Allow validator and serializer functions to have default values

[Viicos]

Fixes #9463

Selected Reviewer: @sydney-runkle

[codspeed-hq]

CodSpeed Performance Report

Merging #9478 will not alter performance

_{Comparing Viicos:inspect-default-arg (9317086) with main (777cff0)}

Summary

✅ 13 untouched benchmarks

[Viicos]

I'm not too happy with how this function grew, hence this "warning" in the docstring. At some point, maybe this inspection of validators/serializers signature could be refactored, e.g.:

@dataclass
class FunctionInfo:
    func: Callable[..., Any]

    @property
    def requires_info_arg(self) -> bool: ...

    ...

[Viicos]

Please review

[sydney-runkle]

@Viicos,

Could you provide some tests with concrete examples of validators / serializers with default values, so that we have some more documented context?

Thanks!

[sydney-runkle]

@Viicos,

Thanks for adding those extra tests. I suppose I still have a bit of a gap in understanding in why we'd want to support these additional args given that you don't call validators or serializers directly.

Could you explain a bit more re the motivation for this change / addition? Thanks!

[Viicos]

Could you explain a bit more re the motivation for this change / addition? Thanks!

While it is really uncommon to use such a function as a validator (or serializer):

pydantic/tests/test_validators.py

Lines 2889 to 2900 in f7f7abd

	def test_validator_with_default_values() -> None:
	def validate_x(v: int, unrelated_arg: int = 1, other_unrelated_arg: int = 2) -> int:
	assert v != -1
	return v
	class Model(BaseModel):
	x: int
	val_x = field_validator('x')(validate_x)
	with pytest.raises(ValidationError):
	Model(x=-1)

It is more common to do:

class Model(BaseModel): 
    x: str

    val_x = field_validator('x', mode='before')(str) 

Or to use str as a plain serializer, etc.

Pydantic uses inspect.signature to check the validity of the validator/serializer callable: number of arguments, whether it takes an info argument or not. On CPython, inspect.signature does not work for some builtins, such as str. As it was done in https://github.com/pydantic/pydantic/pull/9450/files, we assume no info argument is needed in this case, so it works well. On PyPy however, the signature of str is successfully parsed as <Signature (object=<no value>, encoding=<no value>, errors=<no value>)>.

As we can see, str takes up to three arguments. To make it work with validator/serializer callables that do not take an info argument, we need to discard arguments having a default value (apart from the first one which will be the actual value being validated/serialized). After all, there's no reason to define a callable that needs to have the info arg this way:

def my_validator(value: Any, info: ValidationInfo | None = None): ...

If you need the info arg, you should specify it as a required argument.

[davidhewitt]

It seems like the behaviour of this function changed to have the required arguments property, should we rename it to avoid accidents in case anyone is trying to borrow our internals?

[davidhewitt]

Overall this seems very reasonable to me, with just a desire to rename the helper function to help avoid confusion.

[Viicos]

Function renamed

[sydney-runkle]

@Viicos,

Once we get tests passing, we should be all good to merge!

[sydney-runkle]

@Viicos,

Maybe a pull from main would fix those tests?

[sydney-runkle]

Ah with a closer look, looks like the test failure is related to these changes.

[Viicos]

Yep, fixed the recently added test

[sydney-runkle]

Awesome, thanks @Viicos!