asdict filter exclude list matches on name instead of attribute

[henrycjc]

Hi, first of all, love the project and use it nearly every day for my job.

I have come across some behaviour that is unexpected. Please consider this minimal worked example:

import attr
from typing import List

@attr.s(auto_attribs=True)
class Cop:
    my_field: str
    repeated_field_name: str


@attr.s(auto_attribs=True)
class Robber:
    my_other_field: str
    repeated_field_name: str


@attr.s(auto_attribs=True)
class Payload:
    cops: List[Cop]
    robbers: List[Robber]


if __name__ == "__main__":
    cop = Cop("abc", "shared1")
    robber = Robber("def", "shared2")
    payload = Payload([cop], [robber])

    print(attr.asdict(payload, recurse=True))
    """
    {
        'cops': [{'my_field': 'abc', 'repeated_field_name': 'shared1'}], 
        'robbers': [{'my_other_field': 'def', 'repeated_field_name': 'shared2'}]
    }
    """
    print(
        attr.asdict(
            payload, 
            recurse=True, 
            filter=attr.filters.exclude(attr.fields(Cop).repeated_field_name)
        )
    )
    """
    {
        'cops': [{'my_field': 'abc'}], 
        'robbers': [{'my_other_field': 'def'}]
    }
    """

When the asdict filter argument is passed a variadic collection of attributes (available from attr.fields), and you have a range of attrs models that are recursively asdict-ified that may share a field name (in this example repeated_field_name) then the returned dictionary excludes this attribute for ALL classes, not just the one passed to fields. This is understandable to a degree given assert attr.fields(Cop).repeated_field_name == attr.fields(Robber).repeated_field_name holds, but still a little unexpected given the method explicitly does not accept field names as strings.

As it happens, this behaviour is what I wanted 😄 though again I found it unexpected and wanted to confirm it is not a bug before I move forward with the implementation.

If this is a bug, a suggestion would be to permit string field names

[hynek]

OK this is a fun one. Because it's not comparing strings, but Attribute instances. However, Attribute instances are independent from the classes they're defined on – that's clearly a bug.

Try running attr.fields(Robber).repeated_field_name == attr.fields(Cop).repeated_field_name then add e.g. add attr.field(metadata={"foo":42}) to one of the repeated_field_name and run it again.

[hynek]

(to be clearer: technically, this is not a bug. the code does exactly what it says on the tin. however it's absolutely not what people would expect it to do :D)

[henrycjc]

Nice! Thanks @hynek My use case is about dropping a metadata type field from a collection of classes. So e.g. repeated_field_name might be on 10 classes, but I want to optionally drop this from all of my serialised data. It would be nice to somehow support this, though what are your feelings on the use case?

[hynek]

Given how trivial the code currently is, I think it's best if you just implement it yourself for now: https://github.com/python-attrs/attrs/blob/main/src/attr/filters.py