Set operators confusion

[kaste]

Hi 👋 ! I have the following test code:

from collections import defaultdict
from typing import Any, DefaultDict, Dict, Iterable, Set

def xs(xs: Iterable[str]) -> None:
    ...

d: Dict[str, str] = {}
xs(d.keys() - {'2'}) # 1
xs(d.keys() - {2}) # 2
xs(d.keys() - {2} - {'2'}) # 3
xs(d.keys() - {'2'} - {2}) # 4

 10:15  error  mypy  Argument 1 to <set> has incompatible type "int"; expected "str"
 11:1   error  mypy  Argument 1 to "xs" has incompatible type "AbstractSet[Union[str, int]]"; expected
                     "Iterable[str]"

Discussion:

1 is obviously OK; 2 is obviously okay bc a type checker should report here.
3 is not ok although a type checker must report something. The result of the expression is not Union[str, int} and the call to xs is okay. mypy should report the same as in 2 imo
4 surprisingly passes although it's just 3 flipped which is not ok. A type checker should report here with the same error as 2.

Further: If I change the def of xs to

def xs(xs: Set[str]) -> None:
    ...

all 4 calls fail because mypy wants me to use AbstractSet instead of just Set here. This is very unintuitive bc this is just a concrete function over sets, in a code review I would not even know if Set implements something that AbstractSet does not have etc.

[ilevkivskyi]

all 4 calls fail because mypy wants me to use AbstractSet instead of just Set here. This is very unintuitive bc this is just a concrete function over sets, in a code review I would not even know if Set implements something that AbstractSet does not have etc.

What do you mean by "unintuitive"? Accepting KeysView where a Set is expected would be just wrong. Mypy has couple similar exceptions, but only in widely used cases where danger is minimal: for example int is considered a subtype of float although it is not the case at runtime.

[ilevkivskyi]

Anyway, I agree that example 4 should give an error, although I think this is a typeshed issue (there are some Anys used in AbstractSet methods). I would propose to continue a discussion on that tracker, please open an issue there.