is_dataclass() returns True for non-dataclass subclass of dataclass

[JelleZijlstra]

Bug report

Bug description:

If a dataclass has a subclass that is not itself a dataclass, is_dataclass() returns True on the subclass and its instances:

from dataclasses import dataclass, is_dataclass

@dataclass
class X:
    y: int

class Z(X):
    pass

print(is_dataclass(Z))  # True
print(is_dataclass(Z()))  # True

Documentation of is_dataclass() for reference: https://docs.python.org/3.13/library/dataclasses.html#dataclasses.is_dataclass

Intuitively this seems wrong: Z is not itself a dataclass. In pyanalyze I wrote a replacement for is_dataclass() because I needed to check whether the exact class was a dataclass:

def is_dataclass_type(cls: type) -> bool:
    try:
        return "__dataclass_fields__" in cls.__dict__
    except Exception:
        return False

Changing the CPython behavior might be impossible for compatibility reasons, but in that case, we should document and test this edge case.

cc @ericvsmith @carljm for dataclasses.

CPython versions tested on:

CPython main branch

Operating systems tested on:

macOS

Linked PRs

• gh-119260: Clarify is_dataclass Behavior for Subclasses in Documentation and Tests #119480
• [3.13] gh-119260: Clarify is_dataclass Behavior for Subclasses in Documentation and Tests (GH-119480) #119760
• [3.12] gh-119260: Clarify is_dataclass Behavior for Subclasses in Documentation and Tests (GH-119480) #119761

[blhsing]

I think the current behavior makes sense. A subclass of a dataclass still possesses all the characteristics of the dataclass even if the subclass is not decorated with dataclass. Its instances can still be passed to astuple, asdict and replace. It's just that without the dataclass decorator it cannot declare additional fields.

[carljm]

I agree with @blhsing -- I think the current behavior is probably better.

This behavior should be explicitly tested and documented, though; currently it isn't.

[ericvsmith]

Agreed with @carljm, especially the part about adding tests and docs. I'm surprised there's no test for this.

[adiaholic]

I think the current behavior makes sense. A subclass of a dataclass still possesses all the characteristics of the dataclass even if the subclass is not decorated with dataclass. Its instances can still be passed to astuple, asdict and replace. It's just that without the dataclass decorator it cannot declare additional fields.

I was able to declare additional fields in subclasses.

from dataclasses import dataclass, is_dataclass
import unittest

@dataclass
class X:
    y: int

class Z(X):
    z: int = 0  # Adding a default field to demonstrate inheritance and dataclass behavior

class TestDataclassInheritance(unittest.TestCase):
    def test_dataclass_inheritance(self):
        # Check if X is a dataclass
        self.assertTrue(is_dataclass(X), "X should be a dataclass")
        # Check if Z is a dataclass, it should inherit the dataclass behavior from X
        self.assertTrue(is_dataclass(Z), "Z should be a dataclass because it inherits from X")
        # Create an instance of Z and check if the default values are set correctly
        z_instance = Z(y=5)
        self.assertEqual(z_instance.y, 5, "The value of y should be set to 5")
        self.assertEqual(z_instance.z, 0, "The default value of z should be 0")

if __name__ == '__main__':
    unittest.main()

[JelleZijlstra]

Your subclass also has the @dataclass decorator. This issue is about subclasses without the decorator.

[adiaholic]

Your subclass also has the @dataclass decorator. This issue is about subclasses without the decorator.

My bad, corrected the example in the comment above 😃.

Also I have a PR on the way !

[carljm]

I was able to declare additional fields in subclasses.

No, you were able to add a normal Python class attribute to a subclass, which is sufficient to fool your test because of Python's normal behavior where instance attribute access falls back to the class. The class attribute you added in the subclass is not a dataclass field:

>>> Z.__dataclass_fields__.keys()
dict_keys(['y'])
>>> Z(y=1, z=2)
Traceback (most recent call last):
  File "<python-input-6>", line 1, in <module>
    Z(y=1, z=2)
    ~^^^^^^^^^^
TypeError: X.__init__() got an unexpected keyword argument 'z'

[carljm]

Thanks @adiaholic for the PR!