mypy fails to narrow Union[float,int] to float after isinstance(x, float)

[DarwinAwardWinner]

Mypy fails to typecheck the following valid code:

from typing import Union
def maybe_convert_to_int(x: Union[float,int]) -> Union[float,int]:
    '''Convert x from float to int if it can be done losslessly.'''
    if isinstance(x, float):
        # x must be a float here, so this is a legal operation, but
        # mypy still thinks its type is Union[float,int]
        assert isinstance(x, float)
        if x.is_integer():
            x = int(x)
    return x

When I run mypy on it, I get:

$ mypy ~/temp/test_mypy.py
temp/test_mypy.py:8: error: Item "int" of "Union[float, int]" has no attribute "is_integer"
$ python --version
Python 3.7.1
$ mypy --version
mypy 0.650

I would expect this program to pass the type checker, since I am using isinstance to narrow the type to float before performing the float-specific operation float.is_integer. If I replace the assertion with x = float(x), that satisfies the type checker.

[ilevkivskyi]

This is because some special-casing in mypy called "type promotions". The point is that int is not a subclass of float at runtime, but mypy considers int a subtype of float. Because of this, int sneaks into the if statement, where mypy complains about the missing method.

We try to ignore the type promotions while working with isinstance(), but it looks like there is a bug and type promotions are still used. Another possible workaround (apart from the one you mentioned) for you would be to just annotate the functions as (float) -> float. You can probably also change the logic to

def maybe_convert_to_int(x: float) -> float:
    if isinstance(x, int):
        return x
    if x.is_integer():
        x = int(x)
    return x

[ilevkivskyi]

@Michael0x2a was among the last who worked with this code, so probably he has the most context.

[DarwinAwardWinner]

I discovered a workaround that eliminated this code entirely for my specific use case. The reason for this code was to ensure that a float like 1.0 would print as an integer (with no decimal point), but I figured out that '{:g}'.format(x) does that without having to check the type of x.

However, I guess this is still a bug in the general case.