Unsoundness in subclassing Mappings with TypedDicts

[mrkmndz]

In the course of implementing TypedDicts into Pyre, I came across an issue with how mypy is treating the relationship between TypedDicts and Mappings.
It seems like the implemented semantics are to treat a T1 = TypedDict[ {k_1: v_1, k_2: v:2, ... k_n: v:n} ] as a subclass of T2 = Mapping[str, Union[v_1, v_2, ... v_n]]. However, this causes a problem when combined with the structural subtyping rule that T3 = TypedDict[ {k_1: v_1, k_2: v:2, ... k_n: v:n, k': v'} ] is a subclass of T1, which implies that T3 is a subclass of T2. When you call __getitem__ on a T2 that's actually a T3 with key k', you would unexpectedly get a v'. This can lead to an undetected potential AttributeError as in the following example:

from typing import Mapping
from mypy_extensions import TypedDict

class Y():
    def ya(self)-> int:
        return 7

class Z():
    def za(self) -> int:
        return 8

Foo = TypedDict('Foo', {'visible': Y, 'hidden': Z})

Bar = TypedDict('Bar', {'visible': Y})

def access(x:Mapping[str, Y]) -> None:
    if "hidden" in x:
        x["hidden"].ya()

def hide(x:Bar) -> None:
    access(x)

def f(x: Foo) -> None:
    hide(x)

td: Foo = {'visible': Y(), 'hidden': Z()}
f(td)

This passes mypy typechecking but when run gives an attribute error.

I think we're going to end up having TypedDicts just subclass Mapping[str, Any] for this reason.

[JukkaL]

Yeah, we are aware of the problem -- and I actually started working on a fix earlier today :-). I'm trying out what happens if make all typed dicts have Mapping[str, object] as a supertype. This would be safer, but it may require users to add some extra casts.