Adding tagged union serializer 🚀

[sydney-runkle]

Implementing a tagged union serializer to unify patterns for union validation and serialization.

Similar to the validation phase, we attempt to check the discriminator value for the data being serialized. If we find it, we select the serializer corresponding to said tag. In most cases, if we have trouble with this, we fall back to standard union serialization in order to preserve backwards compatibility. This is consistent with our general serialization approach, which is to first try with a designated serializer, then just fallback to inference.

I could see a case for adding support for an explicit serialization discriminator at some point, and this is a good step in the direction of supporting that on the core side of things.

A few notes:

• I've removed the SelfSchema member of the Discriminator enum because its usage was terminated with 5efeaf9
• I've moved the Discriminator structure to a common directory, as it's used across validators and serializers

[codspeed-hq]

CodSpeed Performance Report

Merging #1397 will not alter performance

_{Comparing union-ser-fixes (f1236ff) with main (3d8295e)}

Summary

✅ 155 untouched benchmarks

[sydney-runkle]

I'd love some feedback on my rust here - definitely looking to improve my use of best practices, etc :)

cc @BoxyUwU @davidhewitt

[davidhewitt]

Overall looking solid, a few thoughts / questions!

[Viicos]

Was wondering: will this require any additional changes in pydantic? Or will the following work out of the box?

from typing import Literal, Union

from pydantic import BaseModel, Field, ValidationError


class Cat(BaseModel):
    pet_type: Literal['cat']
    meows: int


class Dog(BaseModel):
    pet_type: Literal['dog']
    barks: float


class Model(BaseModel):
    pet: Union[Cat, Dog] = Field(..., discriminator='pet_type')

assert Model.model_validate({"pet": {"pet_type": "dog", barks: 1.0}}).model_dump() == {"pet": {"pet_type": "dog", barks: 1.0}}

[sydney-runkle]

Was wondering: will this require any additional changes in pydantic? Or will the following work out of the box?

Good question - the above will work out of the box @Viicos! 🚀

[sydney-runkle]

I think this error is fine with the RootModel re the failing integration test. I'm sure we could improve root model / discriminated union behavior - it's currently broken on main, this just more explicitly explains some of the problems we have...

[davidhewitt]

LGTM with one final step...