document ternary/conditional type inference limitations

[petergaultney]

As per the discussion in #4134 , it seems worth documenting that conditional expressions don't always work smoothly with type inference. This PR makes sure to include the word ternary and conditional so that page/doc searches will turn up an explanation for this behavior in case anyone is searching for it.

[ilevkivskyi]

Thanks! I have few suggestions.

[ilevkivskyi]

The example with AnyStr is not the best one, since it combines two tricky points:

• the one you are explaining
• the fact that functions generic in variables with restrictions are type checked multiple times.

I would rather use a single argument function with a Union[str, bytes] argument.

[petergaultney]

Actually, if I try to simplify this down to

def f(x: Union[str, bytes]) -> bytes:
    r = x.encode('ascii') if isinstance(x, str) else x
    # results in 'error: "bytes" has no attribute "encode"; maybe "decode"?'

    if isinstance(x, str):
        r = x.encode('ascii')  # whereas Mypy can narrow within an if statement
    else:
        r = x
    return r

Mypy doesn't error at all, which is quite interesting. Whereas if I use AnyStr on that same argument, there is an error.

It may be that I have oversimplified the explanation for what is going on?

[ilevkivskyi]

To clarify, mypy can restrict types using isinstance() in ternary expressions, but it cannot consider sub-expressions of ternary expression unreachable. Here is even simpler example:

x: bytes

if isinstance(x, str):
    r = x.encode()  # OK, this is unreachable
else:
    r = x

r = x.encode() if isinstance(x, str) else x  # Error here

[JukkaL]

Closing since there has been no activity in a long time. Feel free to repopen if you want to continue working on this.