Frozen dataclasses with slots cannot override __getstate__ or __setstate__

[drhagen]

You are supposed to be able to override how pickle pickles an object with __getstate__ and __setstate__. However, these methods are ignored in a dataclass when specifying both frozen=True and slots=True.

See this example:

import pickle
from dataclasses import dataclass


@dataclass(frozen=True, slots=True)
class Foo:
    bar: int

    def __getstate__(self):
        print("getstate")
        return {"bar": self.bar}

    def __setstate__(self, state):
        print("setstate")
        object.__setattr__(self, "bar", state["bar"])

b = pickle.dumps(Foo(1))
foo = pickle.loads(b)

The expected "getstate" and "setstate" lines are never printed because the supplied methods are never called. If either frozen or slots is removed, the expected lines are printed.

From the source code, it is pretty clear why this is happening. If frozen and slots are both True, then special versions of __getstate__ and __setstate__ are unconditionally attached to the dataclass. The dataclass decorator should probably treat this way that it does other methods—only add a method if it is not already present.

Linked PRs

• gh-104035: Do not ignore user-defined __{get,set}state__ in slotted frozen dataclasses #104041
• [3.11] gh-104035: Do not ignore user-defined __{get,set}state__ in slotted frozen dataclasses (GH-104041) #104044

[ericvsmith]

It does seem reasonable to ignore __getstate__ and __setstate__ if they're already set. Have you tested this? And in particular, does it break any existing tests?

[sobolevn]

Oh, I just went for fixing this in #104041 :)

New logic will allow users to customize __getstate__ and __setstate__ method.
Thanks a lot, @drhagen for finding this :)

It is funny that I've spent the whole day optimizing these two in #103032, but I haven't seen this! 👍