Add a test demonstrating initvar works with inheritance

[dmontagu]

Closes #5496.

It seems that at some point since the 2.0a4 release, this got fixed, but I've added a test written from the code in the issue. (This test fails in 2.0a4.)

This got fixed for python 3.11 after #5760, but not for other python versions..

Selected Reviewer: @adriangb

[cloudflare-pages]

Deploying with    Cloudflare Pages

Latest commit:	a1615bf
Status:	✅  Deploy successful!
Preview URL:	https://ac9f893b.pydantic-docs2.pages.dev
Branch Preview URL:	https://initvar-inheritance.pydantic-docs2.pages.dev

View logs

[dmontagu]

Okay nevermind, it's just resolved in 3.11 :(

For what it's worth, it started working in 3.11 after #5760. I'll keep working through this.

[dmontagu]

Without this, dataclasses.InitVar[<type>] raises errors when returned by typing.get_type_hints:

>           raise TypeError(f"{msg} Got {arg!r:.100}.")
E           TypeError: Forward references must evaluate to types. Got dataclasses.InitVar[int].

The issue is, even if you were willing to rely on dataclasses to tell you what is or isn't an InitVar, you really need this function (or something similar) to allow you to resolve forward references to get what the generic parameter resolves to.

I think monkeypatching InitVar like this is preferable to copying the whole logic of get_type_hints with a tiny tweak to not error if the result isn't callable. But I suppose I could be convinced away from this if we think it's really problematic.

[dmontagu]

please review

[adriangb]

What happens if this gets used before pydantic gets imported? I guess it just doesn't work right?

[dmontagu]

It would error if you tried to use typing.get_type_hints first, but currently I don't think there is a way to get to the point of building a dataclass schema without importing something from pydantic.dataclasses, at least the following works:

from pydantic import BaseModel

initvars = []
@dataclasses.dataclass
class A:
    x: 'dataclasses.InitVar[int]'
    def __post_init__(self, x):
        initvars.append(x)

class B(BaseModel):
    a: A

B.model_validate({'a': {'x': 2}})

assert initvars == [2]

B.model_validate({'a': {'x': 'abc'}})  # errors

Do you think we need to be more careful about ensuring the monkeypatch happens no matter what import path is followed? We could do it in pydantic.__init__ or similar, that would be a little more bulletproof I guess.

[adriangb]

I mean what about:

from dataclasses import dataclass, InitVar

@dataclass
class A:
    x: InitVar[int]  # calls `InitVar.__call__`?

# maybe in another module
from pydantic import BaseModel

class B(BaseModel):
    a: A

[dmontagu]

InitVar.__call__ doesn't actually get called — the typing.get_type_hints function errors if it doesn't think the annotations are actually types. It doesn't think InitVar[int] is a type pre-3.11 I guess, but one thing you can do to get it to make it think something is a type is to make that thing callable (it has a boolean check if callable(maybe_type): in which case it lets you use that annotation and not error). That's why this monkeypatch works — as long as get_type_hints doesn't get called on the type until after the monkeypatch is applied (which I think is the case because that doesn't happen until you are building the schema).

[adriangb]

Please go ahead and merge