Fast, correct Python JSON library supporting dataclasses, datetimes, and numpy

orjson

orjson is a fast, correct JSON library for Python. It benchmarks as the fastest Python library for JSON and is more correct than the standard json library or other third-party libraries. It serializes dataclass, datetime, numpy, and UUID instances natively.

Its features and drawbacks compared to other Python JSON libraries:

• serializes dataclass instances 40-50x as fast as other libraries
• serializes datetime, date, and time instances to RFC 3339 format, e.g., “1970-01-01T00:00:00+00:00”
• serializes numpy.ndarray instances 3-10x as fast as other libraries
• pretty prints 10x to 20x as fast as the standard library
• serializes to bytes rather than str, i.e., is not a drop-in replacement
• serializes str without escaping unicode to ASCII, e.g., “好” rather than “\\u597d”
• serializes float 10x as fast and deserializes twice as fast as other libraries
• serializes arbitrary types using a default hook
• has strict UTF-8 conformance, more correct than the standard library
• has strict JSON conformance in not supporting Nan/Infinity/-Infinity
• has an option for strict JSON conformance on 53-bit integers with default support for 64-bit
• does not support subclasses by default, requiring use of default hook
• does not provide load() or dump() functions for reading from/writing to file-like objects

orjson supports CPython 3.6, 3.7, 3.8, and 3.9. It distributes x86_64/amd64 and aarch64/armv8 wheels for Linux. It distributes x86_64/amd64 wheels for macOS and Windows. orjson does not support PyPy.

orjson is licensed under both the Apache 2.0 and MIT licenses. The repository and issue tracker is github.com/ijl/orjson, and patches may be submitted there. There is a CHANGELOG available in the repository.

1. Usage
   1. Install
   2. Serialize
      1. default
      2. option
   3. Deserialize
2. Types
   1. dataclass
   2. datetime
   3. float
   4. int
   5. numpy
   6. str
   7. UUID
3. Testing
4. Performance
   1. Latency
   2. Memory
   3. Reproducing
5. License

Usage

Install

To install a wheel from PyPI:

pip install --upgrade orjson

To build from source requires Rust on the nightly channel. Package a wheel from a PEP 517 source distribution using pip:

pip wheel --no-binary=orjson orjson

There are no runtime dependencies other than libc. orjson is compatible with systems using glibc earlier than 2.18 if compiled on such a system. Tooling does not currently support musl libc.

Serialize

def dumps(
    __obj: Any,
    default: Optional[Callable[[Any], Any]] = ...,
    option: Optional[int] = ...,
) -> bytes: ...

dumps() serializes Python objects to JSON.

It natively serializes str, dict, list, tuple, int, float, bool, dataclasses.dataclass, typing.TypedDict, datetime.datetime, datetime.date, datetime.time, uuid.UUID, and None instances. It supports arbitrary types through default. It does not serialize subclasses of supported types natively, with the exception of dataclasses.dataclass subclasses.

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

It raises JSONEncodeError on a str that contains invalid UTF-8.

It raises JSONEncodeError on an integer that exceeds 64 bits by default or, with OPT_STRICT_INTEGER, 53 bits.

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

It raises JSONEncodeError if the output of default recurses to handling by default more than 254 levels deep.

It raises JSONEncodeError on circular references.

It raises JSONEncodeError if a tzinfo on a datetime object is incorrect.

JSONEncodeError is a subclass of TypeError. This is for compatibility with the standard library.

default

To serialize a subclass or arbitrary types, specify default as a callable that returns a supported type. default may be a function, lambda, or callable class instance. To specify that a type was not handled by default, raise an exception such as TypeError.

>>> import orjson, decimal
>>>
def default(obj):
    if isinstance(obj, decimal.Decimal):
        return str(obj)
    raise TypeError

>>> orjson.dumps(decimal.Decimal("0.0842389659712649442845"))
JSONEncodeError: Type is not JSON serializable: decimal.Decimal
>>> orjson.dumps(decimal.Decimal("0.0842389659712649442845"), default=default)
b'"0.0842389659712649442845"'
>>> orjson.dumps({1, 2}, default=default)
orjson.JSONEncodeError: Type is not JSON serializable: set

The default callable may return an object that itself must be handled by default up to 254 times before an exception is raised.

It is important that default raise an exception if a type cannot be handled. Python otherwise implicitly returns None, which appears to the caller like a legitimate value and is serialized:

>>> import orjson, json, rapidjson
>>>
def default(obj):
    if isinstance(obj, decimal.Decimal):
        return str(obj)

>>> orjson.dumps({"set":{1, 2}}, default=default)
b'{"set":null}'
>>> json.dumps({"set":{1, 2}}, default=default)
'{"set":null}'
>>> rapidjson.dumps({"set":{1, 2}}, default=default)
'{"set":null}'

option

To modify how data is serialized, specify option. Each option is an integer constant in orjson. To specify multiple options, mask them together, e.g., option=orjson.OPT_STRICT_INTEGER | orjson.OPT_NAIVE_UTC.

OPT_INDENT_2

Pretty-print output with an indent of two spaces. This is equivalent to indent=2 in the standard library. Pretty printing is slower and the output larger. orjson is the fastest compared library at pretty printing and has much less of a slowdown to pretty print than the standard library does. This option is compatible with all other options.

>>> import orjson
>>> orjson.dumps({"a": "b", "c": {"d": True}, "e": [1, 2]})
b'{"a":"b","c":{"d":true},"e":[1,2]}'
>>> orjson.dumps(
    {"a": "b", "c": {"d": True}, "e": [1, 2]},
    option=orjson.OPT_INDENT_2
)
b'{\n  "a": "b",\n  "c": {\n    "d": true\n  },\n  "e": [\n    1,\n    2\n  ]\n}'

If displayed, the indentation and linebreaks appear like this:

{
  "a": "b",
  "c": {
    "d": true
  },
  "e": [
    1,
    2
  ]
}

This measures serializing the github.json fixture as compact (52KiB) or pretty (64KiB):

This measures serializing the citm_catalog.json fixture, more of a worst case due to the amount of nesting and newlines, as compact (489KiB) or pretty (1.1MiB):

rapidjson is blank because it does not support pretty printing. This can be reproduced using the pyindent script.

OPT_NAIVE_UTC

Serialize datetime.datetime objects without a tzinfo as UTC. This has no effect on datetime.datetime objects that have tzinfo set.

>>> import orjson, datetime
>>> orjson.dumps(
        datetime.datetime(1970, 1, 1, 0, 0, 0),
    )
b'"1970-01-01T00:00:00"'
>>> orjson.dumps(
        datetime.datetime(1970, 1, 1, 0, 0, 0),
        option=orjson.OPT_NAIVE_UTC,
    )
b'"1970-01-01T00:00:00+00:00"'

OPT_NON_STR_KEYS

Serialize dict keys of type other than str. This allows dict keys to be one of str, int, float, bool, None, datetime.datetime, datetime.date, datetime.time, and uuid.UUID. For comparison, the standard library serializes str, int, float, bool or None by default. orjson benchmarks as being faster at serializing non-str keys than other libraries. This option is slower for str keys than the default and is not recommended generally.

>>> import orjson, datetime, uuid
>>> orjson.dumps(
        {uuid.UUID("7202d115-7ff3-4c81-a7c1-2a1f067b1ece"): [1, 2, 3]},
        option=orjson.OPT_NON_STR_KEYS,
    )
b'{"7202d115-7ff3-4c81-a7c1-2a1f067b1ece":[1,2,3]}'
>>> orjson.dumps(
        {datetime.datetime(1970, 1, 1, 0, 0, 0): [1, 2, 3]},
        option=orjson.OPT_NON_STR_KEYS | orjson.OPT_NAIVE_UTC,
    )
b'{"1970-01-01T00:00:00+00:00":[1,2,3]}'

These types are generally serialized how they would be as values, e.g., datetime.datetime is still an RFC 3339 string and respects options affecting it. The exception is that int serialization does not respect OPT_STRICT_INTEGER.

This option has the risk of creating duplicate keys. This is because non-str objects may serialize to the same str as an existing key, e.g., {"1": true, 1: false}. The last key to be inserted to the dict will be serialized last and a JSON deserializer will presumably take the last occurrence of a key (in the above, false). The first value will be lost.

This option is compatible with orjson.OPT_SORT_KEYS. If sorting is used, note the sort is unstable and will be unpredictable for duplicate keys.

>>> import orjson, datetime
>>> orjson.dumps(
    {"other": 1, datetime.date(1970, 1, 5): 2, datetime.date(1970, 1, 3): 3},
    option=orjson.OPT_NON_STR_KEYS | orjson.OPT_SORT_KEYS
)
b'{"1970-01-03":3,"1970-01-05":2,"other":1}'

This measures serializing 589KiB of JSON comprising a list of 100 dict in which each dict has both 365 randomly-sorted int keys representing epoch timestamps as well as one str key and the value for each key is a single integer. In “str keys”, the keys were converted to str before serialization, and orjson still specifes option=orjson.OPT_NON_STR_KEYS (which is always somewhat slower).

ujson is blank for sorting because it segfaults. json is blank because it raises TypeError on attempting to sort before converting all keys to str. rapidjson is blank because it does not support non-str keys. This can be reproduced using the pynonstr script.

OPT_OMIT_MICROSECONDS

Do not serialize the microsecond field on datetime.datetime and datetime.time instances.

>>> import orjson, datetime
>>> orjson.dumps(
        datetime.datetime(1970, 1, 1, 0, 0, 0, 1),
    )
b'"1970-01-01T00:00:00.000001"'
>>> orjson.dumps(
        datetime.datetime(1970, 1, 1, 0, 0, 0, 1),
        option=orjson.OPT_OMIT_MICROSECONDS,
    )
b'"1970-01-01T00:00:00"'

OPT_SERIALIZE_DATACLASS

Serialize dataclasses.dataclass instances. For more, see dataclass.

OPT_SERIALIZE_NUMPY

Serialize numpy.ndarray instances. For more, see numpy.

OPT_SERIALIZE_UUID

Serialize uuid.UUID instances. For more, see UUID.

OPT_SORT_KEYS

Serialize dict keys in sorted order. The default is to serialize in an unspecified order. This is equivalent to sort_keys=True in the standard library.

This can be used to ensure the order is deterministic for hashing or tests. It has a substantial performance penalty and is not recommended in general.

>>> import orjson
>>> orjson.dumps({"b": 1, "c": 2, "a": 3})
b'{"b":1,"c":2,"a":3}'
>>> orjson.dumps({"b": 1, "c": 2, "a": 3}, option=orjson.OPT_SORT_KEYS)
b'{"a":3,"b":1,"c":2}'

This measures serializing the twitter.json fixture unsorted and sorted:

The benchmark can be reproduced using the pysort script.

The sorting is not collation/locale-aware:

>>> import orjson
>>> orjson.dumps({"a": 1, "ä": 2, "A": 3}, option=orjson.OPT_SORT_KEYS)
b'{"A":3,"a":1,"\xc3\xa4":2}'

This is the same sorting behavior as the standard library, rapidjson, simplejson, and ujson.

dataclass also serialize as maps but this has no effect on them.

OPT_STRICT_INTEGER

Enforce 53-bit limit on integers. The limit is otherwise 64 bits, the same as the Python standard library. For more, see int.

OPT_UTC_Z

Serialize a UTC timezone on datetime.datetime instances as Z instead of +00:00.

>>> import orjson, datetime
>>> orjson.dumps(
        datetime.datetime(1970, 1, 1, 0, 0, 0, tzinfo=datetime.timezone.utc),
    )
b'"1970-01-01T00:00:00+00:00"'
>>> orjson.dumps(
        datetime.datetime(1970, 1, 1, 0, 0, 0, tzinfo=datetime.timezone.utc),
        option=orjson.OPT_UTC_Z
    )
b'"1970-01-01T00:00:00Z"'

Deserialize

def loads(__obj: Union[bytes, bytearray, str]) -> Any: ...

loads() deserializes JSON to Python objects. It deserializes to dict, list, int, float, str, bool, and None objects.

bytes, bytearray, and str input are accepted. If the input exists as bytes (was read directly from a source), it is recommended to pass bytes. This has lower memory usage and lower latency.

orjson maintains a cache of map keys for the duration of the process. This causes a net reduction in memory usage by avoiding duplicate strings. The keys must be at most 64 chars to be cached and 512 entries are stored.

It raises JSONDecodeError if given an invalid type or invalid JSON. This includes if the input contains NaN, Infinity, or -Infinity, which the standard library allows, but is not valid JSON.

JSONDecodeError is a subclass of json.JSONDecodeError and ValueError. This is for compatibility with the standard library.

Types

dataclass

orjson serializes instances of dataclasses.dataclass natively. It serializes instances 40-50x as fast as other libraries and avoids a severe slowdown seen in other libraries compared to serializing dict. To serialize instances, specify option=orjson.OPT_SERIALIZE_DATACLASS. The option is required so that users may continue to use default until the implementation allows customizing instances' serialization.

It is supported to pass all variants of dataclasses, including dataclasses using __slots__, frozen dataclasses, those with optional or default attributes, and subclasses. There is a performance benefit to not using __slots__.

This measures serializing 555KiB of JSON, orjson natively and other libraries using default to serialize the output of dataclasses.asdict(). This can be reproduced using the pydataclass script.

Dataclasses are serialized as maps, with every attribute serialized and in the order given on class definition:

>>> import dataclasses, orjson, typing

@dataclasses.dataclass
class Member:
    id: int
    active: bool = dataclasses.field(default=False)

@dataclasses.dataclass
class Object:
    id: int
    name: str
    members: typing.List[Member]

>>> orjson.dumps(
        Object(1, "a", [Member(1, True), Member(2)]),
        option=orjson.OPT_SERIALIZE_DATACLASS,
    )
b'{"id":1,"name":"a","members":[{"id":1,"active":true},{"id":2,"active":false}]}'

Users may wish to control how dataclass instances are serialized, e.g., to not serialize an attribute or to change the name of an attribute when serialized. orjson may implement support using the metadata mapping on field attributes, e.g., field(metadata={"json_serialize": False}), if use cases are clear.

datetime

orjson serializes datetime.datetime objects to RFC 3339 format, e.g., “1970-01-01T00:00:00+00:00”. This is a subset of ISO 8601 and compatible with isoformat() in the standard library.

>>> import orjson, datetime, pendulum
>>> orjson.dumps(
    datetime.datetime(2018, 12, 1, 2, 3, 4, 9, tzinfo=pendulum.timezone('Australia/Adelaide'))
)
b'"2018-12-01T02:03:04.000009+10:30"'
>>> orjson.dumps(
    datetime.datetime.fromtimestamp(4123518902).replace(tzinfo=datetime.timezone.utc)
)
b'"2100-09-01T21:55:02+00:00"'
>>> orjson.dumps(
    datetime.datetime.fromtimestamp(4123518902)
)
b'"2100-09-01T21:55:02"'

datetime.datetime supports instances with a tzinfo that is None, datetime.timezone.utc or a timezone instance from the pendulum, pytz, or dateutil/arrow libraries.

datetime.time objects must not have a tzinfo.

>>> import orjson, datetime
>>> orjson.dumps(datetime.time(12, 0, 15, 290))
b'"12:00:15.000290"'

datetime.date objects will always serialize.

>>> import orjson, datetime
>>> orjson.dumps(datetime.date(1900, 1, 2))
b'"1900-01-02"'

Errors with tzinfo result in JSONEncodeError being raised.

It is faster to have orjson serialize datetime objects than to do so before calling dumps(). If using an unsupported type such as pendulum.datetime, use default.

float

orjson serializes and deserializes double precision floats with no loss of precision and consistent rounding. The same behavior is observed in rapidjson, simplejson, and json. ujson is inaccurate in both serialization and deserialization, i.e., it modifies the data.

orjson.dumps() serializes Nan, Infinity, and -Infinity, which are not compliant JSON, as null:

>>> import orjson, ujson, rapidjson, json
>>> orjson.dumps([float("NaN"), float("Infinity"), float("-Infinity")])
b'[null,null,null]'
>>> ujson.dumps([float("NaN"), float("Infinity"), float("-Infinity")])
OverflowError: Invalid Inf value when encoding double
>>> rapidjson.dumps([float("NaN"), float("Infinity"), float("-Infinity")])
'[NaN,Infinity,-Infinity]'
>>> json.dumps([float("NaN"), float("Infinity"), float("-Infinity")])
'[NaN, Infinity, -Infinity]'

int

JSON only requires that implementations accept integers with 53-bit precision. orjson will, by default, serialize 64-bit integers. This is compatible with the Python standard library and other non-browser implementations. For transmitting JSON to a web browser or other strict implementations, dumps() can be configured to raise a JSONEncodeError on values exceeding the 53-bit range.

>>> import orjson
>>> orjson.dumps(9007199254740992)
b'9007199254740992'
>>> orjson.dumps(9007199254740992, option=orjson.OPT_STRICT_INTEGER)
JSONEncodeError: Integer exceeds 53-bit range
>>> orjson.dumps(-9007199254740992, option=orjson.OPT_STRICT_INTEGER)
JSONEncodeError: Integer exceeds 53-bit range

numpy

orjson natively serializes numpy.ndarray instances. Arrays may have a dtype of numpy.bool, numpy.float32, numpy.float64, numpy.int32, numpy.int64, numpy.uint32, numpy.uint64, numpy.uintp, or numpy.intp. orjson is faster than all compared libraries at serializing numpy instances. Serializing numpy data requires specifying option=orjson.OPT_SERIALIZE_NUMPY.

>>> import orjson, numpy
>>> orjson.dumps(
        numpy.array([[1, 2, 3], [4, 5, 6]]),
        option=orjson.OPT_SERIALIZE_NUMPY,
)
b'[[1,2,3],[4,5,6]]'

The array must be a contiguous C array (C_CONTIGUOUS) and one of the supported datatypes. Individual items (e.g., numpy.float64(1)) are not supported.

If an array is not a contiguous C array or contains an supported datatype, orjson falls through to default. In default, obj.tolist() can be specified. If an array is malformed, which is not expected, orjson.JSONEncodeError is raised.

This measures serializing 92MiB of JSON from an numpy.ndarray with dimensions of (50000, 100) and numpy.float64 values:

This measures serializing 100MiB of JSON from an numpy.ndarray with dimensions of (100000, 100) and numpy.int32 values:

This measures serializing 53MiB of JSON from an numpy.ndarray with dimensions of (100000, 100) and numpy.bool values:

In these benchmarks, nujson is used instead of ujson, orjson and nujson serialize natively, and the other libraries use ndarray.tolist(). nujson is blank when it did not roundtrip the data accurately. The RSS column measures peak memory usage during serialization. The odd bool result for nujson is consistent.

orjson does not have an installation or compilation dependency on numpy. The implementation is independent, reading numpy.ndarray using PyArrayInterface.

str

orjson is strict about UTF-8 conformance. This is stricter than the standard library’s json module, which will serialize and deserialize UTF-16 surrogates, e.g., “\ud800”, that are invalid UTF-8.

If orjson.dumps() is given a str that does not contain valid UTF-8, orjson.JSONEncodeError is raised. If loads() receives invalid UTF-8, orjson.JSONDecodeError is raised.

orjson and rapidjson are the only compared JSON libraries to consistently error on bad input.

>>> import orjson, ujson, rapidjson, json
>>> orjson.dumps('\ud800')
JSONEncodeError: str is not valid UTF-8: surrogates not allowed
>>> ujson.dumps('\ud800')
UnicodeEncodeError: 'utf-8' codec ...
>>> rapidjson.dumps('\ud800')
UnicodeEncodeError: 'utf-8' codec ...
>>> json.dumps('\ud800')
'"\\ud800"'
>>> orjson.loads('"\\ud800"')
JSONDecodeError: unexpected end of hex escape at line 1 column 8: line 1 column 1 (char 0)
>>> ujson.loads('"\\ud800"')
''
>>> rapidjson.loads('"\\ud800"')
ValueError: Parse error at offset 1: The surrogate pair in string is invalid.
>>> json.loads('"\\ud800"')
'\ud800'

To make a best effort at deserializing bad input, first decode bytes using the replace or lossy argument for errors:

>>> import orjson
>>> orjson.loads(b'"\xed\xa0\x80"')
JSONDecodeError: str is not valid UTF-8: surrogates not allowed
>>> orjson.loads(b'"\xed\xa0\x80"'.decode("utf-8", "replace"))
'���'

UUID

orjson serializes uuid.UUID instances to RFC 4122 format, e.g., “f81d4fae-7dec-11d0-a765-00a0c91e6bf6”. This requires specifying option=orjson.OPT_SERIALIZE_UUID.

>>> import orjson, uuid
>>> orjson.dumps(
    uuid.UUID('f81d4fae-7dec-11d0-a765-00a0c91e6bf6'),
    option=orjson.OPT_SERIALIZE_UUID,
)
b'"f81d4fae-7dec-11d0-a765-00a0c91e6bf6"'
>>> orjson.dumps(
    uuid.uuid5(uuid.NAMESPACE_DNS, "python.org"),
    option=orjson.OPT_SERIALIZE_UUID,
)
b'"886313e1-3b8a-5372-9b90-0c9aee199e5d"'

Testing

The library has comprehensive tests. There are tests against fixtures in the JSONTestSuite and nativejson-benchmark repositories. It is tested to not crash against the Big List of Naughty Strings. It is tested to not leak memory. It is tested to not crash against and not accept invalid UTF-8. There are integration tests exercising the library’s use in web servers (gunicorn using multiprocess/forked workers) and when multithreaded. It also uses some tests from the ultrajson library.

Performance

Serialization and deserialization performance of orjson is better than ultrajson, rapidjson, simplejson, or json. The benchmarks are done on fixtures of real data:

• twitter.json, 631.5KiB, results of a search on Twitter for “一”, containing CJK strings, dictionaries of strings and arrays of dictionaries, indented.

• github.json, 55.8KiB, a GitHub activity feed, containing dictionaries of strings and arrays of dictionaries, not indented.

• citm_catalog.json, 1.7MiB, concert data, containing nested dictionaries of strings and arrays of integers, indented.

• canada.json, 2.2MiB, coordinates of the Canadian border in GeoJSON format, containing floats and arrays, indented.

Latency

twitter.json serialization

twitter.json deserialization

github.json serialization

github.json deserialization

citm_catalog.json serialization

citm_catalog.json deserialization

canada.json serialization

canada.json deserialization

If a row is blank, the library did not serialize and deserialize the fixture without modifying it, e.g., returning different values for floating point numbers.

Memory

orjson’s memory usage when deserializing is similar to or lower than the standard library and other third-party libraries.

This measures, in the first column, RSS after importing a library and reading the fixture, and in the second column, increases in RSS after repeatedly calling loads() on the fixture.

twitter.json

github.json

citm_catalog.json

canada.json

Reproducing

The above was measured using Python 3.8.1 on Linux with orjson 2.2.1, ujson 1.35, python-rapidson 0.9.1, and simplejson 3.17.0.

The latency results can be reproduced using the pybench and graph scripts. The memory results can be reproduced using the pymem script.

License

orjson was written by ijl «ijl@mailbox.org», copyright 2018 - 2020, licensed under either the Apache 2 or MIT licenses.