Python 3.10 compatibility

[frenzymadness]

Hello.

We have already started testing RPM packages in Fedora with Python 3.10a2 and attrs is currently not compatible.

The problem is caused by the implementation of PEP 563 — type annotations are now stored as strings which cause asserts like this one to fail:

class TestAnnotations:
    """
    Tests for types derived from variable annotations (PEP-526).
    """

    def test_basic_annotations(self):
        """
        Sets the `Attribute.type` attr from basic type annotations.
        """

        @attr.s
        class C:
            x: int = attr.ib()
            y = attr.ib(type=str)
            z = attr.ib()

        assert int is attr.fields(C).x.type

The value stored in attr.fields(C).x.type is a string "int" instead of a class int.

It seems it was discussed before in #424

[Tinche]

Do we need to call attr.resolve_types first in the test?

[frenzymadness]

Do I understand it correctly that we are waiting for #593 here?

[hynek]

Yes. :(

[frenzymadness]

FTR, when issue 593 is fixed now, I can use attr.resolve_types(C) in the test_basic_annotations and that makes the previously mentioned assert int is attr.fields(C).x.type pass but the last assert in the test:

assert C.__init__.__annotations__ == {
    "x": int,
    "y": str,
    "return": None,
}

still complains about {'x': 'int'} != {'x': <class 'int'>}. I'm not sure how much you want to make the attrs behave the same under Python 3.10. If that's the goal, we should probably keep tests untouched.

[hynek]

The test needs to be changed to get_type_hints I'm afraid. There's a bunch more stuff breaking that I'm working on rn.

[frenzymadness]

Great movement. Can we expect a new release soon?

[hynek]

We/I need to write some narrative docs about new features and this issue uncovered that our hooks system is very hard to use with Python 3.10/string types, so we'll have to write some kind of attr.resolve_type() that can convert "str" to str.