arbitrary_types_allowed no longer works with root models in V2

[MarkusSintonen]

Initial Checks

• I confirm that I'm using Pydantic V2

Description

In V2 root models arbitrary_types_allowed no longer works. In V1 it used to work. Is it intended not to work anymore? Example below showing an use case that used to work in V1. I guess the issue is somehow related to generic models handling in V2.

Example Code

import numpy

from pydantic import BaseModel

try:
    from pydantic import ConfigDict, RootModel
    PYDANTIC_VERSION = "v2"
except ImportError:
    PYDANTIC_VERSION = "v1"


def test_arbitrary_root():
    if PYDANTIC_VERSION == 'v2':
        # Now it raises suggesting using arbitrary_types_allowed. But it is set!
        # Exception:
        # pydantic.errors.PydanticSchemaGenerationError: Unable to generate pydantic-core schema for <class 'numpy.ndarray'>.
        # Set `arbitrary_types_allowed=True` in the model_config to ignore this error or implement `__get_pydantic_core_schema__` on your type to fully support it.
        class MyNdarray(RootModel[numpy.ndarray]):
            model_config = ConfigDict(arbitrary_types_allowed=True)
            # defines deserialization of list to ndarray etc...
    else:
        # In v1 it works fine.
        class MyNdarray(BaseModel):
            __root__: numpy.ndarray

            class Config:
                arbitrary_types_allowed = True

            # defines deserialization of list to ndarray etc...

Python, Pydantic & OS Version

pydantic version: 2.0.3
        pydantic-core version: 2.3.0 release build profile
                 install path: /Users/markussintonen/Library/Caches/pypoetry/virtualenvs/leak-test-VvUvdXxD-py3.10/lib/python3.10/site-packages/pydantic
               python version: 3.10.12 (main, Jun 20 2023, 17:00:24) [Clang 14.0.3 (clang-1403.0.22.14.1)]
                     platform: macOS-13.4.1-x86_64-i386-64bit
     optional deps. installed: ['typing-extensions']

Selected Assignee: @lig

[dmontagu]

@MarkusSintonen while I think ideally this wouldn't error, I think it might be hard to get it to work properly, since it's erroring while trying to create RootModel[np.ndarray], which you are inheriting from. Instead, you can do:

from typing import Any

import numpy

from pydantic import ConfigDict, RootModel


class MyNdarray(RootModel[Any]):
    root: numpy.ndarray
    model_config = ConfigDict(arbitrary_types_allowed=True)
    # defines deserialization of list to ndarray etc...

I was worried this might not give the right mypy behavior, but I wasn't able to produce any typing issues using the above, as far as I could tell it behaved the way you'd want it to (same as RootModel[numpy.ndarray], just doesn't give the arbitrary types error). (Let us know if you discover issues with it though.)

I'd be open to PRs fixing it for RootModel but I don't think the right answer is to make arbitrary_types_allowed=True for RootModel by default, and I don't see another good way to achieve this.

Open to debate/other ideas

[lig]

@dmontagu I think that both forms should work identical in general despite what type is being used for RootModel.

class MyNdarray(RootModel[numpy.ndarray]):
    model_config = ConfigDict(arbitrary_types_allowed=True)

and

class MyNdarray(RootModel):
    root = numpy.ndarray
    model_config = ConfigDict(arbitrary_types_allowed=True)

I agree that arbitrary_types_allowed shouldn't be True by default.

I will focus in this issue just on making arbitrary_types_allowed=True work as expected.

[lig]

After thorough review of this issue I'm tempting to call this known behavior and add a note regarding this to the docs.

The issue here is that RootModel[numpy.ndarray] creates a specialized subclass of RootModel that could be used as a full functional model already. Two code examples below produce mostly identical results.

class Model(RootModel):
    root: numpy.ndarray 

and

Model = RootModel[numpy.ndarray]

When creating a RootModel subclass explicitly we can set model_config for it and when creating using an implicit subclass as in RootModel[numpy.ndarray] we cannot.

This means that in the code below we are attempting to create a RootModel subclass without model_config defined before setting model_config for MyNdarray.

class MyNdarray(RootModel[numpy.ndarray]):
    model_config = ConfigDict(arbitrary_types_allowed=True)