os.open should not have AnyStr in its type

[gnprice]

See python/mypy#1943 -- AnyStr shouldn't be used when it only appears once in the type, because it complicates the semantics without adding anything.

[matthiaskramm]

Should type checkers just detect a single usage and optimize for it? AnyStr is a pretty convenient shortcut.

[gnprice]

I'm not convinced that would actually be correct, under the semantics specified in PEP 484.

The semantics of AnyStr is that it's a type variable which has to specifically take on (in Python 3) either the value str or the value bytes. So in code like this:

AnyStr = TypeVar('AnyStr', constraints=[str, bytes])

def f(x: AnyStr) -> None: ...

x = 'a'  # type: Union[str, bytes]

f(x)  # error?

We're trying to use a value of type Union[str, bytes] in a generic function which requires an AnyStr. What value should the type variable AnyStr take there to make that well-typed?

The type Union[str, bytes] isn't equal to either str or bytes, and in fact it isn't usable as either a str or a bytes. So neither possible way to instantiate the type variable makes this code well-typed.

Does that match your interpretation of the semantics? How would you interpret PEP 484 in the context of this code that makes it well-typed?

[matthiaskramm]

I had assumed that TypeVar("T", X, Y) is equivalent to TypeVar("T", Union[X, Y]).

With that, your code sample passes checking, based on Union[str, bytes] <: Union[str, bytes]?

[gvanrossum]

No, TypeVar('T', X, Y) cannot be reduced to TypeVar('T', Union[X, Y]). The latter is actually an error according to PEP 484 (and mypy) -- it requires at least two constraint values if there are any. A single constraint would be indistinguishable from the bare constraint type.

(Constrained type variables are tricky that way. It took me a long time to grok them. And I still haven't found an example besides AnyStr. But that's indispensible -- just not for os.open(). :-)

I'm still not sure how to express what I'd like to see, and maybe it's not worth it. We can't exactly say that Union[str, bytes] <: AnyStr. It's only the case that if there's exactly one parameter of type exactly AnyStr, and no other use of AnyStr in the signature, then Union[str, bytes] should be acceptable. But that seems excessively specific. In the meantime we should just change typeshed.