dataclasses.is_dataclass doesn't always work correctly with Generic types

[Lyle-Alloy]

This was introduced by BPO45663.

>>> import typing
... import dataclasses
... 
... T = typing.TypeVar("T")
... 
... @dataclasses.dataclass
... class A(typing.Generic[T]):
...     x: T
...     
>>> dataclasses.is_dataclass(A)
True
>>> dataclasses.is_dataclass(A[float])
False
>>> dataclasses.is_dataclass(A(0))
True
>>> dataclasses.is_dataclass(A[float](0))
True

That second dataclasses.is_dataclass call should be True

Bug reproduced on cpython 3.9.12 x86_64-linux-gnu

[JelleZijlstra]

I can see how this is confusing, but changing the behavior here may be risky. Right now you can assume that if is_dataclass(o) returns true, you can access o.__dataclass_fields__. But A[float] doesn't have a __dataclass_fields__ attribute. I don't know what you're using is_dataclass() for, but it seems reasonable that a lot of people would use it before checking __dataclass_fields__ for introspection.

So I don't think we should change this in any of the bugfix branches. We could change it on main, but it may be safer to keep the current behavior and document it.

[AlexWaygood]

Right now you can assume that if is_dataclass(o) returns true, you can access o.__dataclass_fields__.

I don't think __dataclass_fields__ is a public attribute (@ericvsmith, correct me if I'm wrong!), so users probably shouldn't be accessing it directly anyway. But you can say the same thing with respect to dataclasses.fields(), which is a public function. Currently, any class for which dataclasses.is_dataclass() returns True can be safely passed to dataclasses.fields(). Changing that might be unfortunate.

[ericvsmith]

Yes, you should use dataclasses.fields() to access the fields, and not inspect o.__dataclass_fields__.

Instead of checking first, you can also just try dataclasses.fields() and catch the TypeError if it's not a dataclass.

[Penchekrak]

Perhaps, something should be done with typing._GenericAlias initialization so it would set __dataclass_fields__ for its instance. The fact that A[float] does not have such information despite logically being a dataclass with specified types, defaults, etc, really sounds like a bug to me.

[JelleZijlstra]

Thinking about this again, I'd prefer to leave the behavior unchanged here. A GenericAlias is not a class, it is a special construct used in typing that proxies some behavior to the class. Trying to make it behave exactly like a class is going to be a never-ending project, and is itself bound to cause problems because at the end of the day, a GenericAlias is not in fact a class.

is_dataclass() is documented as "Returns True if obj is a dataclass or an instance of a dataclass." A GenericAlias of a dataclass is not a dataclass and is definitely not an instance of a dataclass. So I think the current behavior is correct.

[hauntsaninja]

I agree with Jelle and strongly agree with Alex's point about is_dataclass needing to being usable as a guard for fields