Do not count __slots__ as a protocol member

[sobolevn]

Refs #9314
Closes#11884

[AlexWaygood]

It's not so clear-cut, but would it be worth special-casing __class_getitem__ in the same way? It might avoid regressions like this (cc. @hauntsaninja)

[sobolevn]

It's not so clear-cut, but would it be worth special-casing class_getitem in the same way?

Can you please open a new issue for it? Probably we need to discuss / research all special cases separatelly. But, it looks like a good candidate 👍

[AlexWaygood]

It's not so clear-cut, but would it be worth special-casing class_getitem in the same way?

Can you please open a new issue for it? Probably we need to discuss / research all special cases separatelly. But, it looks like a good candidate 👍

Sure!

[AlexWaygood]

New issue filed here: #11886

[A5rocks]

Looks good to me, though #11886 might be easily combined with this if __class_getitem__ is to be ignored (though keeping seperate PRs is admirable and perfectly fine too).

[sobolevn]

Now I am also thinking: what if I want to have a function / interface / contract / whatever that only works with slotted types? How can I express this?

For example:

defpretty_print_slots(klass_with_slots: ???) ->None:
...

I guess what Eric said in #11886 really makes sense. So, proably we can ignore protocols __slots__ = () definition as a "performance hack", we need to treat __slots__: str | Iterable[str] as a part of the protocol.

Do you agree?

[A5rocks]

That sounds like a decent solution too, tried it out a bit and found some weird behavior in current mypy somewhat related to this, just bringing this up because it's a bit funny:

This fails:

importtyping@typing.runtime_checkableclassFoo(typing.Protocol):
__slots__= ()
classBar(Foo):
...
print(issubclass(Bar, Foo)) # this fails

and this works:

importtyping@typing.runtime_checkableclassFoo(typing.Protocol):
__slots__: str|typing.Iterable[str] = ()
classBar(Foo):
...
print(issubclass(Bar, Foo)) # this does not ?

(Anyways, that'll get fixed by this too, it's just some weird behavior...)

Basically yes, I agree that would make sense to me -- though I'm sure any solution would be nice.

[sobolevn]

We are blocked by #11891

[AlexWaygood]

We are blocked by #11891

I wonder if it's even helpful for typeshed to define __slots__ in the stub for object (as it does currently), or if it should be removed. Given that __slots__ are so special, it doesn't seem like it does much for the type-checker.

[sobolevn]

@AlexWaygood well, technically, object does not have __slots__.

>>> object.__slots__
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
AttributeError: type object 'object' has no attribute '__slots__'. Did you mean: '__class__'?
>>> object().__slots__
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
AttributeError: 'object' object has no attribute '__slots__'. Did you mean: '__class__'?

My vote is to remove them (my opinion is pretty simple: typeshed should be as close to source as possible).

[github-actions]

Diff from mypy_primer, showing the effect of this PR on open source code:

discord.py (https://github.com/Rapptz/discord.py)
- discord/ext/commands/core.py:2333: error: Argument 1 to "is_owner" of "BotBase" has incompatible type "Union[discord.user.User, Member]"; expected "discord.abc.User"

[HexDecimal]

What's the status of this?

Instead of hard-coding __slots__, shouldn't this check a collection of what to exclude so that what's excluded can be easily updated? From the discussion on #11886, apparently Python uses this collection to decide what to exclude from Protocol. Perhaps follow exactly what Python excludes rather than debating which ones should be special-cased in Mypy? I assume if this route is taken then Mypy would need a collection for each Python version with differing exclusions.

Maybe I should try making a separate PR with my suggestions?

[sobolevn]

@HexDecimal I've deleted my fork by accident. You can re-submit this if you need this feature :)

[HexDecimal]

I'll go ahead and attempt my own PR. I mostly want to fetch your test to work with.

[sobolevn]

Great, ping me to review your PR :)