType, {PointerType::getUnqual(contextType), PointerType::getUnqual(argsType)}, false);
 
         TCodegenContext ctx(codegen);
         ctx.Func = cast<Function>(module.getOrInsertFunction(name.c_str(), funcType).getCallee());
 
         DISubprogramAnnotator annotator(ctx, ctx.Func);
-        
 
         auto args = ctx.Func->arg_begin();
 
-        const auto resultArg = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? &*args++ : nullptr;
-        if (resultArg) {
-            resultArg->addAttr(Attribute::StructRet);
-            resultArg->addAttr(Attribute::NoAlias);
-        }
-
         ctx.Ctx = &*args;
         const auto argsPtr = &*++args;
 
@@ -140,12 +132,7 @@ class TCallableWrapper : public TCustomValueCodegeneratorNode<TCallableWrapper>
 
         const auto result = GetNodeValue(ResultNode, ctx, block);
 
-        if (resultArg) {
-            new StoreInst(result, resultArg, block);
-            ReturnInst::Create(context, block);
-        } else {
-            ReturnInst::Create(context, result, block);
-        }
+        ReturnInst::Create(context, result, block);
         return ctx.Func;
     }
 
diff --git a/yql/essentials/minikql/comp_nodes/mkql_chain1_map.cpp b/yql/essentials/minikql/comp_nodes/mkql_chain1_map.cpp
index c12d9917c594..56c4b8c9acf6 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_chain1_map.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_chain1_map.cpp
@@ -238,7 +238,7 @@ class TBaseChain1MapWrapper {
             return f;
 
         const auto valueType = Type::getInt128Ty(context);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = IsStream ? Type::getInt32Ty(context) : Type::getInt1Ty(context);
         const auto funcType = FunctionType::get(statusType, {PointerType::getUnqual(contextType), containerType, PointerType::getUnqual(valueType)}, false);
@@ -257,8 +257,7 @@ class TBaseChain1MapWrapper {
         const auto main = BasicBlock::Create(context, "main", ctx.Func);
         auto block = main;
 
-        const auto container = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, containerArg, "load_container", false, block) : static_cast<Value*>(containerArg);
+        const auto container = static_cast<Value*>(containerArg);
 
         const auto good = BasicBlock::Create(context, "good", ctx.Func);
         const auto done = BasicBlock::Create(context, "done", ctx.Func);
@@ -473,20 +472,10 @@ class TListChain1MapWrapper : public TBothWaysCodegeneratorNode<TListChain1MapWr
             const auto doFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TListChain1MapWrapper::MakeLazyList));
             const auto ptrType = PointerType::getUnqual(StructType::get(context));
             const auto self = CastInst::Create(Instruction::IntToPtr, ConstantInt::get(Type::getInt64Ty(context), uintptr_t(this)), ptrType, "self", block);
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                const auto funType = FunctionType::get(list->getType() , {self->getType(), ctx.Ctx->getType(), list->getType()}, false);
-                const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-                const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx, list}, "value", block);
-                map->addIncoming(value, block);
-            } else {
-                const auto resultPtr = new AllocaInst(list->getType(), 0U, "return", block);
-                new StoreInst(list, resultPtr, block);
-                const auto funType = FunctionType::get(Type::getVoidTy(context), {self->getType(), resultPtr->getType(), ctx.Ctx->getType(), resultPtr->getType()}, false);
-                const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-                CallInst::Create(funType, doFuncPtr, {self, resultPtr, ctx.Ctx, resultPtr}, "", block);
-                const auto value = new LoadInst(list->getType(), resultPtr, "value", block);
-                map->addIncoming(value, block);
-            }
+            const auto funType = FunctionType::get(list->getType() , {self->getType(), ctx.Ctx->getType(), list->getType()}, false);
+            const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
+            const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx, list}, "value", block);
+            map->addIncoming(value, block);
             BranchInst::Create(done, block);
         }
 
diff --git a/yql/essentials/minikql/comp_nodes/mkql_chain_map.cpp b/yql/essentials/minikql/comp_nodes/mkql_chain_map.cpp
index e0d9ff1f1303..000387838003 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_chain_map.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_chain_map.cpp
@@ -230,7 +230,7 @@ class TBaseChainMapWrapper {
             return f;
 
         const auto valueType = Type::getInt128Ty(context);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = IsStream ? Type::getInt32Ty(context) : Type::getInt1Ty(context);
         const auto funcType = FunctionType::get(statusType, {PointerType::getUnqual(contextType), containerType, PointerType::getUnqual(valueType), PointerType::getUnqual(valueType)}, false);
@@ -250,8 +250,7 @@ class TBaseChainMapWrapper {
         const auto main = BasicBlock::Create(context, "main", ctx.Func);
         auto block = main;
 
-        const auto container = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, containerArg, "load_container", false, block) : static_cast<Value*>(containerArg);
+        const auto container = static_cast<Value*>(containerArg);
 
         const auto init = BasicBlock::Create(context, "init", ctx.Func);
         const auto work = BasicBlock::Create(context, "work", ctx.Func);
@@ -456,22 +455,10 @@ class TListChainMapWrapper : public TBothWaysCodegeneratorNode<TListChainMapWrap
             const auto doFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TListChainMapWrapper::MakeLazyList));
             const auto ptrType = PointerType::getUnqual(StructType::get(context));
             const auto self = CastInst::Create(Instruction::IntToPtr, ConstantInt::get(Type::getInt64Ty(context), uintptr_t(this)), ptrType, "self", block);
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                const auto funType = FunctionType::get(list->getType() , {self->getType(), ctx.Ctx->getType(), list->getType(), init->getType()}, false);
-                const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-                const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx, list, init}, "value", block);
-                map->addIncoming(value, block);
-            } else {
-                const auto resultPtr = new AllocaInst(list->getType(), 0U, "return", block);
-                const auto initPtr = new AllocaInst(init->getType(), 0U, "init", block);
-                new StoreInst(list, resultPtr, block);
-                new StoreInst(init, initPtr, block);
-                const auto funType = FunctionType::get(Type::getVoidTy(context), {self->getType(), resultPtr->getType(), ctx.Ctx->getType(), resultPtr->getType(), initPtr->getType()}, false);
-                const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-                CallInst::Create(funType, doFuncPtr, {self, resultPtr, ctx.Ctx, resultPtr, initPtr}, "", block);
-                const auto value = new LoadInst(list->getType(), resultPtr, "value", block);
-                map->addIncoming(value, block);
-            }
+            const auto funType = FunctionType::get(list->getType() , {self->getType(), ctx.Ctx->getType(), list->getType(), init->getType()}, false);
+            const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
+            const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx, list, init}, "value", block);
+            map->addIncoming(value, block);
             BranchInst::Create(done, block);
         }
 
diff --git a/yql/essentials/minikql/comp_nodes/mkql_chopper.cpp b/yql/essentials/minikql/comp_nodes/mkql_chopper.cpp
index f3b5aecf0bc3..eb93380e182e 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_chopper.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_chopper.cpp
@@ -554,7 +554,7 @@ class TChopperWrapper : public TCustomValueCodegeneratorNode<TChopperWrapper> {
         MKQL_ENSURE(codegenKeyArg, "Key arg must be codegenerator node.");
 
         const auto valueType = Type::getInt128Ty(context);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = Type::getInt32Ty(context);
         const auto stateType = Type::getInt8Ty(context);
@@ -578,8 +578,7 @@ class TChopperWrapper : public TCustomValueCodegeneratorNode<TChopperWrapper> {
 
         auto block = main;
 
-        const auto container = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, containerArg, "load_container", false, block) : static_cast<Value*>(containerArg);
+        const auto container = static_cast<Value*>(containerArg);
 
         const auto first = new LoadInst(stateType, stateArg, "first", block);
         const auto reload = CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ, first, ConstantInt::get(stateType, ui8(EState::Next)), "reload", block);
@@ -650,7 +649,7 @@ class TChopperWrapper : public TCustomValueCodegeneratorNode<TChopperWrapper> {
         MKQL_ENSURE(codegenInput, "Input arg must be codegenerator node.");
 
         const auto valueType = Type::getInt128Ty(context);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = Type::getInt32Ty(context);
         const auto stateType = Type::getInt8Ty(context);
@@ -680,8 +679,7 @@ class TChopperWrapper : public TCustomValueCodegeneratorNode<TChopperWrapper> {
 
         auto block = main;
 
-        const auto input = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, inputArg, "load_input", false, block) : static_cast<Value*>(inputArg);
+        const auto input = static_cast<Value*>(inputArg);
 
         BranchInst::Create(loop, block);
 
diff --git a/yql/essentials/minikql/comp_nodes/mkql_collect.cpp b/yql/essentials/minikql/comp_nodes/mkql_collect.cpp
index 836bb2726055..ad23aeb44991 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_collect.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_collect.cpp
@@ -42,19 +42,10 @@ using TBaseComputation = TMutableCodegeneratorRootNode<TCollectFlowWrapper>;
 
         const auto list = PHINode::Create(valueType, 2U, "list", work);
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(valueType, {factory->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, empty, PointerType::getUnqual(funType), "empty", block);
-            const auto first = CallInst::Create(funType, funcPtr, {factory}, "init", block);
-            list->addIncoming(first, block);
-        } else {
-            const auto ptr = new AllocaInst(valueType, 0U, "ptr", block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {factory->getType(), ptr->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, empty, PointerType::getUnqual(funType), "empty", block);
-            CallInst::Create(funType, funcPtr, {factory, ptr}, "", block);
-            const auto first = new LoadInst(valueType, ptr, "init", block);
-            list->addIncoming(first, block);
-        }
+        const auto funType = FunctionType::get(valueType, {factory->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, empty, PointerType::getUnqual(funType), "empty", block);
+        const auto first = CallInst::Create(funType, funcPtr, {factory}, "init", block);
+        list->addIncoming(first, block);
 
         BranchInst::Create(work, block);
 
@@ -69,22 +60,10 @@ using TBaseComputation = TMutableCodegeneratorRootNode<TCollectFlowWrapper>;
         {
             block = good;
 
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                const auto funType = FunctionType::get(valueType, {factory->getType(), list->getType(), item->getType()}, false);
-                const auto funcPtr = CastInst::Create(Instruction::IntToPtr, append, PointerType::getUnqual(funType), "append", block);
-                const auto next = CallInst::Create(funType, funcPtr, {factory, list, item}, "next", block);
-                list->addIncoming(next, block);
-            } else {
-                const auto retPtr = new AllocaInst(list->getType(), 0U, "ret_ptr", block);
-                const auto itemPtr = new AllocaInst(item->getType(), 0U, "item_ptr", block);
-                new StoreInst(list, retPtr, block);
-                new StoreInst(item, itemPtr, block);
-                const auto funType = FunctionType::get(Type::getVoidTy(context), {factory->getType(), retPtr->getType(), retPtr->getType(), itemPtr->getType()}, false);
-                const auto funcPtr = CastInst::Create(Instruction::IntToPtr, append, PointerType::getUnqual(funType), "append", block);
-                CallInst::Create(funType, funcPtr, {factory, retPtr, retPtr, itemPtr}, "", block);
-                const auto next = new LoadInst(list->getType(), retPtr, "next", block);
-                list->addIncoming(next, block);
-            }
+            const auto funType = FunctionType::get(valueType, {factory->getType(), list->getType(), item->getType()}, false);
+            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, append, PointerType::getUnqual(funType), "append", block);
+            const auto next = CallInst::Create(funType, funcPtr, {factory, list, item}, "next", block);
+            list->addIncoming(next, block);
             BranchInst::Create(work, block);
         }
 
@@ -154,39 +133,19 @@ class TCollectWrapper : public TMutableCodegeneratorNode<TCollectWrapper<IsList>
             BranchInst::Create(work, done, null, block);
 
             block = work;
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                const auto funType = FunctionType::get(seq->getType(), {factory->getType(), seq->getType()}, false);
-                const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-                const auto res = CallInst::Create(funType, funcPtr, {factory, seq}, "res", block);
-                result->addIncoming(res, block);
-            } else {
-                const auto ptr = new AllocaInst(seq->getType(), 0U, "ptr", block);
-                new StoreInst(seq, ptr, block);
-                const auto funType = FunctionType::get(Type::getVoidTy(context), {factory->getType(), ptr->getType(), ptr->getType()}, false);
-                const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-                CallInst::Create(funType, funcPtr, {factory, ptr, ptr}, "", block);
-                const auto res = new LoadInst(seq->getType(), ptr, "res", block);
-                result->addIncoming(res, block);
-            }
+            const auto funType = FunctionType::get(seq->getType(), {factory->getType(), seq->getType()}, false);
+            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+            const auto res = CallInst::Create(funType, funcPtr, {factory, seq}, "res", block);
+            result->addIncoming(res, block);
             BranchInst::Create(done, block);
 
             block = done;
             return result;
         } else {
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                const auto funType = FunctionType::get(seq->getType(), {factory->getType(), seq->getType()}, false);
-                const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-                const auto res = CallInst::Create(funType, funcPtr, {factory, seq}, "res", block);
-                return res;
-            } else {
-                const auto ptr = new AllocaInst(seq->getType(), 0U, "ptr", block);
-                new StoreInst(seq, ptr, block);
-                const auto funType = FunctionType::get(Type::getVoidTy(context), {factory->getType(), ptr->getType(), ptr->getType()}, false);
-                const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-                CallInst::Create(funType, funcPtr, {factory, ptr, ptr}, "", block);
-                const auto res = new LoadInst(seq->getType(), ptr, "res", block);
-                return res;
-            }
+            const auto funType = FunctionType::get(seq->getType(), {factory->getType(), seq->getType()}, false);
+            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+            const auto res = CallInst::Create(funType, funcPtr, {factory, seq}, "res", block);
+            return res;
         }
     }
 #endif
diff --git a/yql/essentials/minikql/comp_nodes/mkql_combine.cpp b/yql/essentials/minikql/comp_nodes/mkql_combine.cpp
index dffe2e1c5aa4..cc618593bcbf 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_combine.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_combine.cpp
@@ -354,12 +354,7 @@ class TCombineCoreFlowWrapper: public std::conditional_t<IsMultiRowState,
             const auto key = GetNodeValue(Nodes.KeyResultNode, ctx, block);
             codegenKeyArg->CreateSetValue(ctx, block, key);
 
-            const auto keyParam = NYql::NCodegen::ETarget::Windows == ctx.Codegen.GetEffectiveTarget() ?
-                new AllocaInst(key->getType(), 0U, "key_param", &main->back()) : key;
-
-            if (NYql::NCodegen::ETarget::Windows == ctx.Codegen.GetEffectiveTarget()) {
-                new StoreInst(key, keyParam, block);
-            }
+            const auto keyParam = key;
 
             const auto atFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TState::At));
             const auto atType = FunctionType::get(ptrValueType, {stateArg->getType(), keyParam->getType()}, false);
@@ -722,7 +717,7 @@ class TCombineCoreWrapper: public TCustomValueCodegeneratorNode<TCombineCoreWrap
 
         const auto valueType = Type::getInt128Ty(context);
         const auto ptrValueType = PointerType::getUnqual(valueType);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(ptrValueType) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = Type::getInt32Ty(context);
 
@@ -830,8 +825,7 @@ class TCombineCoreWrapper: public TCustomValueCodegeneratorNode<TCombineCoreWrap
 
             const auto used = GetMemoryUsed(MemLimit, ctx, block);
 
-            const auto stream = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-                new LoadInst(valueType, containerArg, "load_container", false, block) : static_cast<Value*>(containerArg);
+            const auto stream = static_cast<Value*>(containerArg);
 
             BranchInst::Create(loop, block);
 
@@ -850,12 +844,7 @@ class TCombineCoreWrapper: public TCustomValueCodegeneratorNode<TCombineCoreWrap
             const auto key = GetNodeValue(Nodes.KeyResultNode, ctx, block);
             codegenKeyArg->CreateSetValue(ctx, block, key);
 
-            const auto keyParam = NYql::NCodegen::ETarget::Windows == ctx.Codegen.GetEffectiveTarget() ?
-                new AllocaInst(key->getType(), 0U, "key_param", &main->back()) : key;
-
-            if (NYql::NCodegen::ETarget::Windows == ctx.Codegen.GetEffectiveTarget()) {
-                new StoreInst(key, keyParam, block);
-            }
+            const auto keyParam = key;
 
             const auto atFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TState::At));
             const auto atType = FunctionType::get(ptrValueType, {stateArg->getType(), keyParam->getType()}, false);
diff --git a/yql/essentials/minikql/comp_nodes/mkql_condense.cpp b/yql/essentials/minikql/comp_nodes/mkql_condense.cpp
index f7c262327357..ff729dbbed65 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_condense.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_condense.cpp
@@ -352,7 +352,7 @@ class TCondenseWrapper : public TCustomValueCodegeneratorNode<TCondenseWrapper<I
             return f;
 
         const auto valueType = Type::getInt128Ty(context);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = Type::getInt32Ty(context);
         const auto stateType = Type::getInt8Ty(context);
@@ -373,8 +373,7 @@ class TCondenseWrapper : public TCustomValueCodegeneratorNode<TCondenseWrapper<I
         const auto main = BasicBlock::Create(context, "main", ctx.Func);
         auto block = main;
 
-        const auto container = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, containerArg, "load_container", false, block) : static_cast<Value*>(containerArg);
+        const auto container = static_cast<Value*>(containerArg);
 
         const auto state = new LoadInst(stateType, statePtr, "state", block);
 
diff --git a/yql/essentials/minikql/comp_nodes/mkql_condense1.cpp b/yql/essentials/minikql/comp_nodes/mkql_condense1.cpp
index 1cf679ed07fe..a573b480720a 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_condense1.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_condense1.cpp
@@ -370,7 +370,7 @@ class TCondense1Wrapper : public TCustomValueCodegeneratorNode<TCondense1Wrapper
             return f;
 
         const auto valueType = Type::getInt128Ty(context);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = Type::getInt32Ty(context);
         const auto stateType = Type::getInt8Ty(context);
@@ -391,8 +391,7 @@ class TCondense1Wrapper : public TCustomValueCodegeneratorNode<TCondense1Wrapper
         const auto main = BasicBlock::Create(context, "main", ctx.Func);
         auto block = main;
 
-        const auto container = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, containerArg, "load_container", false, block) : static_cast<Value*>(containerArg);
+        const auto container = static_cast<Value*>(containerArg);
 
         const auto step = BasicBlock::Create(context, "step", ctx.Func);
         const auto none = BasicBlock::Create(context, "none", ctx.Func);
diff --git a/yql/essentials/minikql/comp_nodes/mkql_decimal_div.cpp b/yql/essentials/minikql/comp_nodes/mkql_decimal_div.cpp
index 394ddbc4f515..be561df7e3fa 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_decimal_div.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_decimal_div.cpp
@@ -46,13 +46,11 @@ class TDecimalDivWrapper : public TMutableCodegeneratorNode<TDecimalDivWrapper<I
         auto& context = ctx.Codegen.GetContext();
 
         const auto valType = Type::getInt128Ty(context);
-        const auto valTypePtr = PointerType::getUnqual(valType);
 
         const auto name = "DecimalMulAndDivNormalMultiplier";
         ctx.Codegen.AddGlobalMapping(name, reinterpret_cast<const void*>(&DecimalMulAndDivNormalMultiplier));
-        const auto fnType = NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget() ?
-            FunctionType::get(valType, { valType, valType, valType }, false):
-            FunctionType::get(Type::getVoidTy(context), { valTypePtr, valTypePtr, valTypePtr, valTypePtr }, false);
+        const auto fnType =
+            FunctionType::get(valType, { valType, valType, valType }, false);
         const auto func = ctx.Codegen.GetModule().getOrInsertFunction(name, fnType);
 
         const auto left = GetNodeValue(Left, ctx, block);
@@ -73,20 +71,7 @@ class TDecimalDivWrapper : public TMutableCodegeneratorNode<TDecimalDivWrapper<I
 
             block = good;
 
-            Value* muldiv;
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                muldiv = CallInst::Create(func, { GetterForInt128(left, block), NDecimal::GenConstant(Divider, context), GetterForInt128(right, block) }, "mul_and_div", block);
-            } else {
-                const auto retPtr = new AllocaInst(valType, 0U, "ret_ptr", block);
-                const auto arg1Ptr = new AllocaInst(valType, 0U, "arg1", block);
-                const auto arg2Ptr = new AllocaInst(valType, 0U, "arg2", block);
-                const auto arg3Ptr = new AllocaInst(valType, 0U, "arg3", block);
-                new StoreInst(GetterForInt128(left, block), arg1Ptr, block);
-                new StoreInst(NDecimal::GenConstant(Divider, context), arg2Ptr, block);
-                new StoreInst(GetterForInt128(right, block), arg3Ptr, block);
-                CallInst::Create(func, { retPtr, arg1Ptr, arg2Ptr, arg3Ptr }, "", block);
-                muldiv = new LoadInst(valType, retPtr, "res", block);
-            }
+            const auto muldiv = CallInst::Create(func, { GetterForInt128(left, block), NDecimal::GenConstant(Divider, context), GetterForInt128(right, block) }, "mul_and_div", block);
 
             const auto ok = NDecimal::GenInBounds(muldiv, NDecimal::GenConstant(-Bound, context), NDecimal::GenConstant(+Bound, context), block);
             const auto nan = NDecimal::GenIsNonComparable(muldiv, context, block);
@@ -102,20 +87,7 @@ class TDecimalDivWrapper : public TMutableCodegeneratorNode<TDecimalDivWrapper<I
             block = done;
             return result;
         } else {
-            Value* muldiv;
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                muldiv = CallInst::Create(func, { GetterForInt128(left, block), NDecimal::GenConstant(Divider, context), GetterForInt128(right, block) }, "mul_and_div", block);
-            } else {
-                const auto retPtr = new AllocaInst(valType, 0U, "ret_ptr", block);
-                const auto arg1Ptr = new AllocaInst(valType, 0U, "arg1", block);
-                const auto arg2Ptr = new AllocaInst(valType, 0U, "arg2", block);
-                const auto arg3Ptr = new AllocaInst(valType, 0U, "arg3", block);
-                new StoreInst(GetterForInt128(left, block), arg1Ptr, block);
-                new StoreInst(NDecimal::GenConstant(Divider, context), arg2Ptr, block);
-                new StoreInst(GetterForInt128(right, block), arg3Ptr, block);
-                CallInst::Create(func, { retPtr, arg1Ptr, arg2Ptr, arg3Ptr }, "", block);
-                muldiv = new LoadInst(valType, retPtr, "res", block);
-            }
+            const auto muldiv = CallInst::Create(func, { GetterForInt128(left, block), NDecimal::GenConstant(Divider, context), GetterForInt128(right, block) }, "mul_and_div", block);
 
             const auto ok = NDecimal::GenInBounds(muldiv, NDecimal::GenConstant(-Bound, context), NDecimal::GenConstant(+Bound, context), block);
             const auto nan = NDecimal::GenIsNonComparable(muldiv, context, block);
@@ -168,13 +140,11 @@ class TDecimalDivIntegralWrapper : public TMutableCodegeneratorNode<TDecimalDivI
         auto& context = ctx.Codegen.GetContext();
 
         const auto valType = Type::getInt128Ty(context);
-        const auto valTypePtr = PointerType::getUnqual(valType);
 
         const auto name = "DecimalDiv";
         ctx.Codegen.AddGlobalMapping(name, reinterpret_cast<const void*>(&DecimalDiv));
-        const auto fnType = NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget() ?
-            FunctionType::get(valType, { valType, valType }, false):
-            FunctionType::get(Type::getVoidTy(context), { valTypePtr, valTypePtr, valTypePtr }, false);
+        const auto fnType =
+            FunctionType::get(valType, { valType, valType }, false);
         const auto func = ctx.Codegen.GetModule().getOrInsertFunction(name, fnType);
 
         const auto left = GetNodeValue(Left, ctx, block);
@@ -199,19 +169,7 @@ class TDecimalDivIntegralWrapper : public TMutableCodegeneratorNode<TDecimalDivI
                 static_cast<CastInst*>(new SExtInst(GetterFor<TRight>(right, context, block), valType, "sext", block)):
                 static_cast<CastInst*>(new ZExtInst(GetterFor<TRight>(right, context, block), valType, "zext", block));
 
-            Value* div;
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                div = CallInst::Create(func, {GetterForInt128(left, block), cast}, "div", block);
-            } else {
-                const auto retPtr = new AllocaInst(valType, 0U, "ret_ptr", block);
-                const auto arg1Ptr = new AllocaInst(valType, 0U, "arg1", block);
-                const auto arg2Ptr = new AllocaInst(valType, 0U, "arg2", block);
-                new StoreInst(GetterForInt128(left, block), arg1Ptr, block);
-                new StoreInst(cast, arg2Ptr, block);
-                CallInst::Create(func, { retPtr, arg1Ptr, arg2Ptr }, "", block);
-                div = new LoadInst(valType, retPtr, "res", block);
-            }
-
+            const auto div = CallInst::Create(func, {GetterForInt128(left, block), cast}, "div", block);
             result->addIncoming(SetterForInt128(div, block), block);
             BranchInst::Create(done, block);
 
@@ -221,18 +179,7 @@ class TDecimalDivIntegralWrapper : public TMutableCodegeneratorNode<TDecimalDivI
             const auto cast = std::is_signed<TRight>() ?
                 static_cast<CastInst*>(new SExtInst(GetterFor<TRight>(right, context, block), valType, "sext", block)):
                 static_cast<CastInst*>(new ZExtInst(GetterFor<TRight>(right, context, block), valType, "zext", block));
-            Value* div;
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                div = CallInst::Create(func, {GetterForInt128(left, block), cast}, "div", block);
-            } else {
-                const auto retPtr = new AllocaInst(valType, 0U, "ret_ptr", block);
-                const auto arg1Ptr = new AllocaInst(valType, 0U, "arg1", block);
-                const auto arg2Ptr = new AllocaInst(valType, 0U, "arg2", block);
-                new StoreInst(GetterForInt128(left, block), arg1Ptr, block);
-                new StoreInst(cast, arg2Ptr, block);
-                CallInst::Create(func, { retPtr, arg1Ptr, arg2Ptr }, "", block);
-                div = new LoadInst(valType, retPtr, "res", block);
-            }
+            const auto div = CallInst::Create(func, {GetterForInt128(left, block), cast}, "div", block);
             return SetterForInt128(div, block);
         }
     }
diff --git a/yql/essentials/minikql/comp_nodes/mkql_decimal_mul.cpp b/yql/essentials/minikql/comp_nodes/mkql_decimal_mul.cpp
index a92f8fa87913..80dd95b79e74 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_decimal_mul.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_decimal_mul.cpp
@@ -47,17 +47,12 @@ class TDecimalMulWrapper : public TMutableCodegeneratorNode<TDecimalMulWrapper<I
         auto& context = ctx.Codegen.GetContext();
 
         const auto valType = Type::getInt128Ty(context);
-        const auto valTypePtr = PointerType::getUnqual(valType);
 
         const bool useMulAddDiv = Divider > 1;
         const auto name = useMulAddDiv ? "DecimalMulAndDivNormalDivider" : "DecimalMul";
         const auto fnType = useMulAddDiv ?
-            NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget() ?
-                    FunctionType::get(valType, { valType, valType, valType }, false):
-                    FunctionType::get(Type::getVoidTy(context), { valTypePtr, valTypePtr, valTypePtr, valTypePtr }, false):
-            NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget() ?
-                FunctionType::get(valType, { valType, valType}, false):
-                FunctionType::get(Type::getVoidTy(context), { valTypePtr, valTypePtr, valTypePtr }, false);
+            FunctionType::get(valType, { valType, valType, valType }, false):
+            FunctionType::get(valType, { valType, valType}, false);
 
         ctx.Codegen.AddGlobalMapping(name,  useMulAddDiv ? reinterpret_cast<const void*>(&DecimalMulAndDivNormalDivider) : reinterpret_cast<const void*>(&DecimalMul));
         const auto func = ctx.Codegen.GetModule().getOrInsertFunction(name, fnType);
@@ -82,31 +77,9 @@ class TDecimalMulWrapper : public TMutableCodegeneratorNode<TDecimalMulWrapper<I
 
             Value* muldiv;
             if (useMulAddDiv) {
-                if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                    muldiv = CallInst::Create(func, { GetterForInt128(left, block), GetterForInt128(right, block), NDecimal::GenConstant(Divider, context) }, "mul_and_div", block);
-                } else {
-                    const auto retPtr = new AllocaInst(valType, 0U, "ret_ptr", block);
-                    const auto arg1Ptr = new AllocaInst(valType, 0U, "arg1", block);
-                    const auto arg2Ptr = new AllocaInst(valType, 0U, "arg2", block);
-                    const auto arg3Ptr = new AllocaInst(valType, 0U, "arg3", block);
-                    new StoreInst(GetterForInt128(left, block), arg1Ptr, block);
-                    new StoreInst(GetterForInt128(right, block), arg2Ptr, block);
-                    new StoreInst(NDecimal::GenConstant(Divider, context), arg3Ptr, block);
-                    CallInst::Create(func, { retPtr, arg1Ptr, arg2Ptr, arg3Ptr }, "", block);
-                    muldiv = new LoadInst(valType, retPtr, "res", block);
-                }
+                muldiv = CallInst::Create(func, { GetterForInt128(left, block), GetterForInt128(right, block), NDecimal::GenConstant(Divider, context) }, "mul_and_div", block);
             } else {
-                if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                    muldiv = CallInst::Create(func, { GetterForInt128(left, block), GetterForInt128(right, block) }, "mul", block);
-                } else {
-                    const auto retPtr = new AllocaInst(valType, 0U, "ret_ptr", block);
-                    const auto arg1Ptr = new AllocaInst(valType, 0U, "arg1", block);
-                    const auto arg2Ptr = new AllocaInst(valType, 0U, "arg2", block);
-                    new StoreInst(GetterForInt128(left, block), arg1Ptr, block);
-                    new StoreInst(GetterForInt128(right, block), arg2Ptr, block);
-                    CallInst::Create(func, { retPtr, arg1Ptr, arg2Ptr }, "", block);
-                    muldiv = new LoadInst(valType, retPtr, "res", block);
-                }
+                muldiv = CallInst::Create(func, { GetterForInt128(left, block), GetterForInt128(right, block) }, "mul", block);
             }
 
             const auto ok = NDecimal::GenInBounds(muldiv, NDecimal::GenConstant(-Bound, context), NDecimal::GenConstant(+Bound, context), block);
@@ -125,31 +98,9 @@ class TDecimalMulWrapper : public TMutableCodegeneratorNode<TDecimalMulWrapper<I
         } else {
             Value* muldiv;
             if (useMulAddDiv) {
-                if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                    muldiv = CallInst::Create(func, { GetterForInt128(left, block), GetterForInt128(right, block), NDecimal::GenConstant(Divider, context) }, "mul_and_div", block);
-                } else {
-                    const auto retPtr = new AllocaInst(valType, 0U, "ret_ptr", block);
-                    const auto arg1Ptr = new AllocaInst(valType, 0U, "arg1", block);
-                    const auto arg2Ptr = new AllocaInst(valType, 0U, "arg2", block);
-                    const auto arg3Ptr = new AllocaInst(valType, 0U, "arg3", block);
-                    new StoreInst(GetterForInt128(left, block), arg1Ptr, block);
-                    new StoreInst(GetterForInt128(right, block), arg2Ptr, block);
-                    new StoreInst(NDecimal::GenConstant(Divider, context), arg3Ptr, block);
-                    CallInst::Create(func, { retPtr, arg1Ptr, arg2Ptr, arg3Ptr }, "", block);
-                    muldiv = new LoadInst(valType, retPtr, "res", block);
-                }
+                muldiv = CallInst::Create(func, { GetterForInt128(left, block), GetterForInt128(right, block), NDecimal::GenConstant(Divider, context) }, "mul_and_div", block);
             } else {
-                if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                    muldiv = CallInst::Create(func, { GetterForInt128(left, block), GetterForInt128(right, block) }, "mul", block);
-                } else {
-                    const auto retPtr = new AllocaInst(valType, 0U, "ret_ptr", block);
-                    const auto arg1Ptr = new AllocaInst(valType, 0U, "arg1", block);
-                    const auto arg2Ptr = new AllocaInst(valType, 0U, "arg2", block);
-                    new StoreInst(GetterForInt128(left, block), arg1Ptr, block);
-                    new StoreInst(GetterForInt128(right, block), arg2Ptr, block);
-                    CallInst::Create(func, { retPtr, arg1Ptr, arg2Ptr }, "", block);
-                    muldiv = new LoadInst(valType, retPtr, "res", block);
-                }
+                muldiv = CallInst::Create(func, { GetterForInt128(left, block), GetterForInt128(right, block) }, "mul", block);
             }
 
             const auto ok = NDecimal::GenInBounds(muldiv, NDecimal::GenConstant(-Bound, context), NDecimal::GenConstant(+Bound, context), block);
@@ -206,13 +157,11 @@ class TDecimalMulIntegralWrapper : public TMutableCodegeneratorNode<TDecimalMulI
         auto& context = ctx.Codegen.GetContext();
 
         const auto valType = Type::getInt128Ty(context);
-        const auto valTypePtr = PointerType::getUnqual(valType);
 
         const auto name = "DecimalMul";
         ctx.Codegen.AddGlobalMapping(name, reinterpret_cast<const void*>(&DecimalMul));
-        const auto fnType = NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget() ?
-            FunctionType::get(valType, { valType, valType }, false):
-            FunctionType::get(Type::getVoidTy(context), { valTypePtr, valTypePtr, valTypePtr }, false);
+        const auto fnType =
+            FunctionType::get(valType, { valType, valType }, false);
         const auto func = ctx.Codegen.GetModule().getOrInsertFunction(name, fnType);
 
         const auto left = GetNodeValue(Left, ctx, block);
@@ -236,18 +185,7 @@ class TDecimalMulIntegralWrapper : public TMutableCodegeneratorNode<TDecimalMulI
             const auto cast = std::is_signed<TRight>() ?
                 static_cast<CastInst*>(new SExtInst(GetterFor<TRight>(right, context, block), valType, "sext", block)):
                 static_cast<CastInst*>(new ZExtInst(GetterFor<TRight>(right, context, block), valType, "zext", block));
-            Value* mul;
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                mul = CallInst::Create(func, {GetterForInt128(left, block), cast}, "div", block);
-            } else {
-                const auto retPtr = new AllocaInst(valType, 0U, "ret_ptr", block);
-                const auto arg1Ptr = new AllocaInst(valType, 0U, "arg1", block);
-                const auto arg2Ptr = new AllocaInst(valType, 0U, "arg2", block);
-                new StoreInst(GetterForInt128(left, block), arg1Ptr, block);
-                new StoreInst(cast, arg2Ptr, block);
-                CallInst::Create(func, { retPtr, arg1Ptr, arg2Ptr }, "", block);
-                mul = new LoadInst(valType, retPtr, "res", block);
-            }
+            const auto mul = CallInst::Create(func, {GetterForInt128(left, block), cast}, "div", block);
 
             const auto ok = NDecimal::GenInBounds(mul, NDecimal::GenConstant(-Bound, context), NDecimal::GenConstant(+Bound, context), block);
             const auto nan = NDecimal::GenIsNonComparable(mul, context, block);
@@ -266,18 +204,7 @@ class TDecimalMulIntegralWrapper : public TMutableCodegeneratorNode<TDecimalMulI
             const auto cast = std::is_signed<TRight>() ?
                 static_cast<CastInst*>(new SExtInst(GetterFor<TRight>(right, context, block), valType, "sext", block)):
                 static_cast<CastInst*>(new ZExtInst(GetterFor<TRight>(right, context, block), valType, "zext", block));
-            Value* mul;
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                mul = CallInst::Create(func, {GetterForInt128(left, block), cast}, "div", block);
-            } else {
-                const auto retPtr = new AllocaInst(valType, 0U, "ret_ptr", block);
-                const auto arg1Ptr = new AllocaInst(valType, 0U, "arg1", block);
-                const auto arg2Ptr = new AllocaInst(valType, 0U, "arg2", block);
-                new StoreInst(GetterForInt128(left, block), arg1Ptr, block);
-                new StoreInst(cast, arg2Ptr, block);
-                CallInst::Create(func, { retPtr, arg1Ptr, arg2Ptr }, "", block);
-                mul = new LoadInst(valType, retPtr, "res", block);
-            }
+            const auto mul = CallInst::Create(func, {GetterForInt128(left, block), cast}, "div", block);
 
             const auto ok = NDecimal::GenInBounds(mul, NDecimal::GenConstant(-Bound, context), NDecimal::GenConstant(+Bound, context), block);
             const auto nan = NDecimal::GenIsNonComparable(mul, context, block);
diff --git a/yql/essentials/minikql/comp_nodes/mkql_dictitems.cpp b/yql/essentials/minikql/comp_nodes/mkql_dictitems.cpp
index 54d839e9c74c..8790089972df 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_dictitems.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_dictitems.cpp
@@ -159,7 +159,7 @@ class TDictItemsWrapper : public TCustomValueCodegeneratorNode<TDictItemsWrapper
         const auto valueType = Type::getInt128Ty(context);
         const auto indexType = Type::getInt32Ty(context);
         const auto pairType = ArrayType::get(valueType, 2U);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = Type::getInt1Ty(context);
         const auto funcType = FunctionType::get(statusType, {PointerType::getUnqual(contextType), containerType, PointerType::getUnqual(valueType)}, false);
@@ -178,8 +178,7 @@ class TDictItemsWrapper : public TCustomValueCodegeneratorNode<TDictItemsWrapper
         const auto main = BasicBlock::Create(context, "main", ctx.Func);
         auto block = main;
 
-        const auto container = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, containerArg, "load_container", false, block) : static_cast<Value*>(containerArg);
+        const auto container = static_cast<Value*>(containerArg);
 
         const auto good = BasicBlock::Create(context, "good", ctx.Func);
         const auto done = BasicBlock::Create(context, "done", ctx.Func);
diff --git a/yql/essentials/minikql/comp_nodes/mkql_discard.cpp b/yql/essentials/minikql/comp_nodes/mkql_discard.cpp
index 1fbbf6ac61a3..969f02973127 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_discard.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_discard.cpp
@@ -167,7 +167,7 @@ class TDiscardWrapper : public TCustomValueCodegeneratorNode<TDiscardWrapper> {
             return f;
 
         const auto valueType = Type::getInt128Ty(context);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = Type::getInt32Ty(context);
         const auto funcType = FunctionType::get(statusType, {PointerType::getUnqual(contextType), containerType, PointerType::getUnqual(valueType)}, false);
@@ -185,8 +185,7 @@ class TDiscardWrapper : public TCustomValueCodegeneratorNode<TDiscardWrapper> {
         const auto main = BasicBlock::Create(context, "main", ctx.Func);
         auto block = main;
 
-        const auto container = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, containerArg, "load_container", false, block) : static_cast<Value*>(containerArg);
+        const auto container = static_cast<Value*>(containerArg);
 
         const auto loop = BasicBlock::Create(context, "loop", ctx.Func);
 
diff --git a/yql/essentials/minikql/comp_nodes/mkql_enumerate.cpp b/yql/essentials/minikql/comp_nodes/mkql_enumerate.cpp
index 9905e97d1ac1..61887fe000e1 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_enumerate.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_enumerate.cpp
@@ -133,20 +133,10 @@ class TEnumerateWrapper : public TMutableCodegeneratorNode<TEnumerateWrapper> {
         const auto ptrType = PointerType::getUnqual(StructType::get(context));
         const auto self = CastInst::Create(Instruction::IntToPtr, ConstantInt::get(Type::getInt64Ty(context), uintptr_t(this)), ptrType, "self", block);
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto signature = FunctionType::get(list->getType(), {self->getType(), ctx.Ctx->getType(), list->getType(), start->getType(), step->getType()}, false);
-            const auto creator = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(signature), "creator", block);
-            const auto output = CallInst::Create(signature, creator, {self, ctx.Ctx, list, start, step}, "output", block);
-            return output;
-        } else {
-            const auto place = new AllocaInst(list->getType(), 0U, "place", block);
-            new StoreInst(list, place, block);
-            const auto signature = FunctionType::get(Type::getVoidTy(context), {self->getType(), place->getType(), ctx.Ctx->getType(), place->getType(), start->getType(), step->getType()}, false);
-            const auto creator = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(signature), "creator", block);
-            CallInst::Create(signature, creator, {self, place, ctx.Ctx, place, start, step}, "", block);
-            const auto output = new LoadInst(list->getType(), place, "output", block);
-            return output;
-        }
+        const auto signature = FunctionType::get(list->getType(), {self->getType(), ctx.Ctx->getType(), list->getType(), start->getType(), step->getType()}, false);
+        const auto creator = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(signature), "creator", block);
+        const auto output = CallInst::Create(signature, creator, {self, ctx.Ctx, list, start, step}, "output", block);
+        return output;
     }
 #endif
 private:
diff --git a/yql/essentials/minikql/comp_nodes/mkql_extend.cpp b/yql/essentials/minikql/comp_nodes/mkql_extend.cpp
index 75033de49dd7..1be59576c0ae 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_extend.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_extend.cpp
@@ -566,19 +566,10 @@ using TBaseComputation = TMutableCodegeneratorNode<TOrderedExtendWrapper<IsStrea
         const auto factory = ctx.GetFactory();
         const auto func = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(IsStream ? &THolderFactory::ExtendStream : &THolderFactory::ExtendList<false>));
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(valueType, {factory->getType(), array->getType(), size->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            const auto res = CallInst::Create(funType, funcPtr, {factory, array, size}, "res", block);
-            return res;
-        } else {
-            const auto retPtr = new AllocaInst(valueType, 0U, "ret_ptr", block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {factory->getType(), retPtr->getType(), array->getType(), size->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            CallInst::Create(funType, funcPtr, {factory, retPtr, array, size}, "", block);
-            const auto res = new LoadInst(valueType, retPtr, "res", block);
-            return res;
-        }
+        const auto funType = FunctionType::get(valueType, {factory->getType(), array->getType(), size->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+        const auto res = CallInst::Create(funType, funcPtr, {factory, array, size}, "res", block);
+        return res;
     }
 #endif
 private:
diff --git a/yql/essentials/minikql/comp_nodes/mkql_filter.cpp b/yql/essentials/minikql/comp_nodes/mkql_filter.cpp
index 5e8993db5781..55edc59f3f1b 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_filter.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_filter.cpp
@@ -280,7 +280,7 @@ class TBaseFilterWrapper {
             return f;
 
         const auto valueType = Type::getInt128Ty(context);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = IsStream ? Type::getInt32Ty(context) : Type::getInt1Ty(context);
         const auto funcType = FunctionType::get(statusType, {PointerType::getUnqual(contextType), containerType, PointerType::getUnqual(valueType)}, false);
@@ -299,8 +299,7 @@ class TBaseFilterWrapper {
         const auto main = BasicBlock::Create(context, "main", ctx.Func);
         auto block = main;
 
-        const auto container = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, containerArg, "load_container", false, block) : static_cast<Value*>(containerArg);
+        const auto container = static_cast<Value*>(containerArg);
 
         const auto loop = BasicBlock::Create(context, "loop", ctx.Func);
         const auto good = BasicBlock::Create(context, "good", ctx.Func);
@@ -488,7 +487,7 @@ class TBaseFilterWithLimitWrapper {
 
         const auto valueType = Type::getInt128Ty(context);
         const auto limitType = Type::getInt64Ty(context);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = IsStream ? Type::getInt32Ty(context) : Type::getInt1Ty(context);
         const auto funcType = FunctionType::get(statusType, {PointerType::getUnqual(contextType), containerType, PointerType::getUnqual(limitType), PointerType::getUnqual(valueType)}, false);
@@ -523,8 +522,7 @@ class TBaseFilterWithLimitWrapper {
         ReturnInst::Create(context, IsStream ? ConstantInt::get(statusType, static_cast<ui32>(NUdf::EFetchStatus::Finish)) : ConstantInt::getFalse(context), block);
 
         block = init;
-        const auto container = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, containerArg, "load_container", false, block) : static_cast<Value*>(containerArg);
+        const auto container = static_cast<Value*>(containerArg);
 
         BranchInst::Create(loop, block);
         block = loop;
@@ -971,20 +969,10 @@ class TListFilterWrapper : public TBothWaysCodegeneratorNode<TListFilterWrapper>
             const auto doFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TListFilterWrapper::MakeLazyList));
             const auto ptrType = PointerType::getUnqual(StructType::get(context));
             const auto self = CastInst::Create(Instruction::IntToPtr, ConstantInt::get(Type::getInt64Ty(context), uintptr_t(this)), ptrType, "self", block);
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                const auto funType = FunctionType::get(list->getType() , {self->getType(), ctx.Ctx->getType(), list->getType()}, false);
-                const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-                const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx, list}, "value", block);
-                out->addIncoming(value, block);
-            } else {
-                const auto resultPtr = new AllocaInst(list->getType(), 0U, "return", block);
-                new StoreInst(list, resultPtr, block);
-                const auto funType = FunctionType::get(Type::getVoidTy(context), {self->getType(), resultPtr->getType(), ctx.Ctx->getType(), resultPtr->getType()}, false);
-                const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-                CallInst::Create(funType, doFuncPtr, {self, resultPtr, ctx.Ctx, resultPtr}, "", block);
-                const auto value = new LoadInst(list->getType(), resultPtr, "value", block);
-                out->addIncoming(value, block);
-            }
+            const auto funType = FunctionType::get(list->getType() , {self->getType(), ctx.Ctx->getType(), list->getType()}, false);
+            const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
+            const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx, list}, "value", block);
+            out->addIncoming(value, block);
             BranchInst::Create(done, block);
         }
 
@@ -1340,20 +1328,10 @@ class TListFilterWithLimitWrapper : public TBothWaysCodegeneratorNode<TListFilte
             const auto doFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TListFilterWithLimitWrapper::MakeLazyList));
             const auto ptrType = PointerType::getUnqual(StructType::get(context));
             const auto self = CastInst::Create(Instruction::IntToPtr, ConstantInt::get(Type::getInt64Ty(context), uintptr_t(this)), ptrType, "self", block);
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                const auto funType = FunctionType::get(list->getType() , {self->getType(), ctx.Ctx->getType(), list->getType(), limit->getType()}, false);
-                const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-                const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx, list, limit}, "value", block);
-                out->addIncoming(value, block);
-            } else {
-                const auto resultPtr = new AllocaInst(list->getType(), 0U, "return", block);
-                new StoreInst(list, resultPtr, block);
-                const auto funType = FunctionType::get(Type::getVoidTy(context), {self->getType(), resultPtr->getType(), ctx.Ctx->getType(), resultPtr->getType(), limit->getType()}, false);
-                const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-                CallInst::Create(funType, doFuncPtr, {self, resultPtr, ctx.Ctx, resultPtr, limit}, "", block);
-                const auto value = new LoadInst(list->getType(), resultPtr, "value", block);
-                out->addIncoming(value, block);
-            }
+            const auto funType = FunctionType::get(list->getType() , {self->getType(), ctx.Ctx->getType(), list->getType(), limit->getType()}, false);
+            const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
+            const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx, list, limit}, "value", block);
+            out->addIncoming(value, block);
             BranchInst::Create(done, block);
         }
 
diff --git a/yql/essentials/minikql/comp_nodes/mkql_flatmap.cpp b/yql/essentials/minikql/comp_nodes/mkql_flatmap.cpp
index dc1d9a7472fe..0e53e650ac6c 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_flatmap.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_flatmap.cpp
@@ -1170,7 +1170,7 @@ class TBaseFlatMapWrapper {
             return f;
 
         const auto valueType = Type::getInt128Ty(context);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = IsInputStream ? Type::getInt32Ty(context) : Type::getInt1Ty(context);
         const auto funcType = FunctionType::get(statusType, {PointerType::getUnqual(contextType), containerType, PointerType::getUnqual(valueType)}, false);
@@ -1189,8 +1189,7 @@ class TBaseFlatMapWrapper {
         const auto main = BasicBlock::Create(context, "main", ctx.Func);
         auto block = main;
 
-        const auto container = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, containerArg, "load_container", false, block) : static_cast<Value*>(containerArg);
+        const auto container = static_cast<Value*>(containerArg);
 
         const auto loop = BasicBlock::Create(context, "loop", ctx.Func);
         const auto good = BasicBlock::Create(context, "good", ctx.Func);
@@ -1242,7 +1241,7 @@ class TBaseFlatMapWrapper {
             return f;
 
         const auto valueType = Type::getInt128Ty(context);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = IsInputStream ? Type::getInt32Ty(context) : Type::getInt1Ty(context);
         const auto stateType = ResultContainerOpt ? Type::getInt32Ty(context) : Type::getInt1Ty(context);
@@ -1263,8 +1262,7 @@ class TBaseFlatMapWrapper {
         const auto main = BasicBlock::Create(context, "main", ctx.Func);
         auto block = main;
 
-        const auto container = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, containerArg, "load_container", false, block) : static_cast<Value*>(containerArg);
+        const auto container = static_cast<Value*>(containerArg);
 
         const auto zero = ConstantInt::get(valueType, 0);
 
@@ -1578,17 +1576,9 @@ class TListFlatMapWrapper : public TBothWaysCodegeneratorNode<TListFlatMapWrappe
 
                 const auto func = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&THolderFactory::ExtendList<ResultContainerOpt>));
 
-                if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                    const auto funType = FunctionType::get(list->getType(), {factory->getType(), vector->getType(), index->getType()}, false);
-                    const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-                    res = CallInst::Create(funType, funcPtr, {factory, vector, index}, "res", block);
-                } else {
-                    const auto retPtr = new AllocaInst(list->getType(), 0U, "ret_ptr", block);
-                    const auto funType = FunctionType::get(Type::getVoidTy(context), {factory->getType(), retPtr->getType(), vector->getType(), index->getType()}, false);
-                    const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-                    CallInst::Create(funType, funcPtr, {factory, retPtr, vector, index}, "", block);
-                    res = new LoadInst(list->getType(), retPtr, "res", block);
-                }
+                const auto funType = FunctionType::get(list->getType(), {factory->getType(), vector->getType(), index->getType()}, false);
+                const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+                res = CallInst::Create(funType, funcPtr, {factory, vector, index}, "res", block);
             }
             map->addIncoming(res, block);
             BranchInst::Create(free, done, heap, block);
@@ -1613,20 +1603,10 @@ class TListFlatMapWrapper : public TBothWaysCodegeneratorNode<TListFlatMapWrappe
             const auto doFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TListFlatMapWrapper::MakeLazyList));
             const auto ptrType = PointerType::getUnqual(StructType::get(context));
             const auto self = CastInst::Create(Instruction::IntToPtr, ConstantInt::get(Type::getInt64Ty(context), uintptr_t(this)), ptrType, "self", block);
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                const auto funType = FunctionType::get(list->getType() , {self->getType(), ctx.Ctx->getType(), list->getType()}, false);
-                const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-                const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx, list}, "value", block);
-                map->addIncoming(value, block);
-            } else {
-                const auto resultPtr = new AllocaInst(list->getType(), 0U, "return", block);
-                new StoreInst(list, resultPtr, block);
-                const auto funType = FunctionType::get(Type::getVoidTy(context), {self->getType(), resultPtr->getType(), ctx.Ctx->getType(), resultPtr->getType()}, false);
-                const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-                CallInst::Create(funType, doFuncPtr, {self, resultPtr, ctx.Ctx, resultPtr}, "", block);
-                const auto value = new LoadInst(list->getType(), resultPtr, "value", block);
-                map->addIncoming(value, block);
-            }
+            const auto funType = FunctionType::get(list->getType() , {self->getType(), ctx.Ctx->getType(), list->getType()}, false);
+            const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
+            const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx, list}, "value", block);
+            map->addIncoming(value, block);
             BranchInst::Create(done, block);
         }
 
diff --git a/yql/essentials/minikql/comp_nodes/mkql_fromstring.cpp b/yql/essentials/minikql/comp_nodes/mkql_fromstring.cpp
index d208343cc2ab..7b8c0fe0a6f7 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_fromstring.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_fromstring.cpp
@@ -78,13 +78,11 @@ class TDecimalFromStringWrapper : public TMutableCodegeneratorNode<TDecimalFromS
 
         const auto valType = Type::getInt128Ty(context);
         const auto psType = Type::getInt8Ty(context);
-        const auto valTypePtr = PointerType::getUnqual(valType);
 
         const auto name = "DecimalFromString";
         ctx.Codegen.AddGlobalMapping(name, reinterpret_cast<const void*>(&DecimalFromString));
-        const auto fnType = NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget() ?
-            FunctionType::get(valType, { valType, psType, psType }, false):
-            FunctionType::get(Type::getVoidTy(context), { valTypePtr, valTypePtr, psType, psType }, false);
+        const auto fnType =
+            FunctionType::get(valType, { valType, psType, psType }, false);
         const auto func = ctx.Codegen.GetModule().getOrInsertFunction(name, fnType);
 
         const auto zero = ConstantInt::get(valType, 0ULL);
@@ -108,15 +106,7 @@ class TDecimalFromStringWrapper : public TMutableCodegeneratorNode<TDecimalFromS
             block = call;
         }
 
-        Value* decimal;
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            decimal = CallInst::Create(func, { value, precision, scale }, "from_string", block);
-        } else {
-            const auto retPtr = new AllocaInst(valType, 0U, "ret_ptr", block);
-            new StoreInst(value, retPtr, block);
-            CallInst::Create(func, { retPtr, retPtr, precision, scale }, "", block);
-            decimal = new LoadInst(valType, retPtr, "res", block);
-        }
+        const auto decimal = CallInst::Create(func, { value, precision, scale }, "from_string", block);
 
         if (Data->IsTemporaryValue())
             ValueCleanup(Data->GetRepresentation(), value, ctx, block);
@@ -198,13 +188,11 @@ class TFromStringWrapper : public TMutableCodegeneratorNode<TFromStringWrapper<I
 
         const auto valType = Type::getInt128Ty(context);
         const auto slotType = Type::getInt32Ty(context);
-        const auto valTypePtr = PointerType::getUnqual(valType);
 
         const auto name = "DataFromString";
         ctx.Codegen.AddGlobalMapping(name, reinterpret_cast<const void*>(&DataFromString));
-        const auto fnType = NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget() ?
-            FunctionType::get(valType, { valType, slotType }, false):
-            FunctionType::get(Type::getVoidTy(context), { valTypePtr, valTypePtr, slotType }, false);
+        const auto fnType =
+            FunctionType::get(valType, { valType, slotType }, false);
         const auto func = ctx.Codegen.GetModule().getOrInsertFunction(name, fnType);
 
         const auto zero = ConstantInt::get(valType, 0ULL);
@@ -224,15 +212,7 @@ class TFromStringWrapper : public TMutableCodegeneratorNode<TFromStringWrapper<I
             block = call;
         }
 
-        Value* data;
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            data = CallInst::Create(func, { value, slot }, "from_string", block);
-        } else {
-            const auto retPtr = new AllocaInst(valType, 0U, "ret_ptr", block);
-            new StoreInst(value, retPtr, block);
-            CallInst::Create(func, { retPtr, retPtr, slot }, "", block);
-            data = new LoadInst(valType, retPtr, "res", block);
-        }
+        Value* data = CallInst::Create(func, { value, slot }, "from_string", block);
 
         if (Data->IsTemporaryValue())
             ValueCleanup(Data->GetRepresentation(), value, ctx, block);
diff --git a/yql/essentials/minikql/comp_nodes/mkql_heap.cpp b/yql/essentials/minikql/comp_nodes/mkql_heap.cpp
index 74fa51806a85..260223a0f36c 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_heap.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_heap.cpp
@@ -84,18 +84,9 @@ class THeapWrapper : public TMutableCodegeneratorNode<THeapWrapper>
         const auto idxType = Type::getInt32Ty(context);
 
         Value* array = nullptr;
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(valueType, {fact->getType(), list->getType(), itemsPtr->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            array = CallInst::Create(funType, funcPtr, {fact, list, itemsPtr}, "array", block);
-        } else {
-            const auto arrayPtr = new AllocaInst(valueType, 0U, "array_ptr", block);
-            new StoreInst(list, arrayPtr, block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {fact->getType(), arrayPtr->getType(), arrayPtr->getType(), itemsPtr->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            CallInst::Create(funType, funcPtr, {fact, arrayPtr, arrayPtr, itemsPtr}, "", block);
-            array = new LoadInst(valueType, arrayPtr, "array", block);
-        }
+        const auto funType = FunctionType::get(valueType, {fact->getType(), list->getType(), itemsPtr->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+        array = CallInst::Create(funType, funcPtr, {fact, list, itemsPtr}, "array", block);
 
         result->addIncoming(array, block);
 
@@ -269,18 +260,9 @@ class TNthWrapper : public TMutableCodegeneratorNode<TNthWrapper>
         const auto idxType = Type::getInt32Ty(context);
 
         Value* array = nullptr;
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(valueType, {fact->getType(), list->getType(), itemsPtr->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            array = CallInst::Create(funType, funcPtr, {fact, list, itemsPtr}, "array", block);
-        } else {
-            const auto arrayPtr = new AllocaInst(valueType, 0U, "array_ptr", block);
-            new StoreInst(list, arrayPtr, block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {fact->getType(), arrayPtr->getType(), arrayPtr->getType(), itemsPtr->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            CallInst::Create(funType, funcPtr, {fact, arrayPtr, arrayPtr, itemsPtr}, "", block);
-            array = new LoadInst(valueType, arrayPtr, "array", block);
-        }
+        const auto funType = FunctionType::get(valueType, {fact->getType(), list->getType(), itemsPtr->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+        array = CallInst::Create(funType, funcPtr, {fact, list, itemsPtr}, "array", block);
 
         result->addIncoming(array, block);
 
diff --git a/yql/essentials/minikql/comp_nodes/mkql_iterator.cpp b/yql/essentials/minikql/comp_nodes/mkql_iterator.cpp
index dcd9d351cfad..b413d1cbca8d 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_iterator.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_iterator.cpp
@@ -30,20 +30,10 @@ class TIteratorWrapper : public TMutableCodegeneratorNode<TIteratorWrapper> {
         const auto factory = ctx.GetFactory();
         const auto func = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&THolderFactory::CreateIteratorOverList));
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto signature = FunctionType::get(value->getType(), {factory->getType(), value->getType()}, false);
-            const auto creator = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(signature), "creator", block);
-            const auto output = CallInst::Create(signature, creator, {factory, value}, "output", block);
-            return output;
-        } else {
-            const auto place = new AllocaInst(value->getType(), 0U, "place", block);
-            new StoreInst(value, place, block);
-            const auto signature = FunctionType::get(Type::getVoidTy(context), {factory->getType(), place->getType(), place->getType()}, false);
-            const auto creator = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(signature), "creator", block);
-            CallInst::Create(signature, creator, {factory, place, place}, "", block);
-            const auto output = new LoadInst(value->getType(), place, "output", block);
-            return output;
-        }
+        const auto signature = FunctionType::get(value->getType(), {factory->getType(), value->getType()}, false);
+        const auto creator = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(signature), "creator", block);
+        const auto output = CallInst::Create(signature, creator, {factory, value}, "output", block);
+        return output;
     }
 #endif
 private:
@@ -76,20 +66,10 @@ class TForwardListWrapper : public TMutableCodegeneratorNode<TForwardListWrapper
         const auto factory = ctx.GetFactory();
         const auto func = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&THolderFactory::CreateForwardList));
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto signature = FunctionType::get(value->getType(), {factory->getType(), value->getType()}, false);
-            const auto creator = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(signature), "creator", block);
-            const auto output = CallInst::Create(signature, creator, {factory, value}, "output", block);
-            return output;
-        } else {
-            const auto place = new AllocaInst(value->getType(), 0U, "place", block);
-            new StoreInst(value, place, block);
-            const auto signature = FunctionType::get(Type::getVoidTy(context), {factory->getType(), place->getType(), place->getType()}, false);
-            const auto creator = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(signature), "creator", block);
-            CallInst::Create(signature, creator, {factory, place, place}, "", block);
-            const auto output = new LoadInst(value->getType(), place, "output", block);
-            return output;
-        }
+        const auto signature = FunctionType::get(value->getType(), {factory->getType(), value->getType()}, false);
+        const auto creator = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(signature), "creator", block);
+        const auto output = CallInst::Create(signature, creator, {factory, value}, "output", block);
+        return output;
     }
 #endif
 private:
diff --git a/yql/essentials/minikql/comp_nodes/mkql_join_dict.cpp b/yql/essentials/minikql/comp_nodes/mkql_join_dict.cpp
index b93970b6466a..127e32a0ebff 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_join_dict.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_join_dict.cpp
@@ -49,27 +49,12 @@ class TJoinDictWrapper : public TMutableCodegeneratorPtrNode<TJoinDictWrapper<Ke
         const auto one = GetNodeValue(Dict1, ctx, block);
         const auto two = GetNodeValue(Dict2, ctx, block);
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto joinFuncType = FunctionType::get(Type::getInt128Ty(context),
-                { joinFuncArg->getType(), ctx.Ctx->getType(), one->getType(), two->getType() }, false);
-            const auto joinFuncPtr = CastInst::Create(Instruction::IntToPtr, joinFunc, PointerType::getUnqual(joinFuncType), "cast", block);
-            const auto join = CallInst::Create(joinFuncType, joinFuncPtr, { joinFuncArg, ctx.Ctx, one, two }, "join", block);
-            AddRefBoxed(join, ctx, block);
-            new StoreInst(join, pointer, block);
-        } else {
-            const auto onePtr = new AllocaInst(one->getType(), 0U, "one_ptr", block);
-            const auto twoPtr = new AllocaInst(two->getType(), 0U, "two_ptr", block);
-            new StoreInst(one, onePtr, block);
-            new StoreInst(two, twoPtr, block);
-
-            const auto joinFuncType = FunctionType::get(Type::getVoidTy(context),
-                { joinFuncArg->getType(), pointer->getType(), ctx.Ctx->getType(), onePtr->getType(), twoPtr->getType() }, false);
-
-            const auto joinFuncPtr = CastInst::Create(Instruction::IntToPtr, joinFunc, PointerType::getUnqual(joinFuncType), "cast", block);
-            CallInst::Create(joinFuncType, joinFuncPtr, { joinFuncArg, pointer, ctx.Ctx, onePtr, twoPtr }, "", block);
-            const auto join = new LoadInst(Type::getInt128Ty(context), pointer, "join", block);
-            AddRefBoxed(join, ctx, block);
-        }
+        const auto joinFuncType = FunctionType::get(Type::getInt128Ty(context),
+            { joinFuncArg->getType(), ctx.Ctx->getType(), one->getType(), two->getType() }, false);
+        const auto joinFuncPtr = CastInst::Create(Instruction::IntToPtr, joinFunc, PointerType::getUnqual(joinFuncType), "cast", block);
+        const auto join = CallInst::Create(joinFuncType, joinFuncPtr, { joinFuncArg, ctx.Ctx, one, two }, "join", block);
+        AddRefBoxed(join, ctx, block);
+        new StoreInst(join, pointer, block);
     }
 #endif
 private:
diff --git a/yql/essentials/minikql/comp_nodes/mkql_lazy_list.cpp b/yql/essentials/minikql/comp_nodes/mkql_lazy_list.cpp
index f7d04055582b..f19ac809a422 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_lazy_list.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_lazy_list.cpp
@@ -60,20 +60,10 @@ class TLazyListWrapper : public TMutableCodegeneratorNode<TLazyListWrapper<IsOpt
 
         block = wrap;
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(list->getType(), {factory->getType(), list->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            const auto res = CallInst::Create(funType, funcPtr, {factory, list}, "res", block);
-            lazy->addIncoming(res, block);
-        } else {
-            const auto retPtr = new AllocaInst(list->getType(), 0U, "ret_ptr", block);
-            new StoreInst(list, retPtr, block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {factory->getType(), retPtr->getType(), retPtr->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            CallInst::Create(funType, funcPtr, {factory, retPtr, retPtr}, "", block);
-            const auto res = new LoadInst(list->getType(), retPtr, "res", block);
-            lazy->addIncoming(res, block);
-        }
+        const auto funType = FunctionType::get(list->getType(), {factory->getType(), list->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+        const auto res = CallInst::Create(funType, funcPtr, {factory, list}, "res", block);
+        lazy->addIncoming(res, block);
 
         BranchInst::Create(done, block);
 
diff --git a/yql/essentials/minikql/comp_nodes/mkql_listfromrange.cpp b/yql/essentials/minikql/comp_nodes/mkql_listfromrange.cpp
index 1e374c4dcca3..fab157dafbfd 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_listfromrange.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_listfromrange.cpp
@@ -279,19 +279,10 @@ class TListFromRangeWrapper : public TMutableCodegeneratorNode<TListFromRangeWra
         const auto timezone = TzDate ? GetterForTimezone(context, startv, block) : ConstantInt::get(Type::getInt16Ty(context), 0);
 
         const auto func = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TListFromRangeWrapper::MakeList));
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto signature = FunctionType::get(valueType, {ctx.Ctx->getType(), start->getType(), end->getType(), step->getType(), timezone->getType()}, false);
-            const auto creator = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(signature), "creator", block);
-            const auto output = CallInst::Create(signature, creator, {ctx.Ctx, start, end, step, timezone}, "output", block);
-            return output;
-        } else {
-            const auto place = new AllocaInst(valueType, 0U, "place", block);
-            const auto signature = FunctionType::get(Type::getVoidTy(context), {place->getType(), ctx.Ctx->getType(), start->getType(), end->getType(), step->getType(), timezone->getType()}, false);
-            const auto creator = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(signature), "creator", block);
-            CallInst::Create(signature, creator, {place, ctx.Ctx, start, end, step, timezone}, "", block);
-            const auto output = new LoadInst(valueType, place, "output", block);
-            return output;
-        }
+        const auto signature = FunctionType::get(valueType, {ctx.Ctx->getType(), start->getType(), end->getType(), step->getType(), timezone->getType()}, false);
+        const auto creator = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(signature), "creator", block);
+        const auto output = CallInst::Create(signature, creator, {ctx.Ctx, start, end, step, timezone}, "output", block);
+        return output;
     }
 #endif
 private:
diff --git a/yql/essentials/minikql/comp_nodes/mkql_map.cpp b/yql/essentials/minikql/comp_nodes/mkql_map.cpp
index 690b68228b2b..ff014286e839 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_map.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_map.cpp
@@ -197,7 +197,7 @@ class TBaseMapWrapper {
             return f;
 
         const auto valueType = Type::getInt128Ty(context);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = IsStream ? Type::getInt32Ty(context) : Type::getInt1Ty(context);
         const auto funcType = FunctionType::get(statusType, {PointerType::getUnqual(contextType), containerType, PointerType::getUnqual(valueType)}, false);
@@ -216,8 +216,7 @@ class TBaseMapWrapper {
         const auto main = BasicBlock::Create(context, "main", ctx.Func);
         auto block = main;
 
-        const auto container = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, containerArg, "load_container", false, block) : static_cast<Value*>(containerArg);
+        const auto container = static_cast<Value*>(containerArg);
 
         const auto good = BasicBlock::Create(context, "good", ctx.Func);
         const auto done = BasicBlock::Create(context, "done", ctx.Func);
@@ -392,20 +391,10 @@ class TListMapWrapper : public TBothWaysCodegeneratorNode<TListMapWrapper>, priv
             const auto doFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TListMapWrapper::MakeLazyList));
             const auto ptrType = PointerType::getUnqual(StructType::get(context));
             const auto self = CastInst::Create(Instruction::IntToPtr, ConstantInt::get(Type::getInt64Ty(context), uintptr_t(this)), ptrType, "self", block);
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                const auto funType = FunctionType::get(list->getType() , {self->getType(), ctx.Ctx->getType(), list->getType()}, false);
-                const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-                const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx, list}, "value", block);
-                map->addIncoming(value, block);
-            } else {
-                const auto resultPtr = new AllocaInst(list->getType(), 0U, "return", block);
-                new StoreInst(list, resultPtr, block);
-                const auto funType = FunctionType::get(Type::getVoidTy(context), {self->getType(), resultPtr->getType(), ctx.Ctx->getType(), resultPtr->getType()}, false);
-                const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-                CallInst::Create(funType, doFuncPtr, {self, resultPtr, ctx.Ctx, resultPtr}, "", block);
-                const auto value = new LoadInst(list->getType(), resultPtr, "value", block);
-                map->addIncoming(value, block);
-            }
+            const auto funType = FunctionType::get(list->getType() , {self->getType(), ctx.Ctx->getType(), list->getType()}, false);
+            const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
+            const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx, list}, "value", block);
+            map->addIncoming(value, block);
             BranchInst::Create(done, block);
         }
 
diff --git a/yql/essentials/minikql/comp_nodes/mkql_map_join.cpp b/yql/essentials/minikql/comp_nodes/mkql_map_join.cpp
index 05de9fafc6de..9cdb16e855f9 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_map_join.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_map_join.cpp
@@ -1544,7 +1544,7 @@ class TMapJoinCoreWrapper : public TMapJoinCoreWrapperBase<IsTuple>, public TCus
         const auto valueType = Type::getInt128Ty(context);
         const auto arrayType = ArrayType::get(valueType, this->OutputRepresentations.size());
         const auto keysType = IsTuple ? ArrayType::get(valueType, this->LeftKeyColumns.size()) : nullptr;
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = Type::getInt32Ty(context);
         const auto funcType = FunctionType::get(statusType, {PointerType::getUnqual(contextType), containerType, containerType, PointerType::getUnqual(valueType)}, false);
@@ -1564,11 +1564,9 @@ class TMapJoinCoreWrapper : public TMapJoinCoreWrapperBase<IsTuple>, public TCus
         const auto main = BasicBlock::Create(context, "main", ctx.Func);
         auto block = main;
 
-        const auto stream = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, streamArg, "load_stream", false, block) : static_cast<Value*>(streamArg);
+        const auto stream = static_cast<Value*>(streamArg);
 
-        const auto dict = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, dictArg, "load_dict", false, block) : static_cast<Value*>(dictArg);
+        const auto dict = static_cast<Value*>(dictArg);
 
         const auto zero = ConstantInt::get(valueType, 0);
         const auto fsok = ConstantInt::get(statusType, static_cast<ui32>(NUdf::EFetchStatus::Ok));
@@ -1701,7 +1699,7 @@ class TMapJoinCoreWrapper : public TMapJoinCoreWrapperBase<IsTuple>, public TCus
         const auto valueType = Type::getInt128Ty(context);
         const auto arrayType = ArrayType::get(valueType, this->OutputRepresentations.size());
         const auto keysType = IsTuple ? ArrayType::get(valueType, this->LeftKeyColumns.size()) : nullptr;
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = Type::getInt32Ty(context);
         const auto funcType = FunctionType::get(statusType, {PointerType::getUnqual(contextType), containerType, containerType, PointerType::getUnqual(valueType), PointerType::getUnqual(valueType), PointerType::getUnqual(valueType)}, false);
@@ -1723,11 +1721,9 @@ class TMapJoinCoreWrapper : public TMapJoinCoreWrapperBase<IsTuple>, public TCus
         const auto main = BasicBlock::Create(context, "main", ctx.Func);
         auto block = main;
 
-        const auto stream = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, streamArg, "load_stream", false, block) : static_cast<Value*>(streamArg);
+        const auto stream = static_cast<Value*>(streamArg);
 
-        const auto dict = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, dictArg, "load_dict", false, block) : static_cast<Value*>(dictArg);
+        const auto dict = static_cast<Value*>(dictArg);
 
         const auto zero = ConstantInt::get(valueType, 0);
         const auto fsok = ConstantInt::get(statusType, static_cast<ui32>(NUdf::EFetchStatus::Ok));
diff --git a/yql/essentials/minikql/comp_nodes/mkql_multimap.cpp b/yql/essentials/minikql/comp_nodes/mkql_multimap.cpp
index c9fad3198fdf..3a3f24e37423 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_multimap.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_multimap.cpp
@@ -282,20 +282,10 @@ class TListMultiMapWrapper : public TBothWaysCodegeneratorNode<TListMultiMapWrap
             const auto doFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TListMultiMapWrapper::MakeLazyList));
             const auto ptrType = PointerType::getUnqual(StructType::get(context));
             const auto self = CastInst::Create(Instruction::IntToPtr, ConstantInt::get(Type::getInt64Ty(context), uintptr_t(this)), ptrType, "self", block);
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                const auto funType = FunctionType::get(list->getType() , {self->getType(), ctx.Ctx->getType(), list->getType()}, false);
-                const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-                const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx, list}, "value", block);
-                map->addIncoming(value, block);
-            } else {
-                const auto resultPtr = new AllocaInst(list->getType(), 0U, "return", block);
-                new StoreInst(list, resultPtr, block);
-                const auto funType = FunctionType::get(Type::getVoidTy(context), {self->getType(), resultPtr->getType(), ctx.Ctx->getType(), resultPtr->getType()}, false);
-                const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-                CallInst::Create(funType, doFuncPtr, {self, resultPtr, ctx.Ctx, resultPtr}, "", block);
-                const auto value = new LoadInst(list->getType(), resultPtr, "value", block);
-                map->addIncoming(value, block);
-            }
+            const auto funType = FunctionType::get(list->getType() , {self->getType(), ctx.Ctx->getType(), list->getType()}, false);
+            const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
+            const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx, list}, "value", block);
+            map->addIncoming(value, block);
             BranchInst::Create(done, block);
         }
 
@@ -335,7 +325,7 @@ class TListMultiMapWrapper : public TBothWaysCodegeneratorNode<TListMultiMapWrap
 
         const auto valueType = Type::getInt128Ty(context);
         const auto positionType = Type::getInt64Ty(context);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = Type::getInt1Ty(context);
         const auto funcType = FunctionType::get(statusType, {PointerType::getUnqual(contextType), containerType, PointerType::getUnqual(positionType), PointerType::getUnqual(valueType)}, false);
@@ -355,8 +345,7 @@ class TListMultiMapWrapper : public TBothWaysCodegeneratorNode<TListMultiMapWrap
         const auto main = BasicBlock::Create(context, "main", ctx.Func);
         auto block = main;
 
-        const auto container = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, containerArg, "load_container", false, block) : static_cast<Value*>(containerArg);
+        const auto container = static_cast<Value*>(containerArg);
 
         const auto position = new LoadInst(positionType, positionArg, "position", false, block);
 
diff --git a/yql/essentials/minikql/comp_nodes/mkql_prepend.cpp b/yql/essentials/minikql/comp_nodes/mkql_prepend.cpp
index 334197d48896..0b29d724f3b8 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_prepend.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_prepend.cpp
@@ -53,44 +53,20 @@ class TPrependWrapper : public TMutableCodegeneratorNode<TPrependWrapper<IsVoid>
             BranchInst::Create(work, done, check, block);
             block = work;
 
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                const auto funType = FunctionType::get(right->getType(), {factory->getType(), left->getType(), right->getType()}, false);
-                const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-                const auto res = CallInst::Create(funType, funcPtr, {factory, left, right}, "res", block);
-                result->addIncoming(res, block);
-            } else {
-                const auto retPtr = new AllocaInst(right->getType(), 0U, "ret_ptr", block);
-                const auto itemPtr = new AllocaInst(left->getType(), 0U, "item_ptr", block);
-                new StoreInst(right, retPtr, block);
-                new StoreInst(left, itemPtr, block);
-                const auto funType = FunctionType::get(Type::getVoidTy(context), {factory->getType(), retPtr->getType(), itemPtr->getType(), retPtr->getType()}, false);
-                const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-                CallInst::Create(funType, funcPtr, {factory, retPtr, itemPtr, retPtr}, "", block);
-                const auto res = new LoadInst(right->getType(), retPtr, "res", block);
-                result->addIncoming(res, block);
-            }
+            const auto funType = FunctionType::get(right->getType(), {factory->getType(), left->getType(), right->getType()}, false);
+            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+            const auto res = CallInst::Create(funType, funcPtr, {factory, left, right}, "res", block);
+            result->addIncoming(res, block);
 
             BranchInst::Create(done, block);
 
             block = done;
             return result;
         } else {
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                const auto funType = FunctionType::get(right->getType(), {factory->getType(), left->getType(), right->getType()}, false);
-                const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-                const auto res = CallInst::Create(funType, funcPtr, {factory, left, right}, "res", block);
-                return res;
-            } else {
-                const auto retPtr = new AllocaInst(right->getType(), 0U, "ret_ptr", block);
-                const auto itemPtr = new AllocaInst(left->getType(), 0U, "item_ptr", block);
-                new StoreInst(right, retPtr, block);
-                new StoreInst(left, itemPtr, block);
-                const auto funType = FunctionType::get(Type::getVoidTy(context), {factory->getType(), retPtr->getType(), itemPtr->getType(), retPtr->getType()}, false);
-                const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-                CallInst::Create(funType, funcPtr, {factory, retPtr, itemPtr, retPtr}, "", block);
-                const auto res = new LoadInst(right->getType(), retPtr, "res", block);
-                return res;
-            }
+            const auto funType = FunctionType::get(right->getType(), {factory->getType(), left->getType(), right->getType()}, false);
+            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+            const auto res = CallInst::Create(funType, funcPtr, {factory, left, right}, "res", block);
+            return res;
         }
     }
 #endif
diff --git a/yql/essentials/minikql/comp_nodes/mkql_reverse.cpp b/yql/essentials/minikql/comp_nodes/mkql_reverse.cpp
index 6549b03dfd75..692595f3a4e3 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_reverse.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_reverse.cpp
@@ -38,20 +38,10 @@ class TReverseWrapper : public TMutableCodegeneratorNode<TReverseWrapper> {
 
         const auto list = GetNodeValue(List, ctx, block);
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(list->getType(), {factory->getType(), builder->getType(), list->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            const auto result = CallInst::Create(funType, funcPtr, {factory, builder, list}, "result", block);
-            return result;
-        } else {
-            const auto retPtr = new AllocaInst(list->getType(), 0U, "ret_ptr", block);
-            new StoreInst(list, retPtr, block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {factory->getType(), retPtr->getType(), builder->getType(), retPtr->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            CallInst::Create(funType, funcPtr, {factory, retPtr, builder, retPtr}, "", block);
-            const auto result = new LoadInst(list->getType(), retPtr, "result", block);
-            return result;
-        }
+        const auto funType = FunctionType::get(list->getType(), {factory->getType(), builder->getType(), list->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+        const auto result = CallInst::Create(funType, funcPtr, {factory, builder, list}, "result", block);
+        return result;
     }
 #endif
 private:
diff --git a/yql/essentials/minikql/comp_nodes/mkql_skip.cpp b/yql/essentials/minikql/comp_nodes/mkql_skip.cpp
index ad9b43ebac63..9f33a1ae1543 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_skip.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_skip.cpp
@@ -262,20 +262,10 @@ class TSkipWrapper : public TMutableCodegeneratorNode<TSkipWrapper> {
         const auto cnt = GetNodeValue(Count, ctx, block);
         const auto count = GetterFor<ui64>(cnt, context, block);
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(list->getType(), {factory->getType(), builder->getType(), list->getType(), count->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            const auto result = CallInst::Create(funType, funcPtr, {factory, builder, list, count}, "result", block);
-            return result;
-        } else {
-            const auto retPtr = new AllocaInst(list->getType(), 0U, "ret_ptr", block);
-            new StoreInst(list, retPtr, block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {factory->getType(), retPtr->getType(), builder->getType(), retPtr->getType(), count->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            CallInst::Create(funType, funcPtr, {factory, retPtr, builder, retPtr, count}, "", block);
-            const auto result = new LoadInst(list->getType(), retPtr, "result", block);
-            return result;
-        }
+        const auto funType = FunctionType::get(list->getType(), {factory->getType(), builder->getType(), list->getType(), count->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+        const auto result = CallInst::Create(funType, funcPtr, {factory, builder, list, count}, "result", block);
+        return result;
     }
 #endif
 
diff --git a/yql/essentials/minikql/comp_nodes/mkql_squeeze_to_list.cpp b/yql/essentials/minikql/comp_nodes/mkql_squeeze_to_list.cpp
index 9a3e257c2c71..ca5e0cf547c1 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_squeeze_to_list.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_squeeze_to_list.cpp
@@ -140,7 +140,7 @@ class TSqueezeToListWrapper: public TStatefulFlowCodegeneratorNode<TSqueezeToLis
 
         const auto push = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TState::Put));
 
-        const auto arg = WrapArgumentForWindows(item, ctx, block);
+        const auto arg = item;
 
         const auto pushType = FunctionType::get(Type::getInt1Ty(context), {stateArg->getType(), arg->getType()}, false);
         const auto pushPtr = CastInst::Create(Instruction::IntToPtr, push, PointerType::getUnqual(pushType), "push", block);
@@ -152,21 +152,11 @@ class TSqueezeToListWrapper: public TStatefulFlowCodegeneratorNode<TSqueezeToLis
 
         const auto pull = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TState::Pull));
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto pullType = FunctionType::get(valueType, {stateArg->getType(), ctx.Ctx->getType()}, false);
-            const auto pullPtr = CastInst::Create(Instruction::IntToPtr, pull, PointerType::getUnqual(pullType), "pull", block);
-            const auto list = CallInst::Create(pullType, pullPtr, {stateArg, ctx.Ctx}, "list", block);
-            UnRefBoxed(state, ctx, block);
-            result->addIncoming(list, block);
-        } else {
-            const auto ptr = new AllocaInst(valueType, 0U, "ptr", block);
-            const auto pullType = FunctionType::get(Type::getVoidTy(context), {stateArg->getType(), ptr->getType(), ctx.Ctx->getType()}, false);
-            const auto pullPtr = CastInst::Create(Instruction::IntToPtr, pull, PointerType::getUnqual(pullType), "pull", block);
-            CallInst::Create(pullType, pullPtr, {stateArg, ptr, ctx.Ctx}, "", block);
-            const auto list = new LoadInst(valueType, ptr, "list", block);
-            UnRefBoxed(state, ctx, block);
-            result->addIncoming(list, block);
-        }
+        const auto pullType = FunctionType::get(valueType, {stateArg->getType(), ctx.Ctx->getType()}, false);
+        const auto pullPtr = CastInst::Create(Instruction::IntToPtr, pull, PointerType::getUnqual(pullType), "pull", block);
+        const auto list = CallInst::Create(pullType, pullPtr, {stateArg, ctx.Ctx}, "list", block);
+        UnRefBoxed(state, ctx, block);
+        result->addIncoming(list, block);
 
         new StoreInst(GetFinish(context), statePtr, block);
         BranchInst::Create(over, block);
diff --git a/yql/essentials/minikql/comp_nodes/mkql_switch.cpp b/yql/essentials/minikql/comp_nodes/mkql_switch.cpp
index 1dab916ca87a..f8c9c2e4602f 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_switch.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_switch.cpp
@@ -304,9 +304,6 @@ class TSwitchFlowWrapper : public TStatefulFlowCodegeneratorNode<TSwitchFlowWrap
 
         auto block = main;
 
-        const auto placeholder = NYql::NCodegen::ETarget::Windows == ctx.Codegen.GetEffectiveTarget() ?
-            new AllocaInst(valueType, 0U, "placeholder", block) : nullptr;
-
         const auto statePtr = GetElementPtrInst::CreateInBounds(valueType, ctx.GetMutables(), {ConstantInt::get(indexType, static_cast<const IComputationNode*>(this)->GetIndex())}, "state_ptr", block);
         const auto state = new LoadInst(valueType, statePtr, "state", block);
         const auto half = CastInst::Create(Instruction::Trunc, state, Type::getInt64Ty(context), "half", block);
@@ -326,17 +323,9 @@ class TSwitchFlowWrapper : public TStatefulFlowCodegeneratorNode<TSwitchFlowWrap
 
         const auto getFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TFlowState::Get));
 
-        Value* input;
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto getType = FunctionType::get(valueType, {stateArg->getType(), pos->getType()}, false);
-            const auto getPtr = CastInst::Create(Instruction::IntToPtr, getFunc, PointerType::getUnqual(getType), "get", block);
-            input = CallInst::Create(getType, getPtr, {stateArg, pos}, "input", block);
-        } else {
-            const auto getType = FunctionType::get(Type::getVoidTy(context), {stateArg->getType(), placeholder->getType(), pos->getType()}, false);
-            const auto getPtr = CastInst::Create(Instruction::IntToPtr, getFunc, PointerType::getUnqual(getType), "get", block);
-            CallInst::Create(getType, getPtr, {stateArg, placeholder, pos}, "", block);
-            input = new LoadInst(valueType, placeholder, "input", block);
-        }
+        const auto getType = FunctionType::get(valueType, {stateArg->getType(), pos->getType()}, false);
+        const auto getPtr = CastInst::Create(Instruction::IntToPtr, getFunc, PointerType::getUnqual(getType), "get", block);
+        const auto input = CallInst::Create(getType, getPtr, {stateArg, pos}, "input", block);
 
         const auto special = SwitchInst::Create(input, good, 2U, block);
         special->addCase(GetYield(context), back);
@@ -469,7 +458,7 @@ class TSwitchFlowWrapper : public TStatefulFlowCodegeneratorNode<TSwitchFlowWrap
             block = good;
 
             const auto addFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TFlowState::Add));
-            const auto addArg = WrapArgumentForWindows(item, ctx, block);
+            const auto addArg = item;
             const auto addType = FunctionType::get(Type::getVoidTy(context), {stateArg->getType(), addArg->getType()}, false);
             const auto addPtr = CastInst::Create(Instruction::IntToPtr, addFunc, PointerType::getUnqual(addType), "add", block);
             CallInst::Create(addType, addPtr, {stateArg, addArg}, "", block);
@@ -812,7 +801,7 @@ class TSwitchWrapper : public TCustomValueCodegeneratorNode<TSwitchWrapper<IsInp
 
         const auto valueType = Type::getInt128Ty(context);
         const auto ptrValueType = PointerType::getUnqual(valueType);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(ptrValueType) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = Type::getInt32Ty(context);
         const auto indexType = Type::getInt32Ty(context);
@@ -883,8 +872,7 @@ class TSwitchWrapper : public TCustomValueCodegeneratorNode<TSwitchWrapper<IsInp
 
             const auto used = GetMemoryUsed(MemLimit, ctx, block);
 
-            const auto stream = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-                new LoadInst(valueType, containerArg, "load_container", false, block) : static_cast<Value*>(containerArg);
+            const auto stream = static_cast<Value*>(containerArg);
 
             BranchInst::Create(loop, block);
 
diff --git a/yql/essentials/minikql/comp_nodes/mkql_take.cpp b/yql/essentials/minikql/comp_nodes/mkql_take.cpp
index 3e19ee2932d8..f2e9bc7023d8 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_take.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_take.cpp
@@ -265,20 +265,10 @@ class TTakeWrapper : public TMutableCodegeneratorNode<TTakeWrapper> {
         const auto cnt = GetNodeValue(Count, ctx, block);
         const auto count = GetterFor<ui64>(cnt, context, block);
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(list->getType(), {factory->getType(), builder->getType(), list->getType(), count->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            const auto result = CallInst::Create(funType, funcPtr, {factory, builder, list, count}, "result", block);
-            return result;
-        } else {
-            const auto retPtr = new AllocaInst(list->getType(), 0U, "ret_ptr", block);
-            new StoreInst(list, retPtr, block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {factory->getType(), retPtr->getType(), builder->getType(), retPtr->getType(), count->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            CallInst::Create(funType, funcPtr, {factory, retPtr, builder, retPtr, count}, "", block);
-            const auto result = new LoadInst(list->getType(), retPtr, "result", block);
-            return result;
-        }
+        const auto funType = FunctionType::get(list->getType(), {factory->getType(), builder->getType(), list->getType(), count->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+        const auto result = CallInst::Create(funType, funcPtr, {factory, builder, list, count}, "result", block);
+        return result;
     }
 #endif
 private:
diff --git a/yql/essentials/minikql/comp_nodes/mkql_todict.cpp b/yql/essentials/minikql/comp_nodes/mkql_todict.cpp
index 0ac6b83a11e5..10d0b4609c06 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_todict.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_todict.cpp
@@ -1065,7 +1065,7 @@ class TSqueezeSetFlowWrapper : public TStatefulFlowCodegeneratorNode<TSqueezeSet
 
         const auto insert = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TState::Insert));
 
-        const auto keyArg = WrapArgumentForWindows(key, ctx, block);
+        const auto keyArg = key;
 
         const auto insType = FunctionType::get(Type::getVoidTy(context), {stateArg->getType(), keyArg->getType()}, false);
         const auto insPtr = CastInst::Create(Instruction::IntToPtr, insert, PointerType::getUnqual(insType), "insert", block);
@@ -1077,21 +1077,11 @@ class TSqueezeSetFlowWrapper : public TStatefulFlowCodegeneratorNode<TSqueezeSet
 
         const auto build = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TState::Build));
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(valueType, {stateArg->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, build, PointerType::getUnqual(funType), "build", block);
-            const auto dict = CallInst::Create(funType, funcPtr, {stateArg}, "dict", block);
-            UnRefBoxed(state, ctx, block);
-            result->addIncoming(dict, block);
-        } else {
-            const auto ptr = new AllocaInst(valueType, 0U, "ptr", block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {stateArg->getType(), ptr->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, build, PointerType::getUnqual(funType), "build", block);
-            CallInst::Create(funType, funcPtr, {stateArg, ptr}, "", block);
-            const auto dict = new LoadInst(valueType, ptr, "dict", block);
-            UnRefBoxed(state, ctx, block);
-            result->addIncoming(dict, block);
-        }
+        const auto funType = FunctionType::get(valueType, {stateArg->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, build, PointerType::getUnqual(funType), "build", block);
+        const auto dict = CallInst::Create(funType, funcPtr, {stateArg}, "dict", block);
+        UnRefBoxed(state, ctx, block);
+        result->addIncoming(dict, block);
 
         new StoreInst(item, statePtr, block);
         BranchInst::Create(over, block);
@@ -1262,7 +1252,7 @@ class TSqueezeSetWideWrapper : public TStatefulFlowCodegeneratorNode<TSqueezeSet
 
         const auto insert = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TState::Insert));
 
-        const auto keyArg = WrapArgumentForWindows(key, ctx, block);
+        const auto keyArg = key;
 
         const auto insType = FunctionType::get(Type::getVoidTy(context), {stateArg->getType(), keyArg->getType()}, false);
         const auto insPtr = CastInst::Create(Instruction::IntToPtr, insert, PointerType::getUnqual(insType), "insert", block);
@@ -1274,21 +1264,11 @@ class TSqueezeSetWideWrapper : public TStatefulFlowCodegeneratorNode<TSqueezeSet
 
         const auto build = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TState::Build));
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(valueType, {stateArg->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, build, PointerType::getUnqual(funType), "build", block);
-            const auto dict = CallInst::Create(funType, funcPtr, {stateArg}, "dict", block);
-            UnRefBoxed(state, ctx, block);
-            result->addIncoming(dict, block);
-        } else {
-            const auto ptr = new AllocaInst(valueType, 0U, "ptr", block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {stateArg->getType(), ptr->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, build, PointerType::getUnqual(funType), "build", block);
-            CallInst::Create(funType, funcPtr, {stateArg, ptr}, "", block);
-            const auto dict = new LoadInst(valueType, ptr, "dict", block);
-            UnRefBoxed(state, ctx, block);
-            result->addIncoming(dict, block);
-        }
+        const auto funType = FunctionType::get(valueType, {stateArg->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, build, PointerType::getUnqual(funType), "build", block);
+        const auto dict = CallInst::Create(funType, funcPtr, {stateArg}, "dict", block);
+        UnRefBoxed(state, ctx, block);
+        result->addIncoming(dict, block);
 
         new StoreInst(GetFinish(context), statePtr, block);
         BranchInst::Create(over, block);
@@ -1580,8 +1560,8 @@ class TSqueezeMapFlowWrapper : public TStatefulFlowCodegeneratorNode<TSqueezeMap
 
         const auto insert = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TState::Insert));
 
-        const auto keyArg = WrapArgumentForWindows(key, ctx, block);
-        const auto payloadArg = WrapArgumentForWindows(payload, ctx, block);
+        const auto keyArg = key;
+        const auto payloadArg = payload;
 
         const auto insType = FunctionType::get(Type::getVoidTy(context), {stateArg->getType(), keyArg->getType(), payloadArg->getType()}, false);
         const auto insPtr = CastInst::Create(Instruction::IntToPtr, insert, PointerType::getUnqual(insType), "insert", block);
@@ -1593,21 +1573,11 @@ class TSqueezeMapFlowWrapper : public TStatefulFlowCodegeneratorNode<TSqueezeMap
 
         const auto build = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TState::Build));
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(valueType, {stateArg->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, build, PointerType::getUnqual(funType), "build", block);
-            const auto dict = CallInst::Create(funType, funcPtr, {stateArg}, "dict", block);
-            UnRefBoxed(state, ctx, block);
-            result->addIncoming(dict, block);
-        } else {
-            const auto ptr = new AllocaInst(valueType, 0U, "ptr", block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {stateArg->getType(), ptr->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, build, PointerType::getUnqual(funType), "build", block);
-            CallInst::Create(funType, funcPtr, {stateArg, ptr}, "", block);
-            const auto dict = new LoadInst(valueType, ptr, "dict", block);
-            UnRefBoxed(state, ctx, block);
-            result->addIncoming(dict, block);
-        }
+        const auto funType = FunctionType::get(valueType, {stateArg->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, build, PointerType::getUnqual(funType), "build", block);
+        const auto dict = CallInst::Create(funType, funcPtr, {stateArg}, "dict", block);
+        UnRefBoxed(state, ctx, block);
+        result->addIncoming(dict, block);
 
         new StoreInst(item, statePtr, block);
         BranchInst::Create(over, block);
@@ -1785,8 +1755,8 @@ class TSqueezeMapWideWrapper : public TStatefulFlowCodegeneratorNode<TSqueezeMap
 
         const auto insert = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TState::Insert));
 
-        const auto keyArg = WrapArgumentForWindows(key, ctx, block);
-        const auto payloadArg = WrapArgumentForWindows(payload, ctx, block);
+        const auto keyArg = key;
+        const auto payloadArg = payload;
 
         const auto insType = FunctionType::get(Type::getVoidTy(context), {stateArg->getType(), keyArg->getType(), payloadArg->getType()}, false);
         const auto insPtr = CastInst::Create(Instruction::IntToPtr, insert, PointerType::getUnqual(insType), "insert", block);
@@ -1798,21 +1768,11 @@ class TSqueezeMapWideWrapper : public TStatefulFlowCodegeneratorNode<TSqueezeMap
 
         const auto build = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TState::Build));
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(valueType, {stateArg->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, build, PointerType::getUnqual(funType), "build", block);
-            const auto dict = CallInst::Create(funType, funcPtr, {stateArg}, "dict", block);
-            UnRefBoxed(state, ctx, block);
-            result->addIncoming(dict, block);
-        } else {
-            const auto ptr = new AllocaInst(valueType, 0U, "ptr", block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {stateArg->getType(), ptr->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, build, PointerType::getUnqual(funType), "build", block);
-            CallInst::Create(funType, funcPtr, {stateArg, ptr}, "", block);
-            const auto dict = new LoadInst(valueType, ptr, "dict", block);
-            UnRefBoxed(state, ctx, block);
-            result->addIncoming(dict, block);
-        }
+        const auto funType = FunctionType::get(valueType, {stateArg->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, build, PointerType::getUnqual(funType), "build", block);
+        const auto dict = CallInst::Create(funType, funcPtr, {stateArg}, "dict", block);
+        UnRefBoxed(state, ctx, block);
+        result->addIncoming(dict, block);
 
         new StoreInst(GetFinish(context), statePtr, block);
         BranchInst::Create(over, block);
diff --git a/yql/essentials/minikql/comp_nodes/mkql_toindexdict.cpp b/yql/essentials/minikql/comp_nodes/mkql_toindexdict.cpp
index 978446be08d9..2bf6be670532 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_toindexdict.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_toindexdict.cpp
@@ -38,20 +38,10 @@ class TToIndexDictWrapper : public TMutableCodegeneratorNode<TToIndexDictWrapper
 
         const auto list = GetNodeValue(List, ctx, block);
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(list->getType(), {factory->getType(), builder->getType(), list->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            const auto result = CallInst::Create(funType, funcPtr, {factory, builder, list}, "result", block);
-            return result;
-        } else {
-            const auto retPtr = new AllocaInst(list->getType(), 0U, "ret_ptr", block);
-            new StoreInst(list, retPtr, block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {factory->getType(), retPtr->getType(), builder->getType(), retPtr->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            CallInst::Create(funType, funcPtr, {factory, retPtr, builder, retPtr}, "", block);
-            const auto result = new LoadInst(list->getType(), retPtr, "result", block);
-            return result;
-        }
+        const auto funType = FunctionType::get(list->getType(), {factory->getType(), builder->getType(), list->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+        const auto result = CallInst::Create(funType, funcPtr, {factory, builder, list}, "result", block);
+        return result;
     }
 #endif
 private:
diff --git a/yql/essentials/minikql/comp_nodes/mkql_tostring.cpp b/yql/essentials/minikql/comp_nodes/mkql_tostring.cpp
index b7155f01ffae..130aaf00c5ea 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_tostring.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_tostring.cpp
@@ -64,13 +64,11 @@ class TDecimalToStringWrapper : public TMutableCodegeneratorNode<TDecimalToStrin
 
         const auto valType = Type::getInt128Ty(context);
         const auto psType = Type::getInt8Ty(context);
-        const auto valTypePtr = PointerType::getUnqual(valType);
 
         const auto name = "DecimalToString";
         ctx.Codegen.AddGlobalMapping(name, reinterpret_cast<const void*>(&DecimalToString));
-        const auto fnType = NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget() ?
-            FunctionType::get(valType, { valType, psType, psType }, false):
-            FunctionType::get(Type::getVoidTy(context), { valTypePtr, valTypePtr, psType, psType }, false);
+        const auto fnType =
+            FunctionType::get(valType, { valType, psType, psType }, false);
         const auto func = ctx.Codegen.GetModule().getOrInsertFunction(name, fnType);
 
         const auto fail = BasicBlock::Create(context, "fail", ctx.Func);
@@ -93,29 +91,14 @@ class TDecimalToStringWrapper : public TMutableCodegeneratorNode<TDecimalToStrin
 
             block = call;
 
-            Value* string;
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                string = CallInst::Create(func, { GetterForInt128(value, block), precision, scale }, "to_string", block);
-            } else {
-                const auto retPtr = new AllocaInst(valType, 0U, "ret_ptr", block);
-                new StoreInst(GetterForInt128(value, block), retPtr, block);
-                CallInst::Create(func, { retPtr, retPtr, precision, scale }, "", block);
-                string = new LoadInst(valType, retPtr, "res", block);
-            }
+            const auto string = CallInst::Create(func, { GetterForInt128(value, block), precision, scale }, "to_string", block);
 
             res->addIncoming(string, block);
 
             const auto test = CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ, string, zero, "test", block);
             BranchInst::Create(fail, nice, test, block);
         } else {
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                result = CallInst::Create(func, { GetterForInt128(value, block), precision, scale }, "to_string", block);
-            } else {
-                const auto retPtr = new AllocaInst(valType, 0U, "ret_ptr", block);
-                new StoreInst(GetterForInt128(value, block), retPtr, block);
-                CallInst::Create(func, { retPtr, retPtr, precision, scale }, "", block);
-                result = new LoadInst(valType, retPtr, "res", block);
-            }
+            result = CallInst::Create(func, { GetterForInt128(value, block), precision, scale }, "to_string", block);
 
             const auto test = CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ, result, zero, "test", block);
             BranchInst::Create(fail, nice, test, block);
@@ -170,13 +153,11 @@ class TToStringWrapper : public TMutableCodegeneratorNode<TToStringWrapper<IsOpt
 
         const auto valType = Type::getInt128Ty(context);
         const auto slotType = Type::getInt32Ty(context);
-        const auto valTypePtr = PointerType::getUnqual(valType);
 
         const auto name = "DataToString";
         ctx.Codegen.AddGlobalMapping(name, reinterpret_cast<const void*>(&DataToString));
-        const auto fnType = NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget() ?
-            FunctionType::get(valType, { valType, slotType }, false):
-            FunctionType::get(Type::getVoidTy(context), { valTypePtr, valTypePtr, slotType }, false);
+        const auto fnType =
+            FunctionType::get(valType, { valType, slotType }, false);
         const auto func = ctx.Codegen.GetModule().getOrInsertFunction(name, fnType);
 
         const auto zero = ConstantInt::get(valType, 0ULL);
@@ -194,15 +175,7 @@ class TToStringWrapper : public TMutableCodegeneratorNode<TToStringWrapper<IsOpt
 
             block = call;
 
-            Value* string;
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                string = CallInst::Create(func, { value, slot }, "to_string", block);
-            } else {
-                const auto retPtr = new AllocaInst(valType, 0U, "ret_ptr", block);
-                new StoreInst(value, retPtr, block);
-                CallInst::Create(func, { retPtr, retPtr, slot }, "", block);
-                string = new LoadInst(valType, retPtr, "res", block);
-            }
+            const auto string = CallInst::Create(func, { value, slot }, "to_string", block);
 
             if (Data->IsTemporaryValue())
                 ValueCleanup(Data->GetRepresentation(), value, ctx, block);
@@ -213,15 +186,7 @@ class TToStringWrapper : public TMutableCodegeneratorNode<TToStringWrapper<IsOpt
             block = done;
             return result;
         } else {
-            Value* string;
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                string = CallInst::Create(func, { value, slot }, "to_string", block);
-            } else {
-                const auto retPtr = new AllocaInst(valType, 0U, "ret_ptr", block);
-                new StoreInst(value, retPtr, block);
-                CallInst::Create(func, { retPtr, retPtr, slot}, "", block);
-                string = new LoadInst(valType, retPtr, "res", block);
-            }
+            const auto string = CallInst::Create(func, { value, slot }, "to_string", block);
 
             if (Data->IsTemporaryValue())
                 ValueCleanup(Data->GetRepresentation(), value, ctx, block);
diff --git a/yql/essentials/minikql/comp_nodes/mkql_while.cpp b/yql/essentials/minikql/comp_nodes/mkql_while.cpp
index 907af8f6252c..233abf7e896d 100644
--- a/yql/essentials/minikql/comp_nodes/mkql_while.cpp
+++ b/yql/essentials/minikql/comp_nodes/mkql_while.cpp
@@ -315,7 +315,7 @@ class TBaseWhileWrapper {
             return f;
 
         const auto valueType = Type::getInt128Ty(context);
-        const auto containerType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ? static_cast<Type*>(PointerType::getUnqual(valueType)) : static_cast<Type*>(valueType);
+        const auto containerType = static_cast<Type*>(valueType);
         const auto contextType = GetCompContextType(context);
         const auto statusType = IsStream ? Type::getInt32Ty(context) : Type::getInt1Ty(context);
         const auto funcType = FunctionType::get(statusType, {PointerType::getUnqual(contextType), containerType, PointerType::getUnqual(valueType), PointerType::getUnqual(valueType)}, false);
@@ -335,8 +335,7 @@ class TBaseWhileWrapper {
         const auto main = BasicBlock::Create(context, "main", ctx.Func);
         auto block = main;
 
-        const auto container = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-            new LoadInst(valueType, containerArg, "load_container", false, block) : static_cast<Value*>(containerArg);
+        const auto container = static_cast<Value*>(containerArg);
 
         const auto good = BasicBlock::Create(context, "good", ctx.Func);
         const auto stop = BasicBlock::Create(context, "stop", ctx.Func);
@@ -610,20 +609,10 @@ class TListWhileWrapper : public TBothWaysCodegeneratorNode<TListWhileWrapper<Sk
             const auto doFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TListWhileWrapper::MakeLazyList));
             const auto ptrType = PointerType::getUnqual(StructType::get(context));
             const auto self = CastInst::Create(Instruction::IntToPtr, ConstantInt::get(Type::getInt64Ty(context), uintptr_t(this)), ptrType, "self", block);
-            if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-                const auto funType = FunctionType::get(list->getType() , {self->getType(), ctx.Ctx->getType(), list->getType()}, false);
-                const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-                const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx, list}, "value", block);
-                out->addIncoming(value, block);
-            } else {
-                const auto resultPtr = new AllocaInst(list->getType(), 0U, "return", block);
-                new StoreInst(list, resultPtr, block);
-                const auto funType = FunctionType::get(Type::getVoidTy(context), {self->getType(), resultPtr->getType(), ctx.Ctx->getType(), resultPtr->getType()}, false);
-                const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-                CallInst::Create(funType, doFuncPtr, {self, resultPtr, ctx.Ctx, resultPtr}, "", block);
-                const auto value = new LoadInst(list->getType(), resultPtr, "value", block);
-                out->addIncoming(value, block);
-            }
+            const auto funType = FunctionType::get(list->getType() , {self->getType(), ctx.Ctx->getType(), list->getType()}, false);
+            const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
+            const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx, list}, "value", block);
+            out->addIncoming(value, block);
             BranchInst::Create(done, block);
         }
 
diff --git a/yql/essentials/minikql/computation/mkql_computation_node_codegen.cpp b/yql/essentials/minikql/computation/mkql_computation_node_codegen.cpp
index 7379efe3f5c6..23be45fa194c 100644
--- a/yql/essentials/minikql/computation/mkql_computation_node_codegen.cpp
+++ b/yql/essentials/minikql/computation/mkql_computation_node_codegen.cpp
@@ -137,13 +137,11 @@ Function* GenerateCompareFunction(NYql::NCodegen::ICodegen& codegen, const TStri
 
     auto& context = codegen.GetContext();
     const auto valueType = Type::getInt128Ty(context);
-    const auto ptrType = PointerType::getUnqual(valueType);
     const auto returnType = Type::getInt1Ty(context);
     const auto contextType = GetCompContextType(context);
 
-    const auto funcType = codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows ?
-        FunctionType::get(returnType, {PointerType::getUnqual(contextType), valueType, valueType}, false):
-        FunctionType::get(returnType, {PointerType::getUnqual(contextType), ptrType, ptrType}, false);
+    const auto funcType =
+        FunctionType::get(returnType, {PointerType::getUnqual(contextType), valueType, valueType}, false);
 
     TCodegenContext ctx(codegen);
     ctx.AlwaysInline = true;
@@ -602,12 +600,10 @@ Function* GenerateEqualsFunction(NYql::NCodegen::ICodegen& codegen, const TStrin
 
     auto& context = codegen.GetContext();
     const auto valueType = Type::getInt128Ty(context);
-    const auto ptrType = PointerType::getUnqual(valueType);
     const auto returnType = Type::getInt1Ty(context);
 
-    const auto funcType = codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows ?
-        FunctionType::get(returnType, {valueType, valueType}, false):
-        FunctionType::get(returnType, {ptrType, ptrType}, false);
+    const auto funcType =
+        FunctionType::get(returnType, {valueType, valueType}, false);
 
     TCodegenContext ctx(codegen);
     ctx.AlwaysInline = true;
@@ -733,12 +729,10 @@ Function* GenerateHashFunction(NYql::NCodegen::ICodegen& codegen, const TString&
 
     auto& context = codegen.GetContext();
     const auto valueType = Type::getInt128Ty(context);
-    const auto ptrType = PointerType::getUnqual(valueType);
     const auto returnType = Type::getInt64Ty(context);
 
-    const auto funcType = codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows ?
-        FunctionType::get(returnType, {valueType}, false):
-        FunctionType::get(returnType, {ptrType}, false);
+    const auto funcType =
+        FunctionType::get(returnType, {valueType}, false);
 
     TCodegenContext ctx(codegen);
     ctx.AlwaysInline = true;
@@ -1069,9 +1063,8 @@ Function* TExternalCodegeneratorRootNode::GenerateGetValue(NYql::NCodegen::ICode
     const auto valueType = Type::getInt128Ty(context);
     const auto contextType = GetCompContextType(context);
 
-    const auto funcType = codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows ?
-        FunctionType::get(valueType, {PointerType::getUnqual(contextType)}, false):
-        FunctionType::get(Type::getVoidTy(context), {PointerType::getUnqual(valueType), PointerType::getUnqual(contextType)}, false);
+    const auto funcType =
+        FunctionType::get(valueType, {PointerType::getUnqual(contextType)}, false);
 
     TCodegenContext ctx(codegen);
     ctx.Func = cast<Function>(module.getOrInsertFunction(name.c_str(), funcType).getCallee());
@@ -1079,25 +1072,13 @@ Function* TExternalCodegeneratorRootNode::GenerateGetValue(NYql::NCodegen::ICode
     DISubprogramAnnotator annotator(ctx, ctx.Func);
 
     auto args = ctx.Func->arg_begin();
-    if (codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows) {
-        auto& firstArg = *args++;
-        firstArg.addAttr(Attribute::StructRet);
-        firstArg.addAttr(Attribute::NoAlias);
-    }
-
     auto main = BasicBlock::Create(context, "main", ctx.Func);
     ctx.Ctx = &*args;
     ctx.Ctx->addAttr(Attribute::NonNull);
 
     const auto get = CreateGetValue(ctx, main);
 
-    if (codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows) {
-        ReturnInst::Create(context, get, main);
-    } else {
-        new StoreInst(get, &*--args, main);
-        ReturnInst::Create(context, main);
-    }
-
+    ReturnInst::Create(context, get, main);
     return ctx.Func;
 }
 
@@ -1110,8 +1091,7 @@ Function* TExternalCodegeneratorRootNode::GenerateSetValue(NYql::NCodegen::ICode
 
     const auto intType = Type::getInt128Ty(context);
     const auto contextType = GetCompContextType(context);
-    const auto valueType = codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
-        (Type*)PointerType::getUnqual(intType) : (Type*)intType;
+    const auto valueType = (Type*)intType;
 
     const auto funcType = FunctionType::get(Type::getVoidTy(context), {PointerType::getUnqual(contextType), valueType}, false);
     TCodegenContext ctx(codegen);
@@ -1127,12 +1107,7 @@ Function* TExternalCodegeneratorRootNode::GenerateSetValue(NYql::NCodegen::ICode
 
     const auto valueArg = &*++args;
 
-    if (codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows) {
-        const auto value = new LoadInst(valueArg->getType(), valueArg, "load_value", main);
-        CreateSetValue(ctx, main, value);
-    } else {
-        CreateSetValue(ctx, main, valueArg);
-    }
+    CreateSetValue(ctx, main, valueArg);
     ReturnInst::Create(context, main);
     return ctx.Func;
 }
@@ -1436,7 +1411,7 @@ template<> Type* GetTypeFor<double>(LLVMContext &context) { return Type::getDoub
 
 void AddRefBoxed(Value* value, const TCodegenContext& ctx, BasicBlock*& block) {
     auto& context = ctx.Codegen.GetContext();
-    const auto load = value->getType()->isPointerTy() ? new LoadInst(value->getType(), value, "load", block) : value;
+    const auto load = value->getType()->isPointerTy() ? new LoadInst(Type::getInt128Ty(context), value, "load", block) : value;
     const auto half = CastInst::Create(Instruction::Trunc, load, Type::getInt64Ty(context), "half", block);
     const auto counterType = Type::getInt32Ty(context);
     const auto type = StructType::get(context, {PointerType::getUnqual(StructType::get(context)), counterType, Type::getInt16Ty(context)});
@@ -1468,7 +1443,7 @@ void AddRefBoxed(Value* value, const TCodegenContext& ctx, BasicBlock*& block) {
 
 void UnRefBoxed(Value* value, const TCodegenContext& ctx, BasicBlock*& block) {
     auto& context = ctx.Codegen.GetContext();
-    const auto load = value->getType()->isPointerTy() ? new LoadInst(value->getType(), value, "load", block) : value;
+    const auto load = value->getType()->isPointerTy() ? new LoadInst(Type::getInt128Ty(context), value, "load", block) : value;
     const auto half = CastInst::Create(Instruction::Trunc, load, Type::getInt64Ty(context), "half", block);
     const auto counterType = Type::getInt32Ty(context);
     const auto type = StructType::get(context, {PointerType::getUnqual(StructType::get(context)), counterType, Type::getInt16Ty(context)});
@@ -1511,7 +1486,7 @@ void UnRefBoxed(Value* value, const TCodegenContext& ctx, BasicBlock*& block) {
 
 void CleanupBoxed(Value* value, const TCodegenContext& ctx, BasicBlock*& block) {
     auto& context = ctx.Codegen.GetContext();
-    const auto load = value->getType()->isPointerTy() ? new LoadInst(value->getType(), value, "load", block) : value;
+    const auto load = value->getType()->isPointerTy() ? new LoadInst(Type::getInt128Ty(context), value, "load", block) : value;
     const auto half = CastInst::Create(Instruction::Trunc, load, Type::getInt64Ty(context), "half", block);
     const auto counterType = Type::getInt32Ty(context);
     const auto type = StructType::get(context, {PointerType::getUnqual(StructType::get(context)), counterType, Type::getInt16Ty(context)});
@@ -1893,20 +1868,10 @@ Value* MakeVariant(Value* item, Value* variant, const TCodegenContext& ctx, Basi
         const auto factory = ctx.GetFactory();
         const auto func = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&THolderFactory::CreateBoxedVariantHolder));
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto signature = FunctionType::get(item->getType(), {factory->getType(), item->getType(), variant->getType()}, false);
-            const auto creator = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(signature), "creator", block);
-            const auto output = CallInst::Create(signature, creator, {factory, item, variant}, "output", block);
-            result->addIncoming(output, block);
-        } else {
-            const auto place = new AllocaInst(item->getType(), 0U, "place", block);
-            new StoreInst(item, place, block);
-            const auto signature = FunctionType::get(Type::getVoidTy(context), {factory->getType(), place->getType(), place->getType(), variant->getType()}, false);
-            const auto creator = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(signature), "creator", block);
-            CallInst::Create(signature, creator, {factory, place, place, variant}, "", block);
-            const auto output = new LoadInst(item->getType(), place, "output", block);
-            result->addIncoming(output, block);
-        }
+        const auto signature = FunctionType::get(item->getType(), {factory->getType(), item->getType(), variant->getType()}, false);
+        const auto creator = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(signature), "creator", block);
+        const auto output = CallInst::Create(signature, creator, {factory, item, variant}, "output", block);
+        result->addIncoming(output, block);
 
         BranchInst::Create(done, block);
     }
@@ -1926,9 +1891,8 @@ Value* GetNodeValue(IComputationNode* node, const TCodegenContext& ctx, BasicBlo
 
     const auto valueType = Type::getInt128Ty(context);
     const auto retPtr = new AllocaInst(valueType, 0U, "return_ptr", &ctx.Func->getEntryBlock().back());
-    const auto funType = ctx.Codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows ?
-        FunctionType::get(Type::getVoidTy(context), {retPtr->getType(), nodeThis->getType(), ctx.Ctx->getType()}, false):
-        FunctionType::get(Type::getVoidTy(context), {nodeThis->getType(), retPtr->getType(), ctx.Ctx->getType()}, false);
+    const auto funType =
+        FunctionType::get(Type::getVoidTy(context), {retPtr->getType(), nodeThis->getType(), ctx.Ctx->getType()}, false);
     const auto ptrFunType = PointerType::getUnqual(funType);
     const auto tableType = PointerType::getUnqual(ptrFunType);
     const auto nodeVTable = CastInst::Create(Instruction::IntToPtr, ptr, PointerType::getUnqual(tableType), "node_vtable", block);
@@ -1937,11 +1901,7 @@ Value* GetNodeValue(IComputationNode* node, const TCodegenContext& ctx, BasicBlo
     const auto elem = GetElementPtrInst::CreateInBounds(ptrFunType, table, {ConstantInt::get(Type::getInt64Ty(context), GetMethodIndex(&IComputationNode::GetValue))}, "element", block);
     const auto func = new LoadInst(ptrFunType, elem, "func", false, block);
 
-    if (ctx.Codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows) {
-        CallInst::Create(funType, func, {retPtr, nodeThis, ctx.Ctx}, "", block);
-    } else {
-        CallInst::Create(funType, func, {nodeThis, retPtr, ctx.Ctx}, "", block);
-    }
+    CallInst::Create(funType, func, {retPtr, nodeThis, ctx.Ctx}, "", block);
 
     ValueRelease(node->GetRepresentation(), retPtr, ctx, block);
     const auto result = new LoadInst(valueType, retPtr, "return", false, block);
@@ -1961,9 +1921,9 @@ void GetNodeValue(Value* value, IComputationNode* node, const TCodegenContext& c
     const auto ptrType = PointerType::getUnqual(StructType::get(context));
     const auto nodeThis = CastInst::Create(Instruction::IntToPtr, ptr, ptrType, "node_this", block);
 
-    const auto funType = ctx.Codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows ?
-        FunctionType::get(Type::getVoidTy(context), {value->getType(), nodeThis->getType(), ctx.Ctx->getType()}, false):
-        FunctionType::get(Type::getVoidTy(context), {nodeThis->getType(), value->getType(), ctx.Ctx->getType()}, false);
+    const auto funType =
+        FunctionType::get(Type::getVoidTy(context), {value->getType(), nodeThis->getType(), ctx.Ctx->getType()}, false);
+
     const auto ptrFunType = PointerType::getUnqual(funType);
     const auto tableType = PointerType::getUnqual(ptrFunType);
     const auto nodeVTable = CastInst::Create(Instruction::IntToPtr, ptr, PointerType::getUnqual(tableType), "node_vtable", block);
@@ -1972,11 +1932,7 @@ void GetNodeValue(Value* value, IComputationNode* node, const TCodegenContext& c
     const auto elem = GetElementPtrInst::CreateInBounds(ptrFunType, table, {ConstantInt::get(Type::getInt64Ty(context), GetMethodIndex(&IComputationNode::GetValue))}, "element", block);
     const auto func = new LoadInst(ptrFunType, elem, "func", false, block);
 
-    if (ctx.Codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows) {
-        CallInst::Create(funType, func, {value, nodeThis, ctx.Ctx}, "", block);
-    } else {
-        CallInst::Create(funType, func, {nodeThis, value, ctx.Ctx}, "", block);
-    }
+    CallInst::Create(funType, func, {value, nodeThis, ctx.Ctx}, "", block);
 }
 
 ICodegeneratorInlineWideNode::TGenerateResult GetNodeValues(IComputationWideFlowNode* node, const TCodegenContext& ctx, BasicBlock*& block) {
@@ -1990,17 +1946,9 @@ Value* GenNewArray(const TCodegenContext& ctx, Value* size, Value* items, BasicB
     const auto fact = ctx.GetFactory();
     const auto func = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&THolderFactory::CreateDirectArrayHolder));
     const auto valueType = Type::getInt128Ty(context);
-    if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-        const auto funType = FunctionType::get(valueType, {fact->getType(), size->getType(), items->getType()}, false);
-        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-        return CallInst::Create(funType, funcPtr, {fact, size, items}, "array", block);
-    } else {
-        const auto resultPtr = new AllocaInst(valueType, 0U, "return", block);
-        const auto funType = FunctionType::get(Type::getVoidTy(context), {fact->getType(), resultPtr->getType(), size->getType(), items->getType()}, false);
-        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-        CallInst::Create(funType, funcPtr, {fact, resultPtr, size, items}, "", block);
-        return new LoadInst(valueType, resultPtr, "array", block);
-    }
+    const auto funType = FunctionType::get(valueType, {fact->getType(), size->getType(), items->getType()}, false);
+    const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+    return CallInst::Create(funType, funcPtr, {fact, size, items}, "array", block);
 }
 
 Value* GetMemoryUsed(ui64 limit, const TCodegenContext& ctx, BasicBlock* block) {
@@ -2064,16 +2012,6 @@ Value* CheckAdjustedMemLimit(ui64 limit, Value* init, const TCodegenContext& ctx
 template Value* CheckAdjustedMemLimit<false>(ui64 limit, Value* init, const TCodegenContext& ctx, BasicBlock*& block);
 template Value* CheckAdjustedMemLimit<true>(ui64 limit, Value* init, const TCodegenContext& ctx, BasicBlock*& block);
 
-Value* WrapArgumentForWindows(Value* arg, const TCodegenContext& ctx, BasicBlock* block) {
-    if (ctx.Codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows) {
-        return arg;
-    }
-
-    const auto newArg = new AllocaInst(arg->getType(), 0, "argument", block);
-    new StoreInst(arg, newArg, block);
-    return newArg;
-}
-
 Value* CallBoxedValueVirtualMethodImpl(uintptr_t methodPtr, Type* returnType, Value* value, NYql::NCodegen::ICodegen& codegen, BasicBlock* block) {
     auto& context = codegen.GetContext();
 
@@ -2101,9 +2039,8 @@ void CallBoxedValueVirtualMethodImpl(uintptr_t methodPtr, Value* output, Value*
     const auto ptrStructType = PointerType::getUnqual(StructType::get(context));
     const auto boxed = CastInst::Create(Instruction::IntToPtr, data, ptrStructType, "boxed", block);
 
-    const auto funType = (codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows) ?
-        FunctionType::get(Type::getVoidTy(context), {output->getType(), boxed->getType()}, false):
-        FunctionType::get(Type::getVoidTy(context), {boxed->getType(), output->getType()}, false);
+    const auto funType =
+        FunctionType::get(Type::getVoidTy(context), {output->getType(), boxed->getType()}, false);
     const auto ptrFunType = PointerType::getUnqual(funType);
     const auto tableType = PointerType::getUnqual(ptrFunType);
     const auto vTable = CastInst::Create(Instruction::IntToPtr, data, PointerType::getUnqual(tableType), "vtable", block);
@@ -2112,11 +2049,7 @@ void CallBoxedValueVirtualMethodImpl(uintptr_t methodPtr, Value* output, Value*
     const auto elem = GetElementPtrInst::CreateInBounds(ptrFunType, table, {ConstantInt::get(Type::getInt64Ty(context), GetMethodPtrIndex(methodPtr))}, "element", block);
     const auto func = new LoadInst(ptrFunType, elem, "func", false, block);
 
-    if (codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows) {
-        CallInst::Create(funType, func, {output, boxed}, "", block);
-    } else {
-        CallInst::Create(funType, func, {boxed, output}, "", block);
-    }
+    CallInst::Create(funType, func, {output, boxed}, "", block);
 }
 
 void CallBoxedValueVirtualMethodImpl(uintptr_t methodPtr, Value* output, Value* value, NYql::NCodegen::ICodegen& codegen, BasicBlock* block, Value* argument) {
@@ -2126,9 +2059,8 @@ void CallBoxedValueVirtualMethodImpl(uintptr_t methodPtr, Value* output, Value*
     const auto ptrStructType = PointerType::getUnqual(StructType::get(context));
     const auto boxed = CastInst::Create(Instruction::IntToPtr, data, ptrStructType, "boxed", block);
 
-    const auto funType = (codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows) ?
-        FunctionType::get(Type::getVoidTy(context), {output->getType(), boxed->getType(), argument->getType()}, false):
-        FunctionType::get(Type::getVoidTy(context), {boxed->getType(), output->getType(), argument->getType()}, false);
+    const auto funType =
+        FunctionType::get(Type::getVoidTy(context), {output->getType(), boxed->getType(), argument->getType()}, false);
     const auto ptrFunType = PointerType::getUnqual(funType);
     const auto tableType = PointerType::getUnqual(ptrFunType);
     const auto vTable = CastInst::Create(Instruction::IntToPtr, data, PointerType::getUnqual(tableType), "vtable", block);
@@ -2137,11 +2069,7 @@ void CallBoxedValueVirtualMethodImpl(uintptr_t methodPtr, Value* output, Value*
     const auto elem = GetElementPtrInst::CreateInBounds(ptrFunType, table, {ConstantInt::get(Type::getInt64Ty(context), GetMethodPtrIndex(methodPtr))}, "element", block);
     const auto func = new LoadInst(ptrFunType, elem, "func", false, block);
 
-    if (codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows) {
-        CallInst::Create(funType, func, {output, boxed, argument}, "", block);
-    } else {
-        CallInst::Create(funType, func, {boxed, output, argument}, "", block);
-    }
+    CallInst::Create(funType, func, {output, boxed, argument}, "", block);
 }
 
 Value* CallBoxedValueVirtualMethodImpl(uintptr_t methodPtr, Type* returnType, Value* value, NYql::NCodegen::ICodegen& codegen, BasicBlock* block, Value* argument) {
@@ -2171,9 +2099,8 @@ void CallBoxedValueVirtualMethodImpl(uintptr_t methodPtr, Value* output, Value*
     const auto ptrStructType = PointerType::getUnqual(StructType::get(context));
     const auto boxed = CastInst::Create(Instruction::IntToPtr, data, ptrStructType, "boxed", block);
 
-    const auto funType = (codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows) ?
-        FunctionType::get(Type::getVoidTy(context), {output->getType(), boxed->getType(), arg1->getType(), arg2->getType()}, false):
-        FunctionType::get(Type::getVoidTy(context), {boxed->getType(), output->getType(), arg1->getType(), arg2->getType()}, false);
+    const auto funType =
+        FunctionType::get(Type::getVoidTy(context), {output->getType(), boxed->getType(), arg1->getType(), arg2->getType()}, false);
     const auto ptrFunType = PointerType::getUnqual(funType);
     const auto tableType = PointerType::getUnqual(ptrFunType);
     const auto vTable = CastInst::Create(Instruction::IntToPtr, data, PointerType::getUnqual(tableType), "vtable", block);
@@ -2182,11 +2109,7 @@ void CallBoxedValueVirtualMethodImpl(uintptr_t methodPtr, Value* output, Value*
     const auto elem = GetElementPtrInst::CreateInBounds(ptrFunType, table, {ConstantInt::get(Type::getInt64Ty(context), GetMethodPtrIndex(methodPtr))}, "element", block);
     const auto func = new LoadInst(ptrFunType, elem, "func", false, block);
 
-    if (codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows) {
-        CallInst::Create(funType, func, {output, boxed, arg1, arg2}, "", block);
-    } else {
-        CallInst::Create(funType, func, {boxed, output, arg1, arg2}, "", block);
-    }
+    CallInst::Create(funType, func, {output, boxed, arg1, arg2}, "", block);
 }
 
 Value* CallBoxedValueVirtualMethodImpl(uintptr_t methodPtr, Type* returnType, Value* value, NYql::NCodegen::ICodegen& codegen, BasicBlock* block, Value* arg1, Value* arg2) {
@@ -2212,59 +2135,19 @@ Value* CallBoxedValueVirtualMethodImpl(uintptr_t methodPtr, Type* returnType, Va
 Value* CallUnaryUnboxedValueFunctionImpl(uintptr_t methodPtr, Type* result, Value* arg, NYql::NCodegen::ICodegen& codegen, BasicBlock* block) {
     auto& context = codegen.GetContext();
     const auto doFunc = ConstantInt::get(Type::getInt64Ty(context), methodPtr);
-    if (NYql::NCodegen::ETarget::Windows != codegen.GetEffectiveTarget()) {
-        const auto funType = FunctionType::get(result, {arg->getType()}, false);
-        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "ptr", block);
-        const auto call = CallInst::Create(funType, funcPtr, {arg}, "call", block);
-        return call;
-    } else {
-        const auto ptrArg = new AllocaInst(arg->getType(), 0U, "arg", block);
-        new StoreInst(arg, ptrArg, block);
-
-        if (Type::getInt128Ty(context) == result) {
-            const auto ptrResult = new AllocaInst(result, 0U, "result", block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {ptrResult->getType(), ptrArg->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "ptr", block);
-            CallInst::Create(funType, funcPtr, {ptrResult, ptrArg}, "", block);
-            const auto res = new LoadInst(result, ptrResult, "res", block);
-            return res;
-        } else {
-            const auto funType = FunctionType::get(result, {ptrArg->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "ptr", block);
-            const auto call = CallInst::Create(funType, funcPtr, {ptrArg}, "call", block);
-            return call;
-        }
-    }
+    const auto funType = FunctionType::get(result, {arg->getType()}, false);
+    const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "ptr", block);
+    const auto call = CallInst::Create(funType, funcPtr, {arg}, "call", block);
+    return call;
 }
 
 Value* CallBinaryUnboxedValueFunctionImpl(uintptr_t methodPtr, Type* result, Value* left, Value* right, NYql::NCodegen::ICodegen& codegen, BasicBlock* block) {
     auto& context = codegen.GetContext();
     const auto doFunc = ConstantInt::get(Type::getInt64Ty(context), methodPtr);
-    if (NYql::NCodegen::ETarget::Windows != codegen.GetEffectiveTarget()) {
-        const auto funType = FunctionType::get(result, {left->getType(), right->getType()}, false);
-        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "ptr", block);
-        const auto call = CallInst::Create(funType, funcPtr, {left, right}, "call", block);
-        return call;
-    } else {
-        const auto ptrLeft = new AllocaInst(left->getType(), 0U, "left", block);
-        const auto ptrRight = new AllocaInst(right->getType(), 0U, "right", block);
-        new StoreInst(left, ptrLeft, block);
-        new StoreInst(right, ptrRight, block);
-
-        if (Type::getInt128Ty(context) == result) {
-            const auto ptrResult = new AllocaInst(result, 0U, "result", block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {ptrResult->getType(), ptrLeft->getType(), ptrRight->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "ptr", block);
-            CallInst::Create(funType, funcPtr, {ptrResult, ptrLeft, ptrRight}, "", block);
-            const auto res = new LoadInst(result, ptrResult, "res", block);
-            return res;
-        } else {
-            const auto funType = FunctionType::get(result, {ptrLeft->getType(), ptrRight->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "ptr", block);
-            const auto call = CallInst::Create(funType, funcPtr, {ptrLeft, ptrRight}, "call", block);
-            return call;
-        }
-    }
+    const auto funType = FunctionType::get(result, {left->getType(), right->getType()}, false);
+    const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "ptr", block);
+    const auto call = CallInst::Create(funType, funcPtr, {left, right}, "call", block);
+    return call;
 }
 
 Y_NO_INLINE Value* TDecoratorCodegeneratorNodeBase::CreateGetValueImpl(IComputationNode* node,
@@ -2541,19 +2424,10 @@ Y_NO_INLINE Value* TMutableCodegeneratorFallbackNodeBase::DoGenerateGetValueImpl
     const auto ptrType = PointerType::getUnqual(StructType::get(context));
     const auto doFunc = ConstantInt::get(Type::getInt64Ty(context), methodPtr);
     const auto self = CastInst::Create(Instruction::IntToPtr, ConstantInt::get(Type::getInt64Ty(context), thisPtr), ptrType, "self", block);
-    if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-        const auto funType = FunctionType::get(type, {self->getType(), ctx.Ctx->getType()}, false);
-        const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-        const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx}, "value", block);
-        return value;
-    } else {
-        const auto resultPtr = new AllocaInst(type, 0U, "return", block);
-        const auto funType = FunctionType::get(Type::getVoidTy(context), {self->getType(), resultPtr->getType(), ctx.Ctx->getType()}, false);
-        const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
-        CallInst::Create(funType, doFuncPtr, {self, resultPtr, ctx.Ctx}, "", block);
-        const auto value = new LoadInst(type, resultPtr, "value", block);
-        return value;
-    }
+    const auto funType = FunctionType::get(type, {self->getType(), ctx.Ctx->getType()}, false);
+    const auto doFuncPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "function", block);
+    const auto value = CallInst::Create(funType, doFuncPtr, {self, ctx.Ctx}, "value", block);
+    return value;
 }
 
 Y_NO_INLINE Function* TCodegeneratorRootNodeBase::GenerateGetValueImpl(
@@ -2567,9 +2441,8 @@ Y_NO_INLINE Function* TCodegeneratorRootNodeBase::GenerateGetValueImpl(
     const auto valueType = Type::getInt128Ty(context);
     const auto contextType = GetCompContextType(context);
 
-    const auto funcType = codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows ?
-        FunctionType::get(valueType, {PointerType::getUnqual(contextType)}, false):
-        FunctionType::get(Type::getVoidTy(context) , {PointerType::getUnqual(valueType), PointerType::getUnqual(contextType)}, false);
+    const auto funcType =
+        FunctionType::get(valueType, {PointerType::getUnqual(contextType)}, false);
 
     TCodegenContext ctx(codegen);
     ctx.Func = cast<Function>(module.getOrInsertFunction(name.c_str(), funcType).getCallee());
@@ -2577,25 +2450,13 @@ Y_NO_INLINE Function* TCodegeneratorRootNodeBase::GenerateGetValueImpl(
     DISubprogramAnnotator annotator(ctx, ctx.Func);
 
     auto args = ctx.Func->arg_begin();
-    if (codegen.GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows) {
-        auto& firstArg = *args++;
-        firstArg.addAttr(Attribute::StructRet);
-        firstArg.addAttr(Attribute::NoAlias);
-    }
-
     auto main = BasicBlock::Create(context, "main", ctx.Func);
     ctx.Ctx = &*args;
     ctx.Ctx->addAttr(Attribute::NonNull);
 
     const auto get = gen->CreateGetValue(ctx, main);
 
-    if (codegen.GetEffectiveTarget() != NYql::NCodegen::ETarget::Windows) {
-        ReturnInst::Create(context, get, main);
-    } else {
-        new StoreInst(get, &*--args, main);
-        ReturnInst::Create(context, main);
-    }
-
+    ReturnInst::Create(context, get, main);
     return ctx.Func;
 }
 
diff --git a/yql/essentials/minikql/computation/mkql_computation_node_codegen.h.txt b/yql/essentials/minikql/computation/mkql_computation_node_codegen.h.txt
index ebd63455ebbe..8835b3b65183 100644
--- a/yql/essentials/minikql/computation/mkql_computation_node_codegen.h.txt
+++ b/yql/essentials/minikql/computation/mkql_computation_node_codegen.h.txt
@@ -173,8 +173,6 @@ public:
     virtual void CreateRun(const TCodegenContext& ctx, BasicBlock*& block, Value* result, Value* args) const = 0;
 };
 
-Value* WrapArgumentForWindows(Value* arg, const TCodegenContext& ctx, BasicBlock* block);
-
 Value* CallBoxedValueVirtualMethodImpl(uintptr_t methodPtr, Type* returnType, Value* value, NYql::NCodegen::ICodegen& codegen, BasicBlock* block);
 
 template<NUdf::TBoxedValueAccessor::EMethod Method>
diff --git a/yql/essentials/minikql/computation/mkql_computation_node_graph.cpp b/yql/essentials/minikql/computation/mkql_computation_node_graph.cpp
index f9209ad23971..e1a79dff5b10 100644
--- a/yql/essentials/minikql/computation/mkql_computation_node_graph.cpp
+++ b/yql/essentials/minikql/computation/mkql_computation_node_graph.cpp
@@ -814,8 +814,8 @@ class TComputationPatternImpl final : public IComputationPattern {
         : Codegen((NYql::NCodegen::ICodegen::IsCodegenAvailable() && opts.OptLLVM != "OFF") || GetEnv(TString("MKQL_FORCE_USE_LLVM")) ? NYql::NCodegen::ICodegen::MakeShared(NYql::NCodegen::ETarget::Native) : NYql::NCodegen::ICodegen::TPtr())
 #endif
     {
-    /// TODO: Enable JIT for AARCH64
-#if defined(__aarch64__)
+    /// TODO: Enable JIT for AARCH64/Win
+#if defined(__aarch64__) || defined(_win_)
         Codegen = {};
 #endif
 
diff --git a/yql/essentials/minikql/computation/mkql_computation_node_holders_codegen.cpp b/yql/essentials/minikql/computation/mkql_computation_node_holders_codegen.cpp
index 3fe7772ba9d1..cead70dec0a9 100644
--- a/yql/essentials/minikql/computation/mkql_computation_node_holders_codegen.cpp
+++ b/yql/essentials/minikql/computation/mkql_computation_node_holders_codegen.cpp
@@ -126,21 +126,12 @@ Value* TContainerCacheOnContext::GenNewArray(ui64 sz, Value* items, const TCodeg
         const auto func = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&THolderFactory::CreateDirectArrayHolder));
         const auto size = ConstantInt::get(Type::getInt64Ty(context), sz);
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(valueType, {fact->getType(), size->getType(), items->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            const auto array = CallInst::Create(funType, funcPtr, {fact, size, items}, "array", block);
-            AddRefBoxed(array, ctx, block);
-            result->addIncoming(array, block);
-            new StoreInst(array, tpsecond, block);
-        } else {
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {fact->getType(), tpsecond->getType(), size->getType(), items->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            CallInst::Create(funType, funcPtr, {fact, tpsecond, size, items}, "", block);
-            const auto array = new LoadInst(valueType, tpsecond, "array", block);
-            AddRefBoxed(array, ctx, block);
-            result->addIncoming(array, block);
-        }
+        const auto funType = FunctionType::get(valueType, {fact->getType(), size->getType(), items->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+        const auto array = CallInst::Create(funType, funcPtr, {fact, size, items}, "array", block);
+        AddRefBoxed(array, ctx, block);
+        result->addIncoming(array, block);
+        new StoreInst(array, tpsecond, block);
 
         BranchInst::Create(exit, block);
     }
@@ -189,19 +180,10 @@ class TEmptyNode : public TMutableCodegeneratorNode<TEmptyNode> {
         const auto factory = ctx.GetFactory();
         const auto func = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&THolderFactory::GetEmptyContainerLazy));
 
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(valueType, {factory->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            const auto res = CallInst::Create(funType, funcPtr, {factory}, "res", block);
-            return res;
-        } else {
-            const auto retPtr = new AllocaInst(valueType, 0U, "ret_ptr", block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {factory->getType(), retPtr->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
-            CallInst::Create(funType, funcPtr, {factory, retPtr}, "", block);
-            const auto res = new LoadInst(valueType, retPtr, "res", block);
-            return res;
-        }
+        const auto funType = FunctionType::get(valueType, {factory->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, func, PointerType::getUnqual(funType), "function", block);
+        const auto res = CallInst::Create(funType, funcPtr, {factory}, "res", block);
+        return res;
     }
 #endif
 private:
diff --git a/yql/essentials/minikql/invoke_builtins/mkql_builtins_find.cpp b/yql/essentials/minikql/invoke_builtins/mkql_builtins_find.cpp
index 1224ecfc59a5..6a9120c302e1 100644
--- a/yql/essentials/minikql/invoke_builtins/mkql_builtins_find.cpp
+++ b/yql/essentials/minikql/invoke_builtins/mkql_builtins_find.cpp
@@ -34,23 +34,10 @@ struct TFind {
                 StaticCast<ui32, std::string_view::size_type>(GetterFor<ui32>(p, context, block), context, block),
             "pos", block):
             StaticCast<ui32, std::string_view::size_type>(GetterFor<ui32>(p, context, block), context, block);
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(string->getType(), {string->getType(), sub->getType(), pos->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "func", block);
-            const auto result = CallInst::Create(funType, funcPtr, {string, sub, pos}, "find", block);
-            return result;
-        } else {
-            const auto ptrArg = new AllocaInst(string->getType(), 0U, "arg", block);
-            const auto ptrSub = new AllocaInst(sub->getType(), 0U, "sub", block);
-            const auto ptrResult = new AllocaInst(string->getType(), 0U, "result", block);
-            new StoreInst(string, ptrArg, block);
-            new StoreInst(sub, ptrSub, block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {ptrResult->getType(), ptrArg->getType(), ptrSub->getType(), pos->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "func", block);
-            CallInst::Create(funType, funcPtr, {ptrResult, ptrArg, ptrSub, pos}, "", block);
-            const auto result = new LoadInst(string->getType(), ptrResult, "find", block);
-            return result;
-        }
+        const auto funType = FunctionType::get(string->getType(), {string->getType(), sub->getType(), pos->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "func", block);
+        const auto result = CallInst::Create(funType, funcPtr, {string, sub, pos}, "find", block);
+        return result;
     }
 #endif
 };
diff --git a/yql/essentials/minikql/invoke_builtins/mkql_builtins_substring.cpp b/yql/essentials/minikql/invoke_builtins/mkql_builtins_substring.cpp
index 2b292f6b36cd..8a0c6f6f586e 100644
--- a/yql/essentials/minikql/invoke_builtins/mkql_builtins_substring.cpp
+++ b/yql/essentials/minikql/invoke_builtins/mkql_builtins_substring.cpp
@@ -35,21 +35,10 @@ struct TSubString {
                 GetterFor<ui32>(cn, context, block), "count", block
             ):
             GetterFor<ui32>(cn, context, block);
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(string->getType(), {string->getType(), start->getType(), count->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "func", block);
-            const auto result = CallInst::Create(funType, funcPtr, {string, start, count}, "substring", block);
-            return result;
-        } else {
-            const auto ptrArg = new AllocaInst(string->getType(), 0U, "arg", block);
-            const auto ptrResult = new AllocaInst(string->getType(), 0U, "result", block);
-            new StoreInst(string, ptrArg, block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {ptrResult->getType(), ptrArg->getType(), start->getType(), count->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "func", block);
-            CallInst::Create(funType, funcPtr, {ptrResult, ptrArg, start, count}, "", block);
-            const auto result = new LoadInst(string->getType(), ptrResult, "substring", block);
-            return result;
-        }
+        const auto funType = FunctionType::get(string->getType(), {string->getType(), start->getType(), count->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "func", block);
+        const auto result = CallInst::Create(funType, funcPtr, {string, start, count}, "substring", block);
+        return result;
     }
 #endif
 };
diff --git a/yql/essentials/minikql/invoke_builtins/mkql_builtins_with.cpp b/yql/essentials/minikql/invoke_builtins/mkql_builtins_with.cpp
index 723bff5cab66..60927c0e187e 100644
--- a/yql/essentials/minikql/invoke_builtins/mkql_builtins_with.cpp
+++ b/yql/essentials/minikql/invoke_builtins/mkql_builtins_with.cpp
@@ -29,23 +29,10 @@ struct TStringWith {
     {
         auto& context = ctx.Codegen.GetContext();
         const auto doFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(StringFunc));
-        if (NYql::NCodegen::ETarget::Windows != ctx.Codegen.GetEffectiveTarget()) {
-            const auto funType = FunctionType::get(string->getType(), {string->getType(), sub->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "func", block);
-            const auto result = CallInst::Create(funType, funcPtr, {string, sub}, "has", block);
-            return result;
-        } else {
-            const auto ptrArg = new AllocaInst(string->getType(), 0U, "arg", block);
-            const auto ptrSub = new AllocaInst(sub->getType(), 0U, "sub", block);
-            const auto ptrResult = new AllocaInst(string->getType(), 0U, "result", block);
-            new StoreInst(string, ptrArg, block);
-            new StoreInst(sub, ptrSub, block);
-            const auto funType = FunctionType::get(Type::getVoidTy(context), {ptrResult->getType(), ptrArg->getType(), ptrSub->getType()}, false);
-            const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "func", block);
-            CallInst::Create(funType, funcPtr, {ptrResult, ptrArg, ptrSub}, "", block);
-            const auto result = new LoadInst(string->getType(), ptrResult, "has", block);
-            return result;
-        }
+        const auto funType = FunctionType::get(string->getType(), {string->getType(), sub->getType()}, false);
+        const auto funcPtr = CastInst::Create(Instruction::IntToPtr, doFunc, PointerType::getUnqual(funType), "func", block);
+        const auto result = CallInst::Create(funType, funcPtr, {string, sub}, "has", block);
+        return result;
     }
 #endif
 };
diff --git a/yql/essentials/minikql/mkql_program_builder.cpp b/yql/essentials/minikql/mkql_program_builder.cpp
index f59e3371fb5e..82c1f604ba1b 100644
--- a/yql/essentials/minikql/mkql_program_builder.cpp
+++ b/yql/essentials/minikql/mkql_program_builder.cpp
@@ -4600,7 +4600,8 @@ TRuntimeNode TProgramBuilder::ToDecimal(TRuntimeNode data, ui8 precision, ui8 sc
         } else if (params.second < scale) {
             return Invoke("ScaleUp_" + ::ToString(scale - params.second), decimal, args);
         } else if (params.second > scale) {
-            return Invoke("ScaleDown_" + ::ToString(params.second - scale), decimal, args);
+            TRuntimeNode scaled = Invoke("ScaleDown_" + ::ToString(params.second - scale), decimal, args);
+            return Invoke("CheckBounds_" + ::ToString(precision), decimal, {{ scaled }});
         } else if (precision < params.first) {
             return Invoke("CheckBounds_" + ::ToString(precision), decimal, args);
         } else if (precision > params.first) {
diff --git a/yql/essentials/tests/sql/minirun/part0/canondata/result.json b/yql/essentials/tests/sql/minirun/part0/canondata/result.json
index 8b7d690d779e..948ea2a8737d 100644
--- a/yql/essentials/tests/sql/minirun/part0/canondata/result.json
+++ b/yql/essentials/tests/sql/minirun/part0/canondata/result.json
@@ -338,6 +338,20 @@
             "uri": "https://{canondata_backend}/1942525/ede9d81525f3cde3c09402fe9435fdbba85f47bc/resource.tar.gz#test.test_expr-as_dict_dict_key-default.txt-Results_/results.txt"
         }
     ],
+    "test.test[expr-cast_decimal-default.txt-Debug]": [
+        {
+            "checksum": "91294751a47cfcb87a5a3af5e0db632c",
+            "size": 1564,
+            "uri": "https://{canondata_backend}/1773845/486531303c5d3ec35ed8056fbdb4be070aadc061/resource.tar.gz#test.test_expr-cast_decimal-default.txt-Debug_/opt.yql"
+        }
+    ],
+    "test.test[expr-cast_decimal-default.txt-Results]": [
+        {
+            "checksum": "7bede4f19295253e7740f658c336afed",
+            "size": 14064,
+            "uri": "https://{canondata_backend}/1773845/486531303c5d3ec35ed8056fbdb4be070aadc061/resource.tar.gz#test.test_expr-cast_decimal-default.txt-Results_/results.txt"
+        }
+    ],
     "test.test[expr-cast_string_implicit-default.txt-Debug]": [
         {
             "checksum": "558f5bb0eaa3ea5ec6716c10389761b5",
diff --git a/yql/essentials/tests/sql/minirun/part1/canondata/result.json b/yql/essentials/tests/sql/minirun/part1/canondata/result.json
index 9db67a2417d7..88857905bb4c 100644
--- a/yql/essentials/tests/sql/minirun/part1/canondata/result.json
+++ b/yql/essentials/tests/sql/minirun/part1/canondata/result.json
@@ -601,6 +601,20 @@
             "uri": "https://{canondata_backend}/1936273/dc5c7fc322621a2828a94006237d3583f38248e5/resource.tar.gz#test.test_join-inmem_by_uncomparable_tuples-default.txt-Results_/results.txt"
         }
     ],
+    "test.test[join-yql-19192-default.txt-Debug]": [
+        {
+            "checksum": "b9c6aea35bc8b70c987653997335369d",
+            "size": 746,
+            "uri": "https://{canondata_backend}/1814674/dfd418faf4299c20acfad523d570f8cfb90f89ea/resource.tar.gz#test.test_join-yql-19192-default.txt-Debug_/opt.yql"
+        }
+    ],
+    "test.test[join-yql-19192-default.txt-Results]": [
+        {
+            "checksum": "6e4f62a8d64ea7bbe1a05464bcb80ba4",
+            "size": 1233,
+            "uri": "https://{canondata_backend}/1814674/dfd418faf4299c20acfad523d570f8cfb90f89ea/resource.tar.gz#test.test_join-yql-19192-default.txt-Results_/results.txt"
+        }
+    ],
     "test.test[library-library_yqls--Debug]": [
         {
             "checksum": "b2bab6a1ef79eb44b255b25d32270957",
diff --git a/yql/essentials/tests/sql/sql2yql/canondata/result.json b/yql/essentials/tests/sql/sql2yql/canondata/result.json
index 3c9da739188d..7cb5c7d39cbb 100644
--- a/yql/essentials/tests/sql/sql2yql/canondata/result.json
+++ b/yql/essentials/tests/sql/sql2yql/canondata/result.json
@@ -2008,6 +2008,27 @@
             "uri": "https://{canondata_backend}/1942173/99e88108149e222741552e7e6cddef041d6a2846/resource.tar.gz#test_sql2yql.test_expr-callable_/sql.yql"
         }
     ],
+    "test_sql2yql.test[expr-cast_decimal]": [
+        {
+            "checksum": "0beea95df85e9ea70b06ffa9d9eac69f",
+            "size": 4965,
+            "uri": "https://{canondata_backend}/1937001/ceef5344342e97b0173a7afba261a6905a152791/resource.tar.gz#test_sql2yql.test_expr-cast_decimal_/sql.yql"
+        }
+    ],
+    "test_sql2yql.test[expr-cast_decimal_implicit_fail_1]": [
+        {
+            "checksum": "a14d03775787d74c68ee4999aba0f644",
+            "size": 1282,
+            "uri": "https://{canondata_backend}/1937001/ceef5344342e97b0173a7afba261a6905a152791/resource.tar.gz#test_sql2yql.test_expr-cast_decimal_implicit_fail_1_/sql.yql"
+        }
+    ],
+    "test_sql2yql.test[expr-cast_decimal_implicit_fail_2]": [
+        {
+            "checksum": "d5c5926076f244e9fb0a01e213a839d3",
+            "size": 1288,
+            "uri": "https://{canondata_backend}/1937001/ceef5344342e97b0173a7afba261a6905a152791/resource.tar.gz#test_sql2yql.test_expr-cast_decimal_implicit_fail_2_/sql.yql"
+        }
+    ],
     "test_sql2yql.test[expr-cast_dynumber]": [
         {
             "checksum": "f8e5ceebd24d7c79be2988ed186d2b90",
@@ -3422,6 +3443,13 @@
             "uri": "https://{canondata_backend}/1942173/99e88108149e222741552e7e6cddef041d6a2846/resource.tar.gz#test_sql2yql.test_join-left_join_with_self_aggr_/sql.yql"
         }
     ],
+    "test_sql2yql.test[join-yql-19192]": [
+        {
+            "checksum": "fffdf1cbb40643da9daf9bdf3edec121",
+            "size": 2305,
+            "uri": "https://{canondata_backend}/1937367/af84d06d0fb3153f82b596e7db07f2186d2ec3a2/resource.tar.gz#test_sql2yql.test_join-yql-19192_/sql.yql"
+        }
+    ],
     "test_sql2yql.test[join-yql-6199]": [
         {
             "checksum": "5421f92ec269e5d76133e1ca54b6d0d0",
@@ -8196,6 +8224,21 @@
             "uri": "file://test_sql_format.test_expr-callable_/formatted.sql"
         }
     ],
+    "test_sql_format.test[expr-cast_decimal]": [
+        {
+            "uri": "file://test_sql_format.test_expr-cast_decimal_/formatted.sql"
+        }
+    ],
+    "test_sql_format.test[expr-cast_decimal_implicit_fail_1]": [
+        {
+            "uri": "file://test_sql_format.test_expr-cast_decimal_implicit_fail_1_/formatted.sql"
+        }
+    ],
+    "test_sql_format.test[expr-cast_decimal_implicit_fail_2]": [
+        {
+            "uri": "file://test_sql_format.test_expr-cast_decimal_implicit_fail_2_/formatted.sql"
+        }
+    ],
     "test_sql_format.test[expr-cast_dynumber]": [
         {
             "uri": "file://test_sql_format.test_expr-cast_dynumber_/formatted.sql"
@@ -9206,6 +9249,11 @@
             "uri": "file://test_sql_format.test_join-left_join_with_self_aggr_/formatted.sql"
         }
     ],
+    "test_sql_format.test[join-yql-19192]": [
+        {
+            "uri": "file://test_sql_format.test_join-yql-19192_/formatted.sql"
+        }
+    ],
     "test_sql_format.test[join-yql-6199]": [
         {
             "uri": "file://test_sql_format.test_join-yql-6199_/formatted.sql"
diff --git a/yql/essentials/tests/sql/sql2yql/canondata/test_sql_format.test_expr-cast_decimal_/formatted.sql b/yql/essentials/tests/sql/sql2yql/canondata/test_sql_format.test_expr-cast_decimal_/formatted.sql
new file mode 100644
index 000000000000..542ccae35641
--- /dev/null
+++ b/yql/essentials/tests/sql/sql2yql/canondata/test_sql_format.test_expr-cast_decimal_/formatted.sql
@@ -0,0 +1,50 @@
+--------------------
+-- Explicit casts --
+--------------------
+$strs = AsList('-1', '0', '1', '2.4', '99.77', '999.5', '-7777');
+
+$data = ListMap(
+    $strs, ($str) -> {
+        RETURN AsStruct($str AS str);
+    }
+);
+
+SELECT
+    dataset.*,
+
+    -- Changing range
+    CAST(dec_3_1 AS Decimal (5, 1)) AS upcast_3_1,
+    CAST(dec_5_2 AS Decimal (4, 2)) AS downcast_5_2,
+
+    -- Changing precision
+    CAST(dec_5_2 AS Decimal (3, 0)) AS rounded_5_2,
+    CAST(dec_7_0 AS Decimal (9, 2)) AS extended_7_0,
+
+    -- Combinations
+    CAST(dec_5_2 AS Decimal (6, 0)) AS upcast_rounded_5_2,
+    CAST(dec_3_1 AS Decimal (7, 3)) AS upcast_extended_3_1,
+    CAST(dec_5_2 AS Decimal (3, 1)) AS downcast_rounded_5_2,
+    CAST(dec_7_0 AS Decimal (5, 2)) AS downcast_extended_7_0
+FROM (
+    SELECT
+        str AS a_str,
+        CAST(str AS Decimal (3, 1)) AS dec_3_1,
+        CAST(str AS Decimal (5, 2)) AS dec_5_2,
+        CAST(str AS Decimal (7, 0)) AS dec_7_0
+    FROM
+        AS_TABLE($data)
+) AS dataset;
+
+--------------------
+-- Implicit  cast --
+--------------------
+$lambda = ($big_dec) -> {
+    RETURN $big_dec;
+};
+
+$func = CALLABLE (Callable<(Decimal (10, 5)) -> Decimal (10, 5)>, $lambda);
+$small_dec = Decimal('3.0', 2, 1);
+
+SELECT
+    $func($small_dec)
+;
diff --git a/yql/essentials/tests/sql/sql2yql/canondata/test_sql_format.test_expr-cast_decimal_implicit_fail_1_/formatted.sql b/yql/essentials/tests/sql/sql2yql/canondata/test_sql_format.test_expr-cast_decimal_implicit_fail_1_/formatted.sql
new file mode 100644
index 000000000000..80c87b460000
--- /dev/null
+++ b/yql/essentials/tests/sql/sql2yql/canondata/test_sql_format.test_expr-cast_decimal_implicit_fail_1_/formatted.sql
@@ -0,0 +1,12 @@
+/* postgres can not */
+/* custom error: Error: Implicit decimal cast would lose precision */
+$lambda = ($small_dec) -> {
+    RETURN $small_dec;
+};
+
+$func_3_1 = CALLABLE (Callable<(Decimal (3, 1)) -> Decimal (3, 1)>, $lambda);
+$dec_4_2 = Decimal('11.1', 4, 2);
+
+SELECT
+    $func_3_1($dec_4_2) AS try_round
+;
diff --git a/yql/essentials/tests/sql/sql2yql/canondata/test_sql_format.test_expr-cast_decimal_implicit_fail_2_/formatted.sql b/yql/essentials/tests/sql/sql2yql/canondata/test_sql_format.test_expr-cast_decimal_implicit_fail_2_/formatted.sql
new file mode 100644
index 000000000000..4691c1a4fd72
--- /dev/null
+++ b/yql/essentials/tests/sql/sql2yql/canondata/test_sql_format.test_expr-cast_decimal_implicit_fail_2_/formatted.sql
@@ -0,0 +1,12 @@
+/* postgres can not */
+/* custom error: Error: Implicit decimal cast would narrow the range */
+$lambda = ($small_dec) -> {
+    RETURN $small_dec;
+};
+
+$func_3_1 = CALLABLE (Callable<(Decimal (3, 1)) -> Decimal (3, 1)>, $lambda);
+$dec_5_1 = Decimal('22.2', 5, 1);
+
+SELECT
+    $func_3_1($dec_5_1) AS try_downcast
+;
diff --git a/yql/essentials/tests/sql/sql2yql/canondata/test_sql_format.test_join-yql-19192_/formatted.sql b/yql/essentials/tests/sql/sql2yql/canondata/test_sql_format.test_join-yql-19192_/formatted.sql
new file mode 100644
index 000000000000..aa600c52ab09
--- /dev/null
+++ b/yql/essentials/tests/sql/sql2yql/canondata/test_sql_format.test_join-yql-19192_/formatted.sql
@@ -0,0 +1,18 @@
+$samples1 = AsList(<|sampl: 1|>, <|sampl: 2|>);
+$sampling_cumulative_prob1 = AsList(<|color: 'blue', cum_prob: 3|>);
+
+$cumulative_bounds = (
+    SELECT
+        LAG(cum_prob) OVER () AS lower_cum_bound
+    FROM
+        AS_TABLE($sampling_cumulative_prob1)
+);
+
+SELECT
+    s.sampl,
+    cb.lower_cum_bound
+FROM
+    AS_TABLE($samples1) AS s
+CROSS JOIN
+    $cumulative_bounds AS cb
+;
diff --git a/yql/essentials/tests/sql/suites/expr/cast_decimal.sql b/yql/essentials/tests/sql/suites/expr/cast_decimal.sql
new file mode 100644
index 000000000000..ed5635406c64
--- /dev/null
+++ b/yql/essentials/tests/sql/suites/expr/cast_decimal.sql
@@ -0,0 +1,40 @@
+--------------------
+-- Explicit casts --
+--------------------
+$strs = AsList("-1", "0", "1", "2.4", "99.77", "999.5", "-7777");
+$data = ListMap($strs, ($str) -> { return AsStruct($str as str); });
+
+SELECT
+    dataset.*,
+
+    -- Changing range
+    CAST(dec_3_1 AS Decimal(5, 1)) AS upcast_3_1,
+    CAST(dec_5_2 AS Decimal(4, 2)) AS downcast_5_2,
+    
+    -- Changing precision
+    CAST(dec_5_2 AS Decimal(3, 0)) AS rounded_5_2,   
+    CAST(dec_7_0 AS Decimal(9, 2)) AS extended_7_0,
+
+    -- Combinations
+    CAST(dec_5_2 AS Decimal(6, 0)) AS upcast_rounded_5_2,
+    CAST(dec_3_1 AS Decimal(7, 3)) AS upcast_extended_3_1,
+    CAST(dec_5_2 AS Decimal(3, 1)) AS downcast_rounded_5_2,
+    CAST(dec_7_0 AS Decimal(5, 2)) AS downcast_extended_7_0
+FROM (
+    SELECT
+        str AS a_str,
+        CAST(str AS Decimal(3, 1)) AS dec_3_1,
+        CAST(str AS Decimal(5, 2)) AS dec_5_2,
+        CAST(str AS Decimal(7, 0)) AS dec_7_0
+    FROM AS_TABLE($data)
+) as dataset;
+
+--------------------
+-- Implicit  cast --
+--------------------
+$lambda = ($big_dec) -> { return $big_dec; };
+$func = Callable(Callable<(Decimal(10, 5))->Decimal(10, 5)>, $lambda);
+
+$small_dec = Decimal("3.0", 2, 1);
+
+SELECT $func($small_dec);
diff --git a/yql/essentials/tests/sql/suites/expr/cast_decimal_implicit_fail_1.cfg b/yql/essentials/tests/sql/suites/expr/cast_decimal_implicit_fail_1.cfg
new file mode 100644
index 000000000000..5dae597903c3
--- /dev/null
+++ b/yql/essentials/tests/sql/suites/expr/cast_decimal_implicit_fail_1.cfg
@@ -0,0 +1 @@
+xfail
diff --git a/yql/essentials/tests/sql/suites/expr/cast_decimal_implicit_fail_1.sql b/yql/essentials/tests/sql/suites/expr/cast_decimal_implicit_fail_1.sql
new file mode 100644
index 000000000000..ab58383d714c
--- /dev/null
+++ b/yql/essentials/tests/sql/suites/expr/cast_decimal_implicit_fail_1.sql
@@ -0,0 +1,8 @@
+/* postgres can not */
+/* custom error: Error: Implicit decimal cast would lose precision */
+$lambda = ($small_dec) -> { return $small_dec; };
+
+$func_3_1 = Callable(Callable<(Decimal(3, 1))->Decimal(3, 1)>, $lambda);
+$dec_4_2 = Decimal("11.1", 4, 2);
+
+SELECT $func_3_1($dec_4_2) AS try_round;
diff --git a/yql/essentials/tests/sql/suites/expr/cast_decimal_implicit_fail_2.cfg b/yql/essentials/tests/sql/suites/expr/cast_decimal_implicit_fail_2.cfg
new file mode 100644
index 000000000000..5dae597903c3
--- /dev/null
+++ b/yql/essentials/tests/sql/suites/expr/cast_decimal_implicit_fail_2.cfg
@@ -0,0 +1 @@
+xfail
diff --git a/yql/essentials/tests/sql/suites/expr/cast_decimal_implicit_fail_2.sql b/yql/essentials/tests/sql/suites/expr/cast_decimal_implicit_fail_2.sql
new file mode 100644
index 000000000000..9da5dd8c4e54
--- /dev/null
+++ b/yql/essentials/tests/sql/suites/expr/cast_decimal_implicit_fail_2.sql
@@ -0,0 +1,8 @@
+/* postgres can not */
+/* custom error: Error: Implicit decimal cast would narrow the range */
+$lambda = ($small_dec) -> { return $small_dec; };
+
+$func_3_1 = Callable(Callable<(Decimal(3, 1))->Decimal(3, 1)>, $lambda);
+$dec_5_1 = Decimal("22.2", 5, 1);
+
+SELECT $func_3_1($dec_5_1) AS try_downcast;
diff --git a/yql/essentials/tests/sql/suites/join/yql-19192.sql b/yql/essentials/tests/sql/suites/join/yql-19192.sql
new file mode 100644
index 000000000000..76cc78833dd0
--- /dev/null
+++ b/yql/essentials/tests/sql/suites/join/yql-19192.sql
@@ -0,0 +1,12 @@
+$samples1 = AsList(<|sampl: 1|>,  <|sampl: 2|>);
+$sampling_cumulative_prob1 = AsList(<|color: "blue", cum_prob: 3|>);
+
+$cumulative_bounds = SELECT
+    LAG(cum_prob) OVER () AS lower_cum_bound
+FROM AS_TABLE($sampling_cumulative_prob1);
+
+SELECT
+    s.sampl,
+    cb.lower_cum_bound
+FROM AS_TABLE($samples1) AS s
+CROSS JOIN $cumulative_bounds AS cb;
\ No newline at end of file
diff --git a/yql/essentials/udfs/common/math/math_udf.cpp b/yql/essentials/udfs/common/math/math_udf.cpp
index 1c2652320bb3..56e86de6288e 100644
--- a/yql/essentials/udfs/common/math/math_udf.cpp
+++ b/yql/essentials/udfs/common/math/math_udf.cpp
@@ -60,12 +60,21 @@ using TTaggedRoundingMode = NYql::NUdf::TTagged<ui32, TagRoundingMode>;
 #define MATH_UDF_MAP_WITHOUT_IR(XX)                                                                 \
     XX(NearbyInt, TOptional<i64>(TAutoMap<double>, TTaggedRoundingMode), 0)
 
+#ifdef DISABLE_IR
+#define MATH_STRICT_UDF(name, signature, optionalArgsCount)                                         \
+    SIMPLE_STRICT_UDF_WITH_OPTIONAL_ARGS(T##name, signature, optionalArgsCount) {                   \
+        TUnboxedValuePod res;                                                                       \
+        name##IR(this, &res, valueBuilder, args);                                                   \
+        return res;                                                                                 \
+    }
+#else
 #define MATH_STRICT_UDF(name, signature, optionalArgsCount)                                                       \
     SIMPLE_STRICT_UDF_WITH_IR(T##name, signature, optionalArgsCount, "/llvm_bc/Math", #name "IR") {               \
         TUnboxedValuePod res;                                                                                     \
         name##IR(this, &res, valueBuilder, args);                                                                 \
         return res;                                                                                               \
     }
+#endif
 
 #define MATH_STRICT_UDF_WITHOUT_IR(name, signature, optionalArgsCount)                                            \
     SIMPLE_STRICT_UDF_WITH_OPTIONAL_ARGS(T##name, signature, optionalArgsCount) {                                                    \
diff --git a/yql/essentials/udfs/common/math/ya.make b/yql/essentials/udfs/common/math/ya.make
index 4c0909d2393a..2e91b467d8ee 100644
--- a/yql/essentials/udfs/common/math/ya.make
+++ b/yql/essentials/udfs/common/math/ya.make
@@ -9,64 +9,70 @@ YQL_UDF_CONTRIB(math_udf)
     SRCS(
         math_udf.cpp
     )
+
+    IF (BUILD_TYPE != "release" AND BUILD_TYPE != "relwithdebinfo")
+        CFLAGS(-DDISABLE_IR)
+    ELSE()
+        USE_LLVM_BC14()
+        
+        LLVM_BC(
+            math_ir.cpp
+            lib/erfinv.cpp
+            NAME Math
+            SYMBOLS
+            PiIR
+            EIR
+            EpsIR
+            RoundDownwardIR
+            RoundToNearestIR
+            RoundTowardZeroIR
+            RoundUpwardIR
+            AbsIR
+            AcosIR
+            AsinIR
+            AsinhIR
+            AtanIR
+            CbrtIR
+            CeilIR
+            CosIR
+            CoshIR
+            ErfIR
+            ErfInvIR
+            ErfcInvIR
+            ExpIR
+            Exp2IR
+            FabsIR
+            FloorIR
+            LgammaIR
+            RintIR
+            SinIR
+            SinhIR
+            SqrtIR
+            TanIR
+            TanhIR
+            TgammaIR
+            TruncIR
+            IsFiniteIR
+            IsInfIR
+            IsNaNIR
+            Atan2IR
+            FmodIR
+            HypotIR
+            RemainderIR
+            PowIR
+            LdexpIR
+            LogIR
+            Log2IR
+            Log10IR
+            SigmoidIR
+            FuzzyEqualsIR
+            RoundIR
+            ModIR
+            RemIR
+        )
+
+    ENDIF()
     
-    USE_LLVM_BC14()
-    
-    LLVM_BC(
-        math_ir.cpp
-        lib/erfinv.cpp
-        NAME Math
-        SYMBOLS
-        PiIR
-        EIR
-        EpsIR
-        RoundDownwardIR
-        RoundToNearestIR
-        RoundTowardZeroIR
-        RoundUpwardIR
-        AbsIR
-        AcosIR
-        AsinIR
-        AsinhIR
-        AtanIR
-        CbrtIR
-        CeilIR
-        CosIR
-        CoshIR
-        ErfIR
-        ErfInvIR
-        ErfcInvIR
-        ExpIR
-        Exp2IR
-        FabsIR
-        FloorIR
-        LgammaIR
-        RintIR
-        SinIR
-        SinhIR
-        SqrtIR
-        TanIR
-        TanhIR
-        TgammaIR
-        TruncIR
-        IsFiniteIR
-        IsInfIR
-        IsNaNIR
-        Atan2IR
-        FmodIR
-        HypotIR
-        RemainderIR
-        PowIR
-        LdexpIR
-        LogIR
-        Log2IR
-        Log10IR
-        SigmoidIR
-        FuzzyEqualsIR
-        RoundIR
-        ModIR
-        RemIR
-    )
     
     PEERDIR(
         yql/essentials/udfs/common/math/lib
diff --git a/yt/cpp/mapreduce/client/batch_request_impl.cpp b/yt/cpp/mapreduce/client/batch_request_impl.cpp
index f45eee319bf2..e7456b918733 100644
--- a/yt/cpp/mapreduce/client/batch_request_impl.cpp
+++ b/yt/cpp/mapreduce/client/batch_request_impl.cpp
@@ -36,11 +36,11 @@ using ::NThreading::NewPromise;
 
 TBatchRequest::TBatchRequest(const TTransactionId& defaultTransaction, ::TIntrusivePtr<TClient> client)
     : DefaultTransaction_(defaultTransaction)
-    , Impl_(MakeIntrusive<THttpRawBatchRequest>(client->GetContext().Config))
+    , Impl_(client->GetRawClient()->CreateRawBatchRequest())
     , Client_(client)
 { }
 
-TBatchRequest::TBatchRequest(THttpRawBatchRequest* impl, ::TIntrusivePtr<TClient> client)
+TBatchRequest::TBatchRequest(IRawBatchRequest* impl, ::TIntrusivePtr<TClient> client)
     : Impl_(impl)
     , Client_(std::move(client))
 { }
@@ -189,7 +189,7 @@ TFuture<TCheckPermissionResponse> TBatchRequest::CheckPermission(
 
 void TBatchRequest::ExecuteBatch(const TExecuteBatchOptions& options)
 {
-    Impl_->ExecuteBatch(Client_->GetRetryPolicy()->CreatePolicyForGenericRequest(), Client_->GetContext(), options);
+    Impl_->ExecuteBatch(options);
 }
 
 ////////////////////////////////////////////////////////////////////////////////
diff --git a/yt/cpp/mapreduce/client/batch_request_impl.h b/yt/cpp/mapreduce/client/batch_request_impl.h
index d53eb196dd55..bbd993a925cc 100644
--- a/yt/cpp/mapreduce/client/batch_request_impl.h
+++ b/yt/cpp/mapreduce/client/batch_request_impl.h
@@ -22,10 +22,6 @@ struct TResponseInfo;
 class TClient;
 using TClientPtr = ::TIntrusivePtr<TClient>;
 
-namespace NRawClient {
-    class THttpRawBatchRequest;
-}
-
 ////////////////////////////////////////////////////////////////////////////////
 
 class TBatchRequest
@@ -119,11 +115,11 @@ class TBatchRequest
     virtual void ExecuteBatch(const TExecuteBatchOptions& executeBatch) override;
 
 private:
-    TBatchRequest(NDetail::NRawClient::THttpRawBatchRequest* impl, ::TIntrusivePtr<TClient> client);
+    TBatchRequest(IRawBatchRequest* impl, ::TIntrusivePtr<TClient> client);
 
 private:
     TTransactionId DefaultTransaction_;
-    ::TIntrusivePtr<NDetail::NRawClient::THttpRawBatchRequest> Impl_;
+    IRawBatchRequestPtr Impl_;
     std::unique_ptr<TBatchRequest> TmpWithTransaction_;
     ::TIntrusivePtr<TClient> Client_;
 
diff --git a/yt/cpp/mapreduce/client/client.cpp b/yt/cpp/mapreduce/client/client.cpp
index 9443ea51fceb..8e3eed7bdc18 100644
--- a/yt/cpp/mapreduce/client/client.cpp
+++ b/yt/cpp/mapreduce/client/client.cpp
@@ -276,7 +276,7 @@ void TClientBase::Concatenate(
 
 TRichYPath TClientBase::CanonizeYPath(const TRichYPath& path)
 {
-    return NRawClient::CanonizeYPath(ClientRetryPolicy_->CreatePolicyForGenericRequest(), Context_, path);
+    return NRawClient::CanonizeYPath(RawClient_, path);
 }
 
 TVector<TTableColumnarStatistics> TClientBase::GetTableColumnarStatistics(
diff --git a/yt/cpp/mapreduce/client/lock.cpp b/yt/cpp/mapreduce/client/lock.cpp
index 46306fa7377c..4eecfdb0d5b2 100644
--- a/yt/cpp/mapreduce/client/lock.cpp
+++ b/yt/cpp/mapreduce/client/lock.cpp
@@ -26,7 +26,7 @@ class TLockPollerItem
         , Acquired_(acquired)
     { }
 
-    void PrepareRequest(THttpRawBatchRequest* batchRequest) override
+    void PrepareRequest(IRawBatchRequest* batchRequest) override
     {
         LockState_ = batchRequest->Get(TTransactionId(), LockStateYPath_, TGetOptions());
     }
diff --git a/yt/cpp/mapreduce/client/operation.cpp b/yt/cpp/mapreduce/client/operation.cpp
index 8ace8a44bb59..ef7015a6f1b7 100644
--- a/yt/cpp/mapreduce/client/operation.cpp
+++ b/yt/cpp/mapreduce/client/operation.cpp
@@ -225,11 +225,10 @@ TStructuredJobTableList ApplyProtobufColumnFilters(
     }
 
     auto isDynamic = NRawClient::BatchTransform(
-        CreateDefaultRequestRetryPolicy(preparer.GetContext().Config),
-        preparer.GetContext(),
+        preparer.GetClient()->GetRawClient(),
         tableList,
-        [&] (NRawClient::THttpRawBatchRequest& batch, const auto& table) {
-            return batch.Get(preparer.GetTransactionId(), table.RichYPath->Path_ + "/@dynamic", TGetOptions());
+        [&] (IRawBatchRequestPtr batch, const auto& table) {
+            return batch->Get(preparer.GetTransactionId(), table.RichYPath->Path_ + "/@dynamic", TGetOptions());
         });
 
     auto newTableList = tableList;
@@ -281,8 +280,8 @@ TSimpleOperationIo CreateSimpleOperationIo(
         structuredJob,
         preparer,
         options,
-        CanonizeStructuredTableList(preparer.GetContext(), GetStructuredInputs(spec)),
-        CanonizeStructuredTableList(preparer.GetContext(), GetStructuredOutputs(spec)),
+        CanonizeStructuredTableList(preparer.GetClient()->GetRawClient(), GetStructuredInputs(spec)),
+        CanonizeStructuredTableList(preparer.GetClient()->GetRawClient(), GetStructuredOutputs(spec)),
         hints,
         nodeReaderFormat,
         GetColumnsUsedInOperation(spec));
@@ -304,8 +303,8 @@ TSimpleOperationIo CreateSimpleOperationIo(
         }
     };
 
-    auto inputs = NRawClient::CanonizeYPaths(/* retryPolicy */ nullptr, preparer.GetContext(), spec.GetInputs());
-    auto outputs = NRawClient::CanonizeYPaths(/* retryPolicy */ nullptr, preparer.GetContext(), spec.GetOutputs());
+    auto inputs = NRawClient::CanonizeYPaths(preparer.GetClient()->GetRawClient(), spec.GetInputs());
+    auto outputs = NRawClient::CanonizeYPaths(preparer.GetClient()->GetRawClient(), spec.GetOutputs());
 
     VerifyHasElements(inputs, "input");
     VerifyHasElements(outputs, "output");
@@ -318,7 +317,6 @@ TSimpleOperationIo CreateSimpleOperationIo(
             inputs,
             outputs,
             preparer.GetClient()->GetRawClient(),
-            preparer.GetContext(),
             preparer.GetClientRetryPolicy(),
             preparer.GetTransactionId()),
         &inputs,
@@ -497,7 +495,6 @@ TSimpleOperationIo CreateSimpleOperationIoHelper(
             structuredInputs,
             structuredOutputs,
             preparer.GetClient()->GetRawClient(),
-            preparer.GetContext(),
             preparer.GetClientRetryPolicy(),
             preparer.GetTransactionId()),
         &structuredInputs,
@@ -644,7 +641,7 @@ TNode BuildAutoMergeSpec(const TAutoMergeSpec& options)
     return result;
 }
 
-TNode BuildJobProfilerSpec(const TJobProfilerSpec& profilerSpec)
+[[maybe_unused]] TNode BuildJobProfilerSpec(const TJobProfilerSpec& profilerSpec)
 {
     TNode result;
     if (profilerSpec.ProfilingBinary_) {
@@ -1635,9 +1632,9 @@ void ExecuteMapReduce(
     TMapReduceOperationSpec spec = spec_;
 
     TMapReduceOperationIo operationIo;
-    auto structuredInputs = CanonizeStructuredTableList(preparer->GetContext(), spec.GetStructuredInputs());
-    auto structuredMapOutputs = CanonizeStructuredTableList(preparer->GetContext(), spec.GetStructuredMapOutputs());
-    auto structuredOutputs = CanonizeStructuredTableList(preparer->GetContext(), spec.GetStructuredOutputs());
+    auto structuredInputs = CanonizeStructuredTableList(preparer->GetClient()->GetRawClient(), spec.GetStructuredInputs());
+    auto structuredMapOutputs = CanonizeStructuredTableList(preparer->GetClient()->GetRawClient(), spec.GetStructuredMapOutputs());
+    auto structuredOutputs = CanonizeStructuredTableList(preparer->GetClient()->GetRawClient(), spec.GetStructuredOutputs());
 
     const bool inferOutputSchema = options.InferOutputSchema_.GetOrElse(preparer->GetContext().Config->InferTableSchema);
 
@@ -1688,7 +1685,6 @@ void ExecuteMapReduce(
                 structuredInputs,
                 mapperOutput,
                 preparer->GetClient()->GetRawClient(),
-                preparer->GetContext(),
                 preparer->GetClientRetryPolicy(),
                 preparer->GetTransactionId()),
             &structuredInputs,
@@ -1758,7 +1754,6 @@ void ExecuteMapReduce(
                     inputs,
                     outputs,
                     preparer->GetClient()->GetRawClient(),
-                    preparer->GetContext(),
                     preparer->GetClientRetryPolicy(),
                     preparer->GetTransactionId()),
                 &inputs,
@@ -1824,7 +1819,6 @@ void ExecuteMapReduce(
                 structuredInputs,
                 structuredOutputs,
                 preparer->GetClient()->GetRawClient(),
-                preparer->GetContext(),
                 preparer->GetClientRetryPolicy(),
                 preparer->GetTransactionId()),
             &structuredInputs,
@@ -1904,9 +1898,9 @@ void ExecuteRawMapReduce(
     YT_LOG_DEBUG("Starting raw map-reduce operation (PreparationId: %v)",
         preparer->GetPreparationId());
     TMapReduceOperationIo operationIo;
-    operationIo.Inputs = NRawClient::CanonizeYPaths(/* retryPolicy */ nullptr, preparer->GetContext(), spec.GetInputs());
-    operationIo.MapOutputs = NRawClient::CanonizeYPaths(/* retryPolicy */ nullptr, preparer->GetContext(), spec.GetMapOutputs());
-    operationIo.Outputs = NRawClient::CanonizeYPaths(/* retryPolicy */ nullptr, preparer->GetContext(), spec.GetOutputs());
+    operationIo.Inputs = NRawClient::CanonizeYPaths(preparer->GetClient()->GetRawClient(), spec.GetInputs());
+    operationIo.MapOutputs = NRawClient::CanonizeYPaths(preparer->GetClient()->GetRawClient(), spec.GetMapOutputs());
+    operationIo.Outputs = NRawClient::CanonizeYPaths(preparer->GetClient()->GetRawClient(), spec.GetOutputs());
 
     VerifyHasElements(operationIo.Inputs, "inputs");
     VerifyHasElements(operationIo.Outputs, "outputs");
@@ -1953,8 +1947,8 @@ void ExecuteSort(
 {
     YT_LOG_DEBUG("Starting sort operation (PreparationId: %v)",
         preparer->GetPreparationId());
-    auto inputs = NRawClient::CanonizeYPaths(/* retryPolicy */ nullptr, preparer->GetContext(), spec.Inputs_);
-    auto output = NRawClient::CanonizeYPath(nullptr, preparer->GetContext(), spec.Output_);
+    auto inputs = NRawClient::CanonizeYPaths(preparer->GetClient()->GetRawClient(), spec.Inputs_);
+    auto output = NRawClient::CanonizeYPath(preparer->GetClient()->GetRawClient(), spec.Output_);
 
     if (options.CreateOutputTables_) {
         CheckInputTablesExist(*preparer, inputs);
@@ -2002,8 +1996,8 @@ void ExecuteMerge(
 {
     YT_LOG_DEBUG("Starting merge operation (PreparationId: %v)",
         preparer->GetPreparationId());
-    auto inputs = NRawClient::CanonizeYPaths(/* retryPolicy */ nullptr, preparer->GetContext(), spec.Inputs_);
-    auto output = NRawClient::CanonizeYPath(nullptr, preparer->GetContext(), spec.Output_);
+    auto inputs = NRawClient::CanonizeYPaths(preparer->GetClient()->GetRawClient(), spec.Inputs_);
+    auto output = NRawClient::CanonizeYPath(preparer->GetClient()->GetRawClient(), spec.Output_);
 
     if (options.CreateOutputTables_) {
         CheckInputTablesExist(*preparer, inputs);
@@ -2052,7 +2046,7 @@ void ExecuteErase(
 {
     YT_LOG_DEBUG("Starting erase operation (PreparationId: %v)",
         preparer->GetPreparationId());
-    auto tablePath = NRawClient::CanonizeYPath(nullptr, preparer->GetContext(), spec.TablePath_);
+    auto tablePath = NRawClient::CanonizeYPath(preparer->GetClient()->GetRawClient(), spec.TablePath_);
 
     TNode specNode = BuildYsonNodeFluently()
     .BeginMap()
@@ -2088,8 +2082,8 @@ void ExecuteRemoteCopy(
 {
     YT_LOG_DEBUG("Starting remote copy operation (PreparationId: %v)",
         preparer->GetPreparationId());
-    auto inputs = NRawClient::CanonizeYPaths(/* retryPolicy */ nullptr, preparer->GetContext(), spec.Inputs_);
-    auto output = NRawClient::CanonizeYPath(nullptr, preparer->GetContext(), spec.Output_);
+    auto inputs = NRawClient::CanonizeYPaths(preparer->GetClient()->GetRawClient(), spec.Inputs_);
+    auto output = NRawClient::CanonizeYPath(preparer->GetClient()->GetRawClient(), spec.Output_);
 
     if (options.CreateOutputTables_) {
         CreateOutputTable(*preparer, output);
@@ -2343,7 +2337,7 @@ class TOperationPollerItem
         : OperationImpl_(std::move(operationImpl))
     { }
 
-    void PrepareRequest(NRawClient::THttpRawBatchRequest* batchRequest) override
+    void PrepareRequest(IRawBatchRequest* batchRequest) override
     {
         auto filter = TOperationAttributeFilter()
             .Add(EOperationAttribute::State)
diff --git a/yt/cpp/mapreduce/client/operation_preparer.cpp b/yt/cpp/mapreduce/client/operation_preparer.cpp
index ac6ea0f9238b..c1907a703781 100644
--- a/yt/cpp/mapreduce/client/operation_preparer.cpp
+++ b/yt/cpp/mapreduce/client/operation_preparer.cpp
@@ -42,7 +42,7 @@ class TWaitOperationStartPollerItem
         , Transaction_(std::move(transaction))
     { }
 
-    void PrepareRequest(NRawClient::THttpRawBatchRequest* batchRequest) override
+    void PrepareRequest(IRawBatchRequest* batchRequest) override
     {
         Future_ = batchRequest->GetOperation(
             OperationId_,
@@ -213,26 +213,26 @@ void TOperationPreparer::LockFiles(TVector<TRichYPath>* paths)
 
     TVector<::NThreading::TFuture<TLockId>> lockIdFutures;
     lockIdFutures.reserve(paths->size());
-    NRawClient::THttpRawBatchRequest lockRequest(GetContext().Config);
+    auto lockRequest = Client_->GetRawClient()->CreateRawBatchRequest();
     for (const auto& path : *paths) {
-        lockIdFutures.push_back(lockRequest.Lock(
+        lockIdFutures.push_back(lockRequest->Lock(
             FileTransaction_->GetId(),
             path.Path_,
             ELockMode::LM_SNAPSHOT,
             TLockOptions().Waitable(true)));
     }
-    lockRequest.ExecuteBatch(ClientRetryPolicy_->CreatePolicyForGenericRequest(), GetContext());
+    lockRequest->ExecuteBatch();
 
     TVector<::NThreading::TFuture<TNode>> nodeIdFutures;
     nodeIdFutures.reserve(paths->size());
-    NRawClient::THttpRawBatchRequest getNodeIdRequest(GetContext().Config);
+    auto getNodeIdRequest = Client_->GetRawClient()->CreateRawBatchRequest();
     for (const auto& lockIdFuture : lockIdFutures) {
-        nodeIdFutures.push_back(getNodeIdRequest.Get(
+        nodeIdFutures.push_back(getNodeIdRequest->Get(
             FileTransaction_->GetId(),
             ::TStringBuilder() << '#' << GetGuidAsString(lockIdFuture.GetValue()) << "/@node_id",
             TGetOptions()));
     }
-    getNodeIdRequest.ExecuteBatch(ClientRetryPolicy_->CreatePolicyForGenericRequest(), GetContext());
+    getNodeIdRequest->ExecuteBatch();
 
     for (size_t i = 0; i != paths->size(); ++i) {
         auto& richPath = (*paths)[i];
@@ -394,7 +394,7 @@ TJobPreparer::TJobPreparer(
 {
 
     CreateStorage();
-    auto cypressFileList = NRawClient::CanonizeYPaths(/* retryPolicy */ nullptr, OperationPreparer_.GetContext(), spec.Files_);
+    auto cypressFileList = NRawClient::CanonizeYPaths(RawClient_, spec.Files_);
 
     for (const auto& file : cypressFileList) {
         UseFileInCypress(file);
diff --git a/yt/cpp/mapreduce/client/prepare_operation.cpp b/yt/cpp/mapreduce/client/prepare_operation.cpp
index 83695ef1e838..31574d7e0e41 100644
--- a/yt/cpp/mapreduce/client/prepare_operation.cpp
+++ b/yt/cpp/mapreduce/client/prepare_operation.cpp
@@ -20,11 +20,9 @@ TOperationPreparationContext::TOperationPreparationContext(
     const TStructuredJobTableList& structuredInputs,
     const TStructuredJobTableList& structuredOutputs,
     const IRawClientPtr& rawClient,
-    const TClientContext& context,
     const IClientRetryPolicyPtr& retryPolicy,
     TTransactionId transactionId)
     : RawClient_(rawClient)
-    , Context_(context)
     , RetryPolicy_(retryPolicy)
     , TransactionId_(transactionId)
     , InputSchemas_(structuredInputs.size())
@@ -44,11 +42,9 @@ TOperationPreparationContext::TOperationPreparationContext(
     TVector<TRichYPath> inputs,
     TVector<TRichYPath> outputs,
     const IRawClientPtr& rawClient,
-    const TClientContext& context,
     const IClientRetryPolicyPtr& retryPolicy,
     TTransactionId transactionId)
     : RawClient_(rawClient)
-    , Context_(context)
     , RetryPolicy_(retryPolicy)
     , TransactionId_(transactionId)
     , InputSchemas_(inputs.size())
@@ -77,17 +73,17 @@ int TOperationPreparationContext::GetOutputCount() const
 const TVector<TTableSchema>& TOperationPreparationContext::GetInputSchemas() const
 {
     TVector<::NThreading::TFuture<TNode>> schemaFutures;
-    NRawClient::THttpRawBatchRequest batch(Context_.Config);
+    auto batch = RawClient_->CreateRawBatchRequest();
     for (int tableIndex = 0; tableIndex < static_cast<int>(InputSchemas_.size()); ++tableIndex) {
         if (InputSchemasLoaded_[tableIndex]) {
             schemaFutures.emplace_back();
             continue;
         }
         Y_ABORT_UNLESS(Inputs_[tableIndex]);
-        schemaFutures.push_back(batch.Get(TransactionId_, Inputs_[tableIndex]->Path_ + "/@schema", TGetOptions{}));
+        schemaFutures.push_back(batch->Get(TransactionId_, Inputs_[tableIndex]->Path_ + "/@schema", TGetOptions{}));
     }
 
-    batch.ExecuteBatch(RetryPolicy_->CreatePolicyForGenericRequest(), Context_);
+    batch->ExecuteBatch();
 
     for (int tableIndex = 0; tableIndex < static_cast<int>(InputSchemas_.size()); ++tableIndex) {
         if (schemaFutures[tableIndex].Initialized()) {
diff --git a/yt/cpp/mapreduce/client/prepare_operation.h b/yt/cpp/mapreduce/client/prepare_operation.h
index 3fde1d1678ee..7b34703e897a 100644
--- a/yt/cpp/mapreduce/client/prepare_operation.h
+++ b/yt/cpp/mapreduce/client/prepare_operation.h
@@ -16,7 +16,6 @@ class TOperationPreparationContext
         const TStructuredJobTableList& structuredInputs,
         const TStructuredJobTableList& structuredOutputs,
         const IRawClientPtr& rawClient,
-        const TClientContext& context,
         const IClientRetryPolicyPtr& retryPolicy,
         TTransactionId transactionId);
 
@@ -24,7 +23,6 @@ class TOperationPreparationContext
         TVector<TRichYPath> inputs,
         TVector<TRichYPath> outputs,
         const IRawClientPtr& rawClient,
-        const TClientContext& context,
         const IClientRetryPolicyPtr& retryPolicy,
         TTransactionId transactionId);
 
@@ -42,7 +40,6 @@ class TOperationPreparationContext
     TVector<TMaybe<TRichYPath>> Outputs_;
 
     const IRawClientPtr RawClient_;
-    const TClientContext& Context_;
     const IClientRetryPolicyPtr RetryPolicy_;
 
     TTransactionId TransactionId_;
diff --git a/yt/cpp/mapreduce/client/py_helpers.cpp b/yt/cpp/mapreduce/client/py_helpers.cpp
index 3072449866ed..79c372daa96c 100644
--- a/yt/cpp/mapreduce/client/py_helpers.cpp
+++ b/yt/cpp/mapreduce/client/py_helpers.cpp
@@ -51,7 +51,7 @@ TStructuredJobTableList NodeToStructuredTablePaths(const TNode& node, const TOpe
             paths.emplace_back(inputNode.AsString());
         }
     }
-    paths = NRawClient::CanonizeYPaths(/* retryPolicy */ nullptr, preparer.GetContext(), paths);
+    paths = NRawClient::CanonizeYPaths(preparer.GetClient()->GetRawClient(), paths);
     TStructuredJobTableList result(intermediateTableCount, TStructuredJobTable::Intermediate(TUnspecifiedTableStructure()));
     for (const auto& path : paths) {
         result.emplace_back(TStructuredJobTable{TUnspecifiedTableStructure(), path});
diff --git a/yt/cpp/mapreduce/client/skiff.cpp b/yt/cpp/mapreduce/client/skiff.cpp
index 499a1308fd16..113dee0743ab 100644
--- a/yt/cpp/mapreduce/client/skiff.cpp
+++ b/yt/cpp/mapreduce/client/skiff.cpp
@@ -279,8 +279,7 @@ TFormat CreateSkiffFormat(const NSkiff::TSkiffSchemaPtr& schema) {
 }
 
 NSkiff::TSkiffSchemaPtr CreateSkiffSchemaIfNecessary(
-    const TClientContext& context,
-    const IClientRetryPolicyPtr& clientRetryPolicy,
+    const IRawClientPtr& rawClient,
     const TTransactionId& transactionId,
     ENodeReaderFormat nodeReaderFormat,
     const TVector<TRichYPath>& tablePaths,
@@ -304,10 +303,9 @@ NSkiff::TSkiffSchemaPtr CreateSkiffSchemaIfNecessary(
     }
 
     auto nodes = NRawClient::BatchTransform(
-        clientRetryPolicy->CreatePolicyForGenericRequest(),
-        context,
-        NRawClient::CanonizeYPaths(clientRetryPolicy->CreatePolicyForGenericRequest(), context, tablePaths),
-        [&] (THttpRawBatchRequest& batch, const TRichYPath& path) {
+        rawClient,
+        NRawClient::CanonizeYPaths(rawClient, tablePaths),
+        [&] (IRawBatchRequestPtr batch, const TRichYPath& path) {
             auto getOptions = TGetOptions()
                 .AttributeFilter(
                     TAttributeFilter()
@@ -315,7 +313,7 @@ NSkiff::TSkiffSchemaPtr CreateSkiffSchemaIfNecessary(
                         .AddAttribute("dynamic")
                         .AddAttribute("type")
                 );
-            return batch.Get(transactionId, path.Path_, getOptions);
+            return batch->Get(transactionId, path.Path_, getOptions);
         });
 
     TVector<NSkiff::TSkiffSchemaPtr> schemas;
diff --git a/yt/cpp/mapreduce/client/skiff.h b/yt/cpp/mapreduce/client/skiff.h
index 82d80a4967a7..5f26dc765630 100644
--- a/yt/cpp/mapreduce/client/skiff.h
+++ b/yt/cpp/mapreduce/client/skiff.h
@@ -59,8 +59,7 @@ void Deserialize(NSkiff::TSkiffSchemaPtr& schema, const TNode& node);
 TFormat CreateSkiffFormat(const NSkiff::TSkiffSchemaPtr& schema);
 
 NSkiff::TSkiffSchemaPtr CreateSkiffSchemaIfNecessary(
-    const TClientContext& context,
-    const IClientRetryPolicyPtr& clientRetryPolicy,
+    const IRawClientPtr& rawClient,
     const TTransactionId& transactionId,
     ENodeReaderFormat nodeReaderFormat,
     const TVector<TRichYPath>& tablePaths,
diff --git a/yt/cpp/mapreduce/client/structured_table_formats.cpp b/yt/cpp/mapreduce/client/structured_table_formats.cpp
index 72b76e6b8054..9da31dcf14e9 100644
--- a/yt/cpp/mapreduce/client/structured_table_formats.cpp
+++ b/yt/cpp/mapreduce/client/structured_table_formats.cpp
@@ -121,8 +121,7 @@ namespace NDetail {
 ////////////////////////////////////////////////////////////////////////////////
 
 NSkiff::TSkiffSchemaPtr TryCreateSkiffSchema(
-    const TClientContext& context,
-    const IClientRetryPolicyPtr& clientRetryPolicy,
+    const IRawClientPtr& rawClient,
     const TTransactionId& transactionId,
     const TVector<TRichYPath>& tables,
     const TOperationOptions& options,
@@ -135,8 +134,7 @@ NSkiff::TSkiffSchemaPtr TryCreateSkiffSchema(
         return nullptr;
     }
     return CreateSkiffSchemaIfNecessary(
-        context,
-        clientRetryPolicy,
+        rawClient,
         transactionId,
         nodeReaderFormat,
         tables,
@@ -212,14 +210,14 @@ TStructuredJobTableList ToStructuredJobTableList(const TVector<TStructuredTableP
     return result;
 }
 
-TStructuredJobTableList CanonizeStructuredTableList(const TClientContext& context, const TVector<TStructuredTablePath>& tableList)
+TStructuredJobTableList CanonizeStructuredTableList(const IRawClientPtr& rawClient, const TVector<TStructuredTablePath>& tableList)
 {
     TVector<TRichYPath> toCanonize;
     toCanonize.reserve(tableList.size());
     for (const auto& table : tableList) {
         toCanonize.emplace_back(table.RichYPath);
     }
-    const auto canonized = NRawClient::CanonizeYPaths(/* retryPolicy */ nullptr, context, toCanonize);
+    const auto canonized = NRawClient::CanonizeYPaths(rawClient, toCanonize);
     Y_ABORT_UNLESS(canonized.size() == tableList.size());
 
     TStructuredJobTableList result;
@@ -434,8 +432,7 @@ std::pair<TFormat, TMaybe<TSmallJobFile>> TFormatBuilder::CreateNodeFormat(
             tableList.emplace_back(*table.RichYPath);
         }
         skiffSchema = TryCreateSkiffSchema(
-            Context_,
-            ClientRetryPolicy_,
+            RawClient_,
             TransactionId_,
             tableList,
             OperationOptions_,
diff --git a/yt/cpp/mapreduce/client/structured_table_formats.h b/yt/cpp/mapreduce/client/structured_table_formats.h
index 162630d41103..64a44c6f4d72 100644
--- a/yt/cpp/mapreduce/client/structured_table_formats.h
+++ b/yt/cpp/mapreduce/client/structured_table_formats.h
@@ -69,7 +69,7 @@ using TStructuredJobTableList = TVector<TStructuredJobTable>;
 TString JobTablePathString(const TStructuredJobTable& jobTable);
 TStructuredJobTableList ToStructuredJobTableList(const TVector<TStructuredTablePath>& tableList);
 
-TStructuredJobTableList CanonizeStructuredTableList(const TClientContext& context, const TVector<TStructuredTablePath>& tableList);
+TStructuredJobTableList CanonizeStructuredTableList(const IRawClientPtr& rawClient, const TVector<TStructuredTablePath>& tableList);
 TVector<TRichYPath> GetPathList(
     const TStructuredJobTableList& tableList,
     const TMaybe<TVector<TTableSchema>>& schemaInferenceResult,
diff --git a/yt/cpp/mapreduce/client/yt_poller.cpp b/yt/cpp/mapreduce/client/yt_poller.cpp
index 13670fddb550..4980d287fafb 100644
--- a/yt/cpp/mapreduce/client/yt_poller.cpp
+++ b/yt/cpp/mapreduce/client/yt_poller.cpp
@@ -92,14 +92,14 @@ void TYtPoller::WatchLoop()
             Y_ABORT_UNLESS(!InProgress_.empty());
         }
 
-        THttpRawBatchRequest rawBatchRequest(Context_.Config);
+        THttpRawBatchRequest rawBatchRequest(Context_, ClientRetryPolicy_->CreatePolicyForGenericRequest());
 
         for (auto& item : InProgress_) {
             item->PrepareRequest(&rawBatchRequest);
         }
 
         try {
-            rawBatchRequest.ExecuteBatch(ClientRetryPolicy_->CreatePolicyForGenericRequest(), Context_);
+            rawBatchRequest.ExecuteBatch();
         } catch (const std::exception& ex) {
             YT_LOG_ERROR("Exception while executing batch request: %v", ex.what());
         }
diff --git a/yt/cpp/mapreduce/client/yt_poller.h b/yt/cpp/mapreduce/client/yt_poller.h
index fe9979a6a91f..c2e680932455 100644
--- a/yt/cpp/mapreduce/client/yt_poller.h
+++ b/yt/cpp/mapreduce/client/yt_poller.h
@@ -15,10 +15,6 @@
 namespace NYT {
 namespace NDetail {
 
-namespace NRawClient {
-    class THttpRawBatchRequest;
-}
-
 ////////////////////////////////////////////////////////////////////////////////
 
 class IYtPollerItem
@@ -33,7 +29,7 @@ class IYtPollerItem
 public:
     virtual ~IYtPollerItem() = default;
 
-    virtual void PrepareRequest(NRawClient::THttpRawBatchRequest* batchRequest) = 0;
+    virtual void PrepareRequest(IRawBatchRequest* batchRequest) = 0;
 
     // Should return PollContinue if poller should continue polling this item.
     // Should return PollBreak if poller should stop polling this item.
diff --git a/yt/cpp/mapreduce/interface/fwd.h b/yt/cpp/mapreduce/interface/fwd.h
index 162a6ee3e9e7..9dcdf3789a91 100644
--- a/yt/cpp/mapreduce/interface/fwd.h
+++ b/yt/cpp/mapreduce/interface/fwd.h
@@ -398,6 +398,13 @@ namespace NYT {
     class IRetryConfigProvider;
     using IRetryConfigProviderPtr = ::TIntrusivePtr<IRetryConfigProvider>;
 
+    ////////////////////////////////////////////////////////////////////////////////
+    // raw_batch_request.h
+    ////////////////////////////////////////////////////////////////////////////////
+
+    class IRawBatchRequest;
+    using IRawBatchRequestPtr = ::TIntrusivePtr<IRawBatchRequest>;
+
     ////////////////////////////////////////////////////////////////////////////////
     // raw_client.h
     ////////////////////////////////////////////////////////////////////////////////
diff --git a/yt/cpp/mapreduce/interface/raw_batch_request.h b/yt/cpp/mapreduce/interface/raw_batch_request.h
new file mode 100644
index 000000000000..ec7b2b542c81
--- /dev/null
+++ b/yt/cpp/mapreduce/interface/raw_batch_request.h
@@ -0,0 +1,134 @@
+#pragma once
+
+#include "client_method_options.h"
+#include "fwd.h"
+#include "operation.h"
+
+#include <library/cpp/threading/future/core/future.h>
+
+namespace NYT {
+
+////////////////////////////////////////////////////////////////////////////////
+
+class IRawBatchRequest
+    : public virtual TThrRefBase
+{
+public:
+    virtual void ExecuteBatch(const TExecuteBatchOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<TNodeId> Create(
+        const TTransactionId& transaction,
+        const TYPath& path,
+        ENodeType type,
+        const TCreateOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<void> Remove(
+        const TTransactionId& transaction,
+        const TYPath& path,
+        const TRemoveOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<bool> Exists(
+        const TTransactionId& transaction,
+        const TYPath& path,
+        const TExistsOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<TNode> Get(
+        const TTransactionId& transaction,
+        const TYPath& path,
+        const TGetOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<void> Set(
+        const TTransactionId& transaction,
+        const TYPath& path,
+        const TNode& value,
+        const TSetOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<TNode::TListType> List(
+        const TTransactionId& transaction,
+        const TYPath& path,
+        const TListOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<TNodeId> Copy(
+        const TTransactionId& transaction,
+        const TYPath& sourcePath,
+        const TYPath& destinationPath,
+        const TCopyOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<TNodeId> Move(
+        const TTransactionId& transaction,
+        const TYPath& sourcePath,
+        const TYPath& destinationPath,
+        const TMoveOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<TNodeId> Link(
+        const TTransactionId& transaction,
+        const TYPath& targetPath,
+        const TYPath& linkPath,
+        const TLinkOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<TLockId> Lock(
+        const TTransactionId& transaction,
+        const TYPath& path,
+        ELockMode mode,
+        const TLockOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<void> Unlock(
+        const TTransactionId& transaction,
+        const TYPath& path,
+        const TUnlockOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<TMaybe<TYPath>> GetFileFromCache(
+        const TTransactionId& transactionId,
+        const TString& md5Signature,
+        const TYPath& cachePath,
+        const TGetFileFromCacheOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<TYPath> PutFileToCache(
+        const TTransactionId& transactionId,
+        const TYPath& filePath,
+        const TString& md5Signature,
+        const TYPath& cachePath,
+        const TPutFileToCacheOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<TCheckPermissionResponse> CheckPermission(
+        const TString& user,
+        EPermission permission,
+        const TYPath& path,
+        const TCheckPermissionOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<TOperationAttributes> GetOperation(
+        const TOperationId& operationId,
+        const TGetOperationOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<void> AbortOperation(const TOperationId& operationId) = 0;
+
+    virtual ::NThreading::TFuture<void> CompleteOperation(const TOperationId& operationId) = 0;
+
+    virtual ::NThreading::TFuture<void> SuspendOperation(
+        const TOperationId& operationId,
+        const TSuspendOperationOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<void> ResumeOperation(
+        const TOperationId& operationId,
+        const TResumeOperationOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<void> UpdateOperationParameters(
+        const TOperationId& operationId,
+        const TUpdateOperationParametersOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<TRichYPath> CanonizeYPath(const TRichYPath& path) = 0;
+
+    virtual ::NThreading::TFuture<TVector<TTableColumnarStatistics>> GetTableColumnarStatistics(
+        const TTransactionId& transaction,
+        const TVector<TRichYPath>& paths,
+        const TGetTableColumnarStatisticsOptions& options = {}) = 0;
+
+    virtual ::NThreading::TFuture<TMultiTablePartitions> GetTablePartitions(
+        const TTransactionId& transaction,
+        const TVector<TRichYPath>& paths,
+        const TGetTablePartitionsOptions& options = {}) = 0;
+};
+
+////////////////////////////////////////////////////////////////////////////////
+
+} // namespace NYT
diff --git a/yt/cpp/mapreduce/interface/raw_client.h b/yt/cpp/mapreduce/interface/raw_client.h
index 1c140d1dd98e..441e738287de 100644
--- a/yt/cpp/mapreduce/interface/raw_client.h
+++ b/yt/cpp/mapreduce/interface/raw_client.h
@@ -327,6 +327,10 @@ class IRawClient
         const TGetTablePartitionsOptions& options = {}) = 0;
 
     virtual ui64 GenerateTimestamp() = 0;
+
+    // Batch
+
+    virtual IRawBatchRequestPtr CreateRawBatchRequest() = 0;
 };
 
 ////////////////////////////////////////////////////////////////////////////////
diff --git a/yt/cpp/mapreduce/raw_client/raw_batch_request.cpp b/yt/cpp/mapreduce/raw_client/raw_batch_request.cpp
index 1a340b8c1700..08fee25b799f 100644
--- a/yt/cpp/mapreduce/raw_client/raw_batch_request.cpp
+++ b/yt/cpp/mapreduce/raw_client/raw_batch_request.cpp
@@ -285,16 +285,14 @@ THttpRawBatchRequest::TBatchItem::TBatchItem(const TBatchItem& batchItem, TInsta
 
 ////////////////////////////////////////////////////////////////////////////////
 
-THttpRawBatchRequest::THttpRawBatchRequest(const TConfigPtr& config)
-    : Config_(config)
+THttpRawBatchRequest::THttpRawBatchRequest(const TClientContext& context, IRequestRetryPolicyPtr retryPolicy)
+    : Context_(context)
+    , RequestRetryPolicy_(std::move(retryPolicy))
 { }
 
 THttpRawBatchRequest::~THttpRawBatchRequest() = default;
 
-void THttpRawBatchRequest::ExecuteBatch(
-    IRequestRetryPolicyPtr retryPolicy,
-    const TClientContext& context,
-    const TExecuteBatchOptions& options)
+void THttpRawBatchRequest::ExecuteBatch(const TExecuteBatchOptions& options)
 {
     if (IsExecuted()) {
         ythrow yexception() << "Cannot execute batch request since it is already executed";
@@ -306,8 +304,8 @@ void THttpRawBatchRequest::ExecuteBatch(
     const auto concurrency = options.Concurrency_.GetOrElse(50);
     const auto batchPartMaxSize = options.BatchPartMaxSize_.GetOrElse(concurrency * 5);
 
-    if (!retryPolicy) {
-        retryPolicy = CreateDefaultRequestRetryPolicy(context.Config);
+    if (!RequestRetryPolicy_) {
+        RequestRetryPolicy_ = CreateDefaultRequestRetryPolicy(Context_.Config);
     }
 
     while (BatchSize()) {
@@ -324,9 +322,9 @@ void THttpRawBatchRequest::ExecuteBatch(
         TResponseInfo result;
         try {
             result = RequestWithRetry<TResponseInfo>(
-                retryPolicy,
-                [&context, &header, &body] (TMutationId& mutationId) {
-                    auto response = RequestWithoutRetry(context, mutationId, header, body);
+                RequestRetryPolicy_,
+                [this, &header, &body] (TMutationId& mutationId) {
+                    auto response = RequestWithoutRetry(Context_, mutationId, header, body);
                     return TResponseInfo{
                         .RequestId = response->GetRequestId(),
                         .Response = response->GetResponse(),
@@ -337,7 +335,7 @@ void THttpRawBatchRequest::ExecuteBatch(
             SetErrorResult(std::current_exception());
             throw;
         }
-        ParseResponse(std::move(result), retryPolicy.Get());
+        ParseResponse(std::move(result), RequestRetryPolicy_.Get());
     }
 }
 
@@ -392,7 +390,7 @@ TFuture<TNodeId> THttpRawBatchRequest::Create(
 {
     return AddRequest<TGuidResponseParser>(
         "create",
-        SerializeParamsForCreate(transaction, Config_->Prefix, path, type, options),
+        SerializeParamsForCreate(transaction, Context_.Config->Prefix, path, type, options),
         Nothing());
 }
 
@@ -403,7 +401,7 @@ TFuture<void> THttpRawBatchRequest::Remove(
 {
     return AddRequest<TVoidResponseParser>(
         "remove",
-        SerializeParamsForRemove(transaction, Config_->Prefix, path, options),
+        SerializeParamsForRemove(transaction, Context_.Config->Prefix, path, options),
         Nothing());
 }
 
@@ -414,7 +412,7 @@ TFuture<bool> THttpRawBatchRequest::Exists(
 {
     return AddRequest<TExistsResponseParser>(
         "exists",
-        SerializeParamsForExists(transaction, Config_->Prefix, path, options),
+        SerializeParamsForExists(transaction, Context_.Config->Prefix, path, options),
         Nothing());
 }
 
@@ -425,7 +423,7 @@ TFuture<TNode> THttpRawBatchRequest::Get(
 {
     return AddRequest<TGetResponseParser>(
         "get",
-        SerializeParamsForGet(transaction, Config_->Prefix, path, options),
+        SerializeParamsForGet(transaction, Context_.Config->Prefix, path, options),
         Nothing());
 }
 
@@ -437,7 +435,7 @@ TFuture<void> THttpRawBatchRequest::Set(
 {
     return AddRequest<TVoidResponseParser>(
         "set",
-        SerializeParamsForSet(transaction, Config_->Prefix, path, options),
+        SerializeParamsForSet(transaction, Context_.Config->Prefix, path, options),
         node);
 }
 
@@ -448,7 +446,7 @@ TFuture<TNode::TListType> THttpRawBatchRequest::List(
 {
     return AddRequest<TListResponseParser>(
         "list",
-        SerializeParamsForList(transaction, Config_->Prefix, path, options),
+        SerializeParamsForList(transaction, Context_.Config->Prefix, path, options),
         Nothing());
 }
 
@@ -460,7 +458,7 @@ TFuture<TNodeId> THttpRawBatchRequest::Copy(
 {
     return AddRequest<TGuidResponseParser>(
         "copy",
-        SerializeParamsForCopy(transaction, Config_->Prefix, sourcePath, destinationPath, options),
+        SerializeParamsForCopy(transaction, Context_.Config->Prefix, sourcePath, destinationPath, options),
         Nothing());
 }
 
@@ -472,7 +470,7 @@ TFuture<TNodeId> THttpRawBatchRequest::Move(
 {
     return AddRequest<TGuidResponseParser>(
         "move",
-        SerializeParamsForMove(transaction, Config_->Prefix, sourcePath, destinationPath, options),
+        SerializeParamsForMove(transaction, Context_.Config->Prefix, sourcePath, destinationPath, options),
         Nothing());
 }
 
@@ -484,7 +482,7 @@ TFuture<TNodeId> THttpRawBatchRequest::Link(
 {
     return AddRequest<TGuidResponseParser>(
         "link",
-        SerializeParamsForLink(transaction, Config_->Prefix, targetPath, linkPath, options),
+        SerializeParamsForLink(transaction, Context_.Config->Prefix, targetPath, linkPath, options),
         Nothing());
 }
 
@@ -496,7 +494,7 @@ TFuture<TLockId> THttpRawBatchRequest::Lock(
 {
     return AddRequest<TGuidResponseParser>(
         "lock",
-        SerializeParamsForLock(transaction, Config_->Prefix, path, mode, options),
+        SerializeParamsForLock(transaction, Context_.Config->Prefix, path, mode, options),
         Nothing());
 }
 
@@ -507,7 +505,7 @@ TFuture<void> THttpRawBatchRequest::Unlock(
 {
     return AddRequest<TVoidResponseParser>(
         "unlock",
-        SerializeParamsForUnlock(transaction, Config_->Prefix, path, options),
+        SerializeParamsForUnlock(transaction, Context_.Config->Prefix, path, options),
         Nothing());
 }
 
@@ -532,7 +530,7 @@ TFuture<TYPath> THttpRawBatchRequest::PutFileToCache(
 {
     return AddRequest<TYPathParser>(
         "put_file_to_cache",
-        SerializeParamsForPutFileToCache(transactionId, Config_->Prefix, filePath, md5Signature, cachePath, options),
+        SerializeParamsForPutFileToCache(transactionId, Context_.Config->Prefix, filePath, md5Signature, cachePath, options),
         Nothing());
 }
 
@@ -544,7 +542,7 @@ TFuture<TCheckPermissionResponse> THttpRawBatchRequest::CheckPermission(
 {
     return AddRequest<TCheckPermissionParser>(
         "check_permission",
-        SerializeParamsForCheckPermission(user, permission, Config_->Prefix, path, options),
+        SerializeParamsForCheckPermission(user, permission, Context_.Config->Prefix, path, options),
         Nothing());
 }
 
@@ -607,7 +605,7 @@ TFuture<TRichYPath> THttpRawBatchRequest::CanonizeYPath(const TRichYPath& path)
     TRichYPath result = path;
     // Out of the symbols in the canonization branch below, only '<' can appear in the beggining of a valid rich YPath.
     if (!result.Path_.StartsWith("<")) {
-        result.Path_ = AddPathPrefix(result.Path_, Config_->Prefix);
+        result.Path_ = AddPathPrefix(result.Path_, Context_.Config->Prefix);
     }
 
     if (result.Path_.find_first_of("<>{}[]") != TString::npos) {
@@ -615,7 +613,7 @@ TFuture<TRichYPath> THttpRawBatchRequest::CanonizeYPath(const TRichYPath& path)
             "parse_ypath",
             SerializeParamsForParseYPath(result),
             Nothing(),
-            MakeIntrusive<TCanonizeYPathResponseParser>(Config_->Prefix, result));
+            MakeIntrusive<TCanonizeYPathResponseParser>(Context_.Config->Prefix, result));
     }
     return NThreading::MakeFuture(result);
 }
diff --git a/yt/cpp/mapreduce/raw_client/raw_batch_request.h b/yt/cpp/mapreduce/raw_client/raw_batch_request.h
index 670b97a84378..2af0e31305f2 100644
--- a/yt/cpp/mapreduce/raw_client/raw_batch_request.h
+++ b/yt/cpp/mapreduce/raw_client/raw_batch_request.h
@@ -2,13 +2,14 @@
 
 #include <yt/cpp/mapreduce/common/fwd.h>
 
-#include <yt/cpp/mapreduce/interface/batch_request.h>
+#include <yt/cpp/mapreduce/http/context.h>
+#include <yt/cpp/mapreduce/http/requests.h>
+
 #include <yt/cpp/mapreduce/interface/fwd.h>
 #include <yt/cpp/mapreduce/interface/node.h>
+#include <yt/cpp/mapreduce/interface/raw_batch_request.h>
 #include <yt/cpp/mapreduce/interface/retry_policy.h>
 
-#include <yt/cpp/mapreduce/http/requests.h>
-
 #include <library/cpp/threading/future/future.h>
 
 #include <util/generic/ptr.h>
@@ -25,7 +26,7 @@ namespace NYT::NDetail::NRawClient {
 ////////////////////////////////////////////////////////////////////////////////
 
 class THttpRawBatchRequest
-    : public TThrRefBase
+    : public IRawBatchRequest
 {
 public:
     struct IResponseItemParser
@@ -38,13 +39,10 @@ class THttpRawBatchRequest
     };
 
 public:
-    THttpRawBatchRequest(const TConfigPtr& config);
+    THttpRawBatchRequest(const TClientContext& context, IRequestRetryPolicyPtr retryPolicy);
     ~THttpRawBatchRequest();
 
-    void ExecuteBatch(
-        IRequestRetryPolicyPtr retryPolicy,
-        const TClientContext& context,
-        const TExecuteBatchOptions& options = {});
+    void ExecuteBatch(const TExecuteBatchOptions& options = {}) override;
 
     bool IsExecuted() const;
     void MarkExecuted();
@@ -68,91 +66,113 @@ class THttpRawBatchRequest
         const TTransactionId& transaction,
         const TYPath& path,
         ENodeType type,
-        const TCreateOptions& options);
+        const TCreateOptions& options = {}) override;
+
     ::NThreading::TFuture<void> Remove(
         const TTransactionId& transaction,
         const TYPath& path,
-        const TRemoveOptions& options);
+        const TRemoveOptions& options = {}) override;
+
     ::NThreading::TFuture<bool> Exists(
         const TTransactionId& transaction,
         const TYPath& path,
-        const TExistsOptions& options);
+        const TExistsOptions& options = {}) override;
+
     ::NThreading::TFuture<TNode> Get(
         const TTransactionId& transaction,
         const TYPath& path,
-        const TGetOptions& options);
+        const TGetOptions& options = {}) override;
+
     ::NThreading::TFuture<void> Set(
         const TTransactionId& transaction,
         const TYPath& path,
         const TNode& value,
-        const TSetOptions& options);
+        const TSetOptions& options = {}) override;
+
     ::NThreading::TFuture<TNode::TListType> List(
         const TTransactionId& transaction,
         const TYPath& path,
-        const TListOptions& options);
+        const TListOptions& options = {}) override;
+
     ::NThreading::TFuture<TNodeId> Copy(
         const TTransactionId& transaction,
         const TYPath& sourcePath,
         const TYPath& destinationPath,
-        const TCopyOptions& options);
+        const TCopyOptions& options = {}) override;
+
     ::NThreading::TFuture<TNodeId> Move(
         const TTransactionId& transaction,
         const TYPath& sourcePath,
         const TYPath& destinationPath,
-        const TMoveOptions& options);
+        const TMoveOptions& options = {}) override;
+
     ::NThreading::TFuture<TNodeId> Link(
         const TTransactionId& transaction,
         const TYPath& targetPath,
         const TYPath& linkPath,
-        const TLinkOptions& options);
+        const TLinkOptions& options = {}) override;
+
     ::NThreading::TFuture<TLockId> Lock(
         const TTransactionId& transaction,
         const TYPath& path,
         ELockMode mode,
-        const TLockOptions& options);
+        const TLockOptions& options = {}) override;
+
     ::NThreading::TFuture<void> Unlock(
         const TTransactionId& transaction,
         const TYPath& path,
-        const TUnlockOptions& options);
+        const TUnlockOptions& options = {}) override;
+
     ::NThreading::TFuture<TMaybe<TYPath>> GetFileFromCache(
         const TTransactionId& transactionId,
         const TString& md5Signature,
         const TYPath& cachePath,
-        const TGetFileFromCacheOptions& options);
+        const TGetFileFromCacheOptions& options = {}) override;
+
     ::NThreading::TFuture<TYPath> PutFileToCache(
         const TTransactionId& transactionId,
         const TYPath& filePath,
         const TString& md5Signature,
         const TYPath& cachePath,
-        const TPutFileToCacheOptions& options);
+        const TPutFileToCacheOptions& options = {}) override;
+
     ::NThreading::TFuture<TCheckPermissionResponse> CheckPermission(
         const TString& user,
         EPermission permission,
         const TYPath& path,
-        const TCheckPermissionOptions& options);
+        const TCheckPermissionOptions& options = {}) override;
+
     ::NThreading::TFuture<TOperationAttributes> GetOperation(
         const TOperationId& operationId,
-        const TGetOperationOptions& options);
-    ::NThreading::TFuture<void> AbortOperation(const TOperationId& operationId);
-    ::NThreading::TFuture<void> CompleteOperation(const TOperationId& operationId);
+        const TGetOperationOptions& options = {}) override;
+
+    ::NThreading::TFuture<void> AbortOperation(const TOperationId& operationId) override;
+
+    ::NThreading::TFuture<void> CompleteOperation(const TOperationId& operationId) override;
+
     ::NThreading::TFuture<void> SuspendOperation(
         const TOperationId& operationId,
-        const TSuspendOperationOptions& options);
+        const TSuspendOperationOptions& options = {}) override;
+
     ::NThreading::TFuture<void> ResumeOperation(
         const TOperationId& operationId,
-        const TResumeOperationOptions& options);
+        const TResumeOperationOptions& options = {}) override;
+
     ::NThreading::TFuture<void> UpdateOperationParameters(
         const TOperationId& operationId,
-        const TUpdateOperationParametersOptions& options);
-    ::NThreading::TFuture<TRichYPath> CanonizeYPath(const TRichYPath& path);
+        const TUpdateOperationParametersOptions& options = {}) override;
+
+    ::NThreading::TFuture<TRichYPath> CanonizeYPath(const TRichYPath& path) override;
+
     ::NThreading::TFuture<TVector<TTableColumnarStatistics>> GetTableColumnarStatistics(
         const TTransactionId& transaction,
         const TVector<TRichYPath>& paths,
-        const TGetTableColumnarStatisticsOptions& options);
+        const TGetTableColumnarStatisticsOptions& options = {}) override;
+
     ::NThreading::TFuture<TMultiTablePartitions> GetTablePartitions(
         const TTransactionId& transaction,
         const TVector<TRichYPath>& paths,
-        const TGetTablePartitionsOptions& options);
+        const TGetTablePartitionsOptions& options = {}) override;
 
 private:
     struct TBatchItem {
@@ -182,7 +202,9 @@ class THttpRawBatchRequest
     void AddRequest(TBatchItem batchItem);
 
 private:
-    TConfigPtr Config_;
+    const TClientContext Context_;
+
+    IRequestRetryPolicyPtr RequestRetryPolicy_;
 
     TDeque<TBatchItem> BatchItemList_;
     bool Executed_ = false;
diff --git a/yt/cpp/mapreduce/raw_client/raw_client.cpp b/yt/cpp/mapreduce/raw_client/raw_client.cpp
index 53d2be114aff..f25fe5a4ee7a 100644
--- a/yt/cpp/mapreduce/raw_client/raw_client.cpp
+++ b/yt/cpp/mapreduce/raw_client/raw_client.cpp
@@ -11,6 +11,7 @@
 #include <yt/cpp/mapreduce/http/retry_request.h>
 
 #include <yt/cpp/mapreduce/interface/fluent.h>
+#include <yt/cpp/mapreduce/interface/fwd.h>
 #include <yt/cpp/mapreduce/interface/operation.h>
 #include <yt/cpp/mapreduce/interface/tvm.h>
 
@@ -925,6 +926,11 @@ ui64 THttpRawClient::GenerateTimestamp()
     return NodeFromYsonString(responseInfo->GetResponse()).AsUint64();
 }
 
+IRawBatchRequestPtr THttpRawClient::CreateRawBatchRequest()
+{
+    return MakeIntrusive<NRawClient::THttpRawBatchRequest>(Context_, /*retryPolicy*/ nullptr);
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 
 } // namespace NYT::NDetail
diff --git a/yt/cpp/mapreduce/raw_client/raw_client.h b/yt/cpp/mapreduce/raw_client/raw_client.h
index e18e32a92ef8..f0e8378db357 100644
--- a/yt/cpp/mapreduce/raw_client/raw_client.h
+++ b/yt/cpp/mapreduce/raw_client/raw_client.h
@@ -335,6 +335,8 @@ class THttpRawClient
 
     ui64 GenerateTimestamp() override;
 
+    IRawBatchRequestPtr CreateRawBatchRequest() override;
+
 private:
     const TClientContext Context_;
 };
diff --git a/yt/cpp/mapreduce/raw_client/raw_requests.cpp b/yt/cpp/mapreduce/raw_client/raw_requests.cpp
index fd5c59ec20c7..30b1619b710d 100644
--- a/yt/cpp/mapreduce/raw_client/raw_requests.cpp
+++ b/yt/cpp/mapreduce/raw_client/raw_requests.cpp
@@ -253,25 +253,26 @@ TCheckPermissionResponse ParseCheckPermissionResponse(const TNode& node)
 }
 
 TRichYPath CanonizeYPath(
-    const IRequestRetryPolicyPtr& retryPolicy,
-    const TClientContext& context,
+    const IRawClientPtr& rawClient,
     const TRichYPath& path)
 {
-    return CanonizeYPaths(retryPolicy, context, {path}).front();
+    return CanonizeYPaths(rawClient, {path}).front();
 }
 
 TVector<TRichYPath> CanonizeYPaths(
-    const IRequestRetryPolicyPtr& retryPolicy,
-    const TClientContext& context,
+    const IRawClientPtr& rawClient,
     const TVector<TRichYPath>& paths)
 {
-    THttpRawBatchRequest batch(context.Config);
+    auto batch = rawClient->CreateRawBatchRequest();
+
     TVector<NThreading::TFuture<TRichYPath>> futures;
     futures.reserve(paths.size());
-    for (int i = 0; i < static_cast<int>(paths.size()); ++i) {
-        futures.push_back(batch.CanonizeYPath(paths[i]));
+    for (const auto& path : paths) {
+        futures.push_back(batch->CanonizeYPath(path));
     }
-    batch.ExecuteBatch(retryPolicy, context);
+
+    batch->ExecuteBatch();
+
     TVector<TRichYPath> result;
     result.reserve(futures.size());
     for (auto& future : futures) {
diff --git a/yt/cpp/mapreduce/raw_client/raw_requests.h b/yt/cpp/mapreduce/raw_client/raw_requests.h
index 860d4ae939c2..148762417141 100644
--- a/yt/cpp/mapreduce/raw_client/raw_requests.h
+++ b/yt/cpp/mapreduce/raw_client/raw_requests.h
@@ -9,6 +9,7 @@
 #include <yt/cpp/mapreduce/interface/client.h>
 #include <yt/cpp/mapreduce/interface/client_method_options.h>
 #include <yt/cpp/mapreduce/interface/operation.h>
+#include <yt/cpp/mapreduce/interface/raw_client.h>
 
 namespace NYT {
 
@@ -33,13 +34,11 @@ TCheckPermissionResponse ParseCheckPermissionResponse(const TNode& node);
 ////////////////////////////////////////////////////////////////////////////////
 
 TRichYPath CanonizeYPath(
-    const IRequestRetryPolicyPtr& retryPolicy,
-    const TClientContext& context,
+    const IRawClientPtr& rawClient,
     const TRichYPath& path);
 
 TVector<TRichYPath> CanonizeYPaths(
-    const IRequestRetryPolicyPtr& retryPolicy,
-    const TClientContext& context,
+    const IRawClientPtr& rawClient,
     const TVector<TRichYPath>& paths);
 
 NHttpClient::IHttpResponsePtr SkyShareTable(
@@ -53,19 +52,18 @@ TAuthorizationInfo WhoAmI(const TClientContext& context);
 
 template<typename TSrc, typename TBatchAdder>
 auto BatchTransform(
-    const IRequestRetryPolicyPtr& retryPolicy,
-    const TClientContext& context,
+    const IRawClientPtr& rawClient,
     const TSrc& src,
     TBatchAdder batchAdder,
     const TExecuteBatchOptions& executeBatchOptions = {})
 {
-    THttpRawBatchRequest batch(context.Config);
+    auto batch = rawClient->CreateRawBatchRequest();
     using TFuture = decltype(batchAdder(batch, *std::begin(src)));
     TVector<TFuture> futures;
     for (const auto& el : src) {
         futures.push_back(batchAdder(batch, el));
     }
-    batch.ExecuteBatch(retryPolicy, context, executeBatchOptions);
+    batch->ExecuteBatch(executeBatchOptions);
     using TDst = decltype(futures[0].ExtractValueSync());
     TVector<TDst> result;
     result.reserve(std::size(src));
diff --git a/yt/cpp/mapreduce/raw_client/ut/raw_batch_request_ut.cpp b/yt/cpp/mapreduce/raw_client/ut/raw_batch_request_ut.cpp
index 36a6935e8201..1871ee80fb47 100644
--- a/yt/cpp/mapreduce/raw_client/ut/raw_batch_request_ut.cpp
+++ b/yt/cpp/mapreduce/raw_client/ut/raw_batch_request_ut.cpp
@@ -73,7 +73,7 @@ TVector<TString> GetAllPathsFromRequestList(const TNode& requestList)
 
 TEST(TBatchRequestImplTest, ParseResponse) {
     TClientContext context;
-    THttpRawBatchRequest batchRequest(context.Config);
+    THttpRawBatchRequest batchRequest(context, /*retryPolicy*/ nullptr);
 
     EXPECT_EQ(batchRequest.BatchSize(), 0u);
 
diff --git a/yt/yql/providers/yt/codec/codegen/yt_codec_cg.cpp b/yt/yql/providers/yt/codec/codegen/yt_codec_cg.cpp
index cdf9588f9488..c81dab97cf76 100644
--- a/yt/yql/providers/yt/codec/codegen/yt_codec_cg.cpp
+++ b/yt/yql/providers/yt/codec/codegen/yt_codec_cg.cpp
@@ -85,7 +85,7 @@ class TYtCodecCgWriter : public IYtCodecCgWriter {
         auto& module = Codegen_->GetModule();
         auto& context = Codegen_->GetContext();
         // input - pointer to struct UnboxedValue as int128 and instance of buffer, output - void
-        const auto funcType = Flat || Codegen_->GetEffectiveTarget() == NYql::NCodegen::ETarget::Windows ?
+        const auto funcType = Flat ?
             FunctionType::get(Type::getVoidTy(context), {PointerType::getUnqual(Type::getInt128Ty(context)), PointerType::getUnqual(Type::getInt8Ty(context))}, false):
             FunctionType::get(Type::getVoidTy(context), {Type::getInt128Ty(context), PointerType::getUnqual(Type::getInt8Ty(context))}, false);
         Func_ = cast<Function>(module.getOrInsertFunction((TStringBuilder() << (Flat ? "YtCodecCgWriterFlat." : "YtCodecCgWriter.") << cookie).data(), funcType).getCallee());
@@ -768,7 +768,7 @@ class TYtCodecCgReader : public IYtCodecCgReader {
         case NUdf::TDataType<NUdf::TTzTimestamp64>::Id: {
             CallInst::Create(module.getFunction("ReadTzTimestamp64"), { buf, velemPtr }, "", Block_);
             break;
-        }        
+        }
 
         default:
             YQL_ENSURE(false, "Unknown data type: " << schemeType);
diff --git a/yt/yql/providers/yt/codec/yt_codec_io.cpp b/yt/yql/providers/yt/codec/yt_codec_io.cpp
index 0530bba5f725..aa47453ea26e 100644
--- a/yt/yql/providers/yt/codec/yt_codec_io.cpp
+++ b/yt/yql/providers/yt/codec/yt_codec_io.cpp
@@ -1655,7 +1655,7 @@ void TMkqlReaderImpl::SetSpecs(const TMkqlIOSpecs& specs, const NKikimr::NMiniKQ
     if (Specs_->UseBlockInput_) {
         Decoder_.Reset(new TArrowDecoder(Buf_, *Specs_, holderFactory, NUdf::GetYqlMemoryPool()));
     } else if (Specs_->UseSkiff_) {
-#ifndef MKQL_DISABLE_CODEGEN
+#if !defined(MKQL_DISABLE_CODEGEN) && !defined(__aarch64__) && !defined(_win_)
         if (Specs_->OptLLVM_ != "OFF" && NCodegen::ICodegen::IsCodegenAvailable()) {
             Decoder_.Reset(new TSkiffLLVMDecoder(Buf_, *Specs_, holderFactory));
         }
diff --git a/yt/yql/providers/yt/comp_nodes/yql_mkql_output.cpp b/yt/yql/providers/yt/comp_nodes/yql_mkql_output.cpp
index a093b6f89be1..1448c77e3745 100644
--- a/yt/yql/providers/yt/comp_nodes/yql_mkql_output.cpp
+++ b/yt/yql/providers/yt/comp_nodes/yql_mkql_output.cpp
@@ -27,7 +27,7 @@ class TYtOutputWrapper : public TDecoratorCodegeneratorNode<TYtOutputWrapper> {
         if (true /*|| TODO: !Writer.GenAddRow(item, ctx, block)*/) {
             const auto addFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TYtOutputWrapper::AddRowImpl));
             const auto selfArg = ConstantInt::get(Type::getInt64Ty(context), ui64(this));
-            const auto arg = WrapArgumentForWindows(item, ctx, block);
+            const auto arg = item;
             const auto addType = FunctionType::get(Type::getVoidTy(context), {selfArg->getType(), arg->getType()}, false);
             const auto addPtr = CastInst::Create(Instruction::IntToPtr, addFunc, PointerType::getUnqual(addType), "write", block);
             CallInst::Create(addType, addPtr, {selfArg, arg}, "", block);
@@ -77,7 +77,7 @@ using TBaseComputation = TStatelessFlowCodegeneratorNode<TYtFlowOutputWrapper>;
 
         const auto addFunc = ConstantInt::get(Type::getInt64Ty(context), GetMethodPtr(&TYtFlowOutputWrapper::AddRowImpl));
         const auto selfArg = ConstantInt::get(Type::getInt64Ty(context), ui64(this));
-        const auto arg = WrapArgumentForWindows(item, ctx, block);
+        const auto arg = item;
         const auto addType = FunctionType::get(Type::getVoidTy(context), {selfArg->getType(), arg->getType()}, false);
         const auto addPtr = CastInst::Create(Instruction::IntToPtr, addFunc, PointerType::getUnqual(addType), "write", block);
         CallInst::Create(addType, addPtr, {selfArg, arg}, "", block);
diff --git a/yt/yql/providers/yt/gateway/file/yql_yt_file_mkql_compiler.cpp b/yt/yql/providers/yt/gateway/file/yql_yt_file_mkql_compiler.cpp
index 7c76f31ad2a6..db7fdd7430da 100644
--- a/yt/yql/providers/yt/gateway/file/yql_yt_file_mkql_compiler.cpp
+++ b/yt/yql/providers/yt/gateway/file/yql_yt_file_mkql_compiler.cpp
@@ -832,54 +832,53 @@ void RegisterYtFileMkqlCompilers(NCommon::TMkqlCallableCompilerBase& compiler) {
             TRuntimeNode mapDirectOutputs;
             if (hasMap) {
                 const auto& mapper = ytMapReduce.Mapper().Cast<TCoLambda>();
-                if (const auto arg = mapper.Args().Arg(0).Raw(); arg != mapper.Body().Raw()) {
-                    values = arg->GetTypeAnn()->GetKind() == ETypeAnnotationKind::Flow ?
-                        ctx.ProgramBuilder.ToFlow(values) : ctx.ProgramBuilder.Iterator(values, {});
+                const auto arg = mapper.Args().Arg(0).Raw();
+                values = arg->GetTypeAnn()->GetKind() == ETypeAnnotationKind::Flow ?
+                    ctx.ProgramBuilder.ToFlow(values) : ctx.ProgramBuilder.Iterator(values, {});
 
-                    if (ETypeAnnotationKind::Multi == GetSeqItemType(*ytMapReduce.Mapper().Cast<TCoLambda>().Args().Arg(0).Ref().GetTypeAnn()).GetKind())
-                        values = ExpandFlow(values, ctx);
+                if (ETypeAnnotationKind::Multi == GetSeqItemType(*ytMapReduce.Mapper().Cast<TCoLambda>().Args().Arg(0).Ref().GetTypeAnn()).GetKind())
+                    values = ExpandFlow(values, ctx);
 
-                    NCommon::TMkqlBuildContext innerCtx(ctx, {{arg, values}}, ytMapReduce.Mapper().Ref().UniqueId());
+                NCommon::TMkqlBuildContext innerCtx(ctx, {{arg, values}}, ytMapReduce.Mapper().Ref().UniqueId());
 
-                    const auto& body = ytMapReduce.Mapper().Cast<TCoLambda>().Body().Ref();
-                    values = NCommon::MkqlBuildExpr(body, innerCtx);
+                const auto& body = ytMapReduce.Mapper().Cast<TCoLambda>().Body().Ref();
+                values = NCommon::MkqlBuildExpr(body, innerCtx);
 
-                    const auto& mapOutItemType = GetSeqItemType(*body.GetTypeAnn());
-                    if (const auto mapOutputTypeSetting = NYql::GetSetting(ytMapReduce.Settings().Ref(), EYtSettingType::MapOutputType)) {
-                        if (ETypeAnnotationKind::Multi == mapOutItemType.GetKind()) {
-                            values = NarrowFlow(values, *AS_TYPE(TStructType, NCommon::BuildType(body, *mapOutputTypeSetting->Tail().GetTypeAnn()->Cast<TTypeExprType>()->GetType(), ctx.ProgramBuilder)), ctx);
-                        }
+                const auto& mapOutItemType = GetSeqItemType(*body.GetTypeAnn());
+                if (const auto mapOutputTypeSetting = NYql::GetSetting(ytMapReduce.Settings().Ref(), EYtSettingType::MapOutputType)) {
+                    if (ETypeAnnotationKind::Multi == mapOutItemType.GetKind()) {
+                        values = NarrowFlow(values, *AS_TYPE(TStructType, NCommon::BuildType(body, *mapOutputTypeSetting->Tail().GetTypeAnn()->Cast<TTypeExprType>()->GetType(), ctx.ProgramBuilder)), ctx);
                     }
+                }
 
-                    values = ToList(values, ctx);
+                values = ToList(values, ctx);
 
-                    if (mapOutItemType.GetKind() == ETypeAnnotationKind::Variant) {
-                        auto tupleType = mapOutItemType.Cast<TVariantExprType>()->GetUnderlyingType()->Cast<TTupleExprType>();
-                        YQL_ENSURE(tupleType->GetSize() > 0);
-                        mapDirectOutputsCount = tupleType->GetSize() - 1;
-                        YQL_ENSURE(mapDirectOutputsCount < outputsCount);
+                if (mapOutItemType.GetKind() == ETypeAnnotationKind::Variant) {
+                    auto tupleType = mapOutItemType.Cast<TVariantExprType>()->GetUnderlyingType()->Cast<TTupleExprType>();
+                    YQL_ENSURE(tupleType->GetSize() > 0);
+                    mapDirectOutputsCount = tupleType->GetSize() - 1;
+                    YQL_ENSURE(mapDirectOutputsCount < outputsCount);
 
-                        values = ctx.ProgramBuilder.Collect(values);
+                    values = ctx.ProgramBuilder.Collect(values);
 
-                        mapDirectOutputs = ctx.ProgramBuilder.OrderedFlatMap(values, [&](TRuntimeNode mapOut) {
-                            return ctx.ProgramBuilder.VisitAll(mapOut, [&](ui32 index, TRuntimeNode varitem) {
-                                if (index == 0) {
-                                    return ctx.ProgramBuilder.NewEmptyOptional(ctx.ProgramBuilder.NewOptionalType(outputItemType));
-                                }
-                                return ctx.ProgramBuilder.NewOptional(ctx.ProgramBuilder.NewVariant(varitem, index - 1, outputItemType));
-                            });
+                    mapDirectOutputs = ctx.ProgramBuilder.OrderedFlatMap(values, [&](TRuntimeNode mapOut) {
+                        return ctx.ProgramBuilder.VisitAll(mapOut, [&](ui32 index, TRuntimeNode varitem) {
+                            if (index == 0) {
+                                return ctx.ProgramBuilder.NewEmptyOptional(ctx.ProgramBuilder.NewOptionalType(outputItemType));
+                            }
+                            return ctx.ProgramBuilder.NewOptional(ctx.ProgramBuilder.NewVariant(varitem, index - 1, outputItemType));
                         });
+                    });
 
-                        auto toReduceType = NCommon::BuildType(body, *tupleType->GetItems().front(), ctx.ProgramBuilder);
-                        values = ctx.ProgramBuilder.OrderedFlatMap(values, [&](TRuntimeNode mapOut) {
-                            return ctx.ProgramBuilder.VisitAll(mapOut, [&](ui32 index, TRuntimeNode varitem) {
-                                if (index == 0) {
-                                    return ctx.ProgramBuilder.NewOptional(varitem);
-                                }
-                                return ctx.ProgramBuilder.NewEmptyOptional(ctx.ProgramBuilder.NewOptionalType(toReduceType));
-                            });
+                    auto toReduceType = NCommon::BuildType(body, *tupleType->GetItems().front(), ctx.ProgramBuilder);
+                    values = ctx.ProgramBuilder.OrderedFlatMap(values, [&](TRuntimeNode mapOut) {
+                        return ctx.ProgramBuilder.VisitAll(mapOut, [&](ui32 index, TRuntimeNode varitem) {
+                            if (index == 0) {
+                                return ctx.ProgramBuilder.NewOptional(varitem);
+                            }
+                            return ctx.ProgramBuilder.NewEmptyOptional(ctx.ProgramBuilder.NewOptionalType(toReduceType));
                         });
-                    }
+                    });
                 }
             }
 
diff --git a/yt/yt/build/ya.make b/yt/yt/build/ya.make
index 3c9ef07a0f4f..5bbd5e3e89a4 100644
--- a/yt/yt/build/ya.make
+++ b/yt/yt/build/ya.make
@@ -3,7 +3,7 @@ LIBRARY()
 INCLUDE(${ARCADIA_ROOT}/yt/ya_cpp.make.inc)
 
 SET(YT_VERSION_MAJOR 25)
-SET(YT_VERSION_MINOR 1)
+SET(YT_VERSION_MINOR 2)
 
 DEFAULT(YT_VERSION_PATCH 0)
 DEFAULT(YT_VERSION_BRANCH "local")
diff --git a/yt/yt/core/actions/future-inl.h b/yt/yt/core/actions/future-inl.h
index 1b5f67d05b1e..090ac8a72749 100644
--- a/yt/yt/core/actions/future-inl.h
+++ b/yt/yt/core/actions/future-inl.h
@@ -2461,7 +2461,7 @@ class TCancelableBoundedConcurrencyRunner
         , ConcurrencyLimit_(concurrencyLimit)
         , Futures_(Callbacks_.size(), VoidFuture)
         , Results_(Callbacks_.size())
-        , CurrentIndex_(std::min<int>(ConcurrencyLimit_, ssize(Callbacks_)))
+        , CurrentIndex_(std::min<int>(ConcurrencyLimit_, std::ssize(Callbacks_)))
         , FailOnFirstError_(failOnError)
     { }
 
@@ -2499,35 +2499,44 @@ class TCancelableBoundedConcurrencyRunner
 
     void RunCallback(int index)
     {
-        auto future = Callbacks_[index]();
+        // ORD-2002: Avoid calling OnResult directly to prevent unbounded
+        // recursive chains like RunCallback -> OnResult -> RunCallback...
+        while (true) {
+            auto future = Callbacks_[index]();
+            if (!future.IsSet()) {
+                {
+                    auto guard = Guard(SpinLock_);
+                    if (Error_) {
+                        guard.Release();
+                        future.Cancel(*Error_);
+                        return;
+                    }
 
-        if (future.IsSet()) {
-            OnResult(index, std::move(future.Get()));
-            return;
-        }
+                    Futures_[index] = future.template As<void>();
+                }
 
-        {
-            auto guard = Guard(SpinLock_);
-            if (Error_) {
-                guard.Release();
-                future.Cancel(*Error_);
-                return;
+                future.Subscribe(
+                    BIND_NO_PROPAGATE(&TCancelableBoundedConcurrencyRunner::OnResult, MakeStrong(this), index)
+                        // NB: Sync invoker protects from unbounded recursion.
+                        .Via(GetSyncInvoker()));
+                break;
             }
 
-            Futures_[index] = future.template As<void>();
-        }
+            auto suggestedIndex = HandleResultAndSuggestNextIndex(index, std::move(future.Get()));
+            if (!suggestedIndex) {
+                break;
+            }
 
-        future.Subscribe(
-            BIND_NO_PROPAGATE(&TCancelableBoundedConcurrencyRunner::OnResult, MakeStrong(this), index)
-                // NB: Sync invoker protects from unbounded recursion.
-                .Via(GetSyncInvoker()));
+            index = *suggestedIndex;
+        }
     }
 
-    void OnResult(int index, const NYT::TErrorOr<T>& result)
+    [[nodiscard]]
+    std::optional<int> HandleResultAndSuggestNextIndex(int index, const TErrorOr<T>& result)
     {
         if (FailOnFirstError_ && !result.IsOK()) {
             OnError(result);
-            return;
+            return std::nullopt;
         }
 
         int newIndex;
@@ -2535,7 +2544,7 @@ class TCancelableBoundedConcurrencyRunner
         {
             auto guard = Guard(SpinLock_);
             if (Error_) {
-                return;
+                return std::nullopt;
             }
 
             newIndex = CurrentIndex_++;
@@ -2543,12 +2552,17 @@ class TCancelableBoundedConcurrencyRunner
             Results_[index] = result;
         }
 
-        if (finishedCount == ssize(Callbacks_)) {
+        if (finishedCount == std::ssize(Callbacks_)) {
             Promise_.TrySet(Results_);
         }
 
-        if (newIndex < ssize(Callbacks_)) {
-            RunCallback(newIndex);
+        return newIndex < std::ssize(Callbacks_) ? std::optional(newIndex) : std::nullopt;
+    }
+
+    void OnResult(int index, const TErrorOr<T>& result)
+    {
+        if (auto suggestedIndex = HandleResultAndSuggestNextIndex(index, result)) {
+            RunCallback(*suggestedIndex);
         }
     }
 
@@ -2616,26 +2630,44 @@ class TBoundedConcurrencyRunner
 
     void RunCallback(int index)
     {
-        auto future = Callbacks_[index]();
-        if (future.IsSet()) {
-            OnResult(index, future.Get());
-        } else {
-            future.Subscribe(
-                BIND_NO_PROPAGATE(&TBoundedConcurrencyRunner::OnResult, MakeStrong(this), index));
+        // ORD-2002: Avoid calling OnResult directly to prevent unbounded
+        // recursive chains like RunCallback -> OnResult -> RunCallback...
+        while (true) {
+            auto future = Callbacks_[index]();
+            if (!future.IsSet()) {
+                future.Subscribe(
+                    BIND_NO_PROPAGATE(&TBoundedConcurrencyRunner::OnResult, MakeStrong(this), index)
+                        // NB: Sync invoker protects from unbounded recursion.
+                        .Via(GetSyncInvoker()));
+                        break;
+            }
+
+            auto suggestedIndex = HandleResultAndSuggestNextIndex(index, future.Get());
+            if (!suggestedIndex) {
+                break;
+            }
+
+            index = *suggestedIndex;
         }
     }
 
-    void OnResult(int index, const NYT::TErrorOr<T>& result)
+    [[nodiscard]]
+    std::optional<int> HandleResultAndSuggestNextIndex(int index, const TErrorOr<T>& result)
     {
         Results_[index] = result;
 
         int newIndex = CurrentIndex_++;
-        if (newIndex < static_cast<ssize_t>(Callbacks_.size())) {
-            RunCallback(newIndex);
+        if (++FinishedCount_ == std::ssize(Callbacks_)) {
+            Promise_.Set(Results_);
         }
 
-        if (++FinishedCount_ == static_cast<ssize_t>(Callbacks_.size())) {
-            Promise_.Set(Results_);
+        return newIndex < std::ssize(Callbacks_) ? std::optional(newIndex) : std::nullopt;
+    }
+
+    void OnResult(int index, const TErrorOr<T>& result)
+    {
+        if (auto suggestedIndex = HandleResultAndSuggestNextIndex(index, result)) {
+            RunCallback(*suggestedIndex);
         }
     }
 };
diff --git a/yt/yt/core/actions/unittests/actions_ut.cpp b/yt/yt/core/actions/unittests/actions_ut.cpp
index bb9a415e970c..272bc2d1951c 100644
--- a/yt/yt/core/actions/unittests/actions_ut.cpp
+++ b/yt/yt/core/actions/unittests/actions_ut.cpp
@@ -94,6 +94,25 @@ TEST(TCancelableRunWithBoundedConcurrencyTest, Cancelation)
     EXPECT_EQ(canceledCount, 4);
 }
 
+TEST(TCancelableRunWithBoundedConcurrencyTest, RecurseRunner)
+{
+    auto threadPool = CreateThreadPool(4, "ThreadPool");
+
+    std::vector<TCallback<TFuture<void>()>> callbacks;
+    for (int i = 0; i < 50'000; ++i) {
+        callbacks.push_back(BIND([] {
+            return VoidFuture;
+        }));
+    }
+
+    auto future = CancelableRunWithBoundedConcurrency<void>(
+        std::move(callbacks),
+        /*concurrencyLimit*/ 5);
+
+    WaitFor(future)
+        .ThrowOnError();
+}
+
 TEST(TAllSucceededBoundedConcurrencyTest, CancelOthers)
 {
     using TCounter = std::atomic<int>;
diff --git a/yt/yt/core/concurrency/new_fair_share_thread_pool.cpp b/yt/yt/core/concurrency/new_fair_share_thread_pool.cpp
deleted file mode 100644
index fec716623def..000000000000
--- a/yt/yt/core/concurrency/new_fair_share_thread_pool.cpp
+++ /dev/null
@@ -1,1425 +0,0 @@
-#include "two_level_fair_share_thread_pool.h"
-#include "new_fair_share_thread_pool.h"
-#include "private.h"
-#include "notify_manager.h"
-#include "profiling_helpers.h"
-#include "scheduler_thread.h"
-#include "thread_pool_detail.h"
-
-#include <yt/yt/core/actions/current_invoker.h>
-
-#include <yt/yt/core/misc/finally.h>
-#include <yt/yt/core/misc/hazard_ptr.h>
-#include <yt/yt/core/misc/heap.h>
-#include <yt/yt/core/misc/ring_queue.h>
-#include <yt/yt/core/misc/mpsc_stack.h>
-#include <yt/yt/core/misc/range_formatters.h>
-
-#include <yt/yt/library/profiling/sensor.h>
-
-#include <library/cpp/yt/containers/intrusive_linked_list.h>
-
-#include <library/cpp/yt/memory/public.h>
-
-#include <library/cpp/yt/misc/tls.h>
-
-#include <util/system/spinlock.h>
-
-#include <util/generic/xrange.h>
-
-namespace NYT::NConcurrency {
-
-using namespace NProfiling;
-
-////////////////////////////////////////////////////////////////////////////////
-
-namespace {
-
-YT_DEFINE_GLOBAL(const NLogging::TLogger, Logger, "FairShareThreadPool");
-
-DECLARE_REFCOUNTED_CLASS(TBucketMapping)
-DECLARE_REFCOUNTED_CLASS(TTwoLevelFairShareQueue)
-DECLARE_REFCOUNTED_CLASS(TBucket)
-
-struct TExecutionPool;
-
-// High 16 bits is thread index and 48 bits for thread pool ptr.
-YT_DEFINE_THREAD_LOCAL(TPackedPtr, ThreadCookie, 0);
-
-constexpr auto LogDurationThreshold = TDuration::Seconds(1);
-
-////////////////////////////////////////////////////////////////////////////////
-
-template <class T>
-class THeapItem
-{
-public:
-    THeapItem(const THeapItem&) = delete;
-    THeapItem& operator=(const THeapItem&) = delete;
-
-    explicit THeapItem(T* ptr)
-        : Ptr_(std::move(ptr))
-    {
-        AdjustBackReference();
-    }
-
-    ~THeapItem()
-    {
-        if (Ptr_) {
-            YT_ASSERT(Ptr_->PositionInHeap_ == this);
-            Ptr_->PositionInHeap_ = nullptr;
-        }
-    }
-
-    T& operator* ()
-    {
-        return *Ptr_;
-    }
-
-    const T& operator*() const
-    {
-        return *Ptr_;
-    }
-
-    T* operator->()
-    {
-        return Ptr_;
-    }
-
-    const T* operator->() const
-    {
-        return Ptr_;
-    }
-
-    bool operator < (const THeapItem<T>& other) const
-    {
-        return *Ptr_ < *other;
-    }
-
-    THeapItem(THeapItem&& other) noexcept
-        : Ptr_(std::move(other.Ptr_))
-    {
-        other.Ptr_ = nullptr;
-        AdjustBackReference();
-    }
-
-    THeapItem& operator=(THeapItem&& other) noexcept
-    {
-        Ptr_ = std::move(other.Ptr_);
-        other.Ptr_ = nullptr;
-        AdjustBackReference();
-
-        return *this;
-    }
-
-private:
-    T* Ptr_;
-
-    void AdjustBackReference()
-    {
-        if (Ptr_) {
-            Ptr_->PositionInHeap_ = this;
-        }
-    }
-};
-
-template <class T>
-class THeapItemBase
-{
-public:
-    friend THeapItem<T>;
-
-    THeapItem<T>* GetPositionInHeap() const
-    {
-        return PositionInHeap_;
-    }
-
-    ~THeapItemBase()
-    {
-        YT_ASSERT(!PositionInHeap_);
-    }
-
-private:
-    THeapItem<T>* PositionInHeap_ = nullptr;
-};
-
-template <class T>
-class TPriorityQueue
-{
-public:
-    void Insert(T* object)
-    {
-        Items_.emplace_back(object);
-        SiftUp(Items_.begin(), Items_.end(), Items_.end() - 1);
-        YT_ASSERT(object->GetPositionInHeap());
-    }
-
-    void Extract(const T* object)
-    {
-        auto* positionInHeap = object->GetPositionInHeap();
-        YT_ASSERT(Items_.data() <= positionInHeap && positionInHeap < Items_.data() + Items_.size());
-
-        std::swap(*positionInHeap, Items_.back());
-        SiftDown(
-            Items_.data(),
-            Items_.data() + std::ssize(Items_) - 1,
-            positionInHeap);
-
-        YT_ASSERT(Items_.back()->GetPositionInHeap() == object->GetPositionInHeap());
-        Items_.pop_back();
-    }
-
-    void AdjustDown(const T* object)
-    {
-        auto* positionInHeap = object->GetPositionInHeap();
-        YT_ASSERT(Items_.data() <= positionInHeap && positionInHeap < Items_.data() + Items_.size());
-        SiftDown(
-            Items_.data(),
-            Items_.data() + std::ssize(Items_),
-            positionInHeap);
-    }
-
-    T* GetFront()
-    {
-        YT_ASSERT(!Empty());
-        return &*Items_.front();
-    }
-
-    bool Empty() const
-    {
-        return Items_.empty();
-    }
-
-    size_t GetSize() const
-    {
-        return Items_.size();
-    }
-
-    T& operator[] (size_t index)
-    {
-        return *Items_[index];
-    }
-
-    template <class F>
-    void ForEach(F&& functor)
-    {
-        for (auto& item : Items_) {
-            functor(&*item);
-        }
-    }
-
-    void Clear()
-    {
-        Items_.clear();
-    }
-
-private:
-    std::vector<THeapItem<T>> Items_;
-};
-
-////////////////////////////////////////////////////////////////////////////////
-
-struct TAction
-{
-    TCpuInstant EnqueuedAt = 0;
-    TCpuInstant StartedAt = 0;
-
-    // Callback keeps raw ptr to bucket to minimize bucket ref count.
-    TClosure Callback;
-    TBucketPtr BucketHolder;
-
-    TPackedPtr EnqueuedThreadCookie = 0;
-};
-
-////////////////////////////////////////////////////////////////////////////////
-
-struct TEnqueuedTime
-    : public THeapItemBase<TEnqueuedTime>
-{
-    TCpuInstant Value = 0;
-};
-
-bool operator < (const TEnqueuedTime& lhs, const TEnqueuedTime& rhs)
-{
-    return lhs.Value < rhs.Value;
-}
-
-////////////////////////////////////////////////////////////////////////////////
-
-// Data for scheduling on the first level.
-struct TExecutionPool final
-    : public THeapItemBase<TExecutionPool>
-{
-    const TString PoolName;
-
-    // Profiling sensors.
-    const NProfiling::TSummary BucketCounter;
-    const NProfiling::TSummary SizeCounter;
-    const NProfiling::TCounter DequeuedCounter;
-    const TEventTimer WaitTimeCounter;
-    const TEventTimer ExecTimeCounter;
-    const TEventTimer TotalTimeCounter;
-    const NProfiling::TTimeCounter CumulativeTimeCounter;
-
-    // Action count is used to decide whether to reset excess time or not.
-    size_t ActionCountInQueue = 0;
-
-    TCpuDuration NextUpdateWeightInstant = 0;
-    double InverseWeight = 1.0;
-    TCpuDuration ExcessTime = 0;
-
-    TPriorityQueue<TBucket> ActiveBucketsHeap;
-    TCpuDuration LastBucketExcessTime = 0;
-
-    TIntrusiveLinkedListNode<TExecutionPool> LinkedListNode;
-    // Execution pool is retained for some after last usage to flush profiling counters.
-    TCpuInstant LastUsageTime = 0;
-
-    TExecutionPool(TString poolName, const TProfiler& profiler)
-        : PoolName(std::move(poolName))
-        , BucketCounter(profiler.Summary("/buckets"))
-        , SizeCounter(profiler.Summary("/size"))
-        , DequeuedCounter(profiler.Counter("/dequeued"))
-        , WaitTimeCounter(profiler.Timer("/time/wait"))
-        , ExecTimeCounter(profiler.Timer("/time/exec"))
-        , TotalTimeCounter(profiler.Timer("/time/total"))
-        , CumulativeTimeCounter(profiler.TimeCounter("/time/cumulative"))
-    { }
-};
-
-bool operator < (const TExecutionPool& lhs, const TExecutionPool& rhs)
-{
-    return lhs.ExcessTime < rhs.ExcessTime;
-}
-
-using TExecutionPoolPtr = ::NYT::TIntrusivePtr<TExecutionPool>;
-
-////////////////////////////////////////////////////////////////////////////////
-
-// Data for scheduling on the second level.
-struct TBucketBase
-{
-    const TString BucketName;
-    const TString PoolName;
-
-    TRingQueue<TAction> ActionQueue;
-    TExecutionPoolPtr Pool = nullptr;
-
-    TCpuDuration ExcessTime = 0;
-
-    TEnqueuedTime EnqueuedTime;
-
-    TBucketBase(TString bucketName, TString poolName)
-        : BucketName(std::move(bucketName))
-        , PoolName(std::move(poolName))
-    { }
-};
-
-bool operator < (const TBucketBase& lhs, const TBucketBase& rhs)
-{
-    return std::tie(lhs.ExcessTime, lhs.EnqueuedTime) < std::tie(rhs.ExcessTime, rhs.EnqueuedTime);
-}
-
-////////////////////////////////////////////////////////////////////////////////
-
-class TBucket
-    : public IInvoker
-    , public THeapItemBase<TBucket>
-    , public TBucketBase
-{
-public:
-    TBucket(TString bucketName, TString poolName, TBucketMappingPtr parent)
-        : TBucketBase(std::move(bucketName), std::move(poolName))
-        , Parent_(std::move(parent))
-    { }
-
-    ~TBucket();
-
-    void RunCallback(const TClosure& callback, TCpuInstant cpuInstant)
-    {
-        YT_LOG_TRACE("Executing callback (EnqueuedAt: %v)", cpuInstant);
-        TCurrentInvokerGuard currentInvokerGuard(this);
-        callback.Run();
-    }
-
-    bool IsSerialized() const override
-    {
-        return false;
-    }
-
-    void Invoke(TClosure callback) override;
-
-    void Invoke(TMutableRange<TClosure> callbacks) override
-    {
-        for (auto& callback : callbacks) {
-            Invoke(std::move(callback));
-        }
-    }
-
-    NThreading::TThreadId GetThreadId() const override
-    {
-        return NThreading::InvalidThreadId;
-    }
-
-    bool CheckAffinity(const IInvokerPtr& invoker) const override
-    {
-        return invoker.Get() == this;
-    }
-
-    void SubscribeWaitTimeObserved(const TWaitTimeObserver& /*callback*/) override
-    { }
-
-    void UnsubscribeWaitTimeObserved(const TWaitTimeObserver& /*callback*/) override
-    { }
-
-private:
-    const TBucketMappingPtr Parent_;
-};
-
-DEFINE_REFCOUNTED_TYPE(TBucket)
-
-////////////////////////////////////////////////////////////////////////////////
-
-class TBucketMapping
-    : public TRefCounted
-{
-public:
-    struct TExecutionPoolToListNode
-    {
-        auto operator() (TExecutionPool* pool) const
-        {
-            return &pool->LinkedListNode;
-        }
-    };
-
-    using TPoolQueue = TIntrusiveLinkedList<TExecutionPool, TExecutionPoolToListNode>;
-
-    explicit TBucketMapping(TDuration poolRetentionTime)
-        : PoolRetentionTime_(poolRetentionTime)
-    { }
-
-    ~TBucketMapping()
-    {
-        auto guard = Guard(MappingLock_);
-
-        while (RetainPoolQueue_.GetSize() > 0) {
-            auto* frontPool = RetainPoolQueue_.GetFront();
-
-            auto poolIt = PoolMapping_.find(frontPool->PoolName);
-            YT_ASSERT(poolIt != PoolMapping_.end() && poolIt->second == frontPool);
-            PoolMapping_.erase(poolIt);
-
-            RetainPoolQueue_.PopFront();
-            NYT::DestroyRefCounted(frontPool);
-        }
-    }
-
-    virtual TProfiler GetPoolProfiler(const TString& poolName) = 0;
-
-    virtual void Invoke(TClosure callback, TBucket* bucket) = 0;
-
-    // GetInvoker is protected by mapping lock (can be sharded).
-    IInvokerPtr GetInvoker(const TString& poolName, const TString& bucketName)
-    {
-        // TODO(lukyan): Use reader guard and update it to writer if needed.
-        auto guard = Guard(MappingLock_);
-
-        auto [bucketIt, bucketInserted] = BucketMapping_.emplace(std::pair(poolName, bucketName), nullptr);
-
-        auto bucket = bucketIt->second ? DangerousGetPtr(bucketIt->second) : nullptr;
-        if (!bucket) {
-            bucket = New<TBucket>(bucketName, poolName, MakeStrong(this));
-            bucketIt->second = bucket.Get();
-            bucket->Pool = GetOrRegisterPool(bucket->PoolName);
-        }
-
-        return bucket;
-    }
-
-    // GetInvoker is protected by mapping lock (can be sharded).
-    void RemoveBucket(TBucket* bucket)
-    {
-        auto guard = Guard(MappingLock_);
-        auto bucketIt = BucketMapping_.find(std::pair(bucket->PoolName, bucket->BucketName));
-
-        if (bucketIt != BucketMapping_.end() && bucketIt->second == bucket) {
-            BucketMapping_.erase(bucketIt);
-        }
-
-        // Detach under lock.
-        auto* rawPool = bucket->Pool.Release();
-
-        // Do not want use NewWithDeleter and keep pointer to TTwoLevelFairShareQueue in each execution pool.
-        if (NYT::GetRefCounter(rawPool)->Unref(1)) {
-            auto poolsToRemove = DetachPool(rawPool);
-            guard.Release();
-
-            while (poolsToRemove.GetSize() > 0) {
-                auto* frontPool = poolsToRemove.GetFront();
-                poolsToRemove.PopFront();
-                NYT::DestroyRefCounted(frontPool);
-            }
-        }
-    }
-
-    TExecutionPoolPtr GetOrRegisterPool(TString poolName)
-    {
-        YT_ASSERT_SPINLOCK_AFFINITY(MappingLock_);
-
-        auto [mappingIt, inserted] = PoolMapping_.emplace(poolName, nullptr);
-        if (!inserted) {
-            YT_ASSERT(mappingIt->second->PoolName == poolName);
-
-            auto* rawPool = mappingIt->second;
-            // If RetainPoolQueue_ contains only one element its LinkedListNode will be null.
-            // Determine that pool is in RetainPoolQueue_ by checking its ref count.
-            if (NYT::GetRefCounter(rawPool)->GetRefCount() == 0) {
-                RetainPoolQueue_.Remove(rawPool);
-                rawPool->LinkedListNode = {};
-
-                YT_LOG_TRACE("Restoring pool (PoolName: %v)", rawPool->PoolName);
-            }
-
-            YT_LOG_TRACE("Reusing pool (PoolName: %v)", rawPool->PoolName);
-
-            // NB: Strong RC could be zero, see above.
-            NYT::GetRefCounter(rawPool)->DangerousRef();
-            return TExecutionPoolPtr(rawPool, /*addReference*/ false);
-        } else {
-            YT_LOG_TRACE("Creating pool (PoolName: %v)", poolName);
-            auto pool = New<TExecutionPool>(poolName, GetPoolProfiler(poolName));
-            mappingIt->second = pool.Get();
-
-            return pool;
-        }
-    }
-
-    TPoolQueue DetachPool(TExecutionPool* pool)
-    {
-        YT_ASSERT_SPINLOCK_AFFINITY(MappingLock_);
-
-        YT_LOG_TRACE("Removing pool (PoolName: %v)", pool->PoolName);
-
-        auto currentInstant = GetCpuInstant();
-        pool->LastUsageTime = currentInstant;
-
-        // Items in RetainPoolQueue_ are ordered by LastUsageTime.
-        // When pool is used again it is removed from RetainPoolQueue_.
-        RetainPoolQueue_.PushBack(pool);
-        return ProceedRetainQueue(currentInstant);
-    }
-
-    TPoolQueue ProceedRetainQueue(TCpuInstant currentInstant)
-    {
-        YT_ASSERT_SPINLOCK_AFFINITY(MappingLock_);
-
-        YT_LOG_TRACE("ProceedRetainQueue (Size: %v)", RetainPoolQueue_.GetSize());
-
-        TPoolQueue poolsToRemove;
-
-        while (RetainPoolQueue_.GetSize() > 0) {
-            auto* frontPool = RetainPoolQueue_.GetFront();
-
-            auto lastUsageTime = frontPool->LastUsageTime;
-            if (CpuDurationToDuration(currentInstant - lastUsageTime) < PoolRetentionTime_) {
-                break;
-            }
-
-            YT_LOG_TRACE("Destroing pool (PoolName: %v)", frontPool->PoolName);
-
-            auto poolIt = PoolMapping_.find(frontPool->PoolName);
-            YT_ASSERT(poolIt != PoolMapping_.end() && poolIt->second == frontPool);
-            PoolMapping_.erase(poolIt);
-
-            RetainPoolQueue_.PopFront();
-            poolsToRemove.PushBack(frontPool);
-        }
-
-        return poolsToRemove;
-    }
-
-    void MaybeProceedRetainQueue(TCpuInstant currentInstant)
-    {
-        if (!MappingLock_.TryAcquire()) {
-            return;
-        }
-
-        auto finally = Finally([&] {
-            MappingLock_.Release();
-        });
-
-        auto poolsToRemove = ProceedRetainQueue(currentInstant);
-        MappingLock_.Release();
-        finally.Release();
-
-        while (poolsToRemove.GetSize() > 0) {
-            auto* frontPool = poolsToRemove.GetFront();
-            poolsToRemove.PopFront();
-            NYT::DestroyRefCounted(frontPool);
-        }
-    }
-
-private:
-    const TDuration PoolRetentionTime_;
-
-    YT_DECLARE_SPIN_LOCK(NThreading::TSpinLock, MappingLock_);
-    THashMap<std::pair<TString, TString>, TBucket*> BucketMapping_;
-    THashMap<TString, TExecutionPool*> PoolMapping_;
-
-    TPoolQueue RetainPoolQueue_;
-};
-
-DEFINE_REFCOUNTED_TYPE(TBucketMapping)
-
-////////////////////////////////////////////////////////////////////////////////
-
-void TBucket::Invoke(TClosure callback)
-{
-    Parent_->Invoke(std::move(callback), this);
-}
-
-TBucket::~TBucket()
-{
-    Parent_->RemoveBucket(this);
-}
-
-////////////////////////////////////////////////////////////////////////////////
-
-DEFINE_ENUM(ERequest,
-    (None)
-    (EndExecute)
-    (FetchNext)
-);
-
-class TTwoLevelFairShareQueue
-    : protected TNotifyManager
-    , public TBucketMapping
-{
-public:
-    using TWaitTimeObserver = ITwoLevelFairShareThreadPool::TWaitTimeObserver;
-
-    TTwoLevelFairShareQueue(
-        TIntrusivePtr<NThreading::TEventCount> callbackEventCount,
-        const TString& threadNamePrefix,
-        const TNewTwoLevelFairShareThreadPoolOptions& options)
-        : TNotifyManager(std::move(callbackEventCount), GetThreadTags(threadNamePrefix), options.PollingPeriod)
-        , TBucketMapping(options.PoolRetentionTime)
-        , ThreadNamePrefix_(threadNamePrefix)
-        , Profiler_(TProfiler{"/fair_share_queue"}
-            .WithHot())
-        , CumulativeSchedulingTimeCounter_(Profiler_
-            .WithTags(GetThreadTags(ThreadNamePrefix_))
-            .TimeCounter("/time/scheduling_cumulative"))
-        , PoolWeightProvider_(options.PoolWeightProvider)
-        , VerboseLogging_(options.VerboseLogging)
-    { }
-
-    ~TTwoLevelFairShareQueue()
-    {
-        Shutdown();
-    }
-
-    void Configure(int threadCount)
-    {
-        ThreadCount_.store(threadCount);
-    }
-
-    void Configure(TDuration pollingPeriod)
-    {
-        TNotifyManager::SetPollingPeriod(pollingPeriod);
-    }
-
-    // (arkady-e1ppa): Explanation of memory orders and fences around Stopped_:
-    /*
-        We have two concurrent actions: Invoke and Shutdown.
-        For our logic Invoke method does 2 things:
-        1) It pushes callback to the queue (we assume it be release RMW operation)
-        In reality right now it is seq_cst, but it this can and should be changed in the future.
-        2) Now, the second (and third) action is seq_cst fence and then seq_cst read of Stopped_.
-        Shutdown also does two thigns:
-        1) It does a seq_cst rmw in Stopped_ (effectively, just write)
-        2) It drains the queue.
-        In order to prevent losing callbacks in queue
-        We need to make sure that either Invoke observes |true| in the read
-        or Shutdown observes placed callback from said Invoke.
-        We care about this, because if callback is stuck in queue it will
-        remain there until process is killed. If said callback MUST be
-        either executed or discarded (e.g. IPollable or Fiber) in order
-        for program to finish correctly (an not get stuck), then such a leak
-        can cause hung shutdown (and it used to do exactly that).
-
-        Relevant parts of execution are written below
-        (letter is the same in the code).
-            T1(Invoke)                  T2(Shutdown)
-        RMW^rel(Queue, 0, 1)   (a)   RMW^rlx(Stopping, false, true) (d)
-        Fence^sc               (b)   Fence^sc                       (e)
-        R^rlx(Stopping, false) (c)   RMW^acq(Queue, 0, 0)           (f)
-
-        We suppose that callback is lost so we haven't
-        seen 1 in (f) nor true in (c).
-        Since (c) reads false it reads from ctor which precedes
-        (d) in modification order. Thus we have (c) -cob-> (d)
-        (coherence-ordered before, https://eel.is/c++draft/atomics.order#3.3).
-        Likewise (f) must have read from some previous
-        dequeue attempt (or ctor) which precedes (a) in modification order.
-        Again, (f) -cob-> (a).
-        Now, for fences we have (sb means sequenced-before):
-        (b) -sb-> (c) -cob-> (d) -sb-> (e) => (b) -S-> (e)
-        (see https://eel.is/c++draft/atomics.order#4.4).
-        We also have
-        (e) -sb-> (f) -cob-> (a) -sb-> (b) => (b) -S-> (e).
-        Thus loop in S found contradicting the assumption
-        that (f) doesn't read 1 and (c) doesn't read true.
-    */
-
-    // Invoke is lock free on a fast path (a.k.a. no shutdown).
-    void Invoke(TClosure callback, TBucket* bucket) override
-    {
-        // We can't guarantee read of |true| in time anyway
-        // So relaxed order is enough.
-        if (Stopped_.load(std::memory_order::relaxed)) {
-            Drain();
-            return;
-        }
-
-        auto cpuInstant = GetCpuInstant();
-
-        YT_LOG_TRACE("Invoking action (EnqueuedAt: %v, Invoker: %v)",
-            cpuInstant,
-            ThreadNamePrefix_);
-
-        TAction action;
-        action.EnqueuedAt = cpuInstant;
-        // Callback keeps raw ptr to bucket to minimize bucket ref count.
-        action.Callback = BIND(&TBucket::RunCallback, Unretained(bucket), std::move(callback), cpuInstant);
-        action.BucketHolder = MakeStrong(bucket);
-        action.EnqueuedThreadCookie = ThreadCookie();
-
-        InvokeQueue_.Enqueue(std::move(action)); // <- (a)
-
-        std::atomic_thread_fence(std::memory_order::seq_cst); // <- (b)
-        if (Stopped_.load(std::memory_order::relaxed)) { // <- (c)
-            // We have encountered stop right after we have enqueued
-            // callback. There is a possibility that it is now stuck
-            // in queue forever. We dequeue it ourselves.
-            Drain();
-        }
-
-        NotifyFromInvoke(cpuInstant, ActiveThreads_.load() == 0);
-    }
-
-    void StopPrologue()
-    {
-        GetEventCount()->NotifyAll();
-    }
-
-    TClosure OnExecute(int index, bool fetchNext, std::function<bool()> isStopping)
-    {
-        while (true) {
-            auto cookie = GetEventCount()->PrepareWait();
-
-            auto& threadState = ThreadStates_[index];
-
-            auto hasAction = threadState.Action.BucketHolder;
-            int activeThreadDelta = hasAction ? -1 : 0;
-
-            auto callback = DoOnExecute(index, fetchNext);
-
-            if (callback) {
-                activeThreadDelta += 1;
-            }
-
-            YT_VERIFY(activeThreadDelta >= -1 && activeThreadDelta <= 1);
-
-            if (activeThreadDelta != 0) {
-                auto activeThreads = ActiveThreads_.fetch_add(activeThreadDelta);
-                auto newActiveThreads = activeThreads + activeThreadDelta;
-                YT_VERIFY(newActiveThreads >= 0 && newActiveThreads <= TThreadPoolBase::MaxThreadCount);
-                activeThreadDelta = 0;
-            }
-
-            if (callback || isStopping()) {
-                CancelWait();
-                return callback;
-            }
-
-            YT_VERIFY(fetchNext);
-            MaybeRunMaintenance(&threadState, GetCpuInstant(), /*flush*/ true);
-            Wait(cookie, isStopping);
-        }
-    }
-
-    void Shutdown()
-    {
-        if (Stopped_.exchange(true, std::memory_order::relaxed)) { // <- (d)
-            return;
-        }
-        std::atomic_thread_fence(std::memory_order::seq_cst); // <- (e)
-
-        Drain(); // <- (f)
-    }
-
-    void Drain()
-    {
-        auto guard = Guard(MainLock_);
-
-        WaitHeap_.Clear();
-
-        std::vector<TBucket*> buckets;
-        ActivePoolsHeap_.ForEach([&] (auto* pool) {
-            pool->ActiveBucketsHeap.ForEach([&] (auto* bucket) {
-                buckets.push_back(bucket);
-            });
-            pool->ActiveBucketsHeap.Clear();
-        });
-        ActivePoolsHeap_.Clear();
-
-        // Actions hold strong references to buckets.
-        // Buckets' ActionQueue must be cleared before destroying actions.
-        std::vector<TAction> actions;
-        for (auto* bucket : buckets) {
-            while (!bucket->ActionQueue.empty()) {
-                actions.push_back(std::move(bucket->ActionQueue.front()));
-                bucket->ActionQueue.pop();
-            }
-        }
-
-        InvokeQueue_.DequeueAll();
-
-        guard.Release();
-        actions.clear();
-
-    }
-
-    void SubscribeWaitTimeObserved(const TWaitTimeObserver& callback)
-    {
-        WaitTimeObservers_.Subscribe(callback);
-    }
-
-    void UnsubscribeWaitTimeObserved(const TWaitTimeObserver& callback)
-    {
-        WaitTimeObservers_.Unsubscribe(callback);
-    }
-
-private:
-    struct TThreadState
-    {
-        std::atomic<ERequest> Request = ERequest::None;
-
-        TCpuInstant AccountedAt = 0;
-        TAction Action;
-
-        // Used to store bucket ref under lock to destroy it outside.
-        TBucketPtr BucketToUnref;
-        int LastActionsInQueue;
-        TDuration TimeFromStart;
-        TDuration TimeFromEnqueue;
-        TCpuInstant LastMaintenanceInstant = {};
-    };
-
-    static_assert(sizeof(TThreadState) >= CacheLineSize);
-
-    const TString ThreadNamePrefix_;
-    const TProfiler Profiler_;
-    const NProfiling::TTimeCounter CumulativeSchedulingTimeCounter_;
-    const IPoolWeightProviderPtr PoolWeightProvider_;
-    const bool VerboseLogging_;
-
-    std::atomic<bool> Stopped_ = false;
-    TMpscStack<TAction> InvokeQueue_;
-    char Padding0_[CacheLineSize - sizeof(TMpscStack<TAction>)];
-
-    // Use simple non adaptive spinlock without complex wait strategies.
-    ::TSpinLock MainLock_;
-    char Padding1_[CacheLineSize - sizeof(::TSpinLock)];
-
-    std::array<TThreadState, TThreadPoolBase::MaxThreadCount> ThreadStates_;
-
-    TPriorityQueue<TExecutionPool> ActivePoolsHeap_;
-    TCpuDuration LastPoolExcessTime_ = 0;
-    TPriorityQueue<TEnqueuedTime> WaitHeap_;
-
-    // Buffer to keep actions during distribution to threads.
-    std::array<TAction, TThreadPoolBase::MaxThreadCount> OtherActions_;
-    std::atomic<int> ThreadCount_ = 0;
-    std::atomic<int> ActiveThreads_ = 0;
-
-    TCallbackList<TWaitTimeObserver::TSignature> WaitTimeObservers_;
-
-    TProfiler GetPoolProfiler(const TString& poolName) override
-    {
-        return Profiler_.WithTags(GetBucketTags(ThreadNamePrefix_, poolName));
-    }
-
-    Y_NO_INLINE void ConsumeInvokeQueue()
-    {
-        YT_ASSERT_SPINLOCK_AFFINITY(MainLock_);
-
-        Y_UNUSED(Padding0_);
-        Y_UNUSED(Padding1_);
-
-        InvokeQueue_.DequeueAll(true, [&] (auto& action) {
-            auto* bucket = action.BucketHolder.Get();
-
-            auto* pool = bucket->Pool.Get();
-
-            YT_VERIFY(!pool->LinkedListNode.Next && !pool->LinkedListNode.Prev);
-
-            if (!pool->GetPositionInHeap()) {
-                // ExcessTime can be greater than last pool excess time
-                // if the pool is "currently executed" and reschedules action.
-                if (pool->ExcessTime < LastPoolExcessTime_) {
-                    // Use last pool excess time to schedule new pool
-                    // after earlier scheduled pools (and not yet executed) in queue.
-
-                    YT_LOG_DEBUG_IF(VerboseLogging_, "Initial pool excess time (Name: %v, ExcessTime: %v -> %v)",
-                        pool->PoolName,
-                        pool->ExcessTime,
-                        LastPoolExcessTime_);
-
-                    pool->ExcessTime = LastPoolExcessTime_;
-                }
-
-                ActivePoolsHeap_.Insert(pool);
-            }
-
-            ++pool->ActionCountInQueue;
-
-            auto enqueuedAt = action.EnqueuedAt;
-
-            bool wasEmpty = bucket->ActionQueue.empty();
-            bucket->ActionQueue.push(std::move(action));
-
-            if (wasEmpty) {
-                bucket->EnqueuedTime.Value = enqueuedAt;
-            }
-
-            YT_ASSERT(wasEmpty == !bucket->GetPositionInHeap());
-
-            if (!bucket->GetPositionInHeap()) {
-                // ExcessTime can be greater than last bucket excess time
-                // if the bucket is "currently executed" and reschedules action.
-                if (bucket->ExcessTime < pool->LastBucketExcessTime) {
-                    // Use last bucket excess time to schedule new bucket
-                    // after earlier scheduled buckets (and not yet executed) in queue.
-
-                    YT_LOG_DEBUG_IF(VerboseLogging_, "Initial bucket excess time (Name: %v, ExcessTime: %v -> %v)",
-                        bucket->BucketName,
-                        bucket->ExcessTime,
-                        pool->LastBucketExcessTime);
-
-                    bucket->ExcessTime = pool->LastBucketExcessTime;
-                }
-
-                pool->ActiveBucketsHeap.Insert(bucket);
-                pool->BucketCounter.Record(pool->ActiveBucketsHeap.GetSize());
-                WaitHeap_.Insert(&bucket->EnqueuedTime);
-            }
-        });
-    }
-
-    void ServeBeginExecute(TThreadState* threadState, TCpuInstant currentInstant, TAction action)
-    {
-        YT_ASSERT_SPINLOCK_AFFINITY(MainLock_);
-
-        YT_ASSERT(!threadState->Action.Callback);
-        YT_ASSERT(!threadState->Action.BucketHolder);
-
-        action.StartedAt = currentInstant;
-
-        threadState->AccountedAt = currentInstant;
-        threadState->Action = std::move(action);
-    }
-
-    void ServeEndExecute(TThreadState* threadState, TCpuInstant /*cpuInstant*/)
-    {
-        YT_ASSERT_SPINLOCK_AFFINITY(MainLock_);
-
-        auto action = std::move(threadState->Action);
-        YT_ASSERT(!threadState->Action.Callback);
-        YT_ASSERT(!action.Callback);
-
-        if (!action.BucketHolder) {
-            // There was no action in begin execute.
-            return;
-        }
-
-        // Try not to change bucket ref count under lock.
-        auto bucket = std::move(action.BucketHolder);
-
-        auto& pool = *bucket->Pool;
-        YT_ASSERT(pool.PoolName == bucket->PoolName);
-
-        // LastActionsInQueue is used to update SizeCounter outside lock.
-        threadState->LastActionsInQueue = --pool.ActionCountInQueue;
-
-        // Do not destroy bucket pointer under lock. Move it in thread state in other place and
-        // destroy in corresponding thread after combiner.
-        threadState->BucketToUnref = std::move(bucket);
-    }
-
-    Y_NO_INLINE void UpdateExcessTime(TBucket* bucket, TCpuDuration duration, TCpuInstant currentInstant)
-    {
-        YT_ASSERT_SPINLOCK_AFFINITY(MainLock_);
-
-        auto* pool = bucket->Pool.Get();
-
-        if (PoolWeightProvider_ && pool->NextUpdateWeightInstant < currentInstant) {
-            pool->NextUpdateWeightInstant = currentInstant + DurationToCpuDuration(TDuration::Seconds(1));
-            pool->InverseWeight = 1.0 / PoolWeightProvider_->GetWeight(pool->PoolName);
-        }
-
-        YT_LOG_DEBUG_IF(VerboseLogging_, "Increment excess time (BucketName: %v, PoolName: %v, ExcessTime: %v -> %v)",
-            bucket->BucketName,
-            bucket->PoolName,
-            bucket->ExcessTime,
-            bucket->ExcessTime + duration);
-
-        pool->ExcessTime += duration * pool->InverseWeight;
-        bucket->ExcessTime += duration;
-
-        if (auto* positionInHeap = pool->GetPositionInHeap()) {
-            ActivePoolsHeap_.AdjustDown(pool);
-        }
-
-        if (auto* positionInHeap = bucket->GetPositionInHeap()) {
-            pool->ActiveBucketsHeap.AdjustDown(bucket);
-        }
-
-        // No need to update wait heap.
-        YT_ASSERT(!bucket->EnqueuedTime.GetPositionInHeap() == !bucket->GetPositionInHeap());
-    }
-
-    Y_NO_INLINE bool GetStarvingBucket(TAction* action)
-    {
-        YT_ASSERT_SPINLOCK_AFFINITY(MainLock_);
-
-        YT_LOG_DEBUG_IF(
-            VerboseLogging_,
-            "Buckets: %v",
-            MakeFormattableView(
-                xrange(size_t(0), ActivePoolsHeap_.GetSize()),
-                [&] (auto* builder, auto index) {
-                    auto& pool = ActivePoolsHeap_[index];
-                    builder->AppendFormat("%v [", CpuDurationToDuration(pool.ExcessTime));
-
-                    for (size_t bucketIndex = 0; bucketIndex < pool.ActiveBucketsHeap.GetSize(); ++bucketIndex) {
-                        const auto& bucket = pool.ActiveBucketsHeap[bucketIndex];
-
-                        builder->AppendFormat("%Qv:%v/%v ",
-                            bucket.BucketName,
-                            CpuDurationToDuration(bucket.ExcessTime),
-                            bucket.ActionQueue.front().EnqueuedAt);
-                    }
-                    builder->AppendFormat("]");
-                }));
-
-        if (ActivePoolsHeap_.Empty()) {
-            return false;
-        }
-
-        auto* pool = ActivePoolsHeap_.GetFront();
-        LastPoolExcessTime_ = pool->ExcessTime;
-
-        auto* bucket = pool->ActiveBucketsHeap.GetFront();
-        pool->LastBucketExcessTime = bucket->ExcessTime;
-
-        YT_ASSERT(!bucket->ActionQueue.empty());
-        *action = std::move(bucket->ActionQueue.front());
-        bucket->ActionQueue.pop();
-
-        YT_ASSERT(bucket == action->BucketHolder);
-
-        if (bucket->ActionQueue.empty()) {
-            bucket->EnqueuedTime.Value = std::numeric_limits<TCpuInstant>::max();
-
-            WaitHeap_.Extract(&bucket->EnqueuedTime);
-
-            pool->ActiveBucketsHeap.Extract(bucket);
-            pool->BucketCounter.Record(pool->ActiveBucketsHeap.GetSize());
-
-            if (pool->ActiveBucketsHeap.Empty()) {
-                ActivePoolsHeap_.Extract(pool);
-            }
-        } else {
-            bucket->EnqueuedTime.Value = bucket->ActionQueue.front().EnqueuedAt;
-            WaitHeap_.AdjustDown(&bucket->EnqueuedTime);
-        }
-
-        return true;
-    }
-
-    Y_NO_INLINE std::tuple<int, int> ServeCombinedRequests(TCpuInstant currentInstant, int currentThreadIndex)
-    {
-        YT_ASSERT_SPINLOCK_AFFINITY(MainLock_);
-
-        auto threadCount = ThreadCount_.load();
-        // Thread pool size can be reconfigures during serving requests.
-        threadCount = std::max(threadCount, currentThreadIndex + 1);
-
-        // Saved thread requests. They must be saved before consuming invoke queue.
-        std::array<bool, TThreadPoolBase::MaxThreadCount> threadRequests{false};
-        std::array<int, TThreadPoolBase::MaxThreadCount> threadIds;
-        int requestCount = 0;
-
-        YT_LOG_TRACE("Updating excess time");
-
-        // Recalculate excess time for all currently evaluating or evaluated recently (end execute) buckets
-        for (int threadIndex = 0; threadIndex < threadCount; ++threadIndex) {
-            auto& threadState = ThreadStates_[threadIndex];
-
-            // TODO(lukyan): Can skip (for threads without requests) or throttle UpdateExcessTime if it happens frequently.
-            // For each currently evaluating buckets recalculate excess time.
-            if (auto* bucket = threadState.Action.BucketHolder.Get()) {
-
-                // TODO(lukyan): Update last excess time for pool without active buckets.
-                UpdateExcessTime(bucket, currentInstant - threadState.AccountedAt, currentInstant);
-                threadState.AccountedAt = currentInstant;
-            }
-
-            auto request = threadState.Request.load(std::memory_order::acquire);
-            if (request != ERequest::None) {
-                ServeEndExecute(&threadState, currentInstant);
-
-                if (request == ERequest::FetchNext) {
-                    // Save requests before ConsumeInvokeQueue. Otherwise some thread can schedule action
-                    // but action can not be fetched (schedule and fetch happens after ConsumeInvokeQueue).
-                    threadRequests[threadIndex] = true;
-                    threadIds[requestCount++] = threadIndex;
-                } else {
-                    threadState.Request.store(ERequest::None, std::memory_order::release);
-                }
-            }
-        }
-
-        YT_LOG_TRACE("Consuming invoke queue");
-
-        ConsumeInvokeQueue();
-
-        int fetchedActions = 0;
-        int otherActionCount = 0;
-
-        // Schedule actions to desired threads.
-        while (fetchedActions < requestCount) {
-            TAction action;
-
-            if (!GetStarvingBucket(&action)) {
-                break;
-            }
-
-            ++fetchedActions;
-
-            int threadIndex = -1;
-
-            auto unpackedCookie = TTaggedPtr<TTwoLevelFairShareQueue>::Unpack(action.EnqueuedThreadCookie);
-            // TODO(lukyan): Check also wait time. If it is too high, no matter where to schedule.
-            if (unpackedCookie.Ptr == this) {
-                threadIndex = unpackedCookie.Tag;
-            }
-
-            if (threadIndex != -1 && threadRequests[threadIndex]) {
-                ServeBeginExecute(&ThreadStates_[threadIndex], currentInstant, std::move(action));
-                threadRequests[threadIndex] = false;
-                ThreadStates_[threadIndex].Request.store(ERequest::None, std::memory_order::release);
-            } else {
-                OtherActions_[otherActionCount++] = std::move(action);
-            }
-        }
-
-        // Schedule other actions.
-        for (int threadIndex : TRange(threadIds.data(), requestCount)) {
-            if (threadRequests[threadIndex]) {
-                if (otherActionCount > 0) {
-                    ServeBeginExecute(&ThreadStates_[threadIndex], currentInstant, std::move(OtherActions_[--otherActionCount]));
-                }
-
-                ThreadStates_[threadIndex].Request.store(ERequest::None, std::memory_order::release);
-            }
-        }
-
-        return {requestCount, fetchedActions};
-    }
-
-    TCpuInstant GetMinEnqueuedAt()
-    {
-        YT_ASSERT_SPINLOCK_AFFINITY(MainLock_);
-
-        return WaitHeap_.Empty()
-            ? std::numeric_limits<TCpuInstant>::max()
-            : WaitHeap_.GetFront()->Value;
-    }
-
-    TClosure DoOnExecute(int index, bool fetchNext)
-    {
-        auto cpuInstant = GetCpuInstant();
-        auto& threadState = ThreadStates_[index];
-
-        const auto& oldAction = threadState.Action;
-        if (oldAction.BucketHolder) {
-            auto waitTime = CpuDurationToDuration(oldAction.StartedAt - oldAction.EnqueuedAt);
-            auto timeFromStart = CpuDurationToDuration(cpuInstant - oldAction.StartedAt);
-            auto timeFromEnqueue = CpuDurationToDuration(cpuInstant - oldAction.EnqueuedAt);
-
-            threadState.TimeFromStart = timeFromStart;
-            threadState.TimeFromEnqueue = timeFromEnqueue;
-
-            if (timeFromStart > LogDurationThreshold) {
-                YT_LOG_DEBUG("Callback execution took too long (Wait: %v, Execution: %v, Total: %v)",
-                    waitTime,
-                    timeFromStart,
-                    timeFromEnqueue);
-            }
-
-            if (waitTime > LogDurationThreshold) {
-                YT_LOG_DEBUG("Callback wait took too long (Wait: %v, Execution: %v, Total: %v)",
-                    waitTime,
-                    timeFromStart,
-                    timeFromEnqueue);
-            }
-        }
-
-        auto finally = Finally([&] {
-            auto bucketToUndef = std::move(threadState.BucketToUnref);
-            if (bucketToUndef) {
-                auto* pool = bucketToUndef->Pool.Get();
-                pool->SizeCounter.Record(threadState.LastActionsInQueue);
-                pool->DequeuedCounter.Increment(1);
-                pool->ExecTimeCounter.Record(threadState.TimeFromStart);
-                pool->TotalTimeCounter.Record(threadState.TimeFromEnqueue);
-                pool->CumulativeTimeCounter.Add(threadState.TimeFromStart);
-                bucketToUndef.Reset();
-            }
-
-            const auto& action = threadState.Action;
-            if (action.BucketHolder) {
-                auto waitTime = CpuDurationToDuration(action.StartedAt - action.EnqueuedAt);
-                action.BucketHolder->Pool->WaitTimeCounter.Record(waitTime);
-                WaitTimeObservers_.Fire(waitTime);
-            }
-
-            MaybeRunMaintenance(&threadState, action.StartedAt, /*flush*/ false);
-
-            CumulativeSchedulingTimeCounter_.Add(CpuDurationToDuration(GetCpuInstant() - cpuInstant));
-
-            if (!fetchNext) {
-                MaybeProceedRetainQueue(cpuInstant);
-            }
-        });
-
-        auto& request = threadState.Request;
-        YT_VERIFY(request == ERequest::None);
-        request.store(fetchNext ? ERequest::FetchNext : ERequest::EndExecute);
-
-        if (MainLock_.IsLocked() || !MainLock_.TryAcquire()) {
-            // Locked here.
-            while (true) {
-                SpinLockPause();
-
-                if (request.load(std::memory_order::acquire) == ERequest::None) {
-                    return std::move(threadState.Action.Callback);
-                } else if (!MainLock_.IsLocked() && MainLock_.TryAcquire()) {
-                    break;
-                }
-            }
-        }
-
-        ResetMinEnqueuedAt();
-
-        YT_LOG_TRACE("Started serving requests");
-        auto [requests, fetchedActions] = ServeCombinedRequests(cpuInstant, index);
-
-        // Evaluate notify condition here, but call NotifyAfterFetch outside lock.
-        auto newMinEnqueuedAt = GetMinEnqueuedAt();
-        MainLock_.Release();
-
-        auto endInstant = GetCpuInstant();
-        YT_LOG_TRACE("Finished serving requests (Duration: %v, Requests: %v, FetchCount: %v, MinEnqueuedAt: %v)",
-            CpuDurationToDuration(endInstant - cpuInstant),
-            requests,
-            fetchedActions,
-            CpuInstantToInstant(newMinEnqueuedAt));
-
-        NotifyAfterFetch(endInstant, newMinEnqueuedAt);
-
-        return std::move(threadState.Action.Callback);
-    }
-
-    static void MaybeRunMaintenance(TThreadState* threadState, TCpuInstant now, bool flush)
-    {
-        YT_ASSERT(threadState);
-
-        constexpr i64 MaintenancePeriod = 1'000'000'000;
-        if (flush || now > threadState->LastMaintenanceInstant + MaintenancePeriod) {
-            ReclaimHazardPointers(false);
-            threadState->LastMaintenanceInstant  = now;
-        }
-    }
-};
-
-DEFINE_REFCOUNTED_TYPE(TTwoLevelFairShareQueue)
-
-////////////////////////////////////////////////////////////////////////////////
-
-class TFairShareThread
-    : public TSchedulerThread
-{
-public:
-    TFairShareThread(
-        TTwoLevelFairShareQueuePtr queue,
-        TIntrusivePtr<NThreading::TEventCount> callbackEventCount,
-        const TString& threadGroupName,
-        const TString& threadName,
-        int index)
-        : TSchedulerThread(
-            std::move(callbackEventCount),
-            threadGroupName,
-            threadName)
-        , Queue_(std::move(queue))
-        , Index_(index)
-    { }
-
-protected:
-    const TTwoLevelFairShareQueuePtr Queue_;
-    const int Index_;
-
-    void OnStart() override
-    {
-        ThreadCookie() = TTaggedPtr(Queue_.Get(), static_cast<ui16>(Index_)).Pack();
-    }
-
-    void StopPrologue() override
-    {
-        Queue_->StopPrologue();
-    }
-
-    TClosure OnExecute() override
-    {
-        bool fetchNext = !TSchedulerThread::IsStopping() || TSchedulerThread::GracefulStop_;
-
-        return Queue_->OnExecute(Index_, fetchNext, [&] {
-            return TSchedulerThread::IsStopping();
-        });
-    }
-
-    TClosure BeginExecute() override
-    {
-        Y_UNREACHABLE();
-    }
-
-    void EndExecute() override
-    {
-        Y_UNREACHABLE();
-    }
-};
-
-DEFINE_REFCOUNTED_TYPE(TFairShareThread)
-
-////////////////////////////////////////////////////////////////////////////////
-
-class TTwoLevelFairShareThreadPool
-    : public ITwoLevelFairShareThreadPool
-    , public TThreadPoolBase
-{
-public:
-    TTwoLevelFairShareThreadPool(
-        int threadCount,
-        const TString& threadNamePrefix,
-        const TNewTwoLevelFairShareThreadPoolOptions& options)
-        : TThreadPoolBase(threadNamePrefix)
-        , Queue_(New<TTwoLevelFairShareQueue>(
-            CallbackEventCount_,
-            ThreadNamePrefix_,
-            options))
-    {
-        SetThreadCount(threadCount);
-    }
-
-    ~TTwoLevelFairShareThreadPool()
-    {
-        Shutdown();
-    }
-
-    void SetThreadCount(int threadCount) override
-    {
-        TThreadPoolBase::SetThreadCount(threadCount);
-    }
-
-    void SetPollingPeriod(TDuration pollingPeriod) override
-    {
-        Queue_->Configure(pollingPeriod);
-    }
-
-    IInvokerPtr GetInvoker(
-        const TString& poolName,
-        const TFairShareThreadPoolTag& bucketName) override
-    {
-        EnsureStarted();
-        return Queue_->GetInvoker(poolName, bucketName);
-    }
-
-    void Shutdown() override
-    {
-        TThreadPoolBase::Shutdown();
-    }
-
-    int GetThreadCount() override
-    {
-        return TThreadPoolBase::GetThreadCount();
-    }
-
-    void SubscribeWaitTimeObserved(const TWaitTimeObserver& callback) override
-    {
-        Queue_->SubscribeWaitTimeObserved(callback);
-    }
-
-    void UnsubscribeWaitTimeObserved(const TWaitTimeObserver& callback) override
-    {
-        Queue_->UnsubscribeWaitTimeObserved(callback);
-    }
-
-private:
-    const TIntrusivePtr<NThreading::TEventCount> CallbackEventCount_ = New<NThreading::TEventCount>();
-    const TTwoLevelFairShareQueuePtr Queue_;
-
-    void DoShutdown() override
-    {
-        Queue_->Shutdown();
-        TThreadPoolBase::DoShutdown();
-        Queue_->Drain();
-    }
-
-    void DoSetThreadCount(int threadCount) override
-    {
-        Queue_->Configure(threadCount);
-        TThreadPoolBase::DoSetThreadCount(threadCount);
-    }
-
-    TSchedulerThreadPtr SpawnThread(int index) override
-    {
-        return New<TFairShareThread>(
-            Queue_,
-            CallbackEventCount_,
-            ThreadNamePrefix_,
-            MakeThreadName(index),
-            index);
-    }
-};
-
-} // namespace
-
-////////////////////////////////////////////////////////////////////////////////
-
-ITwoLevelFairShareThreadPoolPtr CreateNewTwoLevelFairShareThreadPool(
-    int threadCount,
-    const TString& threadNamePrefix,
-    const TNewTwoLevelFairShareThreadPoolOptions& options)
-{
-    return New<TTwoLevelFairShareThreadPool>(
-        threadCount,
-        threadNamePrefix,
-        options);
-}
-
-////////////////////////////////////////////////////////////////////////////////
-
-} // namespace NYT::NConcurrency
diff --git a/yt/yt/core/concurrency/new_fair_share_thread_pool.h b/yt/yt/core/concurrency/new_fair_share_thread_pool.h
deleted file mode 100644
index 8be740969e8d..000000000000
--- a/yt/yt/core/concurrency/new_fair_share_thread_pool.h
+++ /dev/null
@@ -1,26 +0,0 @@
-#pragma once
-
-#include "public.h"
-
-#include <yt/yt/core/actions/public.h>
-
-namespace NYT::NConcurrency {
-
-////////////////////////////////////////////////////////////////////////////////
-
-struct TNewTwoLevelFairShareThreadPoolOptions
-{
-    IPoolWeightProviderPtr PoolWeightProvider = nullptr;
-    bool VerboseLogging = false;
-    TDuration PollingPeriod = TDuration::MilliSeconds(10);
-    TDuration PoolRetentionTime = TDuration::Seconds(30);
-};
-
-ITwoLevelFairShareThreadPoolPtr CreateNewTwoLevelFairShareThreadPool(
-    int threadCount,
-    const TString& threadNamePrefix,
-    const TNewTwoLevelFairShareThreadPoolOptions& options = {});
-
-////////////////////////////////////////////////////////////////////////////////
-
-} // namespace NYT::NConcurrency
diff --git a/yt/yt/core/concurrency/thread_pool_poller.cpp b/yt/yt/core/concurrency/thread_pool_poller.cpp
index 88a32319cb3c..0158c1d77be3 100644
--- a/yt/yt/core/concurrency/thread_pool_poller.cpp
+++ b/yt/yt/core/concurrency/thread_pool_poller.cpp
@@ -2,7 +2,6 @@
 #include "thread_pool_poller.h"
 #include "private.h"
 #include "two_level_fair_share_thread_pool.h"
-#include "new_fair_share_thread_pool.h"
 
 #include <yt/yt/core/misc/collection_helpers.h>
 #include <yt/yt/core/misc/proc.h>
@@ -200,7 +199,7 @@ class TThreadPoolPollerImpl
         // pollables and on shutdown.
         PollerImpl_.Set(nullptr, WakeupHandle_.GetFD(), CONT_POLL_EDGE_TRIGGERED | CONT_POLL_READ);
 
-        FairShareThreadPool_ = CreateNewTwoLevelFairShareThreadPool(
+        FairShareThreadPool_ = CreateTwoLevelFairShareThreadPool(
             threadCount,
             threadNamePrefix + "FS",
             {
diff --git a/yt/yt/core/concurrency/two_level_fair_share_thread_pool.cpp b/yt/yt/core/concurrency/two_level_fair_share_thread_pool.cpp
index 07897b4dc87c..d8dd529326f9 100644
--- a/yt/yt/core/concurrency/two_level_fair_share_thread_pool.cpp
+++ b/yt/yt/core/concurrency/two_level_fair_share_thread_pool.cpp
@@ -1,397 +1,793 @@
 #include "two_level_fair_share_thread_pool.h"
 #include "private.h"
-#include "invoker_queue.h"
+#include "notify_manager.h"
 #include "profiling_helpers.h"
 #include "scheduler_thread.h"
 #include "thread_pool_detail.h"
 
 #include <yt/yt/core/actions/current_invoker.h>
 
+#include <yt/yt/core/misc/finally.h>
+#include <yt/yt/core/misc/hazard_ptr.h>
 #include <yt/yt/core/misc/heap.h>
 #include <yt/yt/core/misc/ring_queue.h>
-
-#include <yt/yt/core/profiling/tscp.h>
+#include <yt/yt/core/misc/mpsc_stack.h>
+#include <yt/yt/core/misc/range_formatters.h>
 
 #include <yt/yt/library/profiling/sensor.h>
 
-#include <library/cpp/yt/memory/weak_ptr.h>
+#include <library/cpp/yt/containers/intrusive_linked_list.h>
 
-#include <util/generic/xrange.h>
+#include <library/cpp/yt/memory/public.h>
+
+#include <library/cpp/yt/misc/tls.h>
+
+#include <util/system/spinlock.h>
 
-#include <util/system/yield.h>
+#include <util/generic/xrange.h>
 
 namespace NYT::NConcurrency {
 
 using namespace NProfiling;
 
-static constexpr auto& Logger = ConcurrencyLogger;
-
 ////////////////////////////////////////////////////////////////////////////////
 
 namespace {
 
-struct THeapItem;
-class TTwoLevelFairShareQueue;
+YT_DEFINE_GLOBAL(const NLogging::TLogger, Logger, "FairShareThreadPool");
 
-DECLARE_REFCOUNTED_STRUCT(TBucket)
+DECLARE_REFCOUNTED_CLASS(TBucketMapping)
+DECLARE_REFCOUNTED_CLASS(TTwoLevelFairShareQueue)
+DECLARE_REFCOUNTED_CLASS(TBucket)
 
-struct TBucket
-    : public IInvoker
+struct TExecutionPool;
+
+// High 16 bits is thread index and 48 bits for thread pool ptr.
+YT_DEFINE_THREAD_LOCAL(TPackedPtr, ThreadCookie, 0);
+
+constexpr auto LogDurationThreshold = TDuration::Seconds(1);
+
+////////////////////////////////////////////////////////////////////////////////
+
+template <class T>
+class THeapItem
 {
-    TBucket(size_t poolId, TFairShareThreadPoolTag tag, TWeakPtr<TTwoLevelFairShareQueue> parent)
-        : PoolId(poolId)
-        , Tag(std::move(tag))
-        , Parent(std::move(parent))
-    { }
+public:
+    THeapItem(const THeapItem&) = delete;
+    THeapItem& operator=(const THeapItem&) = delete;
 
-    void RunCallback(const TClosure& callback)
+    explicit THeapItem(T* ptr)
+        : Ptr_(std::move(ptr))
     {
-        TCurrentInvokerGuard currentInvokerGuard(this);
-        callback();
+        AdjustBackReference();
     }
 
-    void Invoke(TClosure callback) override;
+    ~THeapItem()
+    {
+        if (Ptr_) {
+            YT_ASSERT(Ptr_->PositionInHeap_ == this);
+            Ptr_->PositionInHeap_ = nullptr;
+        }
+    }
 
-    void Invoke(TMutableRange<TClosure> callbacks) override;
+    T& operator* ()
+    {
+        return *Ptr_;
+    }
 
-    void Drain()
+    const T& operator*() const
     {
-        Queue.clear();
+        return *Ptr_;
     }
 
-    NThreading::TThreadId GetThreadId() const override
+    T* operator->()
     {
-        return NThreading::InvalidThreadId;
+        return Ptr_;
     }
 
-    bool CheckAffinity(const IInvokerPtr& invoker) const override
+    const T* operator->() const
     {
-        return invoker.Get() == this;
+        return Ptr_;
     }
 
-    bool IsSerialized() const override
+    bool operator < (const THeapItem<T>& other) const
     {
-        return false;
+        return *Ptr_ < *other;
     }
 
-    void SubscribeWaitTimeObserved(const TWaitTimeObserver& /*callback*/) override
-    { }
+    THeapItem(THeapItem&& other) noexcept
+        : Ptr_(std::move(other.Ptr_))
+    {
+        other.Ptr_ = nullptr;
+        AdjustBackReference();
+    }
 
-    void UnsubscribeWaitTimeObserved(const TWaitTimeObserver& /*callback*/) override
-    { }
+    THeapItem& operator=(THeapItem&& other) noexcept
+    {
+        Ptr_ = std::move(other.Ptr_);
+        other.Ptr_ = nullptr;
+        AdjustBackReference();
 
-    ~TBucket();
+        return *this;
+    }
 
-    const size_t PoolId;
-    const TFairShareThreadPoolTag Tag;
-    TWeakPtr<TTwoLevelFairShareQueue> Parent;
-    TRingQueue<TEnqueuedAction> Queue;
-    THeapItem* HeapIterator = nullptr;
-    NProfiling::TCpuDuration WaitTime = 0;
+private:
+    T* Ptr_;
 
-    TCpuDuration ExcessTime = 0;
-    int CurrentExecutions = 0;
+    void AdjustBackReference()
+    {
+        if (Ptr_) {
+            Ptr_->PositionInHeap_ = this;
+        }
+    }
 };
 
-DEFINE_REFCOUNTED_TYPE(TBucket)
+template <class T>
+class THeapItemBase
+{
+public:
+    friend THeapItem<T>;
+
+    THeapItem<T>* GetPositionInHeap() const
+    {
+        return PositionInHeap_;
+    }
+
+    ~THeapItemBase()
+    {
+        YT_ASSERT(!PositionInHeap_);
+    }
+
+private:
+    THeapItem<T>* PositionInHeap_ = nullptr;
+};
 
-struct THeapItem
+template <class T>
+class TPriorityQueue
 {
-    TBucketPtr Bucket;
+public:
+    void Insert(T* object)
+    {
+        Items_.emplace_back(object);
+        SiftUp(Items_.begin(), Items_.end(), Items_.end() - 1);
+        YT_ASSERT(object->GetPositionInHeap());
+    }
 
-    THeapItem(const THeapItem&) = delete;
-    THeapItem& operator=(const THeapItem&) = delete;
+    void Extract(const T* object)
+    {
+        auto* positionInHeap = object->GetPositionInHeap();
+        YT_ASSERT(Items_.data() <= positionInHeap && positionInHeap < Items_.data() + Items_.size());
 
-    explicit THeapItem(TBucketPtr bucket)
-        : Bucket(std::move(bucket))
+        std::swap(*positionInHeap, Items_.back());
+        SiftDown(
+            Items_.data(),
+            Items_.data() + std::ssize(Items_) - 1,
+            positionInHeap);
+
+        YT_ASSERT(Items_.back()->GetPositionInHeap() == object->GetPositionInHeap());
+        Items_.pop_back();
+    }
+
+    void AdjustDown(const T* object)
     {
-        AdjustBackReference(this);
+        auto* positionInHeap = object->GetPositionInHeap();
+        YT_ASSERT(Items_.data() <= positionInHeap && positionInHeap < Items_.data() + Items_.size());
+        SiftDown(
+            Items_.data(),
+            Items_.data() + std::ssize(Items_),
+            positionInHeap);
     }
 
-    THeapItem(THeapItem&& other) noexcept
-        : Bucket(std::move(other.Bucket))
+    T* GetFront()
     {
-        AdjustBackReference(this);
+        YT_ASSERT(!Empty());
+        return &*Items_.front();
     }
 
-    THeapItem& operator=(THeapItem&& other) noexcept
+    bool Empty() const
     {
-        Bucket = std::move(other.Bucket);
-        AdjustBackReference(this);
+        return Items_.empty();
+    }
 
-        return *this;
+    size_t GetSize() const
+    {
+        return Items_.size();
     }
 
-    void AdjustBackReference(THeapItem* iterator)
+    T& operator[] (size_t index)
     {
-        if (Bucket) {
-            Bucket->HeapIterator = iterator;
-        }
+        return *Items_[index];
     }
 
-    ~THeapItem()
+    template <class F>
+    void ForEach(F&& functor)
     {
-        if (Bucket) {
-            Bucket->HeapIterator = nullptr;
+        for (auto& item : Items_) {
+            functor(&*item);
         }
     }
+
+    void Clear()
+    {
+        Items_.clear();
+    }
+
+private:
+    std::vector<THeapItem<T>> Items_;
+};
+
+////////////////////////////////////////////////////////////////////////////////
+
+struct TAction
+{
+    TCpuInstant EnqueuedAt = 0;
+    TCpuInstant StartedAt = 0;
+
+    // Callback keeps raw ptr to bucket to minimize bucket ref count.
+    TClosure Callback;
+    TBucketPtr BucketHolder;
+
+    TPackedPtr EnqueuedThreadCookie = 0;
+};
+
+////////////////////////////////////////////////////////////////////////////////
+
+struct TEnqueuedTime
+    : public THeapItemBase<TEnqueuedTime>
+{
+    TCpuInstant Value = 0;
 };
 
-bool operator < (const THeapItem& lhs, const THeapItem& rhs)
+bool operator < (const TEnqueuedTime& lhs, const TEnqueuedTime& rhs)
 {
-    return lhs.Bucket->ExcessTime < rhs.Bucket->ExcessTime;
+    return lhs.Value < rhs.Value;
 }
 
 ////////////////////////////////////////////////////////////////////////////////
 
-static constexpr auto LogDurationThreshold = TDuration::Seconds(1);
+// Data for scheduling on the first level.
+struct TExecutionPool final
+    : public THeapItemBase<TExecutionPool>
+{
+    const TString PoolName;
+
+    // Profiling sensors.
+    const NProfiling::TSummary BucketCounter;
+    const NProfiling::TSummary SizeCounter;
+    const NProfiling::TCounter DequeuedCounter;
+    const TEventTimer WaitTimeCounter;
+    const TEventTimer ExecTimeCounter;
+    const TEventTimer TotalTimeCounter;
+    const NProfiling::TTimeCounter CumulativeTimeCounter;
+
+    // Action count is used to decide whether to reset excess time or not.
+    size_t ActionCountInQueue = 0;
+
+    TCpuDuration NextUpdateWeightInstant = 0;
+    double InverseWeight = 1.0;
+    TCpuDuration ExcessTime = 0;
 
-DECLARE_REFCOUNTED_TYPE(TTwoLevelFairShareQueue)
+    TPriorityQueue<TBucket> ActiveBucketsHeap;
+    TCpuDuration LastBucketExcessTime = 0;
+
+    TIntrusiveLinkedListNode<TExecutionPool> LinkedListNode;
+    // Execution pool is retained for some after last usage to flush profiling counters.
+    TCpuInstant LastUsageTime = 0;
+
+    TExecutionPool(TString poolName, const TProfiler& profiler)
+        : PoolName(std::move(poolName))
+        , BucketCounter(profiler.Summary("/buckets"))
+        , SizeCounter(profiler.Summary("/size"))
+        , DequeuedCounter(profiler.Counter("/dequeued"))
+        , WaitTimeCounter(profiler.Timer("/time/wait"))
+        , ExecTimeCounter(profiler.Timer("/time/exec"))
+        , TotalTimeCounter(profiler.Timer("/time/total"))
+        , CumulativeTimeCounter(profiler.TimeCounter("/time/cumulative"))
+    { }
+};
 
-class TTwoLevelFairShareQueue
-    : public TRefCounted
+bool operator < (const TExecutionPool& lhs, const TExecutionPool& rhs)
 {
-public:
-    using TWaitTimeObserver = ITwoLevelFairShareThreadPool::TWaitTimeObserver;
+    return lhs.ExcessTime < rhs.ExcessTime;
+}
 
-    TTwoLevelFairShareQueue(
-        TIntrusivePtr<NThreading::TEventCount> callbackEventCount,
-        const TString& threadNamePrefix,
-        IPoolWeightProviderPtr poolWeightProvider)
-        : CallbackEventCount_(std::move(callbackEventCount))
-        , ThreadNamePrefix_(threadNamePrefix)
-        , Profiler_(TProfiler("/fair_share_queue")
-            .WithHot())
-        , PoolWeightProvider_(std::move(poolWeightProvider))
+using TExecutionPoolPtr = ::NYT::TIntrusivePtr<TExecutionPool>;
+
+////////////////////////////////////////////////////////////////////////////////
+
+// Data for scheduling on the second level.
+struct TBucketBase
+{
+    const TString BucketName;
+    const TString PoolName;
+
+    TRingQueue<TAction> ActionQueue;
+    TExecutionPoolPtr Pool = nullptr;
+
+    TCpuDuration ExcessTime = 0;
+
+    TEnqueuedTime EnqueuedTime;
+
+    TBucketBase(TString bucketName, TString poolName)
+        : BucketName(std::move(bucketName))
+        , PoolName(std::move(poolName))
     { }
+};
 
-    ~TTwoLevelFairShareQueue()
+bool operator < (const TBucketBase& lhs, const TBucketBase& rhs)
+{
+    return std::tie(lhs.ExcessTime, lhs.EnqueuedTime) < std::tie(rhs.ExcessTime, rhs.EnqueuedTime);
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
+class TBucket
+    : public IInvoker
+    , public THeapItemBase<TBucket>
+    , public TBucketBase
+{
+public:
+    TBucket(TString bucketName, TString poolName, TBucketMappingPtr parent)
+        : TBucketBase(std::move(bucketName), std::move(poolName))
+        , Parent_(std::move(parent))
+    { }
+
+    ~TBucket();
+
+    void RunCallback(const TClosure& callback, TCpuInstant cpuInstant)
     {
-        Shutdown();
+        YT_LOG_TRACE("Executing callback (EnqueuedAt: %v)", cpuInstant);
+        TCurrentInvokerGuard currentInvokerGuard(this);
+        callback.Run();
     }
 
-    void Configure(int threadCount)
+    bool IsSerialized() const override
     {
-        ThreadCount_.store(threadCount);
+        return false;
+    }
+
+    void Invoke(TClosure callback) override;
+
+    void Invoke(TMutableRange<TClosure> callbacks) override
+    {
+        for (auto& callback : callbacks) {
+            Invoke(std::move(callback));
+        }
     }
 
-    IInvokerPtr GetInvoker(const TString& poolName, const TFairShareThreadPoolTag& tag)
+    NThreading::TThreadId GetThreadId() const override
     {
-        while (true) {
-            auto guard = Guard(SpinLock_);
+        return NThreading::InvalidThreadId;
+    }
 
-            auto poolIt = NameToPoolId_.find(poolName);
-            if (poolIt == NameToPoolId_.end()) {
-                auto newPoolId = GetLowestEmptyPoolId();
+    bool CheckAffinity(const IInvokerPtr& invoker) const override
+    {
+        return invoker.Get() == this;
+    }
 
-                auto profiler = Profiler_.WithTags(GetBucketTags(ThreadNamePrefix_, poolName));
-                auto newPool = std::make_unique<TExecutionPool>(poolName, profiler);
-                if (newPoolId >= IdToPool_.size()) {
-                    IdToPool_.emplace_back();
-                }
-                IdToPool_[newPoolId] = std::move(newPool);
-                poolIt = NameToPoolId_.emplace(poolName, newPoolId).first;
-            }
+    void SubscribeWaitTimeObserved(const TWaitTimeObserver& /*callback*/) override
+    { }
 
-            auto poolId = poolIt->second;
-            const auto& pool = IdToPool_[poolId];
-            if (PoolWeightProvider_) {
-                pool->Weight = PoolWeightProvider_->GetWeight(poolName);
-            }
+    void UnsubscribeWaitTimeObserved(const TWaitTimeObserver& /*callback*/) override
+    { }
 
-            TBucketPtr bucket;
-            auto bucketIt = pool->TagToBucket.find(tag);
-            if (bucketIt == pool->TagToBucket.end()) {
-                bucket = New<TBucket>(poolId, tag, MakeWeak(this));
-                YT_VERIFY(pool->TagToBucket.emplace(tag, bucket.Get()).second);
-                pool->BucketCounter.Update(pool->TagToBucket.size());
-            } else {
-                bucket = DangerousGetPtr<TBucket>(bucketIt->second);
-                if (!bucket) {
-                    // Bucket is already being destroyed; backoff and retry.
-                    guard.Release();
-                    ThreadYield();
-                    continue;
-                }
-            }
+private:
+    const TBucketMappingPtr Parent_;
+};
 
-            return bucket;
-        }
-    }
+DEFINE_REFCOUNTED_TYPE(TBucket)
+
+////////////////////////////////////////////////////////////////////////////////
 
-    void Invoke(TClosure callback, TBucket* bucket)
+class TBucketMapping
+    : public TRefCounted
+{
+public:
+    struct TExecutionPoolToListNode
     {
-        auto guard = Guard(SpinLock_);
-        // See Shutdown method.
-        if (Stopping_) {
-            return;
+        auto operator() (TExecutionPool* pool) const
+        {
+            return &pool->LinkedListNode;
         }
-        const auto& pool = IdToPool_[bucket->PoolId];
+    };
 
-        pool->SizeCounter.Record(++pool->Size);
+    using TPoolQueue = TIntrusiveLinkedList<TExecutionPool, TExecutionPoolToListNode>;
 
-        if (!bucket->HeapIterator) {
-            // Otherwise ExcessTime will be recalculated in AccountCurrentlyExecutingBuckets.
-            if (bucket->CurrentExecutions == 0 && !pool->Heap.empty()) {
-                bucket->ExcessTime = pool->Heap.front().Bucket->ExcessTime;
-            }
+    explicit TBucketMapping(TDuration poolRetentionTime)
+        : PoolRetentionTime_(poolRetentionTime)
+    { }
+
+    ~TBucketMapping()
+    {
+        auto guard = Guard(MappingLock_);
 
-            pool->Heap.emplace_back(bucket);
-            AdjustHeapBack(pool->Heap.begin(), pool->Heap.end());
-            YT_VERIFY(bucket->HeapIterator);
+        while (RetainPoolQueue_.GetSize() > 0) {
+            auto* frontPool = RetainPoolQueue_.GetFront();
+
+            auto poolIt = PoolMapping_.find(frontPool->PoolName);
+            YT_ASSERT(poolIt != PoolMapping_.end() && poolIt->second == frontPool);
+            PoolMapping_.erase(poolIt);
+
+            RetainPoolQueue_.PopFront();
+            NYT::DestroyRefCounted(frontPool);
         }
+    }
 
-        YT_ASSERT(callback);
+    virtual TProfiler GetPoolProfiler(const TString& poolName) = 0;
 
-        TEnqueuedAction action;
-        action.Finished = false;
-        action.EnqueuedAt = GetCpuInstant();
-        action.Callback = BIND(&TBucket::RunCallback, MakeStrong(bucket), std::move(callback));
-        bucket->Queue.push(std::move(action));
+    virtual void Invoke(TClosure callback, TBucket* bucket) = 0;
 
-        guard.Release();
+    // GetInvoker is protected by mapping lock (can be sharded).
+    IInvokerPtr GetInvoker(const TString& poolName, const TString& bucketName)
+    {
+        // TODO(lukyan): Use reader guard and update it to writer if needed.
+        auto guard = Guard(MappingLock_);
+
+        auto [bucketIt, bucketInserted] = BucketMapping_.emplace(std::pair(poolName, bucketName), nullptr);
+
+        auto bucket = bucketIt->second ? DangerousGetPtr(bucketIt->second) : nullptr;
+        if (!bucket) {
+            bucket = New<TBucket>(bucketName, poolName, MakeStrong(this));
+            bucketIt->second = bucket.Get();
+            bucket->Pool = GetOrRegisterPool(bucket->PoolName);
+        }
 
-        CallbackEventCount_->NotifyOne();
+        return bucket;
     }
 
+    // GetInvoker is protected by mapping lock (can be sharded).
     void RemoveBucket(TBucket* bucket)
     {
-        auto guard = Guard(SpinLock_);
+        auto guard = Guard(MappingLock_);
+        auto bucketIt = BucketMapping_.find(std::pair(bucket->PoolName, bucket->BucketName));
 
-        auto& pool = IdToPool_[bucket->PoolId];
+        if (bucketIt != BucketMapping_.end() && bucketIt->second == bucket) {
+            BucketMapping_.erase(bucketIt);
+        }
+
+        // Detach under lock.
+        auto* rawPool = bucket->Pool.Release();
 
-        auto it = pool->TagToBucket.find(bucket->Tag);
-        YT_VERIFY(it != pool->TagToBucket.end());
-        YT_VERIFY(it->second == bucket);
-        pool->TagToBucket.erase(it);
-        pool->BucketCounter.Update(pool->TagToBucket.size());
+        // Do not want use NewWithDeleter and keep pointer to TTwoLevelFairShareQueue in each execution pool.
+        if (NYT::GetRefCounter(rawPool)->Unref(1)) {
+            auto poolsToRemove = DetachPool(rawPool);
+            guard.Release();
 
-        if (pool->TagToBucket.empty()) {
-            YT_VERIFY(NameToPoolId_.erase(pool->PoolName) == 1);
-            pool.reset();
+            while (poolsToRemove.GetSize() > 0) {
+                auto* frontPool = poolsToRemove.GetFront();
+                poolsToRemove.PopFront();
+                NYT::DestroyRefCounted(frontPool);
+            }
         }
     }
 
-    void Shutdown()
+    TExecutionPoolPtr GetOrRegisterPool(TString poolName)
     {
-        auto guard = Guard(SpinLock_);
-        // NB(arkady-e1ppa): We write/read value under spinlock
-        // instead of atomic in order to ensure that in Invoke method
-        // we either observe |false| and enqueue callback
-        // which will be drained here or we observe |true|
-        // and not enqueue callback at all.
-        Stopping_ = true;
-
-        for (const auto& pool : IdToPool_) {
-            if (pool) {
-                for (const auto& item : pool->Heap) {
-                    item.Bucket->Drain();
-                }
+        YT_ASSERT_SPINLOCK_AFFINITY(MappingLock_);
+
+        auto [mappingIt, inserted] = PoolMapping_.emplace(poolName, nullptr);
+        if (!inserted) {
+            YT_ASSERT(mappingIt->second->PoolName == poolName);
+
+            auto* rawPool = mappingIt->second;
+            // If RetainPoolQueue_ contains only one element its LinkedListNode will be null.
+            // Determine that pool is in RetainPoolQueue_ by checking its ref count.
+            if (NYT::GetRefCounter(rawPool)->GetRefCount() == 0) {
+                RetainPoolQueue_.Remove(rawPool);
+                rawPool->LinkedListNode = {};
+
+                YT_LOG_TRACE("Restoring pool (PoolName: %v)", rawPool->PoolName);
             }
+
+            YT_LOG_TRACE("Reusing pool (PoolName: %v)", rawPool->PoolName);
+
+            // NB: Strong RC could be zero, see above.
+            NYT::GetRefCounter(rawPool)->DangerousRef();
+            return TExecutionPoolPtr(rawPool, /*addReference*/ false);
+        } else {
+            YT_LOG_TRACE("Creating pool (PoolName: %v)", poolName);
+            auto pool = New<TExecutionPool>(poolName, GetPoolProfiler(poolName));
+            mappingIt->second = pool.Get();
+
+            return pool;
         }
     }
 
-    bool BeginExecute(TEnqueuedAction* action, int index)
+    TPoolQueue DetachPool(TExecutionPool* pool)
     {
-        auto& threadState = ThreadStates_[index];
+        YT_ASSERT_SPINLOCK_AFFINITY(MappingLock_);
 
-        YT_ASSERT(!threadState.Bucket);
-        YT_ASSERT(action && action->Finished);
+        YT_LOG_TRACE("Removing pool (PoolName: %v)", pool->PoolName);
 
         auto currentInstant = GetCpuInstant();
+        pool->LastUsageTime = currentInstant;
 
-        TBucketPtr bucket;
-        {
-            auto guard = Guard(SpinLock_);
-            bucket = GetStarvingBucket(action);
+        // Items in RetainPoolQueue_ are ordered by LastUsageTime.
+        // When pool is used again it is removed from RetainPoolQueue_.
+        RetainPoolQueue_.PushBack(pool);
+        return ProceedRetainQueue(currentInstant);
+    }
+
+    TPoolQueue ProceedRetainQueue(TCpuInstant currentInstant)
+    {
+        YT_ASSERT_SPINLOCK_AFFINITY(MappingLock_);
+
+        YT_LOG_TRACE("ProceedRetainQueue (Size: %v)", RetainPoolQueue_.GetSize());
 
-            if (!bucket) {
-                return false;
+        TPoolQueue poolsToRemove;
+
+        while (RetainPoolQueue_.GetSize() > 0) {
+            auto* frontPool = RetainPoolQueue_.GetFront();
+
+            auto lastUsageTime = frontPool->LastUsageTime;
+            if (CpuDurationToDuration(currentInstant - lastUsageTime) < PoolRetentionTime_) {
+                break;
             }
 
-            ++bucket->CurrentExecutions;
+            YT_LOG_TRACE("Destroing pool (PoolName: %v)", frontPool->PoolName);
 
-            threadState.Bucket = bucket;
-            threadState.AccountedAt = currentInstant;
+            auto poolIt = PoolMapping_.find(frontPool->PoolName);
+            YT_ASSERT(poolIt != PoolMapping_.end() && poolIt->second == frontPool);
+            PoolMapping_.erase(poolIt);
 
-            action->StartedAt = currentInstant;
-            bucket->WaitTime = action->StartedAt - action->EnqueuedAt;
+            RetainPoolQueue_.PopFront();
+            poolsToRemove.PushBack(frontPool);
         }
 
-        // Microoptimization: avoid calling CpuDurationToDuration if no observers are registered.
-        if (!WaitTimeObservers_.IsEmpty()) {
-            WaitTimeObservers_.Fire(CpuDurationToDuration(action->StartedAt - action->EnqueuedAt));
+        return poolsToRemove;
+    }
+
+    void MaybeProceedRetainQueue(TCpuInstant currentInstant)
+    {
+        if (!MappingLock_.TryAcquire()) {
+            return;
         }
 
-        YT_ASSERT(action && !action->Finished);
+        auto finally = Finally([&] {
+            MappingLock_.Release();
+        });
 
-        {
-            auto guard = Guard(SpinLock_);
-            auto& pool = IdToPool_[bucket->PoolId];
+        auto poolsToRemove = ProceedRetainQueue(currentInstant);
+        MappingLock_.Release();
+        finally.Release();
 
-            pool->WaitTimeCounter.Record(CpuDurationToDuration(bucket->WaitTime));
+        while (poolsToRemove.GetSize() > 0) {
+            auto* frontPool = poolsToRemove.GetFront();
+            poolsToRemove.PopFront();
+            NYT::DestroyRefCounted(frontPool);
         }
+    }
 
-        return true;
+private:
+    const TDuration PoolRetentionTime_;
+
+    YT_DECLARE_SPIN_LOCK(NThreading::TSpinLock, MappingLock_);
+    THashMap<std::pair<TString, TString>, TBucket*> BucketMapping_;
+    THashMap<TString, TExecutionPool*> PoolMapping_;
+
+    TPoolQueue RetainPoolQueue_;
+};
+
+DEFINE_REFCOUNTED_TYPE(TBucketMapping)
+
+////////////////////////////////////////////////////////////////////////////////
+
+void TBucket::Invoke(TClosure callback)
+{
+    Parent_->Invoke(std::move(callback), this);
+}
+
+TBucket::~TBucket()
+{
+    Parent_->RemoveBucket(this);
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
+DEFINE_ENUM(ERequest,
+    (None)
+    (EndExecute)
+    (FetchNext)
+);
+
+class TTwoLevelFairShareQueue
+    : protected TNotifyManager
+    , public TBucketMapping
+{
+public:
+    using TWaitTimeObserver = ITwoLevelFairShareThreadPool::TWaitTimeObserver;
+
+    TTwoLevelFairShareQueue(
+        TIntrusivePtr<NThreading::TEventCount> callbackEventCount,
+        const TString& threadNamePrefix,
+        const TNewTwoLevelFairShareThreadPoolOptions& options)
+        : TNotifyManager(std::move(callbackEventCount), GetThreadTags(threadNamePrefix), options.PollingPeriod)
+        , TBucketMapping(options.PoolRetentionTime)
+        , ThreadNamePrefix_(threadNamePrefix)
+        , Profiler_(TProfiler{"/fair_share_queue"}
+            .WithHot())
+        , CumulativeSchedulingTimeCounter_(Profiler_
+            .WithTags(GetThreadTags(ThreadNamePrefix_))
+            .TimeCounter("/time/scheduling_cumulative"))
+        , PoolWeightProvider_(options.PoolWeightProvider)
+        , VerboseLogging_(options.VerboseLogging)
+    { }
+
+    ~TTwoLevelFairShareQueue()
+    {
+        Shutdown();
     }
 
-    void EndExecute(TEnqueuedAction* action, int index)
+    void Configure(int threadCount)
     {
-        auto& threadState = ThreadStates_[index];
-        if (!threadState.Bucket) {
+        ThreadCount_.store(threadCount);
+    }
+
+    void Configure(TDuration pollingPeriod)
+    {
+        TNotifyManager::SetPollingPeriod(pollingPeriod);
+    }
+
+    // (arkady-e1ppa): Explanation of memory orders and fences around Stopped_:
+    /*
+        We have two concurrent actions: Invoke and Shutdown.
+        For our logic Invoke method does 2 things:
+        1) It pushes callback to the queue (we assume it be release RMW operation)
+        In reality right now it is seq_cst, but it this can and should be changed in the future.
+        2) Now, the second (and third) action is seq_cst fence and then seq_cst read of Stopped_.
+        Shutdown also does two thigns:
+        1) It does a seq_cst rmw in Stopped_ (effectively, just write)
+        2) It drains the queue.
+        In order to prevent losing callbacks in queue
+        We need to make sure that either Invoke observes |true| in the read
+        or Shutdown observes placed callback from said Invoke.
+        We care about this, because if callback is stuck in queue it will
+        remain there until process is killed. If said callback MUST be
+        either executed or discarded (e.g. IPollable or Fiber) in order
+        for program to finish correctly (an not get stuck), then such a leak
+        can cause hung shutdown (and it used to do exactly that).
+
+        Relevant parts of execution are written below
+        (letter is the same in the code).
+            T1(Invoke)                  T2(Shutdown)
+        RMW^rel(Queue, 0, 1)   (a)   RMW^rlx(Stopping, false, true) (d)
+        Fence^sc               (b)   Fence^sc                       (e)
+        R^rlx(Stopping, false) (c)   RMW^acq(Queue, 0, 0)           (f)
+
+        We suppose that callback is lost so we haven't
+        seen 1 in (f) nor true in (c).
+        Since (c) reads false it reads from ctor which precedes
+        (d) in modification order. Thus we have (c) -cob-> (d)
+        (coherence-ordered before, https://eel.is/c++draft/atomics.order#3.3).
+        Likewise (f) must have read from some previous
+        dequeue attempt (or ctor) which precedes (a) in modification order.
+        Again, (f) -cob-> (a).
+        Now, for fences we have (sb means sequenced-before):
+        (b) -sb-> (c) -cob-> (d) -sb-> (e) => (b) -S-> (e)
+        (see https://eel.is/c++draft/atomics.order#4.4).
+        We also have
+        (e) -sb-> (f) -cob-> (a) -sb-> (b) => (b) -S-> (e).
+        Thus loop in S found contradicting the assumption
+        that (f) doesn't read 1 and (c) doesn't read true.
+    */
+
+    // Invoke is lock free on a fast path (a.k.a. no shutdown).
+    void Invoke(TClosure callback, TBucket* bucket) override
+    {
+        // We can't guarantee read of |true| in time anyway
+        // So relaxed order is enough.
+        if (Stopped_.load(std::memory_order::relaxed)) {
+            Drain();
             return;
         }
 
-        YT_ASSERT(action);
+        auto cpuInstant = GetCpuInstant();
 
-        if (action->Finished) {
-            return;
+        YT_LOG_TRACE("Invoking action (EnqueuedAt: %v, Invoker: %v)",
+            cpuInstant,
+            ThreadNamePrefix_);
+
+        TAction action;
+        action.EnqueuedAt = cpuInstant;
+        // Callback keeps raw ptr to bucket to minimize bucket ref count.
+        action.Callback = BIND(&TBucket::RunCallback, Unretained(bucket), std::move(callback), cpuInstant);
+        action.BucketHolder = MakeStrong(bucket);
+        action.EnqueuedThreadCookie = ThreadCookie();
+
+        InvokeQueue_.Enqueue(std::move(action)); // <- (a)
+
+        std::atomic_thread_fence(std::memory_order::seq_cst); // <- (b)
+        if (Stopped_.load(std::memory_order::relaxed)) { // <- (c)
+            // We have encountered stop right after we have enqueued
+            // callback. There is a possibility that it is now stuck
+            // in queue forever. We dequeue it ourselves.
+            Drain();
         }
 
-        auto currentInstant = GetCpuInstant();
+        NotifyFromInvoke(cpuInstant, ActiveThreads_.load() == 0);
+    }
+
+    void StopPrologue()
+    {
+        GetEventCount()->NotifyAll();
+    }
 
-        action->FinishedAt = currentInstant;
+    TClosure OnExecute(int index, bool fetchNext, std::function<bool()> isStopping)
+    {
+        while (true) {
+            auto cookie = GetEventCount()->PrepareWait();
 
-        auto timeFromStart = CpuDurationToDuration(action->FinishedAt - action->StartedAt);
-        auto timeFromEnqueue = CpuDurationToDuration(action->FinishedAt - action->EnqueuedAt);
+            auto& threadState = ThreadStates_[index];
 
-        {
-            auto guard = Guard(SpinLock_);
-            const auto& pool = IdToPool_[threadState.Bucket->PoolId];
-            pool->SizeCounter.Record(--pool->Size);
-            pool->ExecTimeCounter.Record(timeFromStart);
-            pool->TotalTimeCounter.Record(timeFromEnqueue);
+            auto hasAction = threadState.Action.BucketHolder;
+            int activeThreadDelta = hasAction ? -1 : 0;
+
+            auto callback = DoOnExecute(index, fetchNext);
+
+            if (callback) {
+                activeThreadDelta += 1;
+            }
+
+            YT_VERIFY(activeThreadDelta >= -1 && activeThreadDelta <= 1);
+
+            if (activeThreadDelta != 0) {
+                auto activeThreads = ActiveThreads_.fetch_add(activeThreadDelta);
+                auto newActiveThreads = activeThreads + activeThreadDelta;
+                YT_VERIFY(newActiveThreads >= 0 && newActiveThreads <= TThreadPoolBase::MaxThreadCount);
+                activeThreadDelta = 0;
+            }
+
+            if (callback || isStopping()) {
+                CancelWait();
+                return callback;
+            }
+
+            YT_VERIFY(fetchNext);
+            MaybeRunMaintenance(&threadState, GetCpuInstant(), /*flush*/ true);
+            Wait(cookie, isStopping);
         }
+    }
 
-        if (timeFromStart > LogDurationThreshold) {
-            YT_LOG_DEBUG("Callback execution took too long (Wait: %v, Execution: %v, Total: %v)",
-                CpuDurationToDuration(action->StartedAt - action->EnqueuedAt),
-                timeFromStart,
-                timeFromEnqueue);
+    void Shutdown()
+    {
+        if (Stopped_.exchange(true, std::memory_order::relaxed)) { // <- (d)
+            return;
         }
+        std::atomic_thread_fence(std::memory_order::seq_cst); // <- (e)
 
-        auto waitTime = CpuDurationToDuration(action->StartedAt - action->EnqueuedAt);
+        Drain(); // <- (f)
+    }
 
-        if (waitTime > LogDurationThreshold) {
-            YT_LOG_DEBUG("Callback wait took too long (Wait: %v, Execution: %v, Total: %v)",
-                waitTime,
-                timeFromStart,
-                timeFromEnqueue);
+    void Drain()
+    {
+        auto guard = Guard(MainLock_);
+
+        WaitHeap_.Clear();
+
+        std::vector<TBucket*> buckets;
+        ActivePoolsHeap_.ForEach([&] (auto* pool) {
+            pool->ActiveBucketsHeap.ForEach([&] (auto* bucket) {
+                buckets.push_back(bucket);
+            });
+            pool->ActiveBucketsHeap.Clear();
+        });
+        ActivePoolsHeap_.Clear();
+
+        // Actions hold strong references to buckets.
+        // Buckets' ActionQueue must be cleared before destroying actions.
+        std::vector<TAction> actions;
+        for (auto* bucket : buckets) {
+            while (!bucket->ActionQueue.empty()) {
+                actions.push_back(std::move(bucket->ActionQueue.front()));
+                bucket->ActionQueue.pop();
+            }
         }
 
-        action->Finished = true;
+        InvokeQueue_.DequeueAll();
 
-        // Remove outside lock because of lock inside RemoveBucket.
-        TBucketPtr bucket;
-        {
-            auto guard = Guard(SpinLock_);
-            bucket = std::move(threadState.Bucket);
-
-            UpdateExcessTime(bucket.Get(), currentInstant - threadState.AccountedAt);
-            threadState.AccountedAt = currentInstant;
+        guard.Release();
+        actions.clear();
 
-            YT_VERIFY(bucket->CurrentExecutions-- > 0);
-        }
     }
 
     void SubscribeWaitTimeObserved(const TWaitTimeObserver& callback)
@@ -407,205 +803,459 @@ class TTwoLevelFairShareQueue
 private:
     struct TThreadState
     {
+        std::atomic<ERequest> Request = ERequest::None;
+
         TCpuInstant AccountedAt = 0;
-        TBucketPtr Bucket;
+        TAction Action;
+
+        // Used to store bucket ref under lock to destroy it outside.
+        TBucketPtr BucketToUnref;
+        int LastActionsInQueue;
+        TDuration TimeFromStart;
+        TDuration TimeFromEnqueue;
+        TCpuInstant LastMaintenanceInstant = {};
     };
 
-    struct TExecutionPool
+    static_assert(sizeof(TThreadState) >= CacheLineSize);
+
+    const TString ThreadNamePrefix_;
+    const TProfiler Profiler_;
+    const NProfiling::TTimeCounter CumulativeSchedulingTimeCounter_;
+    const IPoolWeightProviderPtr PoolWeightProvider_;
+    const bool VerboseLogging_;
+
+    std::atomic<bool> Stopped_ = false;
+    TMpscStack<TAction> InvokeQueue_;
+    char Padding0_[CacheLineSize - sizeof(TMpscStack<TAction>)];
+
+    // Use simple non adaptive spinlock without complex wait strategies.
+    ::TSpinLock MainLock_;
+    char Padding1_[CacheLineSize - sizeof(::TSpinLock)];
+
+    std::array<TThreadState, TThreadPoolBase::MaxThreadCount> ThreadStates_;
+
+    TPriorityQueue<TExecutionPool> ActivePoolsHeap_;
+    TCpuDuration LastPoolExcessTime_ = 0;
+    TPriorityQueue<TEnqueuedTime> WaitHeap_;
+
+    // Buffer to keep actions during distribution to threads.
+    std::array<TAction, TThreadPoolBase::MaxThreadCount> OtherActions_;
+    std::atomic<int> ThreadCount_ = 0;
+    std::atomic<int> ActiveThreads_ = 0;
+
+    TCallbackList<TWaitTimeObserver::TSignature> WaitTimeObservers_;
+
+    TProfiler GetPoolProfiler(const TString& poolName) override
     {
-        TExecutionPool(const TString& poolName, const TProfiler& profiler)
-            : PoolName(poolName)
-            , BucketCounter(profiler.Gauge("/buckets"))
-            , SizeCounter(profiler.Summary("/size"))
-            , WaitTimeCounter(profiler.Timer("/time/wait"))
-            , ExecTimeCounter(profiler.Timer("/time/exec"))
-            , TotalTimeCounter(profiler.Timer("/time/total"))
-        { }
+        return Profiler_.WithTags(GetBucketTags(ThreadNamePrefix_, poolName));
+    }
 
-        TBucketPtr GetStarvingBucket(TEnqueuedAction* action)
-        {
-            if (!Heap.empty()) {
-                auto bucket = Heap.front().Bucket;
-                YT_VERIFY(!bucket->Queue.empty());
-                *action = std::move(bucket->Queue.front());
-                bucket->Queue.pop();
-
-                if (bucket->Queue.empty()) {
-                    ExtractHeap(Heap.begin(), Heap.end());
-                    Heap.pop_back();
-                }
+    Y_NO_INLINE void ConsumeInvokeQueue()
+    {
+        YT_ASSERT_SPINLOCK_AFFINITY(MainLock_);
 
-                return bucket;
-            }
+        Y_UNUSED(Padding0_);
+        Y_UNUSED(Padding1_);
 
-            return nullptr;
-        }
+        InvokeQueue_.DequeueAll(true, [&] (auto& action) {
+            auto* bucket = action.BucketHolder.Get();
 
-        const TString PoolName;
+            auto* pool = bucket->Pool.Get();
 
-        TGauge BucketCounter;
-        std::atomic<i64> Size = 0;
-        NProfiling::TSummary SizeCounter;
-        TEventTimer WaitTimeCounter;
-        TEventTimer ExecTimeCounter;
-        TEventTimer TotalTimeCounter;
+            YT_VERIFY(!pool->LinkedListNode.Next && !pool->LinkedListNode.Prev);
 
-        double Weight = 1.0;
+            if (!pool->GetPositionInHeap()) {
+                // ExcessTime can be greater than last pool excess time
+                // if the pool is "currently executed" and reschedules action.
+                if (pool->ExcessTime < LastPoolExcessTime_) {
+                    // Use last pool excess time to schedule new pool
+                    // after earlier scheduled pools (and not yet executed) in queue.
 
-        TCpuDuration ExcessTime = 0;
-        std::vector<THeapItem> Heap;
-        THashMap<TFairShareThreadPoolTag, TBucket*> TagToBucket;
-    };
+                    YT_LOG_DEBUG_IF(VerboseLogging_, "Initial pool excess time (Name: %v, ExcessTime: %v -> %v)",
+                        pool->PoolName,
+                        pool->ExcessTime,
+                        LastPoolExcessTime_);
 
-    const TIntrusivePtr<NThreading::TEventCount> CallbackEventCount_;
-    const TString ThreadNamePrefix_;
-    const TProfiler Profiler_;
-    const IPoolWeightProviderPtr PoolWeightProvider_;
+                    pool->ExcessTime = LastPoolExcessTime_;
+                }
 
-    YT_DECLARE_SPIN_LOCK(NThreading::TSpinLock, SpinLock_);
-    // NB: We set this flag to true so that whomever may spam tasks to queue
-    // will stop doing so after the shutdown.
-    bool Stopping_ = false;
-    std::vector<std::unique_ptr<TExecutionPool>> IdToPool_;
-    THashMap<TString, int> NameToPoolId_;
+                ActivePoolsHeap_.Insert(pool);
+            }
 
-    std::atomic<int> ThreadCount_ = 0;
-    std::array<TThreadState, TThreadPoolBase::MaxThreadCount> ThreadStates_;
+            ++pool->ActionCountInQueue;
 
-    TCallbackList<ITwoLevelFairShareThreadPool::TWaitTimeObserver::TSignature> WaitTimeObservers_;
+            auto enqueuedAt = action.EnqueuedAt;
 
+            bool wasEmpty = bucket->ActionQueue.empty();
+            bucket->ActionQueue.push(std::move(action));
 
-    size_t GetLowestEmptyPoolId()
-    {
-        YT_ASSERT_SPINLOCK_AFFINITY(SpinLock_);
+            if (wasEmpty) {
+                bucket->EnqueuedTime.Value = enqueuedAt;
+            }
 
-        size_t id = 0;
-        while (id < IdToPool_.size() && IdToPool_[id]) {
-            ++id;
-        }
-        return id;
+            YT_ASSERT(wasEmpty == !bucket->GetPositionInHeap());
+
+            if (!bucket->GetPositionInHeap()) {
+                // ExcessTime can be greater than last bucket excess time
+                // if the bucket is "currently executed" and reschedules action.
+                if (bucket->ExcessTime < pool->LastBucketExcessTime) {
+                    // Use last bucket excess time to schedule new bucket
+                    // after earlier scheduled buckets (and not yet executed) in queue.
+
+                    YT_LOG_DEBUG_IF(VerboseLogging_, "Initial bucket excess time (Name: %v, ExcessTime: %v -> %v)",
+                        bucket->BucketName,
+                        bucket->ExcessTime,
+                        pool->LastBucketExcessTime);
+
+                    bucket->ExcessTime = pool->LastBucketExcessTime;
+                }
+
+                pool->ActiveBucketsHeap.Insert(bucket);
+                pool->BucketCounter.Record(pool->ActiveBucketsHeap.GetSize());
+                WaitHeap_.Insert(&bucket->EnqueuedTime);
+            }
+        });
     }
 
-    void AccountCurrentlyExecutingBuckets()
+    void ServeBeginExecute(TThreadState* threadState, TCpuInstant currentInstant, TAction action)
     {
-        YT_ASSERT_SPINLOCK_AFFINITY(SpinLock_);
+        YT_ASSERT_SPINLOCK_AFFINITY(MainLock_);
 
-        auto currentInstant = GetCpuInstant();
-        auto threadCount = ThreadCount_.load();
-        for (int index = 0; index < threadCount; ++index) {
-            auto& threadState = ThreadStates_[index];
-            if (!threadState.Bucket) {
-                continue;
-            }
+        YT_ASSERT(!threadState->Action.Callback);
+        YT_ASSERT(!threadState->Action.BucketHolder);
 
-            auto duration = currentInstant - threadState.AccountedAt;
-            threadState.AccountedAt = currentInstant;
+        action.StartedAt = currentInstant;
 
-            UpdateExcessTime(threadState.Bucket.Get(), duration);
-        }
+        threadState->AccountedAt = currentInstant;
+        threadState->Action = std::move(action);
     }
 
-    void UpdateExcessTime(TBucket* bucket, TCpuDuration duration)
+    void ServeEndExecute(TThreadState* threadState, TCpuInstant /*cpuInstant*/)
     {
-        YT_ASSERT_SPINLOCK_AFFINITY(SpinLock_);
-
-        const auto& pool = IdToPool_[bucket->PoolId];
+        YT_ASSERT_SPINLOCK_AFFINITY(MainLock_);
 
-        pool->ExcessTime += duration / pool->Weight;
-        bucket->ExcessTime += duration;
+        auto action = std::move(threadState->Action);
+        YT_ASSERT(!threadState->Action.Callback);
+        YT_ASSERT(!action.Callback);
 
-        auto positionInHeap = bucket->HeapIterator;
-        if (!positionInHeap) {
+        if (!action.BucketHolder) {
+            // There was no action in begin execute.
             return;
         }
 
-        size_t indexInHeap = positionInHeap - pool->Heap.data();
-        YT_VERIFY(indexInHeap < pool->Heap.size());
-        SiftDown(pool->Heap.begin(), pool->Heap.end(), pool->Heap.begin() + indexInHeap, std::less<>());
+        // Try not to change bucket ref count under lock.
+        auto bucket = std::move(action.BucketHolder);
+
+        auto& pool = *bucket->Pool;
+        YT_ASSERT(pool.PoolName == bucket->PoolName);
+
+        // LastActionsInQueue is used to update SizeCounter outside lock.
+        threadState->LastActionsInQueue = --pool.ActionCountInQueue;
+
+        // Do not destroy bucket pointer under lock. Move it in thread state in other place and
+        // destroy in corresponding thread after combiner.
+        threadState->BucketToUnref = std::move(bucket);
     }
 
-    TBucketPtr GetStarvingBucket(TEnqueuedAction* action)
+    Y_NO_INLINE void UpdateExcessTime(TBucket* bucket, TCpuDuration duration, TCpuInstant currentInstant)
     {
-        YT_ASSERT_SPINLOCK_AFFINITY(SpinLock_);
+        YT_ASSERT_SPINLOCK_AFFINITY(MainLock_);
 
-        // For each currently evaluating buckets recalculate excess time.
-        AccountCurrentlyExecutingBuckets();
+        auto* pool = bucket->Pool.Get();
 
-        // Compute min excess over non-empty queues.
-        auto minExcessTime = std::numeric_limits<NProfiling::TCpuDuration>::max();
+        if (PoolWeightProvider_ && pool->NextUpdateWeightInstant < currentInstant) {
+            pool->NextUpdateWeightInstant = currentInstant + DurationToCpuDuration(TDuration::Seconds(1));
+            pool->InverseWeight = 1.0 / PoolWeightProvider_->GetWeight(pool->PoolName);
+        }
 
-        int minPoolIndex = -1;
-        for (int index = 0; index < std::ssize(IdToPool_); ++index) {
-            const auto& pool = IdToPool_[index];
-            if (pool && !pool->Heap.empty() && pool->ExcessTime < minExcessTime) {
-                minExcessTime = pool->ExcessTime;
-                minPoolIndex = index;
-            }
+        YT_LOG_DEBUG_IF(VerboseLogging_, "Increment excess time (BucketName: %v, PoolName: %v, ExcessTime: %v -> %v)",
+            bucket->BucketName,
+            bucket->PoolName,
+            bucket->ExcessTime,
+            bucket->ExcessTime + duration);
+
+        pool->ExcessTime += duration * pool->InverseWeight;
+        bucket->ExcessTime += duration;
+
+        if (auto* positionInHeap = pool->GetPositionInHeap()) {
+            ActivePoolsHeap_.AdjustDown(pool);
+        }
+
+        if (auto* positionInHeap = bucket->GetPositionInHeap()) {
+            pool->ActiveBucketsHeap.AdjustDown(bucket);
         }
 
-        YT_LOG_TRACE(
+        // No need to update wait heap.
+        YT_ASSERT(!bucket->EnqueuedTime.GetPositionInHeap() == !bucket->GetPositionInHeap());
+    }
+
+    Y_NO_INLINE bool GetStarvingBucket(TAction* action)
+    {
+        YT_ASSERT_SPINLOCK_AFFINITY(MainLock_);
+
+        YT_LOG_DEBUG_IF(
+            VerboseLogging_,
             "Buckets: %v",
             MakeFormattableView(
-                xrange(size_t(0), IdToPool_.size()),
+                xrange(size_t(0), ActivePoolsHeap_.GetSize()),
                 [&] (auto* builder, auto index) {
-                    const auto& pool = IdToPool_[index];
-                    if (!pool) {
-                        builder->AppendString("<null>");
-                        return;
-                    }
-                    builder->AppendFormat("[%v %v ", index, pool->ExcessTime);
-                    for (const auto& [tagId, rawBucket] : pool->TagToBucket) {
-                        if (auto bucket = DangerousGetPtr<TBucket>(rawBucket)) {
-                            auto excess = CpuDurationToDuration(bucket->ExcessTime).MilliSeconds();
-                            builder->AppendFormat("(%v %v) ", tagId, excess);
-                        } else {
-                            builder->AppendFormat("(%v ?) ", tagId);
-                        }
+                    auto& pool = ActivePoolsHeap_[index];
+                    builder->AppendFormat("%v [", CpuDurationToDuration(pool.ExcessTime));
+
+                    for (size_t bucketIndex = 0; bucketIndex < pool.ActiveBucketsHeap.GetSize(); ++bucketIndex) {
+                        const auto& bucket = pool.ActiveBucketsHeap[bucketIndex];
+
+                        builder->AppendFormat("%Qv:%v/%v ",
+                            bucket.BucketName,
+                            CpuDurationToDuration(bucket.ExcessTime),
+                            bucket.ActionQueue.front().EnqueuedAt);
                     }
                     builder->AppendFormat("]");
                 }));
 
-        if (minPoolIndex >= 0) {
-            // Reduce excesses (with truncation).
-            auto delta = IdToPool_[minPoolIndex]->ExcessTime;
-            for (const auto& pool : IdToPool_) {
-                if (pool) {
-                    pool->ExcessTime = std::max<NProfiling::TCpuDuration>(pool->ExcessTime - delta, 0);
-                }
+        if (ActivePoolsHeap_.Empty()) {
+            return false;
+        }
+
+        auto* pool = ActivePoolsHeap_.GetFront();
+        LastPoolExcessTime_ = pool->ExcessTime;
+
+        auto* bucket = pool->ActiveBucketsHeap.GetFront();
+        pool->LastBucketExcessTime = bucket->ExcessTime;
+
+        YT_ASSERT(!bucket->ActionQueue.empty());
+        *action = std::move(bucket->ActionQueue.front());
+        bucket->ActionQueue.pop();
+
+        YT_ASSERT(bucket == action->BucketHolder);
+
+        if (bucket->ActionQueue.empty()) {
+            bucket->EnqueuedTime.Value = std::numeric_limits<TCpuInstant>::max();
+
+            WaitHeap_.Extract(&bucket->EnqueuedTime);
+
+            pool->ActiveBucketsHeap.Extract(bucket);
+            pool->BucketCounter.Record(pool->ActiveBucketsHeap.GetSize());
+
+            if (pool->ActiveBucketsHeap.Empty()) {
+                ActivePoolsHeap_.Extract(pool);
             }
-            return IdToPool_[minPoolIndex]->GetStarvingBucket(action);
+        } else {
+            bucket->EnqueuedTime.Value = bucket->ActionQueue.front().EnqueuedAt;
+            WaitHeap_.AdjustDown(&bucket->EnqueuedTime);
         }
 
-        return nullptr;
+        return true;
     }
-};
 
-DEFINE_REFCOUNTED_TYPE(TTwoLevelFairShareQueue)
+    Y_NO_INLINE std::tuple<int, int> ServeCombinedRequests(TCpuInstant currentInstant, int currentThreadIndex)
+    {
+        YT_ASSERT_SPINLOCK_AFFINITY(MainLock_);
 
-////////////////////////////////////////////////////////////////////////////////
+        auto threadCount = ThreadCount_.load();
+        // Thread pool size can be reconfigures during serving requests.
+        threadCount = std::max(threadCount, currentThreadIndex + 1);
 
-void TBucket::Invoke(TClosure callback)
-{
-    if (auto parent = Parent.Lock()) {
-        parent->Invoke(std::move(callback), this);
+        // Saved thread requests. They must be saved before consuming invoke queue.
+        std::array<bool, TThreadPoolBase::MaxThreadCount> threadRequests{false};
+        std::array<int, TThreadPoolBase::MaxThreadCount> threadIds;
+        int requestCount = 0;
+
+        YT_LOG_TRACE("Updating excess time");
+
+        // Recalculate excess time for all currently evaluating or evaluated recently (end execute) buckets
+        for (int threadIndex = 0; threadIndex < threadCount; ++threadIndex) {
+            auto& threadState = ThreadStates_[threadIndex];
+
+            // TODO(lukyan): Can skip (for threads without requests) or throttle UpdateExcessTime if it happens frequently.
+            // For each currently evaluating buckets recalculate excess time.
+            if (auto* bucket = threadState.Action.BucketHolder.Get()) {
+
+                // TODO(lukyan): Update last excess time for pool without active buckets.
+                UpdateExcessTime(bucket, currentInstant - threadState.AccountedAt, currentInstant);
+                threadState.AccountedAt = currentInstant;
+            }
+
+            auto request = threadState.Request.load(std::memory_order::acquire);
+            if (request != ERequest::None) {
+                ServeEndExecute(&threadState, currentInstant);
+
+                if (request == ERequest::FetchNext) {
+                    // Save requests before ConsumeInvokeQueue. Otherwise some thread can schedule action
+                    // but action can not be fetched (schedule and fetch happens after ConsumeInvokeQueue).
+                    threadRequests[threadIndex] = true;
+                    threadIds[requestCount++] = threadIndex;
+                } else {
+                    threadState.Request.store(ERequest::None, std::memory_order::release);
+                }
+            }
+        }
+
+        YT_LOG_TRACE("Consuming invoke queue");
+
+        ConsumeInvokeQueue();
+
+        int fetchedActions = 0;
+        int otherActionCount = 0;
+
+        // Schedule actions to desired threads.
+        while (fetchedActions < requestCount) {
+            TAction action;
+
+            if (!GetStarvingBucket(&action)) {
+                break;
+            }
+
+            ++fetchedActions;
+
+            int threadIndex = -1;
+
+            auto unpackedCookie = TTaggedPtr<TTwoLevelFairShareQueue>::Unpack(action.EnqueuedThreadCookie);
+            // TODO(lukyan): Check also wait time. If it is too high, no matter where to schedule.
+            if (unpackedCookie.Ptr == this) {
+                threadIndex = unpackedCookie.Tag;
+            }
+
+            if (threadIndex != -1 && threadRequests[threadIndex]) {
+                ServeBeginExecute(&ThreadStates_[threadIndex], currentInstant, std::move(action));
+                threadRequests[threadIndex] = false;
+                ThreadStates_[threadIndex].Request.store(ERequest::None, std::memory_order::release);
+            } else {
+                OtherActions_[otherActionCount++] = std::move(action);
+            }
+        }
+
+        // Schedule other actions.
+        for (int threadIndex : TRange(threadIds.data(), requestCount)) {
+            if (threadRequests[threadIndex]) {
+                if (otherActionCount > 0) {
+                    ServeBeginExecute(&ThreadStates_[threadIndex], currentInstant, std::move(OtherActions_[--otherActionCount]));
+                }
+
+                ThreadStates_[threadIndex].Request.store(ERequest::None, std::memory_order::release);
+            }
+        }
+
+        return {requestCount, fetchedActions};
     }
-}
 
-void TBucket::Invoke(TMutableRange<TClosure> callbacks)
-{
-    if (auto parent = Parent.Lock()) {
-        for (auto& callback : callbacks) {
-            parent->Invoke(std::move(callback), this);
+    TCpuInstant GetMinEnqueuedAt()
+    {
+        YT_ASSERT_SPINLOCK_AFFINITY(MainLock_);
+
+        return WaitHeap_.Empty()
+            ? std::numeric_limits<TCpuInstant>::max()
+            : WaitHeap_.GetFront()->Value;
+    }
+
+    TClosure DoOnExecute(int index, bool fetchNext)
+    {
+        auto cpuInstant = GetCpuInstant();
+        auto& threadState = ThreadStates_[index];
+
+        const auto& oldAction = threadState.Action;
+        if (oldAction.BucketHolder) {
+            auto waitTime = CpuDurationToDuration(oldAction.StartedAt - oldAction.EnqueuedAt);
+            auto timeFromStart = CpuDurationToDuration(cpuInstant - oldAction.StartedAt);
+            auto timeFromEnqueue = CpuDurationToDuration(cpuInstant - oldAction.EnqueuedAt);
+
+            threadState.TimeFromStart = timeFromStart;
+            threadState.TimeFromEnqueue = timeFromEnqueue;
+
+            if (timeFromStart > LogDurationThreshold) {
+                YT_LOG_DEBUG("Callback execution took too long (Wait: %v, Execution: %v, Total: %v)",
+                    waitTime,
+                    timeFromStart,
+                    timeFromEnqueue);
+            }
+
+            if (waitTime > LogDurationThreshold) {
+                YT_LOG_DEBUG("Callback wait took too long (Wait: %v, Execution: %v, Total: %v)",
+                    waitTime,
+                    timeFromStart,
+                    timeFromEnqueue);
+            }
         }
+
+        auto finally = Finally([&] {
+            auto bucketToUndef = std::move(threadState.BucketToUnref);
+            if (bucketToUndef) {
+                auto* pool = bucketToUndef->Pool.Get();
+                pool->SizeCounter.Record(threadState.LastActionsInQueue);
+                pool->DequeuedCounter.Increment(1);
+                pool->ExecTimeCounter.Record(threadState.TimeFromStart);
+                pool->TotalTimeCounter.Record(threadState.TimeFromEnqueue);
+                pool->CumulativeTimeCounter.Add(threadState.TimeFromStart);
+                bucketToUndef.Reset();
+            }
+
+            const auto& action = threadState.Action;
+            if (action.BucketHolder) {
+                auto waitTime = CpuDurationToDuration(action.StartedAt - action.EnqueuedAt);
+                action.BucketHolder->Pool->WaitTimeCounter.Record(waitTime);
+                WaitTimeObservers_.Fire(waitTime);
+            }
+
+            MaybeRunMaintenance(&threadState, action.StartedAt, /*flush*/ false);
+
+            CumulativeSchedulingTimeCounter_.Add(CpuDurationToDuration(GetCpuInstant() - cpuInstant));
+
+            if (!fetchNext) {
+                MaybeProceedRetainQueue(cpuInstant);
+            }
+        });
+
+        auto& request = threadState.Request;
+        YT_VERIFY(request == ERequest::None);
+        request.store(fetchNext ? ERequest::FetchNext : ERequest::EndExecute);
+
+        if (MainLock_.IsLocked() || !MainLock_.TryAcquire()) {
+            // Locked here.
+            while (true) {
+                SpinLockPause();
+
+                if (request.load(std::memory_order::acquire) == ERequest::None) {
+                    return std::move(threadState.Action.Callback);
+                } else if (!MainLock_.IsLocked() && MainLock_.TryAcquire()) {
+                    break;
+                }
+            }
+        }
+
+        ResetMinEnqueuedAt();
+
+        YT_LOG_TRACE("Started serving requests");
+        auto [requests, fetchedActions] = ServeCombinedRequests(cpuInstant, index);
+
+        // Evaluate notify condition here, but call NotifyAfterFetch outside lock.
+        auto newMinEnqueuedAt = GetMinEnqueuedAt();
+        MainLock_.Release();
+
+        auto endInstant = GetCpuInstant();
+        YT_LOG_TRACE("Finished serving requests (Duration: %v, Requests: %v, FetchCount: %v, MinEnqueuedAt: %v)",
+            CpuDurationToDuration(endInstant - cpuInstant),
+            requests,
+            fetchedActions,
+            CpuInstantToInstant(newMinEnqueuedAt));
+
+        NotifyAfterFetch(endInstant, newMinEnqueuedAt);
+
+        return std::move(threadState.Action.Callback);
     }
-}
 
-TBucket::~TBucket()
-{
-    if (auto parent = Parent.Lock()) {
-        parent->RemoveBucket(this);
+    static void MaybeRunMaintenance(TThreadState* threadState, TCpuInstant now, bool flush)
+    {
+        YT_ASSERT(threadState);
+
+        constexpr i64 MaintenancePeriod = 1'000'000'000;
+        if (flush || now > threadState->LastMaintenanceInstant + MaintenancePeriod) {
+            ReclaimHazardPointers(false);
+            threadState->LastMaintenanceInstant  = now;
+        }
     }
-}
+};
+
+DEFINE_REFCOUNTED_TYPE(TTwoLevelFairShareQueue)
 
 ////////////////////////////////////////////////////////////////////////////////
 
@@ -631,16 +1281,33 @@ class TFairShareThread
     const TTwoLevelFairShareQueuePtr Queue_;
     const int Index_;
 
-    TEnqueuedAction CurrentAction_;
+    void OnStart() override
+    {
+        ThreadCookie() = TTaggedPtr(Queue_.Get(), static_cast<ui16>(Index_)).Pack();
+    }
+
+    void StopPrologue() override
+    {
+        Queue_->StopPrologue();
+    }
+
+    TClosure OnExecute() override
+    {
+        bool fetchNext = !TSchedulerThread::IsStopping() || TSchedulerThread::GracefulStop_;
+
+        return Queue_->OnExecute(Index_, fetchNext, [&] {
+            return TSchedulerThread::IsStopping();
+        });
+    }
 
     TClosure BeginExecute() override
     {
-        return BeginExecuteImpl(Queue_->BeginExecute(&CurrentAction_, Index_), &CurrentAction_);
+        Y_UNREACHABLE();
     }
 
     void EndExecute() override
     {
-        Queue_->EndExecute(&CurrentAction_, Index_);
+        Y_UNREACHABLE();
     }
 };
 
@@ -656,15 +1323,14 @@ class TTwoLevelFairShareThreadPool
     TTwoLevelFairShareThreadPool(
         int threadCount,
         const TString& threadNamePrefix,
-        IPoolWeightProviderPtr poolWeightProvider)
+        const TNewTwoLevelFairShareThreadPoolOptions& options)
         : TThreadPoolBase(threadNamePrefix)
         , Queue_(New<TTwoLevelFairShareQueue>(
             CallbackEventCount_,
             ThreadNamePrefix_,
-            std::move(poolWeightProvider)))
+            options))
     {
         SetThreadCount(threadCount);
-        EnsureStarted();
     }
 
     ~TTwoLevelFairShareThreadPool()
@@ -677,20 +1343,17 @@ class TTwoLevelFairShareThreadPool
         TThreadPoolBase::SetThreadCount(threadCount);
     }
 
-    void SetPollingPeriod(TDuration /*pollingPeriod*/) override
-    { }
-
-    int GetThreadCount() override
+    void SetPollingPeriod(TDuration pollingPeriod) override
     {
-        return TThreadPoolBase::GetThreadCount();
+        Queue_->Configure(pollingPeriod);
     }
 
     IInvokerPtr GetInvoker(
         const TString& poolName,
-        const TFairShareThreadPoolTag& tag) override
+        const TFairShareThreadPoolTag& bucketName) override
     {
         EnsureStarted();
-        return Queue_->GetInvoker(poolName, tag);
+        return Queue_->GetInvoker(poolName, bucketName);
     }
 
     void Shutdown() override
@@ -698,6 +1361,11 @@ class TTwoLevelFairShareThreadPool
         TThreadPoolBase::Shutdown();
     }
 
+    int GetThreadCount() override
+    {
+        return TThreadPoolBase::GetThreadCount();
+    }
+
     void SubscribeWaitTimeObserved(const TWaitTimeObserver& callback) override
     {
         Queue_->SubscribeWaitTimeObserved(callback);
@@ -712,11 +1380,11 @@ class TTwoLevelFairShareThreadPool
     const TIntrusivePtr<NThreading::TEventCount> CallbackEventCount_ = New<NThreading::TEventCount>();
     const TTwoLevelFairShareQueuePtr Queue_;
 
-
     void DoShutdown() override
     {
         Queue_->Shutdown();
         TThreadPoolBase::DoShutdown();
+        Queue_->Drain();
     }
 
     void DoSetThreadCount(int threadCount) override
@@ -743,12 +1411,12 @@ class TTwoLevelFairShareThreadPool
 ITwoLevelFairShareThreadPoolPtr CreateTwoLevelFairShareThreadPool(
     int threadCount,
     const TString& threadNamePrefix,
-    IPoolWeightProviderPtr poolWeightProvider)
+    const TNewTwoLevelFairShareThreadPoolOptions& options)
 {
     return New<TTwoLevelFairShareThreadPool>(
         threadCount,
         threadNamePrefix,
-        std::move(poolWeightProvider));
+        options);
 }
 
 ////////////////////////////////////////////////////////////////////////////////
diff --git a/yt/yt/core/concurrency/two_level_fair_share_thread_pool.h b/yt/yt/core/concurrency/two_level_fair_share_thread_pool.h
index 57d616391b58..611b9e7219e6 100644
--- a/yt/yt/core/concurrency/two_level_fair_share_thread_pool.h
+++ b/yt/yt/core/concurrency/two_level_fair_share_thread_pool.h
@@ -40,11 +40,22 @@ struct ITwoLevelFairShareThreadPool
 
 DEFINE_REFCOUNTED_TYPE(ITwoLevelFairShareThreadPool)
 
+////////////////////////////////////////////////////////////////////////////////
+
+struct TNewTwoLevelFairShareThreadPoolOptions
+{
+    IPoolWeightProviderPtr PoolWeightProvider = nullptr;
+    bool VerboseLogging = false;
+    TDuration PollingPeriod = TDuration::MilliSeconds(10);
+    TDuration PoolRetentionTime = TDuration::Seconds(30);
+};
+
+////////////////////////////////////////////////////////////////////////////////
+
 ITwoLevelFairShareThreadPoolPtr CreateTwoLevelFairShareThreadPool(
     int threadCount,
     const TString& threadNamePrefix,
-    IPoolWeightProviderPtr poolWeightProvider = nullptr);
+    const TNewTwoLevelFairShareThreadPoolOptions& options = {});
 
-////////////////////////////////////////////////////////////////////////////////
 
 } // namespace NYT::NConcurrency
diff --git a/yt/yt/core/concurrency/unittests/scheduler_ut.cpp b/yt/yt/core/concurrency/unittests/scheduler_ut.cpp
index 445b98543a83..22ca6eb90fac 100644
--- a/yt/yt/core/concurrency/unittests/scheduler_ut.cpp
+++ b/yt/yt/core/concurrency/unittests/scheduler_ut.cpp
@@ -13,7 +13,6 @@
 #include <yt/yt/core/concurrency/thread_affinity.h>
 #include <yt/yt/core/concurrency/fair_share_thread_pool.h>
 #include <yt/yt/core/concurrency/two_level_fair_share_thread_pool.h>
-#include <yt/yt/core/concurrency/new_fair_share_thread_pool.h>
 
 #include <yt/yt/core/logging/log.h>
 
@@ -1268,7 +1267,7 @@ TEST_P(TFairShareSchedulerTest, TwoLevelFairness)
     YT_VERIFY(numWorkers > numPools);
     YT_VERIFY(numThreads <= numWorkers);
 
-    auto threadPool = CreateNewTwoLevelFairShareThreadPool(
+    auto threadPool = CreateTwoLevelFairShareThreadPool(
         numThreads,
         "MyFairSharePool",
         {
diff --git a/yt/yt/core/concurrency/unittests/two_level_fair_share_thread_pool_ut.cpp b/yt/yt/core/concurrency/unittests/two_level_fair_share_thread_pool_ut.cpp
index c155f00fb893..97a0dbbe83a9 100644
--- a/yt/yt/core/concurrency/unittests/two_level_fair_share_thread_pool_ut.cpp
+++ b/yt/yt/core/concurrency/unittests/two_level_fair_share_thread_pool_ut.cpp
@@ -1,7 +1,6 @@
 #include <yt/yt/core/test_framework/framework.h>
 
 #include <yt/yt/core/concurrency/two_level_fair_share_thread_pool.h>
-#include <yt/yt/core/concurrency/new_fair_share_thread_pool.h>
 
 #include <yt/yt/core/actions/invoker.h>
 #include <yt/yt/core/actions/future.h>
@@ -15,7 +14,7 @@ namespace {
 
 TEST(TTwoLevelFairShareThreadPoolTest, Configure)
 {
-    auto threadPool = CreateNewTwoLevelFairShareThreadPool(1, "Test");
+    auto threadPool = CreateTwoLevelFairShareThreadPool(1, "Test");
     auto counter = std::make_shared<std::atomic<int>>();
     auto callback = BIND([=] { ++*counter; });
     std::vector<TFuture<void>> futures;
diff --git a/yt/yt/core/misc/ref_counted_tracker.h b/yt/yt/core/misc/ref_counted_tracker.h
index 0a3a0de8a443..6cd482da9b7f 100644
--- a/yt/yt/core/misc/ref_counted_tracker.h
+++ b/yt/yt/core/misc/ref_counted_tracker.h
@@ -13,7 +13,7 @@ namespace NYT {
 
 ////////////////////////////////////////////////////////////////////////////////
 
-struct TRefCountedTrackerStatistics
+struct TRefCountedTrackerStatistics final
 {
     struct TStatistics
     {
diff --git a/yt/yt/core/misc/ref_counted_tracker_statistics_producer.cpp b/yt/yt/core/misc/ref_counted_tracker_statistics_producer.cpp
index e6b32125065a..8330e63c5f79 100644
--- a/yt/yt/core/misc/ref_counted_tracker_statistics_producer.cpp
+++ b/yt/yt/core/misc/ref_counted_tracker_statistics_producer.cpp
@@ -2,6 +2,10 @@
 
 #include <yt/yt/core/ytree/fluent.h>
 
+#include <library/cpp/yt/memory/leaky_singleton.h>
+
+#include <library/cpp/yt/threading/spin_lock.h>
+
 namespace NYT {
 
 using namespace NYson;
@@ -9,34 +13,101 @@ using namespace NYTree;
 
 ////////////////////////////////////////////////////////////////////////////////
 
+namespace {
+
+void Produce(IYsonConsumer* consumer, const TRefCountedTrackerStatistics& statistics)
+{
+    BuildYsonFluently(consumer)
+        .BeginMap()
+            .Item("statistics")
+                .DoListFor(statistics.NamedStatistics, [] (TFluentList fluent, const auto& namedSlotStatistics) {
+                    fluent
+                        .Item().BeginMap()
+                            .Item("name").Value(namedSlotStatistics.FullName)
+                            .Item("objects_alive").Value(namedSlotStatistics.ObjectsAlive)
+                            .Item("objects_allocated").Value(namedSlotStatistics.ObjectsAllocated)
+                            .Item("bytes_alive").Value(namedSlotStatistics.BytesAlive)
+                            .Item("bytes_allocated").Value(namedSlotStatistics.BytesAllocated)
+                        .EndMap();
+                })
+            .Item("total").BeginMap()
+                .Item("objects_alive").Value(statistics.TotalStatistics.ObjectsAlive)
+                .Item("objects_allocated").Value(statistics.TotalStatistics.ObjectsAllocated)
+                .Item("bytes_alive").Value(statistics.TotalStatistics.BytesAlive)
+                .Item("bytes_allocated").Value(statistics.TotalStatistics.BytesAllocated)
+            .EndMap()
+        .EndMap();
+}
+
+} // namespace
+
+////////////////////////////////////////////////////////////////////////////////
+
 TYsonProducer CreateRefCountedTrackerStatisticsProducer()
 {
     return BIND([] (IYsonConsumer* consumer) {
-        auto statistics = TRefCountedTracker::Get()->GetStatistics();
-
-        BuildYsonFluently(consumer)
-            .BeginMap()
-                .Item("statistics")
-                    .DoListFor(statistics.NamedStatistics, [] (TFluentList fluent, const auto& namedSlotStatistics) {
-                        fluent
-                            .Item().BeginMap()
-                                .Item("name").Value(namedSlotStatistics.FullName)
-                                .Item("objects_alive").Value(namedSlotStatistics.ObjectsAlive)
-                                .Item("objects_allocated").Value(namedSlotStatistics.ObjectsAllocated)
-                                .Item("bytes_alive").Value(namedSlotStatistics.BytesAlive)
-                                .Item("bytes_allocated").Value(namedSlotStatistics.BytesAllocated)
-                            .EndMap();
-                    })
-                .Item("total").BeginMap()
-                    .Item("objects_alive").Value(statistics.TotalStatistics.ObjectsAlive)
-                    .Item("objects_allocated").Value(statistics.TotalStatistics.ObjectsAllocated)
-                    .Item("bytes_alive").Value(statistics.TotalStatistics.BytesAlive)
-                    .Item("bytes_allocated").Value(statistics.TotalStatistics.BytesAllocated)
-                .EndMap()
-            .EndMap();
+        Produce(consumer, TRefCountedTracker::Get()->GetStatistics());
     });
 }
 
 ////////////////////////////////////////////////////////////////////////////////
 
+class TCachingRefCountedTrackerStatisticsManager
+{
+public:
+    TYsonProducer GetProducer()
+    {
+        return Producer_;
+    }
+
+private:
+    TCachingRefCountedTrackerStatisticsManager()
+        : Producer_(BIND([this] (IYsonConsumer* consumer) {
+            Produce(consumer, *GetCachedStatistics());
+        }))
+    { }
+
+    DECLARE_LEAKY_SINGLETON_FRIEND()
+
+    const TYsonProducer Producer_;
+
+    static constexpr auto CachedStatisticsTtl = TDuration::Seconds(10);
+
+    NThreading::TSpinLock CachedStatisticsLock_;
+    TIntrusivePtr<TRefCountedTrackerStatistics> CachedStatistics_;
+    TInstant CachedStatisticsUpdateTime_;
+
+    TIntrusivePtr<TRefCountedTrackerStatistics> GetCachedStatistics()
+    {
+        auto now = TInstant::Now();
+
+        // Fast path.
+        {
+            auto guard = Guard(CachedStatisticsLock_);
+            if (CachedStatistics_ && now < CachedStatisticsUpdateTime_ + CachedStatisticsTtl) {
+                return CachedStatistics_;
+            }
+        }
+
+        // Slow path.
+        auto statistics = New<TRefCountedTrackerStatistics>(TRefCountedTracker::Get()->GetStatistics());
+        {
+            auto guard = Guard(CachedStatisticsLock_);
+            if (!CachedStatistics_ || now > CachedStatisticsUpdateTime_ + CachedStatisticsTtl) {
+                CachedStatistics_ = std::move(statistics);
+            }
+            return CachedStatistics_;
+        }
+    }
+};
+
+////////////////////////////////////////////////////////////////////////////////
+
+TYsonProducer GetCachingRefCountedTrackerStatisticsProducer()
+{
+    return LeakySingleton<TCachingRefCountedTrackerStatisticsManager>()->GetProducer();
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
 } // namespace NYT
diff --git a/yt/yt/core/misc/ref_counted_tracker_statistics_producer.h b/yt/yt/core/misc/ref_counted_tracker_statistics_producer.h
index 92b2a8b62c2e..cfe7b7ebc9d3 100644
--- a/yt/yt/core/misc/ref_counted_tracker_statistics_producer.h
+++ b/yt/yt/core/misc/ref_counted_tracker_statistics_producer.h
@@ -8,8 +8,12 @@ namespace NYT {
 
 ////////////////////////////////////////////////////////////////////////////////
 
+//! Creates a producer for constructing the most up-to-date state of RCT.
 NYson::TYsonProducer CreateRefCountedTrackerStatisticsProducer();
 
+//! Returns a producer that applies a (reasonable) caching to RCT data.
+NYson::TYsonProducer GetCachingRefCountedTrackerStatisticsProducer();
+
 ////////////////////////////////////////////////////////////////////////////////
 
 } // namespace NYT
diff --git a/yt/yt/core/ya.make b/yt/yt/core/ya.make
index 41e2da5dfa3d..a47b5ddfbf3e 100644
--- a/yt/yt/core/ya.make
+++ b/yt/yt/core/ya.make
@@ -71,7 +71,6 @@ SRCS(
     concurrency/invoker_alarm.cpp
     concurrency/invoker_queue.cpp
     concurrency/lease_manager.cpp
-    concurrency/new_fair_share_thread_pool.cpp
     concurrency/nonblocking_batcher.cpp
     concurrency/notify_manager.cpp
     concurrency/periodic_executor.cpp
diff --git a/yt/yt/core/yson/protobuf_interop.cpp b/yt/yt/core/yson/protobuf_interop.cpp
index 1c09fbe86f8b..d552719d78db 100644
--- a/yt/yt/core/yson/protobuf_interop.cpp
+++ b/yt/yt/core/yson/protobuf_interop.cpp
@@ -808,7 +808,7 @@ TProtobufElement TProtobufField::GetElement(bool insideRepeated) const
         });
     } else {
         return std::make_unique<TProtobufScalarElement>(TProtobufScalarElement{
-            static_cast<TProtobufScalarElement::TType>(GetType()),
+            static_cast<TProtobufElementType>(GetType()),
             GetEnumYsonStorageType()
         });
     }
diff --git a/yt/yt/core/yson/protobuf_interop.h b/yt/yt/core/yson/protobuf_interop.h
index 6327336075e1..c82f5c293b9c 100644
--- a/yt/yt/core/yson/protobuf_interop.h
+++ b/yt/yt/core/yson/protobuf_interop.h
@@ -66,8 +66,7 @@ struct TProtobufMessageElement
 
 struct TProtobufScalarElement
 {
-    YT_DEFINE_STRONG_TYPEDEF(TType, int);
-    TType Type;
+    TProtobufElementType Type;
 
     // Meaningful only when TYPE == TYPE_ENUM.
     EEnumYsonStorageType EnumStorageType;
diff --git a/yt/yt/core/yson/public.h b/yt/yt/core/yson/public.h
index dae0c0b29888..1e757e187429 100644
--- a/yt/yt/core/yson/public.h
+++ b/yt/yt/core/yson/public.h
@@ -96,6 +96,9 @@ DECLARE_REFCOUNTED_CLASS(TProtobufInteropDynamicConfig);
  */
 class TProtobufMessageType;
 
+//! A reflected counterpart of ::google::protobuf::FieldDescriptor::Type.
+YT_DEFINE_STRONG_TYPEDEF(TProtobufElementType, int);
+
 //! An opaque reflected counterpart of ::google::protobuf::EnumDescriptor.
 class TProtobufEnumType;
 
diff --git a/yt/yt/core/yson/unittests/protobuf_yson_ut.cpp b/yt/yt/core/yson/unittests/protobuf_yson_ut.cpp
index 214d748ace78..a369a85275e5 100644
--- a/yt/yt/core/yson/unittests/protobuf_yson_ut.cpp
+++ b/yt/yt/core/yson/unittests/protobuf_yson_ut.cpp
@@ -2113,7 +2113,7 @@ void TestScalarByYPath(const TYPath& path, FieldDescriptor::Type type)
     EXPECT_EQ(path, result.HeadPath);
     EXPECT_EQ("", result.TailPath);
     EXPECT_EQ(
-        static_cast<TProtobufScalarElement::TType>(type),
+        static_cast<TProtobufElementType>(type),
         std::get<std::unique_ptr<TProtobufScalarElement>>(result.Element)->Type);
 }
 
diff --git a/yt/yt/library/tcmalloc/config.cpp b/yt/yt/library/tcmalloc/config.cpp
index b1e7cdf3c099..3219ffcf22d6 100644
--- a/yt/yt/library/tcmalloc/config.cpp
+++ b/yt/yt/library/tcmalloc/config.cpp
@@ -6,6 +6,25 @@ using namespace NYTree;
 
 ////////////////////////////////////////////////////////////////////////////////
 
+void THeapSizeLimitConfig::ApplyDynamicInplace(const TDynamicHeapSizeLimitConfigPtr& dynamicConfig)
+{
+    UpdateYsonStructField(ContainerMemoryRatio, dynamicConfig->ContainerMemoryRatio);
+    UpdateYsonStructField(ContainerMemoryMargin, dynamicConfig->ContainerMemoryMargin);
+    UpdateYsonStructField(Hard, dynamicConfig->Hard);
+    UpdateYsonStructField(DumpMemoryProfileOnViolation, dynamicConfig->DumpMemoryProfileOnViolation);
+    UpdateYsonStructField(MemoryProfileDumpTimeout, dynamicConfig->MemoryProfileDumpTimeout);
+    UpdateYsonStructField(MemoryProfileDumpPath, dynamicConfig->MemoryProfileDumpPath);
+    UpdateYsonStructField(MemoryProfileDumpFilenameSuffix, dynamicConfig->MemoryProfileDumpFilenameSuffix);
+}
+
+THeapSizeLimitConfigPtr THeapSizeLimitConfig::ApplyDynamic(const TDynamicHeapSizeLimitConfigPtr& dynamicConfig) const
+{
+    auto mergedConfig = CloneYsonStruct(MakeStrong(this));
+    mergedConfig->ApplyDynamicInplace(dynamicConfig);
+    mergedConfig->Postprocess();
+    return mergedConfig;
+}
+
 void THeapSizeLimitConfig::Register(TRegistrar registrar)
 {
     registrar.Parameter("container_memory_ratio", &TThis::ContainerMemoryRatio)
@@ -20,7 +39,7 @@ void THeapSizeLimitConfig::Register(TRegistrar registrar)
         .Default(TDuration::Minutes(10));
     registrar.Parameter("memory_profile_dump_path", &TThis::MemoryProfileDumpPath)
         .Default();
-    registrar.Parameter("emory_profile_dump_filename_suffix", &TThis::MemoryProfileDumpFilenameSuffix)
+    registrar.Parameter("memory_profile_dump_filename_suffix", &TThis::MemoryProfileDumpFilenameSuffix)
         .Default();
 
     registrar.Postprocessor([] (THeapSizeLimitConfig* config) {
@@ -30,17 +49,45 @@ void THeapSizeLimitConfig::Register(TRegistrar registrar)
     });
 }
 
-TTCMallocConfigPtr TTCMallocConfig::ApplyDynamic(const TTCMallocConfigPtr& dynamicConfig) const
+////////////////////////////////////////////////////////////////////////////////
+
+void TDynamicHeapSizeLimitConfig::Register(TRegistrar registrar)
 {
-    // TODO(babenko): fix this mess
-    auto mergedConfig = CloneYsonStruct(dynamicConfig);
-    mergedConfig->HeapSizeLimit->MemoryProfileDumpPath = HeapSizeLimit->MemoryProfileDumpPath;
-    mergedConfig->Postprocess();
-    return mergedConfig;
+    registrar.Parameter("container_memory_ratio", &TThis::ContainerMemoryRatio)
+        .Default();
+    registrar.Parameter("container_memory_margin", &TThis::ContainerMemoryMargin)
+        .Default();
+    registrar.Parameter("hard", &TThis::Hard)
+        .Default();
+    registrar.Parameter("dump_memory_profile_on_violation", &TThis::DumpMemoryProfileOnViolation)
+        .Default();
+    registrar.Parameter("memory_profile_dump_timeout", &TThis::MemoryProfileDumpTimeout)
+        .Default();
+    registrar.Parameter("memory_profile_dump_path", &TThis::MemoryProfileDumpPath)
+        .Default();
+    registrar.Parameter("memory_profile_dump_filename_suffix", &TThis::MemoryProfileDumpFilenameSuffix)
+        .Default();
 }
 
 ////////////////////////////////////////////////////////////////////////////////
 
+TTCMallocConfigPtr TTCMallocConfig::ApplyDynamic(const TDynamicTCMallocConfigPtr& dynamicConfig) const
+{
+    auto mergedConfig = CloneYsonStruct(MakeStrong(this));
+    UpdateYsonStructField(mergedConfig->AggressiveReleaseThreshold, dynamicConfig->AggressiveReleaseThreshold);
+    UpdateYsonStructField(mergedConfig->AggressiveReleaseThresholdRatio, dynamicConfig->AggressiveReleaseThresholdRatio);
+    UpdateYsonStructField(mergedConfig->AggressiveReleaseSize, dynamicConfig->AggressiveReleaseSize);
+    UpdateYsonStructField(mergedConfig->AggressiveReleasePeriod, dynamicConfig->AggressiveReleasePeriod);
+    UpdateYsonStructField(mergedConfig->GuardedSamplingRate, dynamicConfig->GuardedSamplingRate);
+    UpdateYsonStructField(mergedConfig->ProfileSamplingRate, dynamicConfig->ProfileSamplingRate);
+    UpdateYsonStructField(mergedConfig->MaxPerCpuCacheSize, dynamicConfig->MaxPerCpuCacheSize);
+    UpdateYsonStructField(mergedConfig->MaxTotalThreadCacheBytes, dynamicConfig->MaxTotalThreadCacheBytes);
+    UpdateYsonStructField(mergedConfig->BackgroundReleaseRate, dynamicConfig->BackgroundReleaseRate);
+    mergedConfig->HeapSizeLimit->ApplyDynamicInplace(dynamicConfig->HeapSizeLimit);
+    mergedConfig->Postprocess();
+    return mergedConfig;
+}
+
 void TTCMallocConfig::Register(TRegistrar registrar)
 {
     registrar.Parameter("aggressive_release_threshold", &TThis::AggressiveReleaseThreshold)
@@ -69,5 +116,33 @@ void TTCMallocConfig::Register(TRegistrar registrar)
 
 ////////////////////////////////////////////////////////////////////////////////
 
+void TDynamicTCMallocConfig::Register(TRegistrar registrar)
+{
+    registrar.Parameter("aggressive_release_threshold", &TThis::AggressiveReleaseThreshold)
+        .Default();
+    registrar.Parameter("aggressive_release_threshold_ratio", &TThis::AggressiveReleaseThresholdRatio)
+        .Optional();
+
+    registrar.Parameter("aggressive_release_size", &TThis::AggressiveReleaseSize)
+        .Default();
+    registrar.Parameter("aggressive_release_period", &TThis::AggressiveReleasePeriod)
+        .Default();
+    registrar.Parameter("guarded_sampling_rate", &TThis::GuardedSamplingRate)
+        .Default();
+    registrar.Parameter("profile_sampling_rate", &TThis::ProfileSamplingRate)
+        .Default();
+    registrar.Parameter("max_per_cpu_cache_size", &TThis::MaxPerCpuCacheSize)
+        .Default();
+    registrar.Parameter("max_total_thread_cache_bytes", &TThis::MaxTotalThreadCacheBytes)
+        .Default();
+    registrar.Parameter("background_release_rate", &TThis::BackgroundReleaseRate)
+        .Default();
+
+    registrar.Parameter("heap_size_limit", &TThis::HeapSizeLimit)
+        .DefaultNew();
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
 } // namespace NYT::NTCMalloc
 
diff --git a/yt/yt/library/tcmalloc/config.h b/yt/yt/library/tcmalloc/config.h
index 1783b1a65cfb..3c140b44308c 100644
--- a/yt/yt/library/tcmalloc/config.h
+++ b/yt/yt/library/tcmalloc/config.h
@@ -34,6 +34,9 @@ struct THeapSizeLimitConfig
     std::optional<TString> MemoryProfileDumpPath;
     std::optional<TString> MemoryProfileDumpFilenameSuffix;
 
+    void ApplyDynamicInplace(const TDynamicHeapSizeLimitConfigPtr& dynamicConfig);
+    THeapSizeLimitConfigPtr ApplyDynamic(const TDynamicHeapSizeLimitConfigPtr& dynamicConfig) const;
+
     REGISTER_YSON_STRUCT(THeapSizeLimitConfig);
 
     static void Register(TRegistrar registrar);
@@ -43,6 +46,30 @@ DEFINE_REFCOUNTED_TYPE(THeapSizeLimitConfig)
 
 ////////////////////////////////////////////////////////////////////////////////
 
+struct TDynamicHeapSizeLimitConfig
+    : public NYTree::TYsonStruct
+{
+    std::optional<double> ContainerMemoryRatio;
+    std::optional<i64> ContainerMemoryMargin;
+
+    std::optional<bool> Hard;
+
+    std::optional<bool> DumpMemoryProfileOnViolation;
+
+    std::optional<TDuration> MemoryProfileDumpTimeout;
+
+    std::optional<TString> MemoryProfileDumpPath;
+    std::optional<TString> MemoryProfileDumpFilenameSuffix;
+
+    REGISTER_YSON_STRUCT(TDynamicHeapSizeLimitConfig);
+
+    static void Register(TRegistrar registrar);
+};
+
+DEFINE_REFCOUNTED_TYPE(TDynamicHeapSizeLimitConfig)
+
+////////////////////////////////////////////////////////////////////////////////
+
 struct TTCMallocConfig
     : public NYTree::TYsonStruct
 {
@@ -66,7 +93,7 @@ struct TTCMallocConfig
 
     THeapSizeLimitConfigPtr HeapSizeLimit;
 
-    TTCMallocConfigPtr ApplyDynamic(const TTCMallocConfigPtr& dynamicConfig) const;
+    TTCMallocConfigPtr ApplyDynamic(const TDynamicTCMallocConfigPtr& dynamicConfig) const;
 
     REGISTER_YSON_STRUCT(TTCMallocConfig);
 
@@ -77,4 +104,32 @@ DEFINE_REFCOUNTED_TYPE(TTCMallocConfig)
 
 ////////////////////////////////////////////////////////////////////////////////
 
+struct TDynamicTCMallocConfig
+    : public NYTree::TYsonStruct
+{
+    std::optional<i64> AggressiveReleaseThreshold;
+
+    std::optional<double> AggressiveReleaseThresholdRatio;
+
+    std::optional<i64> AggressiveReleaseSize;
+    std::optional<TDuration> AggressiveReleasePeriod;
+
+    std::optional<i64> GuardedSamplingRate;
+
+    std::optional<i64> ProfileSamplingRate;
+    std::optional<i64> MaxPerCpuCacheSize;
+    std::optional<i64> MaxTotalThreadCacheBytes;
+    std::optional<i64> BackgroundReleaseRate;
+
+    TDynamicHeapSizeLimitConfigPtr HeapSizeLimit;
+
+    REGISTER_YSON_STRUCT(TDynamicTCMallocConfig);
+
+    static void Register(TRegistrar registrar);
+};
+
+DEFINE_REFCOUNTED_TYPE(TDynamicTCMallocConfig)
+
+////////////////////////////////////////////////////////////////////////////////
+
 } // namespace NYT::NTCMalloc
diff --git a/yt/yt/library/tcmalloc/configure_tcmalloc_manager.cpp b/yt/yt/library/tcmalloc/configure_tcmalloc_manager.cpp
index d5947bf185da..49829b935778 100644
--- a/yt/yt/library/tcmalloc/configure_tcmalloc_manager.cpp
+++ b/yt/yt/library/tcmalloc/configure_tcmalloc_manager.cpp
@@ -14,6 +14,11 @@ void SetupSingletonConfigParameter(TYsonStructParameter<TTCMallocConfigPtr>& par
     parameter.DefaultNew();
 }
 
+void SetupSingletonConfigParameter(TYsonStructParameter<TDynamicTCMallocConfigPtr>& parameter)
+{
+    parameter.DefaultNew();
+}
+
 void ConfigureSingleton(const TTCMallocConfigPtr& config)
 {
     TTCMallocManager::Configure(config);
@@ -21,7 +26,7 @@ void ConfigureSingleton(const TTCMallocConfigPtr& config)
 
 void ReconfigureSingleton(
     const TTCMallocConfigPtr& config,
-    const TTCMallocConfigPtr& dynamicConfig)
+    const TDynamicTCMallocConfigPtr& dynamicConfig)
 {
     TTCMallocManager::Configure(config->ApplyDynamic(dynamicConfig));
 }
@@ -29,7 +34,7 @@ void ReconfigureSingleton(
 YT_DEFINE_RECONFIGURABLE_SINGLETON(
     "tcmalloc",
     TTCMallocConfig,
-    TTCMallocConfig);
+    TDynamicTCMallocConfig);
 
 ////////////////////////////////////////////////////////////////////////////////
 
diff --git a/yt/yt/library/tcmalloc/public.h b/yt/yt/library/tcmalloc/public.h
index f6703f918c38..c9600da99a9d 100644
--- a/yt/yt/library/tcmalloc/public.h
+++ b/yt/yt/library/tcmalloc/public.h
@@ -10,6 +10,8 @@ namespace NYT::NTCMalloc {
 
 DECLARE_REFCOUNTED_STRUCT(TTCMallocConfig)
 DECLARE_REFCOUNTED_STRUCT(THeapSizeLimitConfig)
+DECLARE_REFCOUNTED_STRUCT(TDynamicTCMallocConfig)
+DECLARE_REFCOUNTED_STRUCT(TDynamicHeapSizeLimitConfig)
 
 YT_DECLARE_RECONFIGURABLE_SINGLETON(TTCMallocConfig, TTCMallocConfig);