Architecture

Starting with Pydantic V2, part of the codebase is written in Rust in a separate package called pydantic-core. This was done partly in order to improve validation and serialization performance (with the cost of limited customization and extendibility of the internal logic).

This architecture documentation will first cover how the two pydantic and pydantic-core packages interact together, then will go through the architecture specifics for various patterns (model definition, validation, serialization, JSON Schema).

Usage of the Pydantic library can be divided into two parts:

Model definition, done in the pydantic package.
Model validation and serialization, done in the pydantic-core package.

Model definition

Whenever a Pydantic BaseModel is defined, the metaclass will analyze the body of the model to collect a number of elements:

Defined annotations to build model fields (collected in the model_fields attribute).
Model configuration, set with model_config.
Additional validators/serializers.
Private attributes, class variables, identification of generic parametrization, etc.

Communicating between `pydantic` and `pydantic-core`: the core schema

We then need a way to communicate the collected information from the model definition to pydantic-core, so that validation and serialization is performed accordingly. To do so, Pydantic uses the concept of a core schema: a structured (and serializable) Python dictionary (represented using TypedDict definitions) describing a specific validation and serialization logic. It is the core data structure used to communicate between the pydantic and pydantic-core packages. Every core schema has a required type key, and extra properties depending on this type.

The generation of a core schema is handled in a single place, by the GenerateSchema class (no matter if it is for a Pydantic model or anything else).

In the case of a Pydantic model, a core schema will be constructed and set as the __pydantic_core_schema__ attribute.

To illustrate what a core schema looks like, we will take the example of the bool core schema:

class BoolSchema(TypedDict, total=False):
    type: Required[Literal['bool']]
    strict: bool
    ref: str
    metadata: Any
    serialization: SerSchema

When defining a Pydantic model with a boolean field:

from pydantic import BaseModel, Field


class Model(BaseModel):
    foo: bool = Field(strict=True)

The core schema for the foo field will look like:

{
    'type': 'bool',
    'strict': True,
}

As seen in the BoolSchema definition, the serialization logic is also defined in the core schema. If we were to define a custom serialization function for foo , the serialization key would look like:

For example using the [field_serializer][pydantic.functional_serializers.field_serializer] decorator:

class Model(BaseModel):
foo: bool = Field(strict=True)

@field_serializer('foo', mode='plain')
def serialize_foo(self, value: bool) -> Any:
...

{
    'type': 'function-plain',
    'function': <function Model.serialize_foo at 0x111>,
    'is_field_serializer': True,
    'info_arg': False,
    'return_schema': {'type': 'int'},
}

Note that this is also a core schema definition, just that it is only relevant for pydantic-core during serialization.

Core schemas cover a broad scope, and are used whenever we want to communicate between the Python and Rust side. While the previous examples were related to validation and serialization, it could in theory be used for anything: error management, extra metadata, etc.

JSON Schema generation

You may have noticed that the previous serialization core schema has a return_schema key. This is because the core schema is also used to generate the corresponding JSON Schema.

Similar to how the core schema is generated, the JSON Schema generation is handled by the GenerateJsonSchema class. The generate method is the main entry point and is given the core schema of that model.

Coming back to our bool field example, the bool_schema method will be given the previously generated boolean core schema and will return the following JSON Schema:

{
    {"type": "boolean"}
}

Customizing the core schema and JSON schema

While the GenerateSchema and GenerateJsonSchema classes handle the creation of the corresponding schemas, Pydantic offers a way to customize them in some cases, following a wrapper pattern. This customization is done through the __get_pydantic_core_schema__ and __get_pydantic_json_schema__ methods.

To understand this wrapper pattern, we will take the example of metadata classes used with Annotated, where the __get_pydantic_core_schema__ method can be used:

from typing import Any

from pydantic_core import CoreSchema
from typing_extensions import Annotated

from pydantic import GetCoreSchemaHandler, TypeAdapter


class MyStrict:
  @classmethod
  def __get_pydantic_core_schema__(
      cls, source: Any, handler: GetCoreSchemaHandler
  ) -> CoreSchema:
      schema = handler(source)  # (1)
      schema['strict'] = True
      return schema


class MyGt:
  @classmethod
  def __get_pydantic_core_schema__(
      cls, source: Any, handler: GetCoreSchemaHandler
  ) -> CoreSchema:
      schema = handler(source)  # (2)
      schema['gt'] = 1
      return schema


ta = TypeAdapter(Annotated[int, MyStrict(), MyGt()])

When the GenerateSchema class builds the core schema for Annotated[int, MyStrict(), MyGt()], it will create an instance of a GetCoreSchemaHandler to be passed to the MyGt.__get_pydantic_core_schema__ method.

The source argument depends on the core schema generation pattern. In the case of Annotated, the source will be the type being annotated. When defining a custom type, the source will be the actual class where __get_pydantic_core_schema__ is defined.

Model validation and serialization

While model definition was scoped to the class level (i.e. when defining your model), model validation and serialization happens at the instance level. Both these concepts are handled in pydantic-core (providing a 5 to 20 performance increase compared to Pydantic V1), by using the previously built core schema.

pydantic-core exposes a SchemaValidator and SchemaSerializer class to perform these tasks:

from pydantic import BaseModel


class Model(BaseModel):
  foo: int


model = Model.model_validate({'foo': 1})  # (1)
dumped = model.model_dump()  # (2)