Version 1
VelocyPack (VPack) is a fast and compact serialization format
VPack is (unsigned) byte oriented, so VPack values are simply sequences of bytes and are platform independent. Values are not necessarily aligned, so all access to larger subvalues must be properly organised to avoid alignment assumptions of the CPU.
We describe a single VPack value, which is recursive in nature, but resides (with two exceptions, see below) in a single contiguous block of memory. Assume that the value starts at address A, the first byte V indicates the type (and often the length) of the VPack value at hand:
We first give an overview with a brief but accurate description for reference, for arrays and objects see below for details:
0x00: none - this indicates absence of any type and value, this is not allowed in VPack values0x01: empty array0x02: array without index table (all subitems have the same byte length), 1-byte byte length0x03: array without index table (all subitems have the same byte length), 2-byte byte length0x04: array without index table (all subitems have the same byte length), 4-byte byte length0x05: array without index table (all subitems have the same byte length), 8-byte byte length0x06: array with 1-byte index table offsets, bytelen and # subvals0x07: array with 2-byte index table offsets, bytelen and # subvals0x08: array with 4-byte index table offsets, bytelen and # subvals0x09: array with 8-byte index table offsets, bytelen and # subvals0x0a: empty object0x0b: object with 1-byte index table offsets, sorted by attribute name, 1-byte bytelen and # subvals0x0c: object with 2-byte index table offsets, sorted by attribute name, 2-byte bytelen and # subvals0x0d: object with 4-byte index table offsets, sorted by attribute name, 4-byte bytelen and # subvals0x0e: object with 8-byte index table offsets, sorted by attribute name, 8-byte bytelen and # subvals0x0f: object with 1-byte index table offsets, not sorted by attribute name, 1-byte bytelen and # subvals - OBSOLETE0x10: object with 2-byte index table offsets, not sorted by attribute name, 2-byte bytelen and # subvals - OBSOLETE0x11: object with 4-byte index table offsets, not sorted by attribute name, 4-byte bytelen and # subvals - OBSOLETE0x12: object with 8-byte index table offsets, not sorted by attribute name, 8-byte bytelen and # subvals - OBSOLETE0x13: compact array, no index table0x14: compact object, no index table0x15-0x16: reserved0x17: illegal - this type can be used to indicate a value that is illegal in the embedding application0x18: null0x19: false0x1a: true0x1b: double IEEE-754, 8 bytes follow, stored as little endian uint64 equivalent0x1c: UTC-date in milliseconds since the epoch, stored as 8 byte signed int, little endian, two's complement0x1d: external (only in memory): a char* pointing to the actual place in memory, where another VPack item resides, not allowed in VPack values on disk or on the network0x1e: minKey, nonsensical value that compares < than all other values0x1f: maxKey, nonsensical value that compares > than all other values0x20-0x27: signed int, little endian, 1 to 8 bytes, number is V -0x1f, two's complement0x28-0x2f: uint, little endian, 1 to 8 bytes, number is V -0x270x30-0x39: small integers 0, 1, ... 90x3a-0x3f: small negative integers -6, -5, ..., -10x40-0xbe: UTF-8-string, using V -0x40bytes (not Unicode characters!), length 0 is possible, so0x40is the empty string, maximal length is 126, note that strings here are not zero-terminated and may contain NUL bytes0xbf: long UTF-8-string, next 8 bytes are length of string in bytes (not Unicode characters) as little endian unsigned integer, note that long strings are not zero-terminated and may contain NUL bytes0xc0-0xc7: binary blob, next V -0xbfbytes are the length of blob in bytes, note that binary blobs are not zero-terminated0xc8-0xcf: positive long packed BCD-encoded float, V -0xc7bytes follow that encode in a little endian way the length of the mantissa in bytes. Directly after that follow 4 bytes encoding the (power of 10) exponent, by which the mantissa is to be multiplied, stored as little endian two's complement signed 32-bit integer. After that, as many bytes follow as the length information at the beginning has specified, each byte encodes two digits in big-endian packed BCD. Example: 12345 decimal can be encoded asc8 03 00 00 00 00 01 23 45orc8 03 ff ff ff ff 12 34 500xd0-0xd7: negative long packed BCD-encoded float, V -0xcfbytes follow that encode in a little endian way the length of the mantissa in bytes. After that, same as positive long packed BCD-encoded float above.0xd8-0xed: reserved0xee-0xef: value tagging for logical types0xf0-0xff: custom types
Empty arrays are simply a single byte 0x01.
We next describe the type cases 0x02 to 0x09, see below for the
special compact type 0x13.
Non-empty arrays look like one of the following:
one of 0x02 to 0x05
BYTELENGTH
OPTIONAL UNUSED: padding
sub VPack values
or
0x06
BYTELENGTH in 1 byte
NRITEMS in 1 byte
OPTIONAL UNUSED: 6 bytes of padding
sub VPack values
INDEXTABLE with 1 byte per entry
or
0x07
BYTELENGTH in 2 bytes
NRITEMS in 2 bytes
OPTIONAL UNUSED: 4 bytes of padding
sub VPack values
INDEXTABLE with 4 byte per entry
or
0x08
BYTELENGTH in 4 bytes
NRITEMS in 4 bytes
sub VPack values
INDEXTABLE with 4 byte per entry
or
0x09
BYTELENGTH in 8 bytes
sub VPack values
INDEXTABLE with 8 byte per entry
NRITEMS in 8 bytes
If any optional padding is allowed for a type, the padding must consist of exactly that many bytes that the length of the padding, the length of BYTELENGTH and the length of NRITEMS (if present) sums up to 8. If the length of BYTELENGTH is already 8, there is no padding allowed. The entire padding must consist of zero bytes (ASCII NUL).
Numbers (for byte length, number of subvalues and offsets in the
INDEXTABLE) are little endian unsigned integers, using 1 byte for
types 0x02 and 0x06, 2 bytes for types 0x03 and 0x07, 4 bytes for types
0x04 and 0x08, and 8 bytes for types 0x05 and 0x09.
NRITEMS is a single number as described above.
The INDEXTABLE consists of:
- for types
0x06-0x09an array of offsets (unaligned, in the number format described above) earlier offsets reside at lower addresses. Offsets are measured from the start of the VPack value.
Non-empty arrays of types 0x06 to 0x09 have a small header including
their byte length, the number of subvalues, then all the subvalues and
finally an index table containing offsets to the subvalues. To find the
index table, find the number of subvalues, then the end, and from that
the base of the index table, considering how wide its entries are.
For types 0x02 to 0x05 there is no offset table and no number of items.
The first item begins at address A+2, A+3, A+5 or respectively A+9,
depending on the type and thus the width of the byte length field. Note
the following special rule: The actual position of the first subvalue
is allowed to be further back, after some run of padding zero bytes.
For example, if 2 bytes are used for both the byte length (BYTELENGTH), then an optional padding of 4 zero bytes is then allowed to follow, and the actual VPack subvalues can start at A+9. This is to give a program that builds a VPack value the opportunity to reserve 8 bytes in the beginning and only later find out that fewer bytes suffice to write the byte length. One can determine the number of subvalues by finding the first subvalue, its byte length, and dividing the amount of available space by it.
For types 0x06 to 0x09 the offset table describes where the subvalues
reside. It is not necessary for the subvalues to start immediately after
the number of subvalues field.
As above, it is allowed to include optional padding. Again here, any padding must consist of a run of consecutive zero bytes (ASCII NUL) and must be as long that it fills up the length of BYTELENGTH and the length of NRITEMS to 8.
For example, if both BYTELENGTH and NRITEMS can be expressed using 2 bytes each, the sum of their lengths is 4. It is therefore allowed to add 4 bytes of padding here, so that the first subvalue could be at address A+9.
There is one exception for the 8-byte numbers case (type 0x05):
In this case the number of elements is moved behind the index table.
This is to get away without moving memory when one has reserved 8 bytes
in the beginning and later noticed that all 8 bytes are needed for the
byte length. For this case it is not allowed to include any padding.
All offsets are measured from base A.
Example:
[1,2,3] has the hex dump
02 05 31 32 33
in the most compact representation, but the following are equally possible, though not necessarily advised to use:
Examples:
03 06 00 31 32 33
04 08 00 00 00 31 32 33
05 0c 00 00 00 00 00 00 00 31 32 33
06 09 03 31 32 33 03 04 05
07 0e 00 03 00 31 32 33 05 00 06 00 07 00
08 18 00 00 00 03 00 00 00 31 32 33 09 00 00 00 0a 00 00 00 0b 00 00 00
09
2c 00 00 00 00 00 00 00
31 32 33
09 00 00 00 00 00 00 00
0a 00 00 00 00 00 00 00
0b 00 00 00 00 00 00 00
03 00 00 00 00 00 00 00
Note that it is not recommended to encode short arrays in too long a format.
We now describe the special type 0x13, which is useful for a
particularly compact array representation. Note that to some extent this
goes against the principles of the VelocyPack format, since quick access
to subvalues is no longer possible, all items in the array must be
scanned to find a particular one. However, there are certain use cases
for VelocyPack which only require sequential access (for example JSON
dumping) and have a particular need for compactness.
The overall format of this array type is
0x13 as type byte
BYTELENGTH
sub VPack values
NRITEMS
There is no index table at all, although the sub VelocyPack values can have different byte sizes. The BYTELENGTH and NRITEMS are encoded in a special format, which we describe now.
The BYTELENGTH consists of 1 to 8 bytes, of which all but the last one have their high bit set. Thus, the high bits determine, how many bytes are actually used. The lower 7 bits of all these bits together comprise the actual byte length in a little endian fashion. That is, the byte at address A+1 contains the least significant 7 bits (0 to 6) of the byte length, the following byte at address A+2 contains the bits 7 to 13, and so on. Since the total number of bytes is limited to 8, this encodes unsigned integers of up to 56 bits, which is the overall limit for the size of such a compact array representation.
The NRITEMS entry is encoded essentially the same, except that it is laid out in reverse order in memory. That is, one has to use the BYTELENGTH to find the end of the array value and go back bytes until one finds a byte with high bit reset. The last byte (at the highest memory address) contains the least significant 7 bits of the NRITEMS value, the second one bits 7 to 13 and so on.
Here is an example, the array [1, 16] can be encoded as follows:
13 06
31 28 10
02
Empty objects are simply a single byte 0x0a.
We next describe the type cases 0x0b to 0x12, see below for the
special compact type 0x14.
Non-empty objects look like this:
one of 0x0b - 0x12
BYTELENGTH
optional NRITEMS
sub VPack values as pairs of attribute and value
optional INDEXTABLE
NRITEMS for the 8-byte case
Numbers (for byte length, number of subvalues and offsets in the
INDEXTABLE) are little endian unsigned integers, using 1 byte for
types 0x0b and 0x0f, 2 bytes for types 0x0c and 0x10, 4 bytes for types
0x0d and 0x11, and 8 bytes for types 0x0e and 0x12.
NRITEMS is a single number as described above.
The INDEXTABLE consists of:
- an array of offsets (unaligned, in the number format described above) earlier offsets reside at lower addresses. Offsets are measured from the beginning of the VPack value.
Non-empty objects have a small header including their byte length, the number of subvalues, then all the subvalues and finally an index table containing offsets to the subvalues. To find the index table, find number of subvalues, then the end, and from that the base of the index table, considering how wide its entries are.
For all types the offset table describes where the subvalues reside. It is not necessary for the subvalues to start immediately after the number of subvalues field. For performance reasons when building the value, it could be desirable to reserve 8 bytes for the byte length and the number of subvalues and not fill the gap, even though it turns out later that offsets and thus the byte length only uses 2 bytes, say.
There is one special case: the empty object is simply stored as the
single byte 0x0a.
There is another exception: For 8-byte numbers (0x12) the number of
subvalues is stored behind the INDEXTABLE. This is to get away without
moving memory when one has reserved 8 bytes in the beginning and later
noticed that all 8 bytes are needed for the byte length.
All offsets are measured from base A.
Each entry consists of two parts, the key and the value, they are
encoded as normal VPack values as above, the first is always a short or
long UTF-8 string starting with a byte 0x40-0xbf as described below. The
second is any other VPack value.
There is one extension: For the key it is possible to use the positive
small integer values 0x30-0x39 or an unsigned integer starting with a
type byte of 0x28-0x2f. Any such integer value is an index into an
outside-given table of attribute names. These are convenient when only
very few attribute names occur or some are repeated very often. The
standard way to encode such an attribute name table is as a VPack array
of strings as specified here.
Objects are always stored with sorted key/value pairs, sorted by bytewise comparisons of the keys on each nesting level. Sorting has some overhead but will allow looking up keys in logarithmic time later. Note that only the index table needs to be sorted, it is not required that the offsets in these tables are increasing. Since the index table resides after the actual subvalues, one can build up a complex VPack value by writing linearly.
Example: the object {"a": 12, "b": true, "c": "xyz"} can have the hexdump:
0b
13 03
41 62 1a
41 61 28 0c
41 63 43 78 79 7a
06 03 0a
The same object could have been done with an index table with longer entries, as in this example:
0d
22 00 00 00
03 00 00 00
41 62 1a
41 61 28 0c
41 63 43 78 79 7a
0c 00 00 00 09 00 00 00 10 00 00 00
Similarly with type 0x0c and 2-byte offsets, byte length and number of
subvalues, or with type 0x0e and 8-byte numbers.
Note that it is not recommended to encode short objects with too long index tables.
We now describe the special type 0x14, which is useful for a
particularly compact object representation. Note that to some extent
this goes against the principles of the VelocyPack format, since quick
access to subvalues is no longer possible, all key/value pairs in the
object must be scanned to find a particular one. However, there are
certain use cases for VelocyPack which only require sequential access
(for example JSON dumping) and have a particular need for compactness.
The overall format of this object type is
0x14 as type byte
BYTELENGTH
sub VPack key/value pairs
NRPAIRS
There is no index table at all, although the sub VelocyPack values can
have different byte sizes. The BYTELENGTH and NRPAIRS are encoded in a
special format, which we describe now. It is the same as for the special
compact array type 0x13, which we repeat here for the sake of
completeness.
The BYTELENGTH consists of 1 to 8 bytes, of which all but the last one have their high bit set. Thus, the high bits determine, how many bytes are actually used. The lower 7 bits of all these bits together comprise the actual byte length in a little endian fashion. That is, the byte at address A+1 contains the least significant 7 bits (0 to 6) of the byte length, the following byte at address A+2 contains the bits 7 to 13, and so on. Since the total number of bytes is limited to 8, this encodes unsigned integers of up to 56 bits, which is the overall limit for the size of such a compact array representation.
The NRPAIRS entry is encoded essentially the same, except that it is laid out in reverse order in memory. That is, one has to use the BYTELENGTH to find the end of the array value and go back bytes until one finds a byte with high bit reset. The last byte (at the highest memory address) contains the least significant 7 bits of the NRPAIRS value, the second one bits 7 to 13 and so on.
Here is an example, the object {"a":1, "b":16} can be encoded as follows:
14 0a
41 61 31 42 62 28 10
02
Type 0x1b indicates a double IEEE-754 value using the 8 bytes following
the type byte. To guarantee platform-independentness the details of the
byte order are as follows. Encoding is done by using memcpy to copy the
internal double value to an uint64_t. This 64-bit unsigned integer is
then stored as little endian 8 byte integer in the VPack value. Decoding
works in the opposite direction. This should sort out the undetermined
byte order in IEEE-754 in practice.
Type 0x1c indicates a signed 64-int integer stored in 8 bytes little
endian two's complement notation directly after the type. The value means
a universal UTC-time measured in milliseconds since the epoch, which is
00:00 on 1 January 1970 UTC.
This type is only for use within memory, not for data exchange over disk or network. Therefore, we only need to specify that the following k bytes are the memcpy of a char* on the current architecture. That char* points to the actual VPack value elsewhere in memory.
These values of types 0x1e and 0x1f have no meaning other than comparing
smaller or greater respectively than any other VPack value. The idea is
that these can be used in systems that define a total order on all VPack
values to specify left or right ends of infinite intervals.
There are different ways to specify integers. For small values -6 to 9
inclusively there are specific type bytes in the range 0x30 to 0x3f to
allow for storage in a single byte. After that there are signed and
unsigned integer types that can code in the type byte the number of
bytes used (ranges 0x20-0x27 for signed and 0x28-0x2f for unsigned).
These three values use a single byte to store the corresponding JSON values.
Strings are stored as UTF-8 encoded byte sequences. There are two
variants, a short one and a long one. In the short one, the byte length
(not the number of UTF-8 characters) is directly encoded in the type,
and this works up to and including byte length 126. Types 0x40 to 0xbe
are used for this and the byte length is V - 0x3f, if V is the type
byte. For strings longer than 126 bytes, the type byte is 0xbf and the
byte length of the string is stored in the first 8 bytes after the type
byte, using a little endian unsigned integer representation. The actual
string follows after these 8 bytes. There is no terminating zero byte in
either case and the string may contain zero bytes.
The type bytes 0xc0 to 0xc7 allow to store arbitrary binary byte
sequences as a VPack value. The format is as follows: If V is the type
byte, then V - 0xbf bytes follow it to make a little endian unsigned
integer representing the length of the binary data, which directly
follows these length bytes. No alignment is guaranteed. The content is
entirely up to the user.
These types are used to represent arbitrary precision decimal numbers. There are different types for positive and negative numbers. The overall format of these values is:
one of 0xc8 - 0xcf (positive) or of 0xd0 - 0xd7 (negative)
LENGTH OF MANTISSA in bytes
EXPONENT (as 4-byte little endian signed two's complement integer)
MANTISSA (as packed BCD-encoded integer, big-endian)
The type byte describes the sign of the number as well as the number of
bytes used to specify the byte length of the mantissa. As usual, if V is
the type byte, then V - 0xc7 (in the positive case) or V - 0xcf (in the
negative case) bytes are used for the length of the mantissa, stored as
little endian unsigned integer directly after the byte length. After
this follow exactly 4 bytes (little endian signed two's complement
integer) to specify the exponent. After the exponent, the actual
mantissa bytes follow.
Packed BCD is used, so that each byte stores exactly 2 decimal digits as
in 0x34 for the decimal digits 34. Therefore, the mantissa always has an
even number of decimal digits. Note that the mantissa is stored in big
endian form, to make parsing and dumping efficient. This leads to the
"unholy nibble problem": When a JSON parser sees the beginning of a
longish number, it does not know whether an even or odd number of digits
follow. However, for efficiency reasons it wants to start writing bytes
to the output as it reads the input. This is, where the exponent comes
to the rescue, which is illustrated by the following example.
12345 decimal can be encoded as:
c8 03 00 00 00 00 01 23 45
c8 03 ff ff ff ff 12 34 50
The former encoding puts a leading 0 in the first byte and uses exponent
0, the latter encoding directly starts putting two decimal digits in one
byte and then in the end has to "erase" the trailing 0 by using exponent
-1, encoded by the 4 byte sequence ff ff ff ff.
Therefore, the unholy nibble problem is solved and parsing (and indeed dumping) can be efficient.
Types 0xee-0xef are used for tagging of values to implement logical
types.
For example, if type 0x1c did not exist, the database driver could
serialize a timestamp object (Date in JavaScript, Instant in Java, etc)
into a Unix timestamp, a 64-bit integer. Assuming the lack of schema,
upon deserialization it would not be possible to tell an integer from
a timestamp and deserialize the value accordingly.
Type tagging resolves this by attaching an integer tag to values that can then be read when deserializing the value, e.g. that tag=1 is a timestamp and the relevant timestamp class should be used.
The tag values are specified separately and applications can also specify their own to have the database driver deserialize their specific data types into the appropriate classes (including models).
Essentially this is object-relational mapping for parts of documents.
The format of the type is:
0xee
TAG number in 1 byte
sub VPack value
or
0xef
TAG number in 8 bytes, little-endian encoding
sub VPack value
Note that custom types should usually not be used for data exchange but only internally in systems. Nevertheless, the design of this part of the specification is made such that it is possible by generic methods to derive the byte length of each custom data type.
The following user-defined types exist:
0xf0: 1 byte payload, directly following the type byte0xf1: 2 bytes payload, directly following the type byte0xf2: 4 bytes payload, directly following the type byte0xf3: 8 bytes payload, directly following the type byte0xf4-0xf6: length of the payload is described by a single further unsigned byte directly following the type byte, the payload of that many bytes follows0xf7-0xf9: length of the payload is described by two bytes (little endian unsigned integer) directly following the type byte, the payload of that many bytes follows0xfa-0xfc: length of the payload is described by four bytes (little endian unsigned integer) directly following the type byte, the payload of that many bytes follows0xfd-0xff: length of the payload is described by eight bytes (little endian unsigned integer) directly following the type byte, the payload of that many bytes follows
Note: In types 0xf4 to 0xff the "payload" refers to the actual data not
including the length specification.
Serialized booleans, integers, strings, arrays, objects etc. all have a defined endianess and length, which is platform-independent. These types are fully portable in serialized VelocyPack.
There are still a few caveats when it comes to portability:
It is possible to build up very large values on a 64 bit system, but it may not be possible to read them back on a 32 bit system. This is because the maximum memory allocation size on a 32 bit system may be severely limited compared to a 64 bit system, i.e. a 32 bit OS may simply not allow to allocate buffers larger than 4 GB. This is not a limitation of VelocyPack, but a limitation of 32 bit architectures. If all VelocyPack values are kept small enough so that they are well below the 32 bit length boundaries, this should not matter though.
The VelocyPack type External contains just a raw pointer to memory, which should only be used during the buildup of VelocyPack values in memory. The External type is not supposed to be used in VelocyPack values that are serialized and stored persistently, and then later read back from persistence. Doing it anyway is not portable and will also pose a security risk. Not using the External type for any data that is serialized will avoid this problem entirely.
The VelocyPack type Custom is completely user-defined, and there is no default implementation for them. So it is up to the embedder to make these custom type bindings portable if portability of them is a concern.
VelocyPack Double values are serialized as integer equivalents in a specific way, and unserialized back into integers that overlay a IEEE-754 double-precision floating point value in memory. We found this to be sufficiently portable for our needs, although at least in theory there may be portability issues with some systems.
The following was used as a backing for our "reasonably portable in the real world" assumptions:
It may therefore appear strange that the widespread IEEE 754 floating-point standard does not specify endianness.[17] Theoretically, this means that even standard IEEE floating-point data written by one machine might not be readable by another. However, on modern standard computers (i.e., implementing IEEE 754), one may in practice safely assume that the endianness is the same for floating-point numbers as for integers, making the conversion straightforward regardless of data type.