Extending

Each attrs-decorated class has a __attrs_attrs__ class attribute. It’s a tuple of attrs.Attribute carrying metadata about each attribute.

So it is fairly simple to build your own decorators on top of attrs:

>>> from attr import define
>>> def print_attrs(cls):
...     print(cls.__attrs_attrs__)
...     return cls
>>> @print_attrs
... @define
... class C:
...     a: int
(Attribute(name='a', default=NOTHING, validator=None, repr=True, eq=True, eq_key=None, order=True, order_key=None, hash=None, init=True, metadata=mappingproxy({}), type=<class 'int'>, converter=None, kw_only=False, inherited=False, on_setattr=None, alias='a'),)

Warning

The attrs.define() / attr.s() decorator must be applied first because it puts __attrs_attrs__ in place! That means that is has to come after your decorator because:

@a
@b
def f():
   pass

is just syntactic sugar for:

def original_f():
   pass

f = a(b(original_f))

Wrapping the Decorator

A more elegant way can be to wrap attrs altogether and build a class DSL on top of it.

An example for that is the package environ-config that uses attrs under the hood to define environment-based configurations declaratively without exposing attrs APIs at all.

Another common use case is to overwrite attrs’s defaults.

Mypy

Unfortunately, decorator wrapping currently confuses mypy’s attrs plugin. At the moment, the best workaround is to hold your nose, write a fake Mypy plugin, and mutate a bunch of global variables:

from mypy.plugin import Plugin
from mypy.plugins.attrs import (
   attr_attrib_makers,
   attr_class_makers,
   attr_dataclass_makers,
)

# These work just like `attr.dataclass`.
attr_dataclass_makers.add("my_module.method_looks_like_attr_dataclass")

# This works just like `attr.s`.
attr_class_makers.add("my_module.method_looks_like_attr_s")

# These are our `attr.ib` makers.
attr_attrib_makers.add("my_module.method_looks_like_attrib")

class MyPlugin(Plugin):
    # Our plugin does nothing but it has to exist so this file gets loaded.
    pass


def plugin(version):
    return MyPlugin

Then tell Mypy about your plugin using your project’s mypy.ini:

[mypy]
plugins=<path to file>

Warning

Please note that it is currently impossible to let mypy know that you’ve changed defaults like eq or order. You can only use this trick to tell Mypy that a class is actually an attrs class.

Pyright

Generic decorator wrapping is supported in Pyright via their dataclass_transform specification.

For a custom wrapping of the form:

def custom_define(f):
    return attr.define(f)

This is implemented via a __dataclass_transform__ type decorator in the custom extension’s .pyi of the form:

def __dataclass_transform__(
    *,
    eq_default: bool = True,
    order_default: bool = False,
    kw_only_default: bool = False,
    field_descriptors: Tuple[Union[type, Callable[..., Any]], ...] = (()),
) -> Callable[[_T], _T]: ...

@__dataclass_transform__(field_descriptors=(attr.attrib, attr.field))
def custom_define(f): ...

Warning

dataclass_transform is supported provisionally as of pyright 1.1.135.

Both the Pyright dataclass_transform specification and attrs implementation may change in future versions.

Types

attrs offers two ways of attaching type information to attributes:

This information is available to you:

>>> from attr import attrib, define, field, fields
>>> @define
... class C:
...     x: int = field()
...     y = attrib(type=str)
>>> fields(C).x.type
<class 'int'>
>>> fields(C).y.type
<class 'str'>

Currently, attrs doesn’t do anything with this information but it’s very useful if you’d like to write your own validators or serializers!

Metadata

If you’re the author of a third-party library with attrs integration, you may want to take advantage of attribute metadata.

Here are some tips for effective use of metadata:

  • Try making your metadata keys and values immutable. This keeps the entire Attribute instances immutable too.

  • To avoid metadata key collisions, consider exposing your metadata keys from your modules.:

    from mylib import MY_METADATA_KEY
    
    @define
    class C:
      x = field(metadata={MY_METADATA_KEY: 1})
    

    Metadata should be composable, so consider supporting this approach even if you decide implementing your metadata in one of the following ways.

  • Expose field wrappers for your specific metadata. This is a more graceful approach if your users don’t require metadata from other libraries.

    >>> from attrs import fields, NOTHING
    >>> MY_TYPE_METADATA = '__my_type_metadata'
    >>>
    >>> def typed(
    ...     cls, default=NOTHING, validator=None, repr=True,
    ...     eq=True, order=None, hash=None, init=True, metadata=None,
    ...     converter=None
    ... ):
    ...     metadata = metadata or {}
    ...     metadata[MY_TYPE_METADATA] = cls
    ...     return field(
    ...         default=default, validator=validator, repr=repr,
    ...         eq=eq, order=order, hash=hash, init=init,
    ...         metadata=metadata, converter=converter
    ...     )
    >>>
    >>> @define
    ... class C:
    ...     x: int = typed(int, default=1, init=False)
    >>> fields(C).x.metadata[MY_TYPE_METADATA]
    <class 'int'>
    

Automatic Field Transformation and Modification

attrs allows you to automatically modify or transform the class’ fields while the class is being created. You do this by passing a field_transformer hook to define() (and friends). Its main purpose is to automatically add converters to attributes based on their type to aid the development of API clients and other typed data loaders.

This hook must have the following signature:

your_hook(cls: type, fields: list[attrs.Attribute]) list[attrs.Attribute]
  • cls is your class right before it is being converted into an attrs class. This means it does not yet have the __attrs_attrs__ attribute.

  • fields is a list of all attrs.Attribute instances that will later be set to __attrs_attrs__. You can modify these attributes any way you want: You can add converters, change types, and even remove attributes completely or create new ones!

For example, let’s assume that you really don’t like floats:

>>> def drop_floats(cls, fields):
...     return [f for f in fields if f.type not in {float, 'float'}]
...
>>> @frozen(field_transformer=drop_floats)
... class Data:
...     a: int
...     b: float
...     c: str
...
>>> Data(42, "spam")
Data(a=42, c='spam')

A more realistic example would be to automatically convert data that you, e.g., load from JSON:

>>> from datetime import datetime
>>>
>>> def auto_convert(cls, fields):
...     results = []
...     for field in fields:
...         if field.converter is not None:
...             results.append(field)
...             continue
...         if field.type in {datetime, 'datetime'}:
...             converter = (lambda d: datetime.fromisoformat(d) if isinstance(d, str) else d)
...         else:
...             converter = None
...         results.append(field.evolve(converter=converter))
...     return results
...
>>> @frozen(field_transformer=auto_convert)
... class Data:
...     a: int
...     b: str
...     c: datetime
...
>>> from_json = {"a": 3, "b": "spam", "c": "2020-05-04T13:37:00"}
>>> Data(**from_json)  # ****
Data(a=3, b='spam', c=datetime.datetime(2020, 5, 4, 13, 37))

Or, perhaps you would prefer to generate dataclass-compatible __init__ signatures via a default field alias. Note, field_transformer operates on attrs.Attribute instances before the default private-attribute handling is applied so explicit user-provided aliases can be detected.

>>> def dataclass_names(cls, fields):
...     return [
...         field.evolve(alias=field.name)
...         if not field.alias
...         else field
...         for field in fields
...     ]
...
>>> @frozen(field_transformer=dataclass_names)
... class Data:
...     public: int
...     _private: str
...     explicit: str = field(alias="aliased_name")
...
>>> Data(public=42, _private="spam", aliased_name="yes")
Data(public=42, _private='spam', explicit='yes')

Customize Value Serialization in asdict()

attrs allows you to serialize instances of attrs classes to dicts using the attrs.asdict() function. However, the result can not always be serialized since most data types will remain as they are:

>>> import json
>>> import datetime
>>> from attrs import asdict
>>>
>>> @frozen
... class Data:
...    dt: datetime.datetime
...
>>> data = asdict(Data(datetime.datetime(2020, 5, 4, 13, 37)))
>>> data
{'dt': datetime.datetime(2020, 5, 4, 13, 37)}
>>> json.dumps(data)
Traceback (most recent call last):
   ...
TypeError: Object of type datetime is not JSON serializable

To help you with this, asdict() allows you to pass a value_serializer hook. It has the signature

your_hook(inst: type, field: attrs.Attribute, value: Any) Any
>>> from attr import asdict
>>> def serialize(inst, field, value):
...     if isinstance(value, datetime.datetime):
...         return value.isoformat()
...     return value
...
>>> data = asdict(
...     Data(datetime.datetime(2020, 5, 4, 13, 37)),
...     value_serializer=serialize,
... )
>>> data
{'dt': '2020-05-04T13:37:00'}
>>> json.dumps(data)
'{"dt": "2020-05-04T13:37:00"}'