Skip to content

a8m/djson

Repository files navigation

DJSON

GoDoc Build Status LICENSE

DJSON is a JSON decoder for Go that is 2~ to 3~ times faster than the standard encoding/json and the existing solutions, when dealing with arbitrary JSON payload. See benchmarks below.
It is a good approach for people who are using json.Unmarshal together with interface{}, don't know what the schema is, and still want good performance with minimal changes.

Motivation

While searching for a JSON parser solution for my projects, that is faster than the standard library, with zero reflection tests, allocates less memory and is still safe(I didn't want the "unsafe" package in my production code, in order to reduce memory consumption).
I found that almost all implemtations are just wrappers around the standard library and aren't fast enough for my needs.
I encountered two projects: ujson that is the UltraJSON implementation and jsonparser, that is a pretty awesome project.
ujson seems to be faster than encoding/json but still doesn't meet my requirements.
jsonparser seems to be really fast, and I even use it for some of my new projects.
However, its API is different, and I would need to change too much of my code in order to work with it.
Also, for my processing work that involves ETL, changing and setting new fields on the JSON object, I need to transform the jsonparser result to map[string]interface{} and it seems that it loses its power.

Advantages and Stability

As you can see in the benchmark below, DJSON is faster and allocates less memory than the other alternatives.
The current version is 1.0.0-alpha.1, and I'm waiting to hear from you if there are any issues or bug reports, to make it stable.
(comment: there is a test file named decode_test that contains a test case that compares the results to encoding/json - feel free to add more values if you find they are important)
I'm also plaining to add the DecodeStream(io.ReaderCloser) method(or NewDecoder(io.ReaderCloser)), to support stream decoding without breaking performance.

Benchmark

There are 3 benchmark types: small, medium and large payloads.
All the 3 are taken from the jsonparser project, and they try to simulate a real-life usage. Each result from the different benchmark types is shown in a metric table below. The lower the metrics are, the better the result is. Time/op is in nanoseconds, B/op is how many bytes were allocated per op and allocs/op is the total number of memory allocations.
Benchmark results that are better than encoding/json are marked in bold text.
The Benchmark tests run on AWS EC2 instance(c4.xlarge). see: screenshots

Compared libraries:

Small payload

Each library in the test gets a small payload to process that weighs 134 bytes.
You can see the payload here, and the test screenshot here.

Library Time/op B/op allocs/op
encoding/json 8646 1993 60
ugorji/go/codec 9272 4513 41
antonholmquist/jason 7336 3201 49
bitly/go-simplejson 5253 2241 36
Jeffail/gabs 4788 1409 33
mreiferson/go-ujson 3897 1393 35
a8m/djson 2534 1137 25
a8m/djson.AllocString 2195 1169 13

Medium payload

Each library in the test gets a medium payload to process that weighs 1.7KB.
You can see the payload here, and the test screenshot here.

Library Time/op B/op allocs/op
encoding/json 42029 10652 218
ugorji/go/codec 65007 15267 313
antonholmquist/jason 45676 17476 224
bitly/go-simplejson 45164 17156 219
Jeffail/gabs 41045 10515 211
mreiferson/go-ujson 33213 11506 267
a8m/djson 22871 10100 195
a8m/djson.AllocString 19296 10619 87

Large payload

Each library in the test gets a large payload to process that weighs 28KB.
You can see the payload here, and the test screenshot here.

Library Time/op B/op allocs/op
encoding/json 717882 212827 3247
ugorji/go/codec 1052347 239130 4426
antonholmquist/jason 751910 277931 3257
bitly/go-simplejson 753663 277628 3252
Jeffail/gabs 714304 212740 3241
mreiferson/go-ujson 599868 235789 4057
a8m/djson 437031 210997 2932
a8m/djson.AllocString 372382 214053 1413

LICENSE

MIT