Awesome

Installation

Jettison uses Go modules. Releases are tagged according to the SemVer format, prefixed with a v, starting from 0.2.0. You can get the latest release using the following command.

$ go get github.com/wI2L/jettison@latest

:warning: From version v0.7.4, the package requires Go 1.17+ to build, due to the usage of the new build constraints.

Key features

• Fast, see benchmarks
• No dynamic memory allocations in hot paths
• Behavior identical to the standard library by default
• No code generation required
• Clear and concise API
• Configurable with opt-in functional options
• Native support for many standard library types, see improvements
• Custom AppendMarshaler interface to avoid allocations
• Extensive testsuite that compares its output against encoding/json

Overview

The goal of Jettision is to take up the idea introduced by the bet365/jingo package and build a fully-featured JSON encoder around it, that comply with the behavior of the encoding/json package. Unlike the latter, Jettison does not use reflection during marshaling, but only once to create the instruction set for a given type ahead of time. The drawback to this approach requires to instantiate an instruction-set once for each type that needs to be marshaled, but that is overcomed with a package cache.

The package aims to have a behavior similar to that of the standard library for all types encoding and struct tags, meaning that the documentation of the json.Marshal function is applicable for Jettison, with a few exceptions described in this section. As such, most of the tests compare their output against it to guarantee that.

Implementation details

The main concept of Jettison consists of using pre-build instructions-set to reduce the cost of using the reflect package at runtime. When marshaling a value, a set of instructions is recursively generated for its type, which defines how to iteratively encode it. An instruction is a function or a closure, that have all the information required to read the data from memory using unsafe operations (pointer type conversion, arithmetic...) during the instruction set execution.

Differences with encoding/json

All notable differences with the standard library behavior are listed below. Please note that these might evolve with future versions of the package.

Improvements

• The time.Time and time.Duration types are handled natively. For time values, the encoder doesn't invoke MarshalJSON or MarshalText, but use the time.AppendFormat function instead, and write the result to the stream. Similarly, for durations, it isn't necessary to implements the json.Marshaler or encoding.TextMarshaler interfaces on a custom wrapper type, the encoder uses the result of one of the methods Minutes, Seconds, Nanoseconds or String, based on the duration format configured.

• The sync.Map type is handled natively. The marshaling behavior is similar to the one of a standard Go map. The option UnsortedMap can also be used in cunjunction with this type to disable the default keys sort.

• The omitnil field tag's option can be used to specify that a field with a nil pointer should be omitted from the encoding. This option has precedence over the omitempty option. Note that struct fields that implement the json.Marshaler interface will be omitted too, if they return the literal JSON null value.

Bugs

Go1.13 and backward

• Nil map keys values implementing the encoding.TextMarshaler interface are encoded as empty strings, while the encoding/json package currently panic because of that. See this issue for more details.^[1]

• Nil struct fields implementing the encoding.TextMarshaler interface are encoded as null, while the encoding/json package currently panic because of that. See this issue for more details.^[1]

^{1: The issues mentioned above have had their associated CL merged, and was released with Go 1.14.}

Usage

Basic

As stated above, the library behave similarly to the encoding/json package. You can simply replace the json.Marshal function with jettison.Marshal, and expect the same output with better performances.

type X struct {
   A string `json:"a"`
   B int64  `json:"b"`
}
b, err := jettison.Marshal(X{
   A: "Loreum",
   B: 42,
})
if err != nil {
   log.Fatal(err)
}
os.Stdout.Write(b)

Result

{"a":"Loreum","b":42}

Advanced

If more control over the encoding behavior is required, use the MarshalOpts function instead. The second parameter is variadic and accept a list of functional opt-in options described below:

Take a look at the examples to see these options in action.

Benchmarks

If you'd like to run the benchmarks yourself, use the following command.

go get github.com/cespare/prettybench
go test -bench=. | prettybench

Results -short

These benchmarks were run 10x (statistics computed with benchstat) on a MacBook Pro 15", with the following specs:

OS:  macOS Catalina (10.15.7)
CPU: 2.6 GHz Intel Core i7
Mem: 16GB
Go:  go version go1.17 darwin/amd64
Tag: v0.7.2

name speed Simple/standard-8 236MB/s ± 1% Simple/jsoniter-8 247MB/s ± 0% Simple/segmentj-8 516MB/s ± 1% Simple/jettison-8 331MB/s ± 1% Complex/standard-8 72.9MB/s ± 0% Complex/jsoniter-8 70.6MB/s ± 0% Complex/segmentj-8 108MB/s ± 0% Complex/jettison-8 144MB/s ± 1% CodeMarshal/standard-8 289MB/s ± 0% CodeMarshal/jsoniter-8 306MB/s ± 1% CodeMarshal/segmentj-8 443MB/s ± 1% CodeMarshal/jettison-8 349MB/s ± 1% Map/standard-8 46.6MB/s ± 1% Map/jsoniter-8 51.5MB/s ± 0% Map/segmentj-8 52.8MB/s ± 0% Map/jettison-8 110MB/s ± 1% Map/jettison-nosort-8 168MB/s ± 1%

name alloc/op Simple/standard-8 144B ± 0% Simple/jsoniter-8 152B ± 0% Simple/segmentj-8 144B ± 0% Simple/jettison-8 144B ± 0% Complex/standard-8 4.05kB ± 0% Complex/jsoniter-8 3.95kB ± 0% Complex/segmentj-8 2.56kB ± 0% Complex/jettison-8 935B ± 0% CodeMarshal/standard-8 1.97MB ± 0% CodeMarshal/jsoniter-8 2.00MB ± 0% CodeMarshal/segmentj-8 1.98MB ± 2% CodeMarshal/jettison-8 1.98MB ± 2% Map/standard-8 888B ± 0% Map/jsoniter-8 884B ± 0% Map/segmentj-8 576B ± 0% Map/jettison-8 96.0B ± 0% Map/jettison-nosort-8 160B ± 0%

name allocs/op Simple/standard-8 1.00 ± 0% Simple/jsoniter-8 2.00 ± 0% Simple/segmentj-8 1.00 ± 0% Simple/jettison-8 1.00 ± 0% Complex/standard-8 79.0 ± 0% Complex/jsoniter-8 71.0 ± 0% Complex/segmentj-8 52.0 ± 0% Complex/jettison-8 8.00 ± 0% CodeMarshal/standard-8 1.00 ± 0% CodeMarshal/jsoniter-8 2.00 ± 0% CodeMarshal/segmentj-8 1.00 ± 0% CodeMarshal/jettison-8 1.00 ± 0% Map/standard-8 19.0 ± 0% Map/jsoniter-8 14.0 ± 0% Map/segmentj-8 18.0 ± 0% Map/jettison-8 1.00 ± 0% Map/jettison-nosort-8 2.00 ± 0% </pre></details>

Simple [source]

Basic payload with fields of type string, int and bool.

Complex [source]

Large payload with a variety of composite Go types, such as struct, map, interface, multi-dimensions array and slice, with pointer and non-pointer value types.

Please note that this test is somewhat positively influenced by the performances of map marshaling.

CodeMarshal [source]

Borrowed from the encoding/json tests. See testdata/code.json.gz.

Map [source]

Simple map[string]int with 6 keys.

Credits

This library and its design has been inspired by the work of others at @bet365 and @segmentio. See the following projects for reference:

• bet365/jingo
• segmentio/encoding

License

Jettison is licensed under the MIT license. See the LICENSE file.

This package also uses some portions of code from the Go encoding/json package. The associated license can be found in LICENSE.golang.