Add support for attrs v21.3.0+

[jacobbogdanov]

Description

Since version 21.3.0
you can now import attrs instead of just import attr.

This patch adds support so that astroid doesn't barf on classes created using @attrs.define.

NOTE: Not not all names, most notably attr.ib and attr.s are available in attrs.

>>> import attrs
>>> [name for name in dir(attrs) if not name.startswith("_")]
['Attribute', 'Factory', 'NOTHING', 'asdict', 'assoc', 'astuple', 'cmp_using', 'converters', 'defi
ne', 'evolve', 'exceptions', 'field', 'fields', 'fields_dict', 'filters', 'frozen', 'has', 'make_c
lass', 'mutable', 'resolve_types', 'setters', 'validate', 'validators']

The only decorators in that list are @frozen, @mutable and @define. The rest are classes or functions for interacting with attrs classes.

I added support for @define because the first example from the release notes has it,

from attrs import define

@define
class HelloWorld:
    modern: bool = True

but was unsure about whether @mutable and @frozen should be done too.

Type of Changes

	Type
✓	🐛 Bug fix

Related Issue

Closes #1330

[DanielNoord]

I wonder if we should add this.

Sadly this brain does not infer the actual function behind the decorator but only looks at the name. However, if anybody uses define as a decorator name in non-attr code they will now also get picked up by this brain. Fort some reason we allowed attrib and field here, see #1187 (review).

Do we want to add define as well @Pierre-Sassoulas? I can't imagine it being a very often used name for decorators, but it might create troubles down the line..

[Pierre-Sassoulas]

Well I don't like assuming that something is part of a lib only because of its name especially since we have inference capability in astroid. I know we're doing it for np in some place. But inference would permit to infer the real name of the origin lib. I don't know enough about brains to suggest something proper instead. What do you think @hippo91 ?

[hippo91]

@Pierre-Sassoulas i agree with you and @DanielNoord as well.
IMHO define should not be part of this set.

[jacobbogdanov]

okay, I've removed @define from the list of supported decorator names.

[DanielNoord]

Thanks for making the PR as well 😄

Just one question about what we should actually add to the list.

[Pierre-Sassoulas]

If you want black to format properly small line, you need to remove the ending comma (it's black magic comma)

[jacobbogdanov]

If you want black to format properly small line, you need to remove the ending comma (it's black magic comma)

Thanks for the pointer! I think the issue is that I have an older black installed on my system. I omitted the formatting changes in my first commit but on the second I forgot to only stage the relevant portions

[DanielNoord]

Thanks @jacobbogdanov!

One recommendation for a little test docstring but rest LGTM!

[DanielNoord]

I think we could discuss if attrib and field should be in this list but that ship has sailed I think.

Thanks for removing define! 👍

[Pierre-Sassoulas]

I think you're right, it's also a problem.

Suggested change

	("attr.ib", "attrib", "attr.attrib", "attr.field", "attrs.field", "field")
	("attr.ib", "attr.attrib", "attr.field", "attrs.field")

Suggestion done in order to check if it's causing test fails. Imo we can most probably not infer the value as aliasing of attrs must be rare considering it's a small name, but we also should infer.

[jacobbogdanov]

looks like it breaks the existing test_attr_transform as well as the newly added test_attrs_transform.

        @attrs
        class Bah:
            d = attrib(attr.Factory(dict))

Also there's a regression test in pylint that also relies on the existing attrib; https://github.com/PyCQA/pylint/blob/main/tests/functional/r/regression/regression_4439.py

[DanielNoord]

@Pierre-Sassoulas Let's revert then. Thanks @jacobbogdanov for checking this out.

[Pierre-Sassoulas]

Thanks a lot for the fix @jacobbogdanov ! It will be available in astroid 2.10 !

[DanielNoord]

We forgot to add a changelog entry 😅