Description
Feature or enhancement
It would be nice if certain abstract methods / dunder methods would be defined with positional-only arguments, so that when one subclasses them an appropriate name can be chosen, or they can be defined positional-only in the subclass.1
Code sample in pyright playground1
from typing import overload
from collections.abc import Sequence
class MySequence[T](Sequence[T]):
def __len__(self): return 0
@overload
def __getitem__(self, idx: int, /) -> T: ...
@overload
def __getitem__(self, idx: slice, /) -> Sequence[T]: ...
def __getitem__(self, idx: int | slice, /) -> T | Sequence[T]: # ❌ override
return MySequence[T]()
In some cases, the stubs already fake that the argument is positional-only, such as for Container.__contains__
:
def __contains__(self, x: object, /) -> bool: ...
intypeshed/stdlib/typing.pyi
def __contains__(self, x):
incpython/Lib/_collections_abc.py
Personally, I think only changing the stubs would be a pragmatic choice, but I was told by the maintainers that they believe this change should be done in Cpython first.
Backward Compatibility
-
Changing non-abstract methods can potentially break user code. For example, this is currently legal code at runtime
from collections.abc import Container. Container.__subclasshook__(C=list) # True
-
Changing the signature of abstract methods can introduce type-checker errors Code sample in pyright playground1
from collections.abc import Sequence def get(x: Sequence[float], index: int) -> float: return x.__getitem__(index=index)
With this in mind, I believe changing abstrac dunder methods to use positional only arguments should be relatively safe, as calling them directly with keyword is very uncommon, and at worst it will introduce type errors.
Personally, I'd also like so see methods like Container.__subclasshook__
use positional-only arguments, but I also understand if this is not feasible due to bc-conerns.