feat: add session affinity to k8s TS (#2519)

* feat: add session affinity to k8s TS Signed-off-by: jagadeesh <[email protected]> * fix spell check Signed-off-by: jagadeesh <[email protected]> * fix docs Signed-off-by: jagadeesh <[email protected]> --------- Signed-off-by: jagadeesh <[email protected]> Co-authored-by: Geeta Chauhan <[email protected]> Co-authored-by: Ankith Gunapal <[email protected]>
pytorch · Aug 24, 2023 · d47b14d · d47b14d
1 parent 448aad3
commit d47b14d
Show file tree

Hide file tree

Showing 9 changed files with 132 additions and 11 deletions.
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
@@ -25,12 +25,12 @@ repos:
       - id: python-no-log-warn
       - id: python-use-type-annotations
   - repo: https://github.com/hadialqattan/pycln
-    rev: v2.1.3
+    rev: v2.1.5
     hooks:
       - id: pycln
         args: [--all]
   - repo: https://github.com/psf/black
-    rev: 23.1.0
+    rev: 23.7.0
     hooks:
       - id: black
         additional_dependencies: ['click==8.0.4']

diff --git a/kubernetes/Helm/templates/torchserve.yaml b/kubernetes/Helm/templates/torchserve.yaml
@@ -20,7 +20,9 @@ spec:
   - name: metrics
     port: {{ .Values.torchserve.metrics_port }}
     targetPort: ts-metrics
-  type: LoadBalancer
+  - name: grpc
+    port: {{ .Values.torchserve.grpc_inference_port }}
+    targetPort: ts-grpc
   selector:
     app: torchserve
 ---
@@ -55,6 +57,8 @@ spec:
           containerPort: {{ .Values.torchserve.management_port }}
         - name: ts-metrics
           containerPort: {{ .Values.torchserve.metrics_port }}
+        - name: ts-grpc
+          containerPort: {{ .Values.torchserve.grpc_inference_port }}
         imagePullPolicy: IfNotPresent
         volumeMounts:
           - mountPath: {{ .Values.torchserve.pvd_mount }}

diff --git a/kubernetes/Helm/values.yaml b/kubernetes/Helm/values.yaml
@@ -8,13 +8,15 @@ torchserve:
   management_port: 8081
   inference_port: 8080
   metrics_port: 8082
+  grpc_inference_port: 7070
+
   pvd_mount: /home/model-server/shared/
   n_gpu: 4
   n_cpu: 16
   memory_limit: 32Gi
 
 deployment:
-  replicas: 1
+  replicas: 2
 
 persistentVolume:
   name: efs-claim
diff --git a/kubernetes/README.md b/kubernetes/README.md
@@ -53,6 +53,7 @@ torchserve:
   management_port: 8081
   inference_port: 8080
   metrics_port: 8082
+  grpc_inference_port: 7070
   pvd_mount: /home/model-server/shared/
   n_gpu: 1
   n_cpu: 1
@@ -290,6 +291,51 @@ Follow the link for log aggregation with EFK Stack.\
 ## Autoscaling
   [Autoscaling with torchserve metrics](autoscale.md)
 
+## Session Affinity with Multiple Torchserve pods
+
+### Pre-requisites
+
+ - Follow the instructions above and deploy Torchserve with more than 1 replica to the kubernetes cluster
+ - Download Istio and add to path as shown [here](https://istio.io/latest/docs/setup/getting-started/#download)
+ - Install Istio with below command
+   - `istioctl install --set meshConfig.accessLogFile=/dev/stdout`
+
+### Steps
+
+Now we have multiple replicas of Torchserve running and istio installed. We can apply gateway, virtual service and destination rule to enable session affinity to the user requests.
+
+ - Apply the istio gateway via `kubectl apply -f gateway.yaml`
+   - This gateway exposes all the host behind it via port 80 as defined in the yaml file.
+ - Apply the virtual service with command `kubectl apply -f virtual_service.yaml`
+   - This with look for header named `protocol` in the incoming request and forward the request to Torchserve service. If the `protocol` header has a value `rest` then the request is forwarded to port `8080` of Torchserve service and if the `protocol` header has a value `grpc` then the request is forwarded to port `7070` for Torchserve service.
+ - Apply the destination Rule using the command `kubectl apply -f destination_rule.yaml`.
+   - The destination rule look for a http cookie with a key `session_id`. The request with `session_id` is served by the same pod that served the previous request with the same `session_id`
+
+### HTTP Inference
+
+- Fetch the external IP from istio-ingress gateway using the below command
+
+```bash
+ubuntu@ubuntu$ kubectl get svc -n istio-system
+NAME                   TYPE           CLUSTER-IP      EXTERNAL-IP                                                               PORT(S)                                                   AGE
+istio-ingressgateway   LoadBalancer   10.100.84.243   a918b2zzzzzzzzzzzzzzzzzzzzzz-1466623565.us-west-2.elb.amazonaws.com   15021:32270/TCP,80:31978/TCP,443:31775/TCP,70:31778/TCP   2d6h
+```
+
+- Make Request as shown below
+
+```bash
+curl -v -H "protocol: REST" --cookie "session_id="12345" http://a918b2d70dbddzzzzzzzzzzz49ec8cf03b-1466623565.us-west-2.elb.amazonaws.com:80/predictions/<model_name> -d "data=<input-string>"
+```
+
+### gRPC Inference
+
+- Refer [grpc_api](../docs/grpc_api.md) to generate python files and run
+
+```bash
+python ts_scripts/torchserve_grpc_client.py infer <model_name> <input-string>
+```
+
+
 ## Roadmap
 
 * [] Log / Metrics Aggregation using [AWS Container Insights](https://docs.aws.amazon.com/AmazonCloudWatch/latest/monitoring/ContainerInsights.html)

diff --git a/kubernetes/destination_rule.yaml b/kubernetes/destination_rule.yaml
@@ -0,0 +1,13 @@
+apiVersion: networking.istio.io/v1alpha3
+kind: DestinationRule
+metadata:
+  name: torchserve-dr
+spec:
+  host: torchserve.default.svc.cluster.local # <ts-service-name>.<namespace>.svc.cluster.local
+  trafficPolicy:
+    loadBalancer:
+      consistentHash:
+        # httpHeaderName: x-user
+        httpCookie:
+          name: session_id
+          ttl: 60s
diff --git a/kubernetes/gateway.yaml b/kubernetes/gateway.yaml
@@ -0,0 +1,14 @@
+apiVersion: networking.istio.io/v1beta1
+kind: Gateway
+metadata:
+  name: torchserve-gw
+spec:
+  selector:
+    istio: ingressgateway
+  servers:
+  - hosts:
+    - '*'
+    port:
+      name: http
+      number: 80
+      protocol: HTTP
diff --git a/kubernetes/virtual_service.yaml b/kubernetes/virtual_service.yaml
@@ -0,0 +1,36 @@
+apiVersion: networking.istio.io/v1alpha3
+kind: VirtualService
+metadata:
+  name: torchserve-vs
+spec:
+  hosts:
+    - "*"
+  gateways:
+    - torchserve-gw
+  http:
+    - match:
+        - uri:
+            prefix: /metrics
+      route:
+        - destination:
+            host: torchserve.default.svc.cluster.local
+            port:
+              number: 8082
+    - match:
+        - headers:
+            protocol:
+              exact: REST
+      route:
+        - destination:
+            host: torchserve.default.svc.cluster.local # <ts-service-name>.<namespace>.svc.cluster.local
+            port:
+              number: 8080
+    - match:
+        - headers:
+            protocol:
+              exact: gRPC
+      route:
+        - destination:
+            host: torchserve.default.svc.cluster.local # <ts-service-name>.<namespace>.svc.cluster.local
+            port:
+              number: 7070
diff --git a/ts_scripts/spellcheck_conf/wordlist.txt b/ts_scripts/spellcheck_conf/wordlist.txt
@@ -1065,6 +1065,9 @@ ActionSLAM
 statins
 ci
 chatGPT
+accessLogFile
+istioctl
+meshConfig
 baseimage
 cuDNN
 Xformer

diff --git a/ts_scripts/torchserve_grpc_client.py b/ts_scripts/torchserve_grpc_client.py
@@ -19,13 +19,14 @@ def get_management_stub():
     return stub
 
 
-def infer(stub, model_name, model_input):
+def infer(stub, model_name, model_input, metadata):
     with open(model_input, "rb") as f:
         data = f.read()
 
     input_data = {"data": data}
     response = stub.Predictions(
-        inference_pb2.PredictionsRequest(model_name=model_name, input=input_data)
+        inference_pb2.PredictionsRequest(model_name=model_name, input=input_data),
+        metadata=metadata,
     )
 
     try:
@@ -35,13 +36,14 @@ def infer(stub, model_name, model_input):
         exit(1)
 
 
-def infer_stream(stub, model_name, model_input):
+def infer_stream(stub, model_name, model_input, metadata):
     with open(model_input, "rb") as f:
         data = f.read()
 
     input_data = {"data": data}
     responses = stub.StreamPredictions(
-        inference_pb2.PredictionsRequest(model_name=model_name, input=input_data)
+        inference_pb2.PredictionsRequest(model_name=model_name, input=input_data),
+        metadata=metadata,
     )
 
     try:
@@ -92,7 +94,6 @@ def unregister(stub, model_name):
 
 
 if __name__ == "__main__":
-
     parent_parser = argparse.ArgumentParser(add_help=False)
     parent_parser.add_argument(
         "model_name",
@@ -141,10 +142,12 @@ def unregister(stub, model_name):
 
     args = parser.parse_args()
 
+    metadata = (("protocol", "gRPC"), ("session_id", "12345"))
+
     if args.action == "infer":
-        infer(get_inference_stub(), args.model_name, args.model_input)
+        infer(get_inference_stub(), args.model_name, args.model_input, metadata)
     elif args.action == "infer_stream":
-        infer_stream(get_inference_stub(), args.model_name, args.model_input)
+        infer_stream(get_inference_stub(), args.model_name, args.model_input, metadata)
     elif args.action == "register":
         register(get_management_stub(), args.model_name, args.mar_set)
     elif args.action == "unregister":