Typing: Expose a cast variant to remove Optional

[cjerdonek]

I recently noticed that it would be useful to have a cast() variant that removes Optional (e.g. for DRY purposes). This would be useful in cases where earlier code has ruled out the None case and e.g. mypy's type inference isn't able to detect it. I believe this could also help prevent errors because as it is people can pass the wrong typ into cast() (or code can change over time, etc).

This is related to (or a special case of) issue #565. My apologies in advance if this feature already exists or if the idea has already been rejected.

[gvanrossum]

I don't think it exists. Let me see if I understand you correctly.

def is_null(arg: object) -> bool:
    # A silly function that dynamically checks for null and empty values
    return not arg

class C:
    # Some class with some method
    def meth(self) -> int: ...

def foo(arg: Optional[C]) -> int:
    if is_null(arg):
        return 0
    return arg.meth()  # error: Item "None" of "Optional[C]" has no attribute "meth

The error on the last line is a false positive since is_null(arg) checks for None.

So you would want to write something like

    arg = cast_away_optional(arg)
    reveal_type(arg)  # Should print C, not Optional[C]
    return arg.meth()  # Should not be an error

Am I right?

[cjerdonek]

Yep, that's right!

[JelleZijlstra]

It's not quite the same, but mypy already supports doing this with assertions:

$ mypy -c '''                          
from typing import Optional
def f(x: Optional[int]) -> None:
    reveal_type(x)
    assert x is not None
    reveal_type(x)
'''
<string>:4: error: Revealed type is 'Union[builtins.int, None]'
<string>:6: error: Revealed type is 'builtins.int'

Of course, unlike the proposed cast operation, this also does a runtime check.

[gvanrossum]

Likely the runtime overhead of a single is None comparison is less than that of a single no-op function call. So it sounds like the assert is the better solution. @cjerdonek Do you agree?

[cjerdonek]

The assert is good to know (I didn't know about that when I filed the issue), and I agree is good for those cases. However, there is at least one case where it seems like it might not be as good (and was one of the reasons that led me to filing this). Take this snippet from pip's code base:

for line in data.split(b'\n')[:2]:
    if line[0:1] == b'#' and ENCODING_RE.search(line):
        encoding = ENCODING_RE.search(line).groups()[0].decode('ascii')
        return data.decode(encoding)

(from https://github.com/pypa/pip/blob/53d57a2accf994dccb5aee563368fa2ac8f55fa2/src/pip/_internal/utils/encoding.py#L33-L36 )

Here the issue is that after the if check we know that the return value of the method call (ENCODING_RE.search(line)) won't be None. A cast function like I suggested would let one do the assignment inline, like so:

encoding = cast_away_optional(
  ENCODING_RE.search(line),
).groups()[0].decode('ascii')

It seems using an assert would require adding more.

[gvanrossum]

Okay, though the desire to do this in-line rather than using a separate statement makes the desire a lot weaker. It's a slippery slope; Python has inline conditionals, loops (comprehensions) and now (in 3.8) assignments, but not other things like try/except, while, del, or, indeed, assert...

[cjerdonek]

but not other things like try/except, while, del, or, indeed, assert...

Indeed.. Really, the suggestion is just to add a light-weight way to tell the type-checker that something isn't None (which doesn't automatically require using assert). Using cast() as is is kind of annoying because e.g. the caller shouldn't have to know that ENCODING_RE.search() returns Match[bytes]:

encoding = cast(
    Match[bytes],
    ENCODING_RE.search(line),
).groups()[0].decode('ascii')

And using assert requires adding two lines including an assignment (an assignment expression can't be used because pip needs to support older versions):

for line in data.split(b'\n')[:2]:
    if line[0:1] == b'#' and ENCODING_RE.search(line):
        match = ENCODING_RE.search(line)
        assert match is not None
        encoding = match.groups()[0].decode('ascii')
        return data.decode(encoding)

But if you want to close this, I won't feel bad because I had a chance to make my case. :)

[gvanrossum]

I guess you can write your own:

T = TypeVar('T')
def cast_away_optional(arg: Optional[T]) -> T:
    assert arg is not None
    return arg

(Untested.)

[cjerdonek]

Thanks. Good to know.

[dkgi]

FWIW we use something very similar internally:
https://github.com/facebook/pyre-check/blob/master/pyre_extensions/__init__.py#L7-L8