Invariance not enforced for class-scoped TypeVar used in base class expression

[erictraut]

Mypy does not detect the case where a covariant or contravariant class (i.e. a class that is parameterized by a class-scoped covariant or contravariant TypeVar) derives from an invariant class. This creates a type correctness hole as shown in the following code sample.

from typing import TypeVar

T = TypeVar("T", covariant=True)

class MyList(list[T]): pass

def append_float(y: MyList[float]):
 y.append(1.0)

x: MyList[int] = MyList([1, 2, 3])
append_float(x)

Pyright likewise did not previously catch this case. I just added the check here.

[hauntsaninja]

It's gotten late for me, so I'm now confused by variance. @erictraut I noticed you didn't implement the check for whether a covariant class derives from a contravariant class, and vice versa? I think we should do that as well, because it breaks transitivity of subtyping.

from typing import TypeVar, Generic
T = TypeVar('T')
T_co = TypeVar('T_co', covariant=True)
T_contra = TypeVar('T_contra', contravariant=True)

class A: ...
class B(A): ...

class Parent(Generic[T_co]): ...
class Child(Parent[T_contra], Generic[T_contra]): ...

def takes_parent_a(f: Parent[A]): ...
def takes_parent_b(f: Parent[B]): ...
def takes_child_a(f: Child[A]): ...
def takes_child_b(f: Child[B]): ...

def give_child_a(t: Child[A]):
 takes_parent_a(t)
 takes_parent_b(t)
 takes_child_a(t)
 takes_child_b(t)

def give_child_b(t: Child[B]):
 takes_parent_a(t)
 takes_parent_b(t)
 takes_child_a(t)
 takes_child_b(t)

I think both mypy and pyright agree that Child[A] is a Child[B] and Child[B] is a Parent[B], but that Child[A] is not a Parent[B].

It looks like typeshed breaks this with Container and SupportsGetItem (in the following example pyright gives something different because pyright narrows as per #2008 (comment) , but you could trick it into showing the same behaviour)...

from _typeshed import SupportsGetItem
from typing import Container

class A: ...
class B(A): ...

sgia: SupportsGetItem[A, None]
sgib: SupportsGetItem[B, None] = sgia
conb: Container[B] = sgib # no error
conb = sgia # error

cc @JelleZijlstra

[ilevkivskyi]

Isn't this a duplicate of a (super-old) #736?

[erictraut]

@hauntsaninja, yes, I agree that mismatched covariance/contravariance mismatch is also a problem. I think it's unlikely that anyone would attempt to use those combinations, but for completeness that check could be added. [Update: I modified pyright to include this check as well.] [Update 2: After claiming that this would be unlikely to happen, I see that typeshed violates this in one place — in the SupportsGetItem protocol class. Looks like you also discovered this when you were making your change to mypy.]

@ilevkivskyi, yes, this looks like a duplicate. Apologies, I didn't notice the existing bug. I'll close this one accordingly.