Fix asyncio.gather regression

[AlexWaygood]

Adding the empty-tuple overload caused major problems for pyright, and that overload only deals with an unlikely edge case anyway. Get rid of it, and replace the fallback overload with a more general overload.

Fixes #8270.

[hauntsaninja]

Worth a regression test?

[AlexWaygood]

Worth a regression test?

Good job you suggested that...

[AlexWaygood]

(The reason why this PR is not a simple reversion of 0740d1c is that the added tests do not pass with a simple revert PR.)

[hauntsaninja]

Thanks! The test looks good, and helps make changes here much more sane.

That said, it might be nice to get rid of the type ignores. What goes wrong with adding the empty overload at the bottom?

    ...
    @overload
     def gather(
         __coro_or_future1: _FutureLike[Any],
         __coro_or_future2: _FutureLike[Any],
         __coro_or_future3: _FutureLike[Any],
         __coro_or_future4: _FutureLike[Any],
         __coro_or_future5: _FutureLike[Any],
         __coro_or_future6: _FutureLike[Any],
         *coros_or_futures: _FutureLike[Any],
         return_exceptions: bool = ...,
     ) -> Future[list[Any]]: ...
     @overload
     def gather(*, return_exceptions: bool = ...) -> Future[tuple[()]]: ...

Edit: minor, but maybe the clearest thing to do would be to move the big overload to the top, since that matches what type checkers are doing when they type check gather(*[...])

[github-actions]

According to mypy_primer, this change has no effect on the checked open source code. 🤖🎉

[AlexWaygood]

What goes wrong with adding the empty overload at the bottom?

That's what I tried at first: https://github.com/python/typeshed/pull/8271/files/a345b43e68b56359e4decead1a261e283f8081a4.

But the test cases failed: https://github.com/python/typeshed/runs/7299574031?check_suite_focus=true

It seems like pyright picks the empty overload for asyncio.gather(*[...]) even if that overload is the last overload!

minor, but maybe the clearest thing to do would be to move the big overload to the top, since that matches what type checkers are doing when they type check gather(*[...])

Meh, not sure I agree that that would improve readability :-)

[hauntsaninja]

Hmm, not sure I understand what pyright is doing then, so you may want someone else to review this. Test case looks good though!

[AlexWaygood]

Hmm, not sure I understand what pyright is doing then, so you may want someone else to review this. Test case looks good though!

I think I understand what pyright's doing (though it did surprise me at first, for sure), and the test cases make me confident that this fixes the regression, so I'm inclined to merge. If somebody figures a way to get all the test cases to pass without the type ignores, they can always file another PR in the future. Does that sound okay?

[hauntsaninja]

Wait, why is pyright doing what it's doing? @erictraut also mentioned moving the empty overload to the bottom as something that's likely expected to work in microsoft/pyright#3685 (comment)

Yeah, fine to merge, I was just clarifying what my "approve" of the PR promised :-)