Allow subclassing of `ValidationError`

[sydney-runkle]

Probably belongs in pydantic-core, but keeping this here for tracking purposes.

This would allow support for more custom errors, ex: changing the error loc as requested in #8092.

[Youssefares]

I'm going to look into this, feel free to assign to me, very keen to see this feature ✨

[sydney-runkle]

Great! I think the implementation will largely be in pydantic-core.

[Youssefares]

Still looking at this.. Not sure what to do with .from_exception_data yet. I would love for it to know to use the subclass out of the box so CustomValidationError.from_exception_data returns CustomValidationError without the need for an override.

I maybe wrong but just looking at python for now (pyo3 asside) and factories. There's a way to do this:

class Super():
    def __init__(self, msg):
        self.msg = msg

    @classmethod
    def factory(cls, msg) -> Self:
        return cls(msg)

class Sub(Super):
    pass

def test_sub():
    assert isinstance(Sub.factory(msg='la'), Sub)

# success

so long as factory is a class method, I don't think there's an equivalent with a static method which it is at the moment https://github.com/pydantic/pydantic-core/blob/main/python/pydantic_core/_pydantic_core.pyi#L746-L765.

Now, I think that's a non breaking change.. the problem is: that's not so easy to implement with pyo3, as far as I can see, there's no easy way to downcast to a PyType, especially with no constructor defined for the parent ValidationError.

I actually think there may not be a way to access

Would appreciate some guidance if you've encountered something like this before. Neither a rust nor a pyo3 guru here.

[sydney-runkle]

Now, I think that's a non breaking change

Nice find - I think I agree with you here!

the problem is: that's not so easy to implement with pyo3, as far as I can see, there's no easy way to downcast to a PyType, especially with no constructor defined for the parent ValidationError.

Huh, interesting question. Most likely for @davidhewitt or @BoxyUwU. What do you folks think here?

[davidhewitt]

If I understand correctly, I think you want the rust implementation of from_exception_data to have this signature:

    #[classmethod]
    #[pyo3(signature = (title, line_errors, input_type="python", hide_input=false))]
    fn from_exception_data(
        cls: Bound<'_, PyType>,
        py: Python,
        title: PyObject,
        line_errors: Bound<'_, PyList>,
        input_type: &str,
        hide_input: bool,
    ) -> PyResult<Py<Self>> {

[Youssefares]

@davidhewitt appreciate your guidance for this, thanks for getting back!
So I did implement something similar in this PR pydantic/pydantic-core#1413 but couldn't implement this without adding a constructor to ValidationError.

To achieve the goal of this PR, from_exception_data should be returning an object of type cls, so that the test below passes. Is there a way to do that without defining a new method and calling cls's new?

def test_validation_error_subclassable():
    """Assert subclassable and inheritance hierarchy as expected"""

    class CustomValidationError(ValidationError):
        pass

    with pytest.raises(ValidationError) as exception_info:
        raise CustomValidationError.from_exception_data(
            'My CustomError',
            [
                InitErrorDetails(
                    type='value_error',
                    loc=('myField',),
                    msg='This is my custom error.',
                    input='something invalid',
                    ctx={
                        'myField': 'something invalid',
                        'error': "'something invalid' is not a valid value for 'myField'",
                    },
                )
            ],
        )
    assert isinstance(exception_info.value, CustomValidationError)

I don't think it's possible, because rust and pyo3 have no pointers to the subclass, and will only know how to construct the parent. We have to make a python call to cls (cls.call1?), which will have no constructor unless we define it on the ValidationError pyclass..

Hope this makes sense.

[sydney-runkle]

Hey @Youssefares,

Thanks for your help here! @davidhewitt ping - hopefully we can move this associated PR forward for v2.10!