Bytedance Sonic Save

A blazingly fast JSON serializing & deserializing library

Project README


A blazingly fast JSON serializing & deserializing library, accelerated by JIT (just-in-time compiling) and SIMD (single-instruction-multiple-data).


  • Go 1.15/1.16/1.17/1.18/1.19
  • Linux/MacOS/Windows
  • Amd64 ARCH


  • Runtime object binding without code generation
  • Complete APIs for JSON value manipulation
  • Fast, fast, fast!


For all sizes of json and all scenarios of usage, Sonic performs best.

  • Medium (13KB, 300+ key, 6 layers)
goversion: 1.17.1
goos: darwin
goarch: amd64
cpu: Intel(R) Core(TM) i9-9880H CPU @ 2.30GHz
BenchmarkEncoder_Generic_Sonic-16                      32393 ns/op         402.40 MB/s       11965 B/op          4 allocs/op
BenchmarkEncoder_Generic_Sonic_Fast-16                 21668 ns/op         601.57 MB/s       10940 B/op          4 allocs/op
BenchmarkEncoder_Generic_JsonIter-16                   42168 ns/op         309.12 MB/s       14345 B/op        115 allocs/op
BenchmarkEncoder_Generic_GoJson-16                     65189 ns/op         199.96 MB/s       23261 B/op         16 allocs/op
BenchmarkEncoder_Generic_StdLib-16                    106322 ns/op         122.60 MB/s       49136 B/op        789 allocs/op
BenchmarkEncoder_Binding_Sonic-16                       6269 ns/op        2079.26 MB/s       14173 B/op          4 allocs/op
BenchmarkEncoder_Binding_Sonic_Fast-16                  5281 ns/op        2468.16 MB/s       12322 B/op          4 allocs/op
BenchmarkEncoder_Binding_JsonIter-16                   20056 ns/op         649.93 MB/s        9488 B/op          2 allocs/op
BenchmarkEncoder_Binding_GoJson-16                      8311 ns/op        1568.32 MB/s        9481 B/op          1 allocs/op
BenchmarkEncoder_Binding_StdLib-16                     16448 ns/op         792.52 MB/s        9479 B/op          1 allocs/op
BenchmarkEncoder_Parallel_Generic_Sonic-16              6681 ns/op        1950.93 MB/s       12738 B/op          4 allocs/op
BenchmarkEncoder_Parallel_Generic_Sonic_Fast-16         4179 ns/op        3118.99 MB/s       10757 B/op          4 allocs/op
BenchmarkEncoder_Parallel_Generic_JsonIter-16           9861 ns/op        1321.84 MB/s       14362 B/op        115 allocs/op
BenchmarkEncoder_Parallel_Generic_GoJson-16            18850 ns/op         691.52 MB/s       23278 B/op         16 allocs/op
BenchmarkEncoder_Parallel_Generic_StdLib-16            45902 ns/op         283.97 MB/s       49174 B/op        789 allocs/op
BenchmarkEncoder_Parallel_Binding_Sonic-16              1480 ns/op        8810.09 MB/s       13049 B/op          4 allocs/op
BenchmarkEncoder_Parallel_Binding_Sonic_Fast-16         1209 ns/op        10785.23 MB/s      11546 B/op          4 allocs/op
BenchmarkEncoder_Parallel_Binding_JsonIter-16           6170 ns/op        2112.58 MB/s        9504 B/op          2 allocs/op
BenchmarkEncoder_Parallel_Binding_GoJson-16             3321 ns/op        3925.52 MB/s        9496 B/op          1 allocs/op
BenchmarkEncoder_Parallel_Binding_StdLib-16             3739 ns/op        3486.49 MB/s        9480 B/op          1 allocs/op

BenchmarkDecoder_Generic_Sonic-16                      66812 ns/op         195.10 MB/s       57602 B/op        723 allocs/op
BenchmarkDecoder_Generic_Sonic_Fast-16                 54523 ns/op         239.07 MB/s       49786 B/op        313 allocs/op
BenchmarkDecoder_Generic_StdLib-16                    124260 ns/op         104.90 MB/s       50869 B/op        772 allocs/op
BenchmarkDecoder_Generic_JsonIter-16                   91274 ns/op         142.81 MB/s       55782 B/op       1068 allocs/op
BenchmarkDecoder_Generic_GoJson-16                     88569 ns/op         147.17 MB/s       66367 B/op        973 allocs/op
BenchmarkDecoder_Binding_Sonic-16                      32557 ns/op         400.38 MB/s       28302 B/op        137 allocs/op
BenchmarkDecoder_Binding_Sonic_Fast-16                 28649 ns/op         455.00 MB/s       24999 B/op         34 allocs/op
BenchmarkDecoder_Binding_StdLib-16                    111437 ns/op         116.97 MB/s       10576 B/op        208 allocs/op
BenchmarkDecoder_Binding_JsonIter-16                   35090 ns/op         371.48 MB/s       14673 B/op        385 allocs/op
BenchmarkDecoder_Binding_GoJson-16                     28738 ns/op         453.59 MB/s       22039 B/op         49 allocs/op
BenchmarkDecoder_Parallel_Generic_Sonic-16             12321 ns/op        1057.91 MB/s       57233 B/op        723 allocs/op
BenchmarkDecoder_Parallel_Generic_Sonic_Fast-16        10644 ns/op        1224.64 MB/s       49362 B/op        313 allocs/op
BenchmarkDecoder_Parallel_Generic_StdLib-16            57587 ns/op         226.35 MB/s       50874 B/op        772 allocs/op
BenchmarkDecoder_Parallel_Generic_JsonIter-16          38666 ns/op         337.12 MB/s       55789 B/op       1068 allocs/op
BenchmarkDecoder_Parallel_Generic_GoJson-16            30259 ns/op         430.79 MB/s       66370 B/op        974 allocs/op
BenchmarkDecoder_Parallel_Binding_Sonic-16              5965 ns/op        2185.28 MB/s       27747 B/op        137 allocs/op
BenchmarkDecoder_Parallel_Binding_Sonic_Fast-16         5170 ns/op        2521.31 MB/s       24715 B/op         34 allocs/op
BenchmarkDecoder_Parallel_Binding_StdLib-16            27582 ns/op         472.58 MB/s       10576 B/op        208 allocs/op
BenchmarkDecoder_Parallel_Binding_JsonIter-16          13571 ns/op         960.51 MB/s       14685 B/op        385 allocs/op
BenchmarkDecoder_Parallel_Binding_GoJson-16            10031 ns/op        1299.51 MB/s       22111 B/op         49 allocs/op

BenchmarkGetOne_Sonic-16                               11650 ns/op        1117.81 MB/s          29 B/op          1 allocs/op
BenchmarkGetOne_Gjson-16                                9431 ns/op        1380.81 MB/s           0 B/op          0 allocs/op
BenchmarkGetOne_Jsoniter-16                            51178 ns/op         254.46 MB/s       27936 B/op        647 allocs/op
BenchmarkGetOne_Parallel_Sonic-16                       1955 ns/op        6659.94 MB/s         125 B/op          1 allocs/op
BenchmarkGetOne_Parallel_Gjson-16                       1076 ns/op        12098.62 MB/s          0 B/op          0 allocs/op
BenchmarkGetOne_Parallel_Jsoniter-16                   17741 ns/op         734.06 MB/s       27945 B/op        647 allocs/op
BenchmarkSetOne_Sonic-16                               16124 ns/op         807.70 MB/s        1787 B/op         17 allocs/op
BenchmarkSetOne_Sjson-16                               36456 ns/op         357.22 MB/s       52180 B/op          9 allocs/op
BenchmarkSetOne_Jsoniter-16                            79475 ns/op         163.86 MB/s       45862 B/op        964 allocs/op
BenchmarkSetOne_Parallel_Sonic-16                       2383 ns/op        5465.02 MB/s        2186 B/op         17 allocs/op
BenchmarkSetOne_Parallel_Sjson-16                      18194 ns/op         715.77 MB/s       52247 B/op          9 allocs/op
BenchmarkSetOne_Parallel_Jsoniter-16                   33560 ns/op         388.05 MB/s       45892 B/op        964 allocs/op
  • Small (400B, 11 keys, 3 layers) small benchmarks
  • Large (635KB, 10000+ key, 6 layers) large benchmarks

See for benchmark codes.

How it works




Default behaviors are mostly consistent with encoding/json, except HTML escaping form (see Escape HTML) and SortKeys feature (optional support see Sort Keys) that is NOT in conformity to RFC8259.

import ""

var data YourSchema
// Marshal
output, err := sonic.Marshal(&data)
// Unmarshal
err := sonic.Unmarshal(output, &data)

Streaming IO

Sonic supports to decode json from io.Reader or encode objects into io.Writer, aiming at handling multiple values as well as reducing memory consuming.

  • encoder
import ""

var o1 =  map[string]interface{}{
         "a": "b"
var o2 = 1
var w = bytes.NewBuffer(nil)
var enc = encoder.NewStreamEncoder(w)
println(w1.String()) // "{\"a\":\"b\"}\n1"
  • decoder
import ""

var o =  map[string]interface{}{}
var r = strings.NewReader(`{"a":"b"}{"1":"2"}`)
var dec = decoder.NewStreamDecoder(r)
fmt.Printf("%+v", o) // map[1:2 a:b]

Use Number/Use Int64

import ""

var input = `1`
var data interface{}

// default float64
dc := decoder.NewDecoder(input)
dc.Decode(&data) // data == float64(1)
// use json.Number
dc = decoder.NewDecoder(input)
dc.Decode(&data) // data == json.Number("1")
// use int64
dc = decoder.NewDecoder(input)
dc.Decode(&data) // data == int64(1)

root, err := sonic.GetFromString(input)
// Get json.Number
jn := root.Number()
jm := root.InterfaceUseNumber().(json.Number) // jn == jm
// Get float64
fn := root.Float64()
fm := root.Interface().(float64) // jn == jm

Sort Keys

On account of the performance loss from sorting (roughly 10%), sonic doesn't enable this feature by default. If your component depends on it to work (like zstd), Use it like this:

import ""
import ""

// Binding map only
m := map[string]interface{}{}
v, err := encoder.Encode(m, encoder.SortMapKeys)

// Or ast.Node.SortKeys() before marshal
var root := sonic.Get(JSON)
err := root.SortKeys()

Escape HTML

On account of the performance loss (roughly 15%), sonic doesn't enable this feature by default. You can use encoder.EscapeHTML option to open this feature (align with encoding/json.HTMLEscape).

import ""

v := map[string]string{"&&":"<>"}
ret, err := Encode(v, EscapeHTML) // ret == `{"\u0026\u0026":{"X":"\u003c\u003e"}}`

Compact Format

Sonic encodes primitive objects (struct/map...) as compact-format JSON by default, except marshaling json.RawMessage or json.Marshaler: sonic ensures validating their output JSON but DONOT compacting them for performance concerns. We provide the option encoder.CompactMarshaler to add compacting process.

import ""
import ""

var data interface{}
err := sonic.Unmarshal("[[[}]]", &data)
if err != nil {
    /*one line by default*/
    println(e.Error())) // "Syntax error at index 3: invalid char\n\n\t[[[}]]\n\t...^..\n"
    /*pretty print*/
    if e, ok := err.(decoder.SyntaxError); ok {
        /*Syntax error at index 3: invalid char



Sonic/ast.Node is a completely self-contained AST for JSON. It implements serialization and deserialization both, and provides robust APIs for obtaining and modification of generic data.


Search partial JSON by given paths, which must be non-negative integer or string or nil

import ""

input := []byte(`{"key1":[{},{"key2":{"key3":[1,2,3]}}]}`)

// no path, returns entire json
root, err := sonic.Get(input)
raw := root.Raw() // == string(input)

// multiple paths
root, err := sonic.Get(input, "key1", 1, "key2")
sub := root.Get("key3").Index(2).Int64() // == 3

Tip: since Index() uses offset to locate data, which is much faster than scanning like Get(), we suggest you use it as much as possible. And sonic also provides another API IndexOrGet() to underlying use offset as well as ensuring the key is matched.


Modify the json content by Set()/Unset()

import ""

// Set
exist, err := root.Set("key4", NewBool(true)) // exist == false
alias1 := root.Get("key4")
println(alias1.Valid()) // true
alias2 := root.Index(1)
println(alias1 == alias2) // true

// Unset
exist, err := root.UnsetByIndex(1) // exist == true
println(root.Get("key4").Check()) // "value not exist"


To encode ast.Node as json, use MarshalJson() or json.Marshal() (MUST pass the node's pointer)

import (

buf, err := root.MarshalJson()
println(string(buf))                // {"key1":[{},{"key2":{"key3":[1,2,3]}}]}
exp, err := json.Marshal(&root)     // WARN: use pointer
println(string(buf) == string(exp)) // true


  • validation: Check(), Error(), Valid(), Exist()
  • searching: Index(), Get(), IndexPair(), IndexOrGet(), GetByPath()
  • go-type casting: Int64(), Float64(), String(), Number(), Bool(), Map[UseNumber|UseNode](), Array[UseNumber|UseNode](), Interface[UseNumber|UseNode]()
  • go-type packing: NewRaw(), NewNumber(), NewNull(), NewBool(), NewString(), NewObject(), NewArray()
  • iteration: Values(), Properties(), ForEach(), SortKeys()
  • modification: Set(), SetByIndex(), Add()


Sonic DOES NOT ensure to support all environments, due to the difficulty of developing high-performance codes. For developers who use sonic to build their applications in different environments, we have the following suggestions:

  • Developing on Mac M1: Make sure you have Rosetta 2 installed on your machine, and set GOARCH=amd64 when building your application. Rosetta 2 can automatically translate x86 binaries to arm64 binaries and run x86 applications on Mac M1.
  • Developing on Linux arm64: You can install qemu and use the qemu-x86_64 -cpu max command to convert x86 binaries to amr64 binaries for applications built with sonic. The qemu can achieve similar transfer effect to Rosetta 2 on Mac M1.

For developers who want to use sonic on Linux arm64 without qemu, or those who want to handle JSON strictly consistent with encoding/json, we provide some compatible APIs as sonic.API

  • ConfigDefault: the sonic's default config (EscapeHTML=false,SortKeys=false...) to run on sonic-supporting environment. It will fall back to encoding/json with the corresponding config, and some options like SortKeys=false will be invalid.
  • ConfigStd: the std-compatible config (EscapeHTML=true,SortKeys=true...) to run on sonic-supporting environment. It will fall back to encoding/json.
  • ConfigFastest: the fastest config (NoQuoteTextMarshaler=true) to run on sonic-supporting environment. It will fall back to encoding/json with corresponding config, and some options will be invalid.



Since Sonic uses golang-asm as a JIT assembler, which is NOT very suitable for runtime compiling, first-hit running of a huge schema may cause request-timeout or even process-OOM. For better stability, we advise to use Pretouch() for huge-schema or compact-memory application before Marshal()/Unmarshal().

import (

func init() {
    var v HugeStruct

    // For most large types (nesting depth <= option.DefaultMaxInlineDepth)
    err := sonic.Pretouch(reflect.TypeOf(v))

    // with more CompileOption...
    err := sonic.Pretouch(reflect.TypeOf(v), 
        // If the type is too deep nesting (nesting depth > option.DefaultMaxInlineDepth),
        // you can set compile recursive loops in Pretouch for better stability in JIT.
        // For large nested struct, try to set smaller depth to reduce compiling time.

Copy string

When decoding string values without any escaped characters, sonic references them from the origin JSON buffer instead of mallocing a new buffer to copy. This helps a lot for CPU performance but may leave the whole JSON buffer in memory as long as the decoded objects are being used. In practice, we found the extra memory introduced by referring JSON buffer is usually 20% ~ 80% of decoded objects. Once an application holds these objects for a long time (for example, cache the decoded objects for reusing), its in-use memory on the server may go up. We provide the option decoder.CopyString() for users to choose not to reference the JSON buffer, which may cause a decline in CPU performance to some degree.

Pass string or []byte?

For alignment to encoding/json, we provide API to pass []byte as an argument, but the string-to-bytes copy is conducted at the same time considering safety, which may lose performance when origin JSON is huge. Therefore, you can use UnmarshalString() and GetFromString() to pass a string, as long as your origin data is a string or nocopy-cast is safe for your []byte. We also provide API MarshalString() for convenient nocopy-cast of encoded JSON []byte, which is safe since sonic's output bytes is always duplicated and unique.

Accelerate encoding.TextMarshaler

To ensure data security, sonic.Encoder quotes and escapes string values from encoding.TextMarshaler interfaces by default, which may degrade performance much if most of your data is in form of them. We provide encoder.NoQuoteTextMarshaler to skip these operations, which means you MUST ensure their output string escaped and quoted in accordance with RFC8259.

Better performance for generic data

In fully-parsed scenario, Unmarshal() performs better than Get()+Node.Interface(). But if you only have a part of schema for specific json, you can combine Get() and Unmarshal() together:

import ""

node, err := sonic.GetFromString(_TwitterJson, "statuses", 3, "user")
var user User // your partial schema...
err = sonic.UnmarshalString(node.Raw(), &user)

Even if you don't have any schema, use ast.Node as the container of generic values instead of map or interface:

import ""

root, err := sonic.GetFromString(_TwitterJson)
user := root.GetByPath("statuses", 3, "user")  // === root.Get("status").Index(3).Get("user")
err = user.Check()

// err = user.LoadAll() // only call this when you want to use 'user' concurrently...
go someFunc(user)

Why? Because ast.Node stores its children using array:

  • Array's performance is much better than Map when Inserting (Deserialize) and Scanning (Serialize) data;
  • Hashing (map[x]) is not as efficient as Indexing (array[x]), which ast.Node can conduct on both array and object;
  • Using Interface()/Map() means Sonic must parse all the underlying values, while ast.Node can parse them on demand.

CAUTION: ast.Node DOESN'T ensure concurrent security directly, due to its lazy-load design. However, your can call Node.Load()/Node.LoadAll() to achieve that, which may bring performance reduction while it still works faster than converting to map or interface{}


Sonic is a subproject of CloudWeGo. We are committed to building a cloud native ecosystem.

Open Source Agenda is not affiliated with "Bytedance Sonic" Project. README Source: bytedance/sonic
Open Issues
Last Commit
2 weeks ago

Open Source Agenda Badge

Open Source Agenda Rating