Bug when mixing metaclasses w/ `dataclassy` (Python 3.6 only) about dataclassy

[biqqles]

As I see it there are three possibilities that explain this (ranked in order of my guesses at likelihood):

1. The implementation of typing.Generic changed between 3.6 and 3.7. I think this is most likely because the typing module underwent a lot of changes between these versions IIRC
2. The implementation of abc.ABCMeta changed
3. Python became somehow "better" at resolving metaclass mixing

I will take some time later to try to prove or disprove these theories, then hopefully fix the problem if it's with dataclassy, or find a workaround.

[biqqles]

MRE:

from typing import Generic, TypeVar
from dataclassy import dataclass

X = TypeVar('X')

@dataclass
class T(Generic[X]):
    ...

[biqqles]

Diff: python/cpython@v3.6.14...v3.7.10 Unfortunately I can't find a way to link to the single file typing.py.

[fubuloubu]

Diff: python/[email protected] Unfortunately I can't find a way to link to the single file typing.py.

These seems relevant (in typing.py)

So, seems like in Python <=3.6 only you'd need to integrate GenericMeta into dataclassy's meta to support Generics

[fubuloubu]

So, seems like in Python <=3.6 only you'd need to integrate GenericMeta into dataclassy's meta to support Generics

Or... I can just drop 3.6 support from my library 😓 (which I've been meaning to anyways)

[biqqles]

Great find. The 3.7 release notes are a little more blunt:

the generic types can be used without metaclass conflicts

This kind of makes me want to say it's a typing problem, but I'm not against a fix in dataclassy if it's easy.

Anyway, this made me curious what the same paragraph looked like in 3.6. There I read:

The metaclass used by Generic is a subclass of abc.ABCMeta.

This made me try implementing class AbstractDataClassMeta(DataClassMeta, GenericMeta): as a sneaky workaround. It seems like it should maybe work but instead of (or before, quite possibly) the conflict TypeError you get TypeError: Cannot inherit from plain Generic when actually using the decorator.

Edit: trying your suggestion of basing DataClassMeta on GenericMeta causes the same error.

[fubuloubu]

@biqqles thanks for looking into it, I've decided to just deprecate support for Python 3.6 in my framework. IPython was also giving me grief, so it made sense to do.