pydantic-gubbins

Table of Contents

1. Overview
2. Typing
   1. DiscriminatedUnion
   2. SubclassOf
   3. Union
   4. FrozenDict
3. Descriptor Support

Overview

This project contains various utils for working with pydantic.

Typing

DiscriminatedUnion

A common pattern pydantic users encounter is how to serialise/deserialise a field whose type is a union of BaseModel types. This improved in pydantic v2 and various ways of implementing a discriminated union are detailed here. However, this approach is a bit unsatisfying as:

• One is forced to explicitly implement a literal for each affected type
• The type literal will be serialised regardless of whether it's needed or not. E.g, if there is a field explicitly of such a type (and not a union), the type literal will still be serialised

Stack Overflow and other forums have many long discussions on this topic, without apparently offering a solution, so I have included my own implementation of DiscriminatedUnion. What it does:

1. Creates tagged union of the types: Union[Annotated[t1, Tag("t1")], Annotated[t1, Tag("t2")], ...]
2. Adds a WrapSerializer to include the tag name in the serialised output
3. Adds a Discrimator with a callable to retrieve the tag name from the serialised form
4. Uses the type's __name__ by default but type.TYPE_KEY if present

SubclassOf

SubclassOf takes a single type as a parameter and returns a DiscriminatedUnion of all the (recursive) subclasses of that type

Union

This can be used in place of typing.Union. It converts unions of BaseModel and dataclass types into a DiscriminatedUnion. It also separates such "model" types from other types. In the event that both are encounted, it returns Union[Union[<other types>], DiscriminatedUnion[<model types>]]

FrozenDict

I found this implementation (and I can't remember where!) and included it. This is because I have some upcoming changes which will convert collection types into immutable equivalents, to be combined with frozen models.

Descriptor Support

pydantic does not support using descriptors for model fields. I have raised an issue for this and submitted PRs for pydantic (some further discussion on that thread) and pydantic-core but the maintainers are correct in that the whole descriptor issue really needs more discussion.

In the interim, this project supplies an implementation of BaseModel, which can be used in place of the standard pydantic offering and which supports descriptors for model fields. Absent descriptor fields, it will perform exactly as the original. It is not a large amount of code and the changes are summarised below. The intent of these changes is the descriptor model fields should behave as closely as possible to descriptors in dataclasses. Please note that property or cached_property passed as annotations will be ignored. This is because pydantic already has special-case logic for them.

1. The metaclass adds the descriptors onto the the returned type, calls __set_name__ on them, and populates __pydantic_descriptor_fields__
2. The methods on BaseModel which access __dict__ directly have been overridden to extend their functionality to include descriptor fields
3. Access to __dict__ itself is now controlled by a descriptor. This implementation is rather low-level and possibly inadvisable. The same result might be achieved by using a model validator, however, there are many places in pydantic and pydantic-core where __dict__ is accessed directly and I'm not convinced all would be covered by a validator

Using the BaseModel supplied by this project, the below works:

from pydantic_gubbins import BaseModel
from typing import Any

_field_descriptor_undefined = object()


class FieldDescriptor:
    """ Example descriptor, just to show storage somewhere other than __dict__ """

    def __init__(self, default: Any = _field_descriptor_undefined):
        self.__default = default
        self.__name = None
        self.__values = {}

    def __get__(self, instance, owner):
        if instance is not None:
            try:
                return self.__values[id(instance)][self.__name]
            except KeyError:
                pass

        if self.__default is _field_descriptor_undefined:
            raise AttributeError

        return self.__default

    def __set__(self, instance, value):
        self.__values.setdefault(id(instance), {})[self.__name] = value

    def __set_name__(self, owner, name):
        self.__name = name



class Foo(BaseModel):
    s: str
    i: int = FieldDescriptor(-1)