github.com/arloliu/jsonpack

Fast and space efficiency JSON serialization golang library. It is a schema oriented design which leverages schema definition to encode JSON document into compact binary encoded format, and decodes back into JSON document.

Introduction

When we want to exchange data between services or over network, the JSON is most popular format to do it. In golang world, the most convenient way is using official encoding/json package to marshal and unmarshal JSON document from/to struct or map. For usual RESTful web service scenario, JSON format is quite convenience and representative, but for real-time message exchanging or other scenarios that has small footprint data and low-latency requirement, JSON is a bit too heavy weight, not only data footprint is not space saving, but also has heavy loading in encode/decode procedure.

So if we want to a compact, small footprint data for exchanging over network, and also leverages the convenience of JSON, we need an encoding format that removes "property name" and other notations likes ':', '[', '{'...etc from original JSON document, and leaves "value" only.

To achieve this goal, we need a schematic to define our JSON document and provide enough information for serialization engine to know sequence and data type of every properties in document. it's the reason why JSONPack is a schema oriented design library.

Key Features

• Similar Marshal / Unmarshal API to standard encoding/json package.
• Space saving encoded format, the size of encoded data is similar to Protocol Buffers, can be 30-80% of original JSON document, depends on data.
• Blazing fast, provides about 3.x decoding speed compared to protobuf and many times than other JSON packages.
• Memory saving design, avoids any un-necessary memory allocations, suitable for embedded environment.
• Has production ready javascript implementation Buffer Plus, can be used in node.js and Web browser environment.
• No need to write schema definition by hand, jsonpack will generate schema definition from golang struct automatically.

How to Get

go get github.com/arloliu/jsonpack

Usage

Example of add schema definition

import "github.com/arloliu/jsonpack"

type Info struct {
	Name string `json:"name"`
	Area uint32 `json:"area"`
	// omit this field
	ExcludeField string `-`
}

jsonPack := jsonpack.NewJSONPack()
sch, err := jsonPack.AddSchema("Info", Info{}, jsonpack.LittleEndian)

Example of encoding data with Info struct

infoStruct := &Info{
    Name: "example name",
    Area: 888,
}
// encodedResult1 contains encoded data,
encodedResult1, err := jsonPack.Marshal("Info", infoStruct)

Example of encoding data with golang map

infoMap := map[string]interface{} {
	"name": "example name",
	"area": 888,
}

encodedResult2, err := jsonPack.Marshal("Info", infoMap)

Example of decoding data

decodeInfoStruct = Info{}
err := jsonPack.Decode("Info", encodedResult1, &decodeInfoStruct)

decodeInfoMap = make(map[string]interface{})
err := jsonPack.Decode("Info", encodedResult2, &decodeInfoMap)

Benchmark

The benchmark result is a important reference but not always suitable for every scenarios.

Test environment: Intel i7-9700K CPU@3.60GHz.

Benchmark code is here

Sorts from fastest to slowest in the following.

Encode from golang map

	ns/op	allocation bytes	allocation times
jsonpack	1933 ns/op	752 B/op	2 allocs/op
jsoniter	10134 ns/op	3320 B/op	46 allocs/op
std. json	23560 ns/op	8610 B/op	171 allocs/op
goccy	75298 ns/op	82639 B/op	651 allocs/op

Decode into golang map

	ns/op	allocation bytes	allocation times
jsonpack	6461 ns/op	6512 B/op	96 allocs/op
jsoniter	17436 ns/op	9666 B/op	290 allocs/op
std. json	18949 ns/op	8864 B/op	228 allocs/op
goccy	19985 ns/op	15900 B/op	316 allocs/op

Encode from golang struct

	ns/op	allocation bytes	allocation times
jsonpack	1834 ns/op	800 B/op	3 allocs/op
protobuf	1972 ns/op	896 B/op	1 allocs/op
goccy	2166 ns/op	1280 B/op	1 allocs/op
jsoniter	3372 ns/op	1296 B/op	3 allocs/op
std. json	3578 ns/op	1280 B/op	1 allocs/op

Decode into golang struct

	ns/op	allocation bytes	allocation times
jsonpack	1475 ns/op	96 B/op	2 allocs/op
goccy	3284 ns/op	2215 B/op	5 allocs/op
jsoniter	4680 ns/op	1072 B/op	79 allocs/op
protobuf	5075 ns/op	3152 B/op	84 allocs/op
std. json	18378 ns/op	1232 B/op	69 allocs/op

The benchmark result indicates jsonpack keeps constant performance on both encoding and encoding side, and keeps very low memory allocation size and times.

Explanation of Benchmark Result

The benchmark result also delivers some important messages.

1. The performance of operating with golang map sucks.
2. The pointer of type likes *string, *int32 cost twice of time.
3. Golang reflection is expensive, reflect.ValueOf(...) is very expensive.

Performance Tips

According to the explanation of benchmarking and some tests during the implementation stage of this library, there are some suggestions in the following.

1. It's better to use struct instead of map for encode/decode if possible.
2. When declares struct fields, declares it as value instead of pointer to value gives performance gain.
3. Designs a small and compacted structure, uses number type to represent enum. fields instead of declares it as string type.