isinstance Does not Narrow Generic Parameters

[mheripsos]

Directed to new issue from microsoft/pylance-release#2128

Environment data

• Pylance version: v2024.9.103
• OS and version: Windows 11 Pro 23H2 build 22631.4169
• Python version (& distribution if applicable, e.g. Anaconda): 3.12.4

Code Snippet

class Parent[T, U]:
    def __init__(self, value: T, other: U) -> None:
        self.value = value
        self.other = other

class Sibling[T, U](Parent[T, U]):
    ...

class Child[T](Parent[T, T]):
    def meth[U](self, other: Parent[T, U]) -> None:
        print(other.value) # other: Parent[T@Child, U@meth]; .value: T@Child
        if isinstance(other, Child):
            print(other.value) # other: Child[Unknown]; .value: Unknown
        elif isinstance(other, Sibling):
            print(other.value) # other: Sibling[T@Child, U@meth]; .value: T@Child
        else:
            print(other.other)

Repro Steps

1. Just let Pylance parse the above code

Expected behavior

in the if isinstance(other, Child) the type of other should be narrowed to Child[T] / Parent[T, T] and U should be infered to T.

From my understanding, unless there is some slip in my logic I'm not catching, this is the only valid inference that matches both other: Parent[T, U] and other: Child[Unknown]

Actual behavior

Inferences left inline above in the comments, I wrote out Pylance's inferences.

Particularly, again in that if block, other: Child[Unknown] instead of other: Child[T]

Logs

The "Python" log looks uninteresting, and the "Python Language Server" log blew up with a ton of stuff. Not sure what I would be looking for in here, but I think it's a bad idea to just dump it in here.

[rchiodo]

Typechecking issue, transferring to Pyright.

[mheripsos]

I Initially marked this as "Bug" because the inference is correct for Sibling class but not for Child. That distinction appears to be a bug, but I guess it could be considered an enhancement as well if the Sibling case works only incidentally.

[erictraut]

Pyright does try to retain type arguments when an isinstance type guard is used. You can see that it does so in the Sibling case, although it doesn't in the Child case. By comparison, mypy (another popular Python type checker) doesn't preserve type arguments in either case.

class Parent[T, U]:
    def __init__(self, value: T, other: U) -> None: ...

class Sibling[T, U](Parent[T, U]): ...

class Child[T](Parent[T, T]):
    def meth[U](self, other: Parent[T, U]) -> None:
        if isinstance(other, Sibling):
            reveal_type(other) # Pyright reveals Sibling[T, U], mypy reveals Sibling[Any, Any]

        if isinstance(other, Child):
            reveal_type(other) # Pyright reveals Child[Unknown], mypy reveals Child[Any]

The challenge with the Child isinstance check is that the un-narrowed type of other is Parent[T, U]. The question is whether a type checker can identify a valid type X such that Child[X] is a subtype of Parent[T, U] given that Child[X] derives from Parent[X, X]. This is tricky because of variance rules. In your example, there is a valid solution for X: T | U, so pyright could safely narrow its type to Child[T | U]. Pyright's algorithm is not currently sophisticated enough to handle this case, so it falls back to Unknown.

I've enhanced the algorithm so it can handle this condition. This will be included in the next release.