Performant Entity Serialization: BSON vs MessagePack (vs JSON)

Question

// Please note that I'm author of MessagePack. This answer may be biased.

Format design

1. Compatibility with JSON

   In spite of its name, BSON's compatibility with JSON is not so good compared with MessagePack.

   BSON has special types like "ObjectId", "Min key", "UUID" or "MD5" (I think these types are required by MongoDB). These types are not compatible with JSON. That means some type information can be lost when you convert objects from BSON to JSON, but of course only when these special types are in the BSON source. It can be a disadvantage to use both JSON and BSON in single service.

   MessagePack is designed to be transparently converted from/to JSON.

2. MessagePack is smaller than BSON

   MessagePack's format is less verbose than BSON. As the result, MessagePack can serialize objects smaller than BSON.

   For example, a simple map {"a":1, "b":2} is serialized in 7 bytes with MessagePack, while BSON uses 19 bytes.

3. BSON supports in-place updating

   With BSON, you can modify part of stored object without re-serializing the whole of the object. Let's suppose a map {"a":1, "b":2} is stored in a file and you want to update the value of "a" from 1 to 2000.

   With MessagePack, 1 uses only 1 byte but 2000 uses 3 bytes. So "b" must be moved backward by 2 bytes, while "b" is not modified.

   With BSON, both 1 and 2000 use 5 bytes. Because of this verbosity, you don't have to move "b".

4. MessagePack has RPC

   MessagePack, Protocol Buffers, Thrift and Avro support RPC. But BSON doesn't.

These differences imply that MessagePack is originally designed for network communication while BSON is designed for storages.

Implementation and API design

1. MessagePack has type-checking APIs (Java, C++ and D)

   MessagePack supports static-typing.

   Dynamic-typing used with JSON or BSON are useful for dynamic languages like Ruby, Python or JavaScript. But troublesome for static languages. You must write boring type-checking codes.

   MessagePack provides type-checking API. It converts dynamically-typed objects into statically-typed objects. Here is a simple example (C++):

    #include <msgpack.hpp>

    class myclass {
    private:
        std::string str;
        std::vector<int> vec;
    public:
        // This macro enables this class to be serialized/deserialized
        MSGPACK_DEFINE(str, vec);
    };

    int main(void) {
        // serialize
        myclass m1 = ...;

        msgpack::sbuffer buffer;
        msgpack::pack(&buffer, m1);

        // deserialize
        msgpack::unpacked result;
        msgpack::unpack(&result, buffer.data(), buffer.size());

        // you get dynamically-typed object
        msgpack::object obj = result.get();

        // convert it to statically-typed object
        myclass m2 = obj.as<myclass>();
    }

2. MessagePack has IDL

   It's related to the type-checking API, MessagePack supports IDL. (specification is available from: http://wiki.msgpack.org/display/MSGPACK/Design+of+IDL)

   Protocol Buffers and Thrift require IDL (don't support dynamic-typing) and provide more mature IDL implementation.

3. MessagePack has streaming API (Ruby, Python, Java, C++, ...)

   MessagePack supports streaming deserializers. This feature is useful for network communication. Here is an example (Ruby):

    require 'msgpack'

    # write objects to stdout
    $stdout.write [1,2,3].to_msgpack
    $stdout.write [1,2,3].to_msgpack

    # read objects from stdin using streaming deserializer
    unpacker = MessagePack::Unpacker.new($stdin)
    # use iterator
    unpacker.each {|obj|
      p obj
    }

I think it's very important to mention that it depends on what your client/server environment look like.

If you are passing bytes multiple times without inspection, such as with a message queue system or streaming log entries to disk, then you may well prefer a binary encoding to emphasize the compact size. Otherwise it's a case by case issue with different environments.

Some environments can have very fast serialization and deserialization to/from msgpack/protobuf's, others not so much. In general, the more low-level the language/environment the better binary serialization will work. In higher level languages (node.js, .Net, JVM) you will often see that JSON serialization is actually faster. The question then becomes is your network overhead more or less constrained than your memory/cpu?

With regards to msgpack vs bson vs protocol buffers... msgpack is the least bytes of the group, protocol buffers being about the same. BSON defines more broad native types than the other two, and may be a better match to your object model, but this makes it more verbose. Protocol buffers have the advantage of being designed to stream... which makes it a more natural format for a binary transfer/storage format.

Personally, I would lean towards the transparency that JSON offers directly, unless there is a clear need for lighter traffic. Over HTTP with gzipped data, the difference in network overhead are even less of an issue between the formats.

Quick test shows minified JSON is deserialized faster than binary MessagePack. In the tests Article.json is 550kb minified JSON, Article.mpack is 420kb MP-version of it. May be an implementation issue of course.

MessagePack:

//test_mp.js
var msg = require('msgpack');
var fs = require('fs');

var article = fs.readFileSync('Article.mpack');

for (var i = 0; i < 10000; i++) {
    msg.unpack(article);    
}

JSON:

// test_json.js
var msg = require('msgpack');
var fs = require('fs');

var article = fs.readFileSync('Article.json', 'utf-8');

for (var i = 0; i < 10000; i++) {
    JSON.parse(article);
}

So times are:

Anarki:Downloads oleksii$ time node test_mp.js 

real    2m45.042s
user    2m44.662s
sys     0m2.034s

Anarki:Downloads oleksii$ time node test_json.js 

real    2m15.497s
user    2m15.458s
sys     0m0.824s

So space is saved, but faster? No.

Tested versions:

Anarki:Downloads oleksii$ node --version
v0.8.12
Anarki:Downloads oleksii$ npm list msgpack
/Users/oleksii
└── [email protected]  

A key difference not yet mentioned is that BSON contains size information in bytes for the entire document and further nested sub-documents.

document    ::=     int32 e_list

This has two major benefits for restricted environments (e.g. embedded) where size and performance is important.

1. You can immediately check if the data you're going to parse represents a complete document or if you're going to need to request more at some point (be it from some connection or storage). Since this is most likely an asynchronous operation you might already send a new request before parsing.
2. Your data might contain entire sub-documents with irrelevant information for you. BSON allows you to easily traverse to the next object past the sub-document by using the size information of the sub-document to skip it. msgpack on the other hands contains the number of elements inside whats called a map (similar to BSON's sub-documents). While this is undoubtedly useful information it doesn't help the parser. You'd still have to parse every single object inside the map and can't just skip it. Depending on the structure of your data this might have a huge impact on performance.

I made quick benchmark to compare encoding and decoding speed of MessagePack vs BSON. BSON is faster at least if you have large binary arrays:

BSON writer: 2296 ms (243487 bytes)
BSON reader: 435 ms
MESSAGEPACK writer: 5472 ms (243510 bytes)
MESSAGEPACK reader: 1364 ms

Using C# Newtonsoft.Json and MessagePack by neuecc:

    public class TestData
    {
        public byte[] buffer;
        public bool foobar;
        public int x, y, w, h;
    }

    static void Main(string[] args)
    {
        try
        {
            int loop = 10000;

            var buffer = new TestData();
            TestData data2;
            byte[] data = null;
            int val = 0, val2 = 0, val3 = 0;

            buffer.buffer = new byte[243432];

            var sw = new Stopwatch();

            sw.Start();
            for (int i = 0; i < loop; i++)
            {
                data = SerializeBson(buffer);
                val2 = data.Length;
            }

            var rc1 = sw.ElapsedMilliseconds;

            sw.Restart();
            for (int i = 0; i < loop; i++)
            {
                data2 = DeserializeBson(data);
                val += data2.buffer[0];
            }
            var rc2 = sw.ElapsedMilliseconds;

            sw.Restart();
            for (int i = 0; i < loop; i++)
            {
                data = SerializeMP(buffer);
                val3 = data.Length;
                val += data[0];
            }

            var rc3 = sw.ElapsedMilliseconds;

            sw.Restart();
            for (int i = 0; i < loop; i++)
            {
                data2 = DeserializeMP(data);
                val += data2.buffer[0];
            }
            var rc4 = sw.ElapsedMilliseconds;

            Console.WriteLine("Results:", val);
            Console.WriteLine("BSON writer: {0} ms ({1} bytes)", rc1, val2);
            Console.WriteLine("BSON reader: {0} ms", rc2);
            Console.WriteLine("MESSAGEPACK writer: {0} ms ({1} bytes)", rc3, val3);
            Console.WriteLine("MESSAGEPACK reader: {0} ms", rc4);
        }
        catch (Exception e)
        {
            Console.WriteLine(e);
        }

        Console.ReadLine();
    }

    static private byte[] SerializeBson(TestData data)
    {
        var ms = new MemoryStream();

        using (var writer = new Newtonsoft.Json.Bson.BsonWriter(ms))
        {
            var s = new Newtonsoft.Json.JsonSerializer();
            s.Serialize(writer, data);
            return ms.ToArray();
        }
    }

    static private TestData DeserializeBson(byte[] data)
    {
        var ms = new MemoryStream(data);

        using (var reader = new Newtonsoft.Json.Bson.BsonReader(ms))
        {
            var s = new Newtonsoft.Json.JsonSerializer();
            return s.Deserialize<TestData>(reader);
        }
    }

    static private byte[] SerializeMP(TestData data)
    {
        return MessagePackSerializer.Typeless.Serialize(data);
    }

    static private TestData DeserializeMP(byte[] data)
    {
        return (TestData)MessagePackSerializer.Typeless.Deserialize(data);
    }