Tuple length checking doesn't work with `tuple()` constructor

[xavdid]

I searched for other, similar issues, but none seemed to cover this exact case.

Environment data

• Language Server version: 2022.2.1
• OS and version: darwin arm64
• Python version (and distribution if applicable, e.g. Anaconda): 3.10.0
• python.analysis.indexing: null
• python.analysis.typeCheckingMode: basic

Expected behaviour

An error should be raised if tuple() is assigned to a tuple type of defined length

Actual behaviour

No error is raised

Code Snippet / Additional information

from typing import Tuple
Point3D = Tuple[int, int, int]

t1: Point3D = 1, 2, 3  # correct no error
t2: Point3D = 1, 2, 3, 4  # correct error; Element size mismatch; expected 3 but received 4
t3: Point3D = () # correct error; Element size mismatch; expected 3 but received 0
t4: Point3D = tuple() # expected error, none thrown

Given that tuple() returns an empty tuple, I'd expect t3 and t4 to behave identically. But, only one throws an error.

[erictraut]

This is expected behavior, although I admit that it's less than ideal.

The type returned by the constructor call tuple() is dictated by the typeshed stub file builtins.pyi. Here's the specific line that declares the __new__ method. You'll see that the parameter __iterable has a default value. If you call the constructor and don't provide an argument for this parameter, the type variable _T_co is not populated, so the resulting type is tuple[_T_co@tuple], which is effectively tuple[Any].

A potential fix is for the typeshed stub to provide an overload that specifies a return type of tuple[()] if the __iterable argument isn't provided. You're welcome to file a bug (or a PR) in the typeshed repo. Pylance regularly picks up new versions of the typeshed stubs.

In the meantime, I recommend using the () form rather than tuple(). Not only is it more terse, but it also avoids this problem from a type checking standpoint.

[xavdid]

Fair enough, thank you!

[xavdid]

I'm happy to file a PR. While I was playing with it, I noticed that mypy does handle this as expected. t4 above throws main.py:7: error: Incompatible types in assignment (expression has type "Tuple[<nothing>, ...]", variable has type "Tuple[int, int, int]").

I think the typeshed PR makes sense in this case, but I wanted to point that out in case it was useful.

[erictraut]

Yes, I noticed that mypy handles this case. It probably overrides the type stub in this particular case. I'm reluctant to ignore / override the information in the type stub because that often leads to other problems as the type stubs evolve or improve. It's preferable to solve this problem in the stubs.