Performance tips

In most cases Pydantic won't be your bottle neck, only follow this if you're sure it's necessary.

In general, use model_validate_json() not model_validate(json.loads(...))

On model_validate(json.loads(...)), the JSON is parsed in Python, then converted to a dict, then it's validated internally. On the other hand, model_validate_json() already performs the validation internally.

There are a few cases where model_validate(json.loads(...)) may be faster. Specifically, when using a 'before' or 'wrap' validator on a model, validation may be faster with the two step method. You can read more about these special cases in this discussion.

Many performance improvements are currently in the works for pydantic-core, as discussed here. Once these changes are merged, we should be at the point where model_validate_json() is always faster than model_validate(json.loads(...)).

TypeAdapter instantiated once

The idea here is to avoid constructing validators and serializers more than necessary. Each time a TypeAdapter is instantiated, it will construct a new validator and serializer. If you're using a TypeAdapter in a function, it will be instantiated each time the function is called. Instead, instantiate it once, and reuse it.

Bad Good

from typing import List

from pydantic import TypeAdapter


def my_func():
    adapter = TypeAdapter(List[int])
    # do something with adapter

from typing import List

from pydantic import TypeAdapter

adapter = TypeAdapter(List[int])

def my_func():
    ...
    # do something with adapter

Sequence vs list or tuple - Mapping vs dict

When using Sequence, Pydantic calls isinstance(value, Sequence) to check if the value is a sequence. Also, Pydantic will try to validate against different types of sequences, like list and tuple. If you know the value is a list or tuple, use list or tuple instead of Sequence.

The same applies to Mapping and dict. If you know the value is a dict, use dict instead of Mapping.

Don't do validation when you don't have to - use Any to keep the value unchanged

If you don't need to validate a value, use Any to keep the value unchanged.

from typing import Any

from pydantic import BaseModel


class Model(BaseModel):
    a: Any


model = Model(a=1)

Avoid extra information via subclasses of primitives

Don't do thisDo this

class CompletedStr(str):
    def __init__(self, s: str):
        self.s = s
        self.done = False

from pydantic import BaseModel


class CompletedModel(BaseModel):
    s: str
    done: bool = False

Use tagged union, not union

Tagged union (or discriminated union) is a union with a field that indicates which type it is.

from typing import Any

from typing_extensions import Literal

from pydantic import BaseModel, Field


class DivModel(BaseModel):
    el_type: Literal['div'] = 'div'
    class_name: str | None = None
    children: list[Any] | None = None


class SpanModel(BaseModel):
    el_type: Literal['span'] = 'span'
    class_name: str | None = None
    contents: str | None = None


class ButtonModel(BaseModel):
    el_type: Literal['button'] = 'button'
    class_name: str | None = None
    contents: str | None = None


class InputModel(BaseModel):
    el_type: Literal['input'] = 'input'
    class_name: str | None = None
    value: str | None = None


class Html(BaseModel):
    contents: DivModel | SpanModel | ButtonModel | InputModel = Field(
        discriminator='el_type'
    )

See Discriminated Unions for more details.

Use TypedDict over nested models

Instead of using nested models, use TypedDict to define the structure of the data.

Performance comparison

With a simple benchmark, TypedDict is about ~2.5x faster than nested models:

from timeit import timeit

from typing_extensions import TypedDict

from pydantic import BaseModel, TypeAdapter


class A(TypedDict):
    a: str
    b: int


class TypedModel(TypedDict):
    a: A


class B(BaseModel):
    a: str
    b: int


class Model(BaseModel):
    b: B


ta = TypeAdapter(TypedModel)
result1 = timeit(
    lambda: ta.validate_python({'a': {'a': 'a', 'b': 2}}), number=10000
)
result2 = timeit(
    lambda: Model.model_validate({'b': {'a': 'a', 'b': 2}}), number=10000
)
print(result2 / result1)

Avoid wrap validators if you really care about performance

Wrap validators are generally slower than other validators. This is because they require that data is materialized in Python during validation. Wrap validators can be incredibly useful for complex validation logic, but if you're looking for the best performance, you should avoid them.

Failing early with FailFast

Starting in v2.8+, you can apply the FailFast annotation to sequence types to fail early if any item in the sequence fails validation. If you use this annotation, you won't get validation errors for the rest of the items in the sequence if one fails, so you're effectively trading off visibility for performance.

from typing import List

from typing_extensions import Annotated

from pydantic import FailFast, TypeAdapter, ValidationError

ta = TypeAdapter(Annotated[List[bool], FailFast()])
try:
    ta.validate_python([True, 'invalid', False, 'also invalid'])
except ValidationError as exc:
    print(exc)
    """
    1 validation error for list[bool]
    1
      Input should be a valid boolean, unable to interpret input [type=bool_parsing, input_value='invalid', input_type=str]
    """

Read more about FailFast here.