Assigning to an attribute with the same name as a dataclass `InitVar` raises `has no attribute` on 0.960

[paw-lu]

Bug Report

On 0.960, if an attribute is typed as a dataclasses.InitVar and later tries to assign an attribute with the same name, mypy will complain on the assignment statement that the dataclass has no such attribute.

This does not occur on 0.950.

To Reproduce

import dataclasses
from dataclasses import InitVar


@dataclasses.dataclass
class Foo:
    a: int
    b: InitVar[int]

    def __post_init__(self, b: int) -> None:
        self.b = self.a + 1

$ pipx run --spec mypy==0.950 mypy --strict script.py
Success: no issues found in 1 source file

$ pipx run --spec mypy==0.960 mypy --strict script.py
script.py: note: In member "__post_init__" of class "Foo":
script.py:11:9: error: "Foo" has no attribute "b"  [attr-defined]
            self.b = self.a + 1
            ^
Found 1 error in 1 file (checked 1 source file)

Expected Behavior

I expected that mypy would not complain about an assignment just because it shares the same name as an InitVar.

Actual Behavior

mypy warns the user that the attribute does not exists in an assignment expression.

Your Environment

• Mypy version used: 0.960
• Mypy command-line flags: --strict
• Python version used: 3.10
• Operating system and version: macOS 12.3.1

[erictraut]

PEP 557 refers to these as "init-only" fields. An init-only field is provided as arguments to the __post_init__ method but is otherwise not used as an attribute for the dataclass. I think mypy is correct in emitting an error here. Pyright does the same.

You should use a different name for the instance variable assigned in the __post_init__ method, a name that doesn't conflict with the name of the InitVar field.

[JelleZijlstra]

I agree with Eric that this behavior is correct, but I don't recall a change that was meant to introduce this check. Perhaps we should add a new unit test to ensure mypy's behavior remains consistent.

[AlexWaygood]

A unit test was added for disallowing this behaviour in #12798, so the error being emitted is expected behaviour, not a bug.

Note that if you use slotted dataclasses on Python 3.10+, this will also fail at runtime:

>>> from dataclasses import dataclass, InitVar
>>> @dataclass(slots=True)
... class Foo:
...     bar: InitVar[str]
...     baz: InitVar[int] = 0
...
>>> f = Foo('hi', 5)
>>> f.__slots__
()
>>> f.bar = 'bye'
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: 'Foo' object has no attribute 'bar'
>>> f.baz = 6
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: 'Foo' object attribute 'baz' is read-only

[AlexWaygood]

Closing this since it's a feature-not-a-bug, expected behaviour, and has a unit test.

[paw-lu]

Thanks @AlexWaygood for linking #12798. Appreciated! Seems like a straighforward case!

[JukkaL]

Here's an example of how to work around this -- you can define two attributes, one is a regular one and the other is an InitVar:

from dataclasses import dataclass, InitVar, field

@dataclass
class D:
    _x: InitVar[int]
    x: str = field(init=False)

    def __post_init__(self, _x: int) -> None:
        self.x = str(_x)

[ksamuel]

@JukkaL : this workaround mean you can do D(1), but not D(x=1), which is making things hard to read.

This means right now there is no clean way in dataclass to both have a transformed attribute you transform from initial values and slot at the same time.

It's not a huge problem, but I don't think this ticket should be closed until an alternative is proposed for that use case.

[AlexWaygood]

@JukkaL : this workaround mean you can do D(1), but not D(x=1), which is making things hard to read.

@ksamuel does this not work?

from dataclasses import dataclass

@dataclass(init=False)
class D:
    x: str
    def __init__(self, x: int) -> None:
        self.x = str(x)