Annotation Inheritance

Let's talk about annotations.

Type annotations in Python are mostly a static declaration to a type-checker like mypy or pyright about the expected types. However, they are also a dynamic data structure which a growing number of libraries such as the original attrs and dataclasses in the standard library, and even sqlalchemy use at runtime.

>>> from dataclasses import dataclass
>>>
>>> @dataclass
... class C:
...    a: int
...    b: str
...
>>> C(1, "a")
C(a=1, b='a')

These libraries inspect the annotations of a class to generate __init__ and __eq__, saving a lot of boilerplate code. You could call this type of API named tuple without the tuple. (To get meta, the typing module has added dataclass_transform which libraries can use to properly annotate new class decorators with this API.)

These libraries support inheritance of fields.

>>> @dataclass
... class D(C):
...    e: int
...
>>> D(1, "a", 2)
D(a=1, b='a', e=2)

Type checkers also consider class annotations to be inherited. For example, mypy considers this to be correct:

class A:
    a: int

class B(A): pass

B().a

That code fails at runtime, because nothing is actually setting a on the B instance. But, what if B was a dataclass?

>>> class A:
...    a: int
...
>>> @dataclass
... class B:
...    pass
...
>>> B(1)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: __init__() takes 1 positional argument but 2 were given

It doesn't work, because annotations are not inherited.

>>> A.__annotations__
{'a': <class 'int'>}
>>> B.__annotations__
{}

It's up to the library to look up it's inheritance tree and decide to include the annotations of parents or not when generating code. As it happens, dataclasses has made the design decision to only inherit annotations from other dataclasses.

As an aside, class variables which are used to represent default values are inherited.

>>> class A:
...    a = 1
...
>>> @dataclass
... class B(A):
...    a: int
...
>>> B()
B(a=1)

We can write another decorator which grabs annotations from parents and adds them in method resolution order, as if they were inherited.

def inherit_annotations(cls):
    annotations = {}
    for parent in cls.__mro__[::-1]:
    	# reverse order so children override parents
        annotations.update(getattr(parent, "__annotations__", {}))
        # use getattr(): not everything has __annotations__
    cls.__annotations__.update(annotations)
    return cls

Since all dataclasses sees is the __annotations__ dict at runtime, any modifications made before the class decorator runs will be reflected in the generated fields.

>>> @dataclass
... @inherit_annotations
... class B(A): pass
...
>>> B(1)
B(a=1)

Here's a robustified version of the function.

I know what you're thinking though: why not just use multiple class decorators? Sure, all but one of the generated __init__s will be overwritten, but that's fine because they all have the same behavior anyway.

import attr
from dataclasses import dataclass


@dataclass
@attr.define
class DualCitizen:
    a: int

@dataclass
class Dataclassified(DualCitizen):
    pass

@attr.define
class Attrodofined(DualCitizen):
    pass

Looks like perfectly normal class definitions.

>>> DualCitizen(1)
DualCitizen(a=1)
>>> Dataclassified(1)
Dataclassified(a=1)
>>> Attrodofined(1)
Attrodofined(1)

And it works.

So, type-checkers consider annotations to be inherited, but class decorators which use annotations at runtime only inherit annotations from ancestors with the same decorator. We can work around this either by multiply decorating the ancestors, or by pulling annotations from ancestors into __annotations__.