jsonstar
extends Python's standard JSON encoder and decoder to easily handle your custom types.
This means you won't have to transform your custom types into dictionaries with primitive types before encoding them to JSON. And you won't have to parse back the encoded strings into your custom types after decoding them from JSON.
pip install jsonstar
The jsonstar
module provides the same API as the standard json
module, so you can use it as a drop in replacement.
Simply change your import from import json
to import jsonstar as json
and you're good to go.
Consider you have a pydantic Employee
class that you want to serialize to JSON.
from decimal import Decimal
from datetime import date
from pydantic import BaseModel
class Employee(BaseModel):
name: str
salary: Decimal
birthday: date
roles: set
employee = Employee(
name="John Doe",
salary=Decimal("1000.00"),
birthday=date(1990, 1, 1),
roles={"A", "B", "C"},
)
The standard json
module can't serialize the employee
instance, requiring you to call its dict
method.
This will not suffice, because the standard json
module don't know how to encode Decimal
, date
and set
.
Your solution would include some transfomation of the employee
instance and its attributes before encoding it to JSON.
That is where jsonstar
shines by providing default encoder for common types like pydantic.BaseModel
,
decimal.Decimal
, datetime.date
and set
. And allowing you to easily add your own encoders.
from jsonstar as json
print(json.dumps(employee))
# {"name": "John Doe", "salary": "1000.00", "birthday": "1990-01-01", "roles": ["A", "B", "C"]}
By default, jsonstar
provides encoders for the following types:
attrs
classesdataclasses.dataclass
classesdatetime.date
datetime.datetime
datetime.time
datetime.timedelta
decimal.Decimal
frozenset
pydantic.BaseModel
set
uuid.UUID
Yes. If you think that a default encoder for a common type is missing, please open an issue or a pull request. See the How to contribute section for more details.
First you need to decide where you want your encoder to be available:
- Class default encoders happen when your
MyEncoder
class inherits fromJSONEncoderStar
and you add encoders to it. - Library-wide default encoder are added directly to
JSONEncoderStar
class and is available everywhere in your project.
Also you have two types of encoders to choose from:
- Typed encoders are used to encode a specific type identified by
isinstance
. - Functional encoders are used to encode an object based on arbitraty logic.
Note: From experience we find that class encoders are the most common use case.
import jsonstar as json
from decimal import Decimal
from datetime import date
# You can declare it on the special class attributes
class MyEncoder(json.JSONEncoderStar):
_default_typed_encoders = {Decimal: lambda o: str(o.quantize(Decimal("1.00")))}
# Or you can register it after the class is declared
MyEncoder.register_default_encoder(lambda o: o.strftime("%Y-%m-%d"), date)
import jsonstar as json
from decimal import Decimal
def two_decimals_encoder(obj):
"""Encodes a decimal with only two decimal places."""
return str(obj.quantize(Decimal("1.00")))
json.register_default_encoder(Decimal, two_decimals_encoder)
Typed encoders are specific to a type and it's inherited types.
When registering a typed encoder, you simply pass the encoder and the type to the chosen registration method.
When you add a typed encoder, jsonstar
will check if any base class already has a registered encoder make sure the more generic encoder is used last, respecting Python's Method Resolution Order (MRO).
Functional encoders are used to encode an object based on arbitraty logic and not specific to a type.
To register a functional encoder, you simply pass the encoder to the chosen registration method omiting the type.
All functional encoders are called only for objects that do not have a registered typed encoder.
Pull requests are welcome and must have associated tests.
For major changes, please open an issue first to discuss what you would like to change.
Henrique Bastos henrique@bastos.net