orjson is a fast JSON library for Python. It benchmarks as the fastest Python library for JSON serialization, with 1.6x to 2.6x the performance of the nearest other library, and deserialization performance of 0.95x to 1.2x the nearest other library.

It supports CPython 3.5, 3.6, and 3.7. It is not intended as a drop-in replacement for the standard library's json module.

To install a manylinux wheel from PyPI:

pip install --upgrade orjson

To build a release wheel from source, assuming a Rust nightly toolchain and Python environment:

git clone --recurse-submodules https://github.com/ijl/orjson.git && cd orjson
virtualenv .venv && source .venv/bin/activate
pip install --upgrade pyo3-pack
pyo3-pack build --release --strip --interpreter python3.7

There is no runtime dependency other than a manylinux environment (i.e., deploying this does not require Rust or non-libc type libraries.)

def dumps(obj: Union[str, bytes, dict, list, tuple, int, float, None]) -> bytes: ...

dumps() serializes Python objects to JSON.

It has no options, does not support hooks for custom objects, and does not support subclasses. It supports 64-bit integers and 64-bit floats, which is the same as the standard library json module.

It raises TypeError on an unsupported type. This exception message describes the invalid object.

It raises TypeError on an integer that is too large.

It raises TypeError if a dict has a key of a type other than str.

import orjson

try:
    val = orjson.dumps(...)
except TypeError:
    raise

def loads(obj: Union[bytes, str]) -> Union[dict, list, int, float, str]: ...

loads() deserializes JSON to Python objects.

It raises orjson.JSONDecodeError if given an invalid type or invalid JSON. This exception is a subclass of ValueError.

import orjson

try:
    val = orjson.loads(...)
except orjson.JSONDecodeError:
    raise

There are slight differences in output between libraries. The differences are not an issue for interoperability. Note orjson returns bytes. Its output is slightly smaller as well.

>>> import orjson, ujson, rapidjson, json
>>> data = {'bool': True, '🐈':'哈哈', 'int': 9223372036854775807, 'float': 1.337e+40}
>>> orjson.dumps(data)
b'{"bool":true,"\xf0\x9f\x90\x88":"\xe5\x93\x88\xe5\x93\x88","int":9223372036854775807,"float":1.337e40}'
>>> ujson.dumps(data)
'{"bool":true,"\\ud83d\\udc08":"\\u54c8\\u54c8","int":9223372036854775807,"float":1.337000000000000e+40}'
>>> rapidjson.dumps(data)
'{"bool":true,"\\uD83D\\uDC08":"\\u54C8\\u54C8","int":9223372036854775807,"float":1.337e+40}'
>>> json.dumps(data)
'{"bool": true, "\\ud83d\\udc08": "\\u54c8\\u54c8", "int": 9223372036854775807, "float": 1.337e+40}'

The library has comprehensive tests. There are unit tests against the roundtrip, jsonchecker, and fixtures files of the nativejson-benchmark repository. It is tested to not crash against the Big List of Naughty Strings. There are integration tests exercising the library's use in web servers (uwsgi and gunicorn, using multiprocess/forked workers) and when multithreaded. It also uses some tests from the ultrajson library.

Serialization performance of orjson is better than ultrajson, rapidjson, or json. Deserialization performance is better to about the same as ultrajson.

This was measured using orjson 1.2.0 on Python 3.7.1 and Linux. The above can be reproduced using the pybench and graph scripts.

orjson is dual licensed under the Apache 2.0 and MIT licenses. It contains code from the hyperjson and ultrajson libraries. It is implemented using the serde_json and pyo3 libraries.