msgpackr

The msgpackr package is an extremely fast MessagePack NodeJS/JavaScript implementation. Currently, it is significantly faster than any other known implementations, faster than Avro (for JS), and generally faster than native V8 JSON.stringify/parse, on NodeJS. It also includes an optional record extension (the r in msgpackr), for defining record structures that makes MessagePack even faster and more compact, often over twice as fast as even native JSON functions, several times faster than other JS implementations, and 15-50% more compact. See the performance section for more details. Structured cloning (with support for cyclical references) is also supported through optional extensions.

Basic Usage

Install with:

npm i msgpackr

And import or require it for basic standard serialization/encoding (pack) and deserialization/decoding (unpack) functions:

import { unpack, pack } from 'msgpackr';
let serializedAsBuffer = pack(value);
let data = unpack(serializedAsBuffer);

This pack function will generate standard MessagePack without any extensions that should be compatible with any standard MessagePack parser/decoder. It will serialize JavaScript objects as MessagePack maps by default. The unpack function will deserialize MessagePack maps as an Object with the properties from the map.

Node Usage

The msgpackr package runs on any modern JS platform, but is optimized for NodeJS usage (and will use a node addon for performance boost as an optional dependency).

Streams

We can use the including streaming functionality (which further improves performance). The PackrStream is a NodeJS transform stream that can be used to serialize objects to a binary stream (writing to network/socket, IPC, etc.), and the UnpackrStream can be used to deserialize objects from a binary sream (reading from network/socket, etc.):

import { PackrStream } from 'msgpackr';
let stream = new PackrStream();
stream.write(myData);

Or for a full example of sending and receiving data on a stream:

import { PackrStream, UnpackrStream } from 'msgpackr';
let sendingStream = new PackrStream();
let receivingStream = new UnpackrStream();
// we are just piping to our own stream, but normally you would send and
// receive over some type of inter-process or network connection.
sendingStream.pipe(receivingStream);
sendingStream.write(myData);
receivingStream.on('data', (data) => {
	// received data
});

The PackrStream and UnpackrStream instances will have also the record structure extension enabled by default (see below).

Deno and Bun Usage

Msgpackr modules are standard ESM modules and can be loaded directly from the deno.land registry for msgpackr for use in Deno or using the NPM module loader with import { unpack } from 'npm:msgpackr'. The standard pack/encode and unpack/decode functionality is available on Deno, like other platforms. msgpackr can be used like any other package on Bun.

Browser Usage

Msgpackr works as standalone JavaScript as well, and runs on modern browsers. It includes a bundled script, at dist/index.js for ease of direct loading:

<script src="node_modules/msgpackr/dist/index.js"></script>

This is UMD based, and will register as a module if possible, or create a msgpackr global with all the exported functions.

For module-based development, it is recommended that you directly import the module of interest, to minimize dependencies that get pulled into your application:

import { unpack } from 'msgpackr/unpack' // if you only need to unpack

The package also includes a minified bundle in index.min.js. Additionally, the package includes a version that excludes dynamic code evaluation called index-no-eval.js, for situations where Content Security Policy (CSP) forbids eval/Function in code. The dynamic evaluation provides important performance optimizations (for records), so is not recommended unless required by CSP policy.

Structured Cloning

You can also use msgpackr for structured cloning. By enabling the structuredClone option, you can include references to other objects or cyclic references, and object identity will be preserved. Structured cloning also enables preserving certain typed objects like Error, Set, RegExp and TypedArray instances. For example:

let obj = {
	set: new Set(['a', 'b']),
	regular: /a\spattern/
};
obj.self = obj;
let packr = new Packr({ structuredClone: true });
let serialized = packr.pack(obj);
let copy = packr.unpack(serialized);
copy.self === copy // true
copy.set.has('a') // true

This option is disabled by default because it uses extensions and reference checking degrades performance (by about 25-30%). (Note this implementation doesn't serialize every class/type specified in the HTML specification since not all of them make sense for storing across platforms.)

Alternate Terminology

If you prefer to use encoder/decode terminology, msgpackr exports aliases, so decode is equivalent to unpack, encode is pack, Encoder is Packr, Decoder is Unpackr, and EncoderStream and DecoderStream can be used as well.

Record / Object Structures

There is a critical difference between maps (or dictionaries) that hold an arbitrary set of keys and values (JavaScript Map is designed for these), and records or object structures that have a well-defined set of fields. Typical JS objects/records may have many instances re(use) the same structure. By using the record extension, this distinction is preserved in MessagePack and the encoding can reuse structures and not only provides better type preservation, but yield much more compact encodings and increase decoding performance by 2-3x. Msgpackr automatically generates record definitions that are reused and referenced by objects with the same structure. There are a number of ways to use this to our advantage. For large object structures with repeating nested objects with similar structures, simply serializing with the record extension can yield significant benefits. To use the record structures extension, we create a new Packr instance. By default a new Packr instance will have the record extension enabled:

import { Packr } from 'msgpackr';
let packr = new Packr();
packr.pack(bigDataWithLotsOfObjects);

Another way to further leverage the benefits of the msgpackr record structures is to use streams that naturally allow for data to reuse based on previous record structures. The stream classes have the record structure extension enabled by default and provide excellent out-of-the-box performance.

When creating a new Packr, Unpackr, PackrStream, or UnpackrStream instance, we can enable or disable the record structure extension with the useRecords property. When this is false, the record structure extension will be disabled (standard/compatibility mode), and all objects will revert to being serialized using MessageMap maps, and all maps will be deserialized to JS Objects as properties (like the standalone pack and unpack functions).

Streaming with record structures works by encoding a structure the first time it is seen in a stream and referencing the structure in later messages that are sent across that stream. When an encoder can expect a decoder to understand previous structure references, this can be configured using the sequential: true flag, which is auto-enabled by streams, but can also be used with Packr instances.

Shared Record Structures

Another useful way of using msgpackr, and the record extension, is for storing data in a databases, files, or other storage systems. If a number of objects with common data structures are being stored, a shared structure can be used to greatly improve data storage and deserialization efficiency. In the simplest form, provide a structures array, which is updated if any new object structure is encountered:

import { Packr } from 'msgpackr';
let packr = new Packr({
	structures: [... structures that were last generated ...]
});

If you are working with persisted data, you will need to persist the structures data when it is updated. Msgpackr provides an API for loading and saving the structures on demand (which is robust and can be used in multiple-process situations where other processes may be updating this same structures array), we just need to provide a way to store the generated shared structure so it is available to deserialize stored data in the future:

import { Packr } from 'msgpackr';
let packr = new Packr({
	getStructures() {
		// storing our data in file (but we could also store in a db or key-value store)
		return unpack(readFileSync('my-shared-structures.mp')) || [];
	},
	saveStructures(structures) {
		writeFileSync('my-shared-structures.mp', pack(structures));
	}
});

Msgpackr will automatically add and saves structures as it encounters any new object structures (up to a limit of 32, by default). It will always add structures in an incremental/compatible way: Any object encoded with an earlier structure can be decoded with a later version (as long as it is persisted).

Shared Structures Options

By default there is a limit of 32 shared structures. This default is designed to record common shared structures, but also be resilient against sharing too many structures if there are many objects with dynamic properties that are likely to be repeated. This also allows for slightly more efficient one byte encoding. However, if your application has more structures that are commonly repeated, you can increase this limit by setting maxSharedStructures to a higher value. The maximum supported shared structures is 8160.

You can also provide a shouldShareStructure function in the options if you want to specifically indicate which structures should be shared. This is called during the encoding process with the array of keys for a structure that is being considered for addition to the shared structure. For example, you might want:

	maxSharedStructures: 100,
	shouldShareStructure(keys) {
		return !(keys[0] > 1) // don't share structures that consist of numbers as keys
	}

Reading Multiple Values

If you have a buffer with multiple values sequentially encoded, you can choose to parse and read multiple values. This can be done using the unpackMultiple function/method, which can return an array of all the values it can sequentially parse within the provided buffer. For example:

let data = new Uint8Array([1, 2, 3]) // encodings of values 1, 2, and 3
let values = unpackMultiple(data) // [1, 2, 3]

Alternately, you can provide a callback function that is called as the parsing occurs with each value, and can optionally terminate the parsing by returning false:

let data = new Uint8Array([1, 2, 3]) // encodings of values 1, 2, and 3
unpackMultiple(data, (value) => {
	// called for each value
	// return false if you wish to end the parsing
})

If you need to know the start and end offsets of the unpacked values, these are provided as optional parameters in the callback:

let data = new Uint8Array([1, 2, 3]) // encodings of values 1, 2, and 3
unpackMultiple(data, (value,start,end) => {
	// called for each value
	// `start` is the data buffer offset where the value was read from
	// `end` is `start` plus the byte length of the encoded value
	// return false if you wish to end the parsing
})

Options

The following options properties can be provided to the Packr or Unpackr constructor:

• useRecords - Setting this to false disables the record extension and stores JavaScript objects as MessagePack maps, and unpacks maps as JavaScript Objects, which ensures compatibilty with other decoders. Setting this to a function will use records for objects where useRecords(object) returns true.
• structures - Provides the array of structures that is to be used for record extension, if you want the structures saved and used again. This array will be modified in place with new record structures that are serialized (if less than 32 structures are in the array).
• moreTypes - Enable serialization of additional built-in types/classes including typed arrays, Sets, Maps, and Errors.
• structuredClone - This enables the structured cloning extensions that will encode object/cyclic references. moreTypes is enabled by default when this is enabled.
• mapsAsObjects - If true, this will decode MessagePack maps and JS Objects with the map entries decoded to object properties. If false, maps are decoded as JavaScript Maps. This is disabled by default if useRecords is enabled (which allows Maps to be preserved), and is enabled by default if useRecords is disabled.
• useFloat32 - This will enable msgpackr to encode non-integer numbers as float32. See next section for possible values.
• variableMapSize - This will use varying map size definition (fixmap, map16, map32) based on the number of keys when encoding objects, which yields slightly more compact encodings (for small objects), but is typically 5-10% slower during encoding. This is necessary if you need to use objects with more than 65535 keys. This is only relevant when record extension is disabled.
• bundleStrings - If true this uses a custom extension that bundles strings together, so that they can be decoded more quickly on browsers and Deno that do not have access to the NodeJS addon. This a custom extension, so both encoder and decoder need to support this. This can yield significant decoding performance increases on browsers (30%-50%).
• copyBuffers - When decoding a MessagePack with binary data (Buffers are encoded as binary data), copy the buffer rather than providing a slice/view of the buffer. If you want your input data to be collected or modified while the decoded embedded buffer continues to live on, you can use this option (there is extra overhead to copying).
• useTimestamp32 - Encode JS Dates in 32-bit format when possible by dropping the milliseconds. This is a more efficient encoding of dates. You can also cause dates to use 32-bit format by manually setting the milliseconds to zero (date.setMilliseconds(0)).
• sequential - Encode structures in serialized data, and reference previously encoded structures with expectation that decoder will read the encoded structures in the same order as encoded, with unpackMultiple.
• largeBigIntToFloat - If a bigint needs to be encoded that is larger than will fit in 64-bit integers, it will be encoded as a float-64 (otherwise will throw a RangeError).
• encodeUndefinedAsNil - Encodes a value of undefined as a MessagePack nil, the same as a null.
• int64AsType - This will decode uint64 and int64 numbers as the specified type. The type can be bigint (default), number, string, or auto (where range [-2^53...2^53] is represented by number and everything else by a bigint).
• onInvalidDate - This can be provided as function that will be called when an invalid date is provided. The function can throw an error, or return a value that will be encoded in place of the invalid date. If not provided, an invalid date will be encoded as an invalid timestamp (which decodes with msgpackr back to an invalid date).
• mapAsEmptyObject - Encodes JS Maps as empty objects (for back-compat with older libraries).
• setAsEmptyObject - Encodes JS Sets as empty objects (for back-compat with older libraries).

32-bit Float Options

By default all non-integer numbers are serialized as 64-bit float (double). This is fast, and ensures maximum precision. However, often real-world data doesn't not need 64-bits of precision, and using 32-bit encoding can be much more space efficient. There are several options that provide more efficient encodings. Using the decimal rounding options for encoding and decoding provides lossless storage of common decimal representations like 7.99, in more efficient 32-bit format (rather than 64-bit). The useFloat32 property has several possible options, available from the module as constants:

import { FLOAT32_OPTIONS } from 'msgpackr';
const { ALWAYS, DECIMAL_ROUND, DECIMAL_FIT } = FLOAT32_OPTIONS;

• ALWAYS (1) - Always will encode non-integers (absolute less than 2147483648) as 32-bit float.
• DECIMAL_ROUND (3) - Always will encode non-integers as 32-bit float, and when decoding 32-bit float, round to the significant decimal digits (usually 7, but 6 or 8 digits for some ranges).
• DECIMAL_FIT (4) - Only encode non-integers as 32-bit float if all significant digits (usually up to 7) can be unambiguously encoded as a 32-bit float, and decode/unpack with decimal rounding (same as above). This will ensure round-trip encoding/decoding without loss in precision and uses 32-bit when possible.

Note, that the performance is decreased with decimal rounding by about 20-25%, although if only 5% of your values are floating point, that will only have about a 1% impact overall.

In addition, msgpackr exports a roundFloat32(number) function that can be used to round floating point numbers to the maximum significant decimal digits that can be stored in 32-bit float, just as DECIMAL_ROUND does when decoding. This can be useful for determining how a number will be decoded prior to encoding it.

Performance

Native Acceleration

Msgpackr employs an optional native node-addon to accelerate the parsing of strings. This should be automatically installed and utilized on NodeJS. However, you can verify this by checking the isNativeAccelerationEnabled property that is exported from msgpackr. If this is false, the msgpackr-extract package may not have been properly installed, and you may want to verify that it is installed correctly:

import { isNativeAccelerationEnabled } from 'msgpackr'
if (!isNativeAccelerationEnabled)
	console.warn('Native acceleration not enabled, verify that install finished properly')

Benchmarks

Msgpackr is fast. Really fast. Here is comparison with the next fastest JS projects using the benchmark tool from msgpack-lite (and the sample data is from some clinical research data we use that has a good mix of different value types and structures). It also includes comparison to V8 native JSON functionality, and JavaScript Avro (avsc, a very optimized Avro implementation):

operation | op | ms | op/s ---------------------------------------------------------- | ------: | ----: | -----: buf = Buffer(JSON.stringify(obj)); | 81600 | 5002 | 16313 obj = JSON.parse(buf); | 90700 | 5004 | 18125 require("msgpackr").pack(obj); | 169700 | 5000 | 33940 require("msgpackr").unpack(buf); | 109700 | 5003 | 21926 msgpackr w/ shared structures: packr.pack(obj); | 190400 | 5001 | 38072 msgpackr w/ shared structures: packr.unpack(buf); | 422900 | 5000 | 84580 buf = require("msgpack-lite").encode(obj); | 31300 | 5005 | 6253 obj = require("msgpack-lite").decode(buf); | 15700 | 5007 | 3135 buf = require("@msgpack/msgpack").encode(obj); | 103100 | 5003 | 20607 obj = require("@msgpack/msgpack").decode(buf); | 59100 | 5004 | 11810 buf = require("notepack").encode(obj); | 65500 | 5007 | 13081 obj = require("notepack").decode(buf); | 33400 | 5009 | 6667 obj = require("msgpack-unpack").decode(buf); | 6900 | 5036 | 1370 require("avsc")...make schema/type...type.toBuffer(obj); | 89300 | 5005 | 17842 require("avsc")...make schema/type...type.fromBuffer(obj); | 108400 | 5001 | 21675

All benchmarks were performed on Node 15 / V8 8.6 (Windows i7-4770 3.4Ghz). (avsc is schema-based and more comparable in style to msgpackr with shared structures).

Here is a benchmark of streaming data (again borrowed from msgpack-lite's benchmarking), where msgpackr is able to take advantage of the structured record extension and really demonstrate its performance capabilities:

operation (1000000 x 2) | op | ms | op/s ------------------------------------------------ | ------: | ----: | -----: new PackrStream().write(obj); | 1000000 | 372 | 2688172 new UnpackrStream().write(buf); | 1000000 | 247 | 4048582 stream.write(msgpack.encode(obj)); | 1000000 | 2898 | 345065 stream.write(msgpack.decode(buf)); | 1000000 | 1969 | 507872 stream.write(notepack.encode(obj)); | 1000000 | 901 | 1109877 stream.write(notepack.decode(buf)); | 1000000 | 1012 | 988142 msgpack.Encoder().on("data",ondata).encode(obj); | 1000000 | 1763 | 567214 msgpack.createDecodeStream().write(buf); | 1000000 | 2222 | 450045 msgpack.createEncodeStream().write(obj); | 1000000 | 1577 | 634115 msgpack.Decoder().on("data",ondata).decode(buf); | 1000000 | 2246 | 445235

See the benchmark.md for more benchmarks and information about benchmarking.

Custom Extensions

You can add your own custom extensions, which can be used to encode specific types/classes in certain ways. This is done by using the addExtension function, and specifying the class, extension type code (should be a number from 1-100, reserving negatives for MessagePack, 101-127 for msgpackr), and your pack and unpack functions (or just the one you need).

import { addExtension, Packr } from 'msgpackr';

class MyCustomClass {...}

let extPackr = new Packr();
addExtension({
	Class: MyCustomClass,
	type: 11, // register your own extension code (a type code from 1-100)
	pack(instance) {
		// define how your custom class should be encoded
		return Buffer.from([instance.myData]); // return a buffer
	}
	unpack(buffer) {
		// define how your custom class should be decoded
		let instance = new MyCustomClass();
		instance.myData = buffer[0];
		return instance; // decoded value from buffer
	}
});

If you want to use msgpackr to encode and decode the data within your extensions, you can use the read and write functions and read and write data/objects that will be encoded and decoded by msgpackr, which can be easier and faster than creating and receiving separate buffers:

import { addExtension, Packr } from 'msgpackr';

class MyCustomClass {...}

let extPackr = new Packr();
addExtension({
	Class: MyCustomClass,
	type: 11, // register your own extension code (a type code from 1-100)
	write(instance) {
		// define how your custom class should be encoded
		return instance.myData; // return some data to be encoded
	}
	read(data) {
		// define how your custom class should be decoded,
		// data will already be unpacked/decoded
		let instance = new MyCustomClass();
		instance.myData = data;
		return instance; // return decoded value
	}
});

Note that you can just return the same object from write, and in this case msgpackr will encode it using the default object/array encoding:

addExtension({
	Class: MyCustomClass,
	type: 12,
	read: function(data) {
		Object.setPrototypeOf(data, MyCustomClass.prototype)
		return data
	},
	write: function(data) {
		return data
	}
})

You can also create an extension with Class and write methods, but no type (or read), if you just want to customize how a class is serialized without using MessagePack extension encoding.

Additional Performance Optimizations

Msgpackr is already fast, but here are some tips for making it faster:

Buffer Reuse

Msgpackr is designed to work well with reusable buffers. Allocating new buffers can be relatively expensive, so if you have Node addons, it can be much faster to reuse buffers and use memcpy to copy data into existing buffers. Then msgpackr unpack can be executed on the same buffer, with new data, and optionally take a second paramter indicating the effective size of the available data in the buffer.

Arena Allocation (useBuffer())

During the serialization process, data is written to buffers. Again, allocating new buffers is a relatively expensive process, and the useBuffer method can help allow reuse of buffers that will further improve performance. With useBuffer method, you can provide a buffer, serialize data into it, and when it is known that you are done using that buffer, you can call useBuffer again to reuse it. The use of useBuffer is never required, buffers will still be handled and cleaned up through GC if not used, it just provides a small performance boost.

Record Structure Extension Definition

The record struction extension uses extension id 0x72 ("r") to declare the use of this functionality. The extension "data" byte (or bytes) identifies the byte or bytes used to identify the start of a record in the subsequent MessagePack block or stream. The identifier byte (or the first byte in a sequence) must be from 0x40 - 0x7f (and therefore replaces one byte representations of positive integers 64 - 127, which can alternately be represented with int or uint types). The extension declaration must be immediately follow by an MessagePack array that defines the field names of the record structure.

Once a record identifier and record field names have been defined, the parser/decoder should proceed to read the next value. Any subsequent use of the record identifier as a value in the block or stream should parsed as a record instance, and the next n values, where is n is the number of fields (as defined in the array of field names), should be read as the values of the fields. For example, here we have defined a structure with fields "foo" and "bar", with the record identifier 0x40, and then read a record instance that defines the field values of 4 and 2, respectively:

+--------+--------+--------+~~~~~~~~~~~~~~~~~~~~~~~~~+--------+--------+
|  0xd4  |  0x72  |  0x40  | array: [ "foo", "bar" ] |  0x04  |  0x02  |
+--------+--------+--------+~~~~~~~~~~~~~~~~~~~~~~~~~+--------+--------+

Which should generate an object that would correspond to JSON:

{ "foo": 4, "bar": 2}

Additional value types

msgpackr supports undefined (using fixext1 + type: 0 + data: 0 to match other JS implementations), NaN, Infinity, and -Infinity (using standard IEEE 754 representations with doubles/floats).

Dates

msgpackr saves all JavaScript Dates using the standard MessagePack date extension (type -1), using the smallest of 32-bit, 64-bit or 96-bit format needed to store the date without data loss (or using 32-bit if useTimestamp32 options is specified).

Structured Cloning

With structured cloning enabled, msgpackr will also use extensions to store Set, Map, Error, RegExp, ArrayBufferView objects and preserve their types.

Alternate Encoding/Package

The high-performance serialization and deserialization algorithms in the msgpackr package are also available in the cbor-x for the CBOR format, with the same API and design. A quick summary of the pros and cons of using MessagePack vs CBOR are:

• MessagePack has wider adoption, and, at least with this implementation is slightly more efficient (by roughly 1%).
• CBOR has an official IETF standardization track, and the record extensions is conceptually/philosophically a better fit for CBOR tags.

License

MIT

Browser Consideration

MessagePack can be a great choice for high-performance data delivery to browsers, as reasonable data size is possible without compression. And msgpackr works very well in modern browsers. However, it is worth noting that if you want highly compact data, brotli or gzip are most effective in compressing, and MessagePack's character frequency tends to defeat Huffman encoding used by these standard compression algorithms, resulting in less compact data than compressed JSON.

Credits

Various projects have been inspirations for this, and code has been borrowed from https://github.com/msgpack/msgpack-javascript and https://github.com/mtth/avsc.