pydantic-v2.md

Pydantic V2 Plan


**Samuel Colvin** •  [@samuelcolvin](https://github.com/samuelcolvin) •  :octicons-calendar-24: Jul 6, 2022 •  :octicons-clock-24: 10 min read

I've spoken to quite a few people about pydantic V2, and mention it in passing even more.

I think I owe people a proper explanation of the plan for V2:

• What will be added
• What will be removed
• What will change
• How I'm intending to go about completing it and getting it released
• Some idea of timeframe 😨

Here goes...

Plan & Timeframe

I'm currently taking a kind of sabbatical after leaving my last job to get pydantic V2 released. Why? I ask myself that question quite often. I'm very proud of how much pydantic is used, but I'm less proud of its internals. Since it's something people seem to care about and use quite a lot (26m downloads a month, used by 72k public repos, 10k stars). I want it to be as good as possible.

While I'm on the subject of why, how and my odd sabbatical: if you work for a large company who use pydantic a lot, you should encourage the company to sponsor me a meaningful amount, like Salesforce did. This is not charity, recruitment or marketing - the argument should be about how much the company will save if pydantic is 10x faster, more stable and more powerful - it would be worth paying me 10% of that to make it happen.

The plan is to have pydantic V2 released within 3 months of full-time work (again, that'll be sooner if I can continue to work on it full-time :face_with_raised_eyebrow:).

Before pydantic V2 can be released, we need to released pydantic V1.10 - there are lots of changes in the main branch of pydantic contributed by the community, it's only fair to provide a release including those changes, many of them will remain unchanged for V2, the rest will act as a requirement to make sure pydantic V2 includes the capabilities they implemented.

The basic road map for me is as follows:

1. Implement a few more critical features in pydantic-core
2. Release V0.1 of pydantic-core
3. Work on getting pydantic V1.10 out - basically merge all open PRs that are finished
4. Release pydantic V1.10
5. Delete all stale PRs which didn't make it into V1.10, apologise profusely to their authors who put their valuable time into pydantic only to have their PRs closed 🙏
6. Rename master to main, seems like a good time to do this
7. Change the main branch of pydantic to target V2
8. Start tearing pydantic code apart and see how many existing tests can be made to pass
9. Rinse, repeat
10. Release pydantic V2 🎉

Plan is to have all this done by the end of October, definitely by the end of the year.

Introduction

Pydantic began life as an experiment in some code in a long dead project. I ended up making the code into a package and releasing it. It got a bit of attention on hacker news when it was first released, but started to get really popular when Sebastián Ramírez used it in FastAPI. Since then the package and its usage have grown enormously. The core logic however has remained relatively unchanged since the initial experiment. It's old, it smells, it needs to be rebuilt.

The release of version 2 is an opportunity to rebuild pydantic and correct many things that don't make sense - to make pydantic amazing 🚀.

Much of the work on V2 is already done, but there's still a lot to do. Now seems a good opportunity to explain what V2 is going to look like and get feedback from users.

Headlines

For good and bad, here are some of the biggest changes expected in V2.

The core validation logic of pydantic V2 will be performed by a separate package pydantic-core which I've been building over the last few months.

pydantic-core is written in Rust using the excellent pyo3 library which provides rust bindings for python.

!!! note The python interface to pydantic shouldn't change as a result of using pydantic-core, instead pydantic will use type annotations to build a schema for pydantic-core to use.

pydantic-core is usable now, albeit with a fairly unintuitive API, if you're interested, please give it a try.

pydantic-core provides validators for all common data types, see a list here. Other, less commonly used data types will be supported via validator functions.

Performance 👍

As a result of the move to rust for the validation logic (and significant improvements in how validation objects are structured) pydantic V2 will be significantly faster than pydantic V1.

Looking at the pydantic-core benchmarks today, pydantic V2 is between 4x and 50x faster than pydantic V1.9.1.

In general, pydantic V2 is about 17x faster than V1 when validating a model containing a range of common fields.

Strict Mode 👍

People have long complained about pydantic for coercing data instead of throwing an error. E.g. input to an int field could be 123 or the string "123" which would be converted to 123.

pydantic-core comes with "strict mode" built in. With this only the exact data type is allowed, e.g. passing "123" to an int field would result in a validation error.

This will allow pydantic V2 to offer a strict switch which can be set on either a model or a field.

IsInstance checks 👍

Strict mode also means it makes sense to provide an is_instance method on validators which effectively run validation then throw away the result while avoiding the (admittedly small) overhead of creating and raising and error or returning the validation result.

Formalised Conversion Table 👍

As well as complaints about coercion, another (legitimate) complaint was inconsistency around data conversion.

In pydantic V2, the following principle will govern when data should be converted in "lax mode" (strict=False):

If the input data has a SINGLE and INTUITIVE representation, in the field's type, AND no data is lost during the conversion, then the data will be converted, Otherwise a validation error is raised. There is one exception to this rule: string fields - virtually all data has an intuitive representation as a string (e.g. repr() and str()), therefore a custom rule is required: only str, bytes and bytearray are valid as inputs to string fields.

Some examples of what that means in practice:

Field Type	Input	Single & Intuitive R.	data Loss	Result
int	"123"	:material-check:	:material-close:	Convert
int	123.0	:material-check:	:material-close:	Convert
int	123.1	:material-check:	:material-check:	Error
date	"2020-01-01"	:material-check:	:material-close:	Convert
date	"2020-01-01T12:00:00"	:material-check:	:material-check:	Error
int	b"1"	:material-close:	:material-close:	Error

(For the last case converting bytes to an int could reasonably mean int(bytes_data.decode()) or int.from_bytes(b'1', 'big'), hence an error)

In addition to the general rule, we'll provide a conversion table which defines exactly what data will be allowed to which field types. See the table below for a start on this.

Built in JSON support 👍

pydantic-core can parse JSON directly into a model or output type, this both improves performance and avoids issue with strictness - e.g. if you have a strict model with a datetime field, the input must be a datetime object, but clearly that makes no sense when parsing JSON which has no datatime type. Same with bytes and many other types.

Pydantic V2 will therefore allow some conversion when validating JSON directly, even in strict mode (e.g. ISO8601 string -> datetime, str -> bytes) even though this would not be allowed when validating a python object.

In future direct validation of JSON will also allow:

• parsing in a separate thread while starting validation in the main thread
• line numbers from JSON to be included in the validation errors

!!! note Pydantic has always had special support for JSON, that is not going to change.

While in theory other formats
could be specifically supported, the overheads are significant and I don't think there's another format that's
used widely enough to be worth specific logic. Other formats can be parsed to python then validated, similarly
when serialising, data can be exported to a python object, then serialised, 
see [below](#improvements-to-dumpingserializationexport).

Validation without a Model 👍

In pydantic V1 the core of all validation was a pydantic model, this led to significant overheads and complexity when the output data type was not a model.

pydantic-core operates on a tree of validators with no "model" type required at the base of the tree. It can therefore validate a single string or datetime value, a TypeDict or Model equally easily.

This feature will provide significant addition performance improvements in scenarios like:

• adding validation to dataclasses
• validating URL arguments, query strings, headers, etc. in FastAPI
• adding validation to TypedDict
• function argument validation

Basically anywhere were you don't care about a traditional model class instance.

We'll need to add standalone methods for generating json schema and dumping these objects to JSON etc.

Required vs. Nullable Cleanup 👍

Pydantic previously had a somewhat confused idea about "required" vs. "nullable". This mostly resulted from my misgivings about marking a field as Optional[int] but requiring a value to be provided but allowing it to be None.

In pydantic V2, pydantic will move to match dataclasses, thus:

from pydantic import BaseModel

class Foo(BaseModel):
    f1: str  # required, cannot be None
    f2: str | None  # required, can be None - same as Optional[str] / Union[str, None]
    f3: str | None = None  # optional, can be None
    f4: str = None  # optional, but cannot be None  

Validator Function Improvements 👍 👍 👍

This is one of the changes in pydantic V2 that I'm most excited about, I've been talking about something like this for a long time, see #1984, but couldn't find a way to do this until now.

Fields which use a function for validation can be any of the following types:

• function before mode - where the function is called before the inner validator is called
• function after mode - where the function is called after the inner validator is called
• plan mode - where there's no inner validator
• wrap mode - where the function takes a reference to a function which calls the inner validator, and can therefore modify the input before inner validation, modify the output after inner validation, conditionally not call the inner validator or catch errors from the inner validator and return a default value, or change the error

An example how a wrap validator might look:

from datetime import datetime
from pydantic import BaseModel, ValidationError, validator

class MyModel(BaseModel):
    timestamp: datetime

    @validator('timestamp', mode='wrap')
    def validate_timestamp(cls, v, handler):
        if v == 'now':
            # we don't want to bother with further validation, just return the value
            return datetime.now()
        try:
            return handler(v)
        except ValidationError:
            # validation failed, in this case we want to return a default value
            return datetime(2000, 1, 1)

As well as being powerful, this provides a great "escape hatch" when pydantic validation doesn't do what you want.

Improvements to Dumping/Serialization/Export 👍 😕

(I haven't worked on this yet, so these ideas are only provisional)

There has long been a debate about how to handle converting data when extracting it from a model. One of the features people have long requested is the ability to convert data to JSON compliant types while converting a model to a dict.

My plan is to move data export into pydantic-core, with that one implementation can support all export modes without compromising (and hopefully significantly improving) performance.

I see four different export/serialisation scenarios:

1. Extracting the field values of a model with no conversion, effectively model.__dict__ but with the current filtering logic provided by .dict()
2. Extracting the field values of a model recursively (effectively what .dict() does now) - sub-models are converted to dicts, but other fields remain unchanged.
3. Extracting data and converting at the same time (e.g. to JSON compliant types)
4. Serialising data straight to JSON

I think all 4 modes can be supported in a single implementation, with a kind of "3.5" mode where a python function is used to convert the data as the use wishes.

The current include and exclude logic is extremely complicated, hopefully it won't be too hard to translate it to rust.

We should also add support for validate_alias and dump_alias to allow for customising field keys.

Model namespace cleanup 👍

For years I've wanted to clean up the model namespace, see #1001. This would avoid confusing gotchas when field names clash with methods on a model, it would also make it safer to add more methods to a model without risking new clashes.

After much deliberation (and even giving a lightning talk at the python language submit about alternatives, see here) I've decided to go with the simplest and clearest approach, the expense of a bit more typing:

All methods on models with start with model_, fields names will not be allowed to start with "model", aliases can be used if required.

This will mean the following methods and attributes on a model:

• .__dict__ as currently, holds a dict of validated data
• .__fields_set__ as currently, set containing which fields were set (vs. populated from defaults)
• .model_validate() (currently .parse_obj()) - validate data
• .model_validate_json() (currently parse_raw(j..., content_type='application/json')) - validate data from JSON
• .model_dump() (currently .dict()) - as above, with new mode argument
• .model_json() (currently .json()) - alias of .model_dump(mode='json')
• .model_schema() (currently .schema())
• .model_schema_json() (currently .schema_json())
• .model_update_forward_refs() (currently .update_forward_refs()) - update forward references
• .model_copy() (currently .copy()) - copy a model
• .model_construct() (currently .construct()) - construct a model with no validation
• ._model_validator attribute holding the internal pydantic SchemaValidator
• .model_fields (currently .__fields__) - although the format will have to change a lot, might be an alias of ._model_validator.schema
• ModelConfig (currently Config) - configuration class for models

The following methods will be removed:

• .parse_file()
• .parse_raw()
• .from_orm() - the functionality has been configured as a Config property call from_attributes which can be set either on ModelConfig or on a specific ModelClass or TypedDict field

Strict API & API documentation 👍

When preparing a pydantic V2, we'll make a strict distinction between the public API and private functions & classes. Private objects clearly identified as private via _internal sub package to discourage use.

The public API will have API documentation. I've recently been working with the wonderful mkdocstrings for both dirty-equals and watchfiles documentation. I intend to use mkdocstrings to generate complete API documentation.

This wouldn't replace the current example-based somewhat informal documentation style byt instead will augment it.

Error descriptions 👍

The way line errors (the individual errors with a ValidationError) are built has become much more sophisticated in pydantic-core.

There's a well=defined set of error codes and messages.

More will be added when other type are validated via pure python validators in pydantic.

I would like to add a dedicated section to the documentation with extra information for each type of error.

This would be another key in a line error documentation, which would link to the appropriate section in the docs.

Thus, errors might look like:

[
    {
        'kind': 'greater_than_equal',
        'loc': ['age'],
        'message': 'Value must be greater than or equal to 18',
        'input_value': 11,
        'context': {'ge': 18},
        'documentation': 'https://pydantic.dev/errors/#greater_than_equal',
    },
    {
        'kind': 'bool_parsing',
        'loc': ['is_developer'],
        'message': 'Value must be a valid boolean, unable to interpret input',
        'input_value': 'foobar',
        'documentation': 'https://pydantic.dev/errors/#bool_parsing',
    },
]

(I own the pydantic.dev domain and will use it for at least these errors so that even if the docs URL changes, the error will still link to the correct place.)

No pure python implementation 😦

Since pydantic-core is written in Rust, and I have absolutely no intention of rewriting it in python, pydantic V2 will only work where a binary package can be installed.

pydantic-core will provide binaries in PyPI for (at least):

• Linux: x86_64, aarch64, i686, armv7l, musl-x86_64 & musl-aarch64
• MacOS: x86_64 & arm64 (except python 3.7)
• Windows: amd64 & win32
• Web Assembly: wasm32 (pydantic-core is already compiled for wasm32 using emscripten and unit tests pass, except where cpython itself has problems)

Other binaries can be added provided they can be (cross-)compiled on github actions. If no binary is available from PyPI, pydantic-core can be compiled from source if Rust stable is available.

The only place where I know this will cause problems is Raspberry Pi, which is a bit of mess when it comes to packages written in rust for python. Effectively, until that's fixed you'll likely have to install pydantic-core with pip install -i https://pypi.org/simple/ pydantic-core.

Pydantic becomes a pure python package 😐

Pydantic V1.X is a pure python code base but is compiled with cython to provide some performance improvements. Since the "hot" code is moved to pydantic-core, pydantic itself can go back to being a pure python package.

This should significantly reduce the size of the pydantic package and make unit tests of pydantic much faster. In addition, some constraints on pydantic code can be removed once it no-longer has to be compilable with cython.

Some pieces of edge logic could get a little slower as they're no longer compiled.

I'm dropping the word "parse" and just using "validate" 😐

Partly due to the issues with the lack of strict mode. I've previously gone back and forth between using "parse" and "validate" for what pydantic does.

While pydantic is not simply a validation library (and I'm sure some would argue validation is not strictly what it does), most people use the validation.

it's time to stop fighting that, use consistent names.

The word "parse" will no longer be used except when talking about JSON parsing.

Changes to custom field types 😐

Since the core structure of validators has changed from "a list of validators to call on each field" to "a tree of validators which call each other", the __get_validators__ way of defining custom field types no longer makes sense.

Instead we'll look for the attribute __pydantic_schema__ which must be a pydantic-core compliant schema for validating data to this field type, except for the function item which can be a string which should be the name of a class function - validate below.

Here's an example of how a custom field type could be defined:

from pydantic import ValiationSchema

class Foobar:
    __pydantic_schema__: ValiationSchema = {
        'type': 'function',
        'mode': 'after',
        'function': 'validate',
        'schema': {'type': 'str'}
    }
    
    def __init__(self, value: str):
        self.value = value

    @classmethod
    def validate(cls, value):
        if 'foobar' in value:
            return Foobar(value)
        else:
            raise ValueError('expected foobar')

What's going on here: __pydantic_schema__ defines a schema which effectively says:

Validate input data as a string, then call the validate with that string, use the output of validate as the final result of validation.

ValiationSchema is just an alias to pydantic_core.Schema which is a type defining the schema for validation schemas.

!!! note pydantic-core schema has full type definitions although since the type is recursive, mypy can't provide static type analysis, pyright however can.

Other Improvements 👍

1. Recursive models with cyclic references - although recursive models were supported by pydantic V1, data with cyclic references cause recursion errors, in pydantic-core code is correctly detected and a validation error is raised
2. The reason I've been so keen to get pydantic-core to compile and run with wasm is that I want all examples in the docs of pydantic V2 to be editable and runnable in the browser
3. Full (pun intended) support for TypedDict, including full=False - e.g. omitted keys
4. from_orm has become from_attributes and is now defined at schema generation time (either via model config or field config)
5. input_value has been added to each line error in a ValidationError which should provide.
6. on_error logic in a schema which allows either a default value to be used in the event of an error, or that value to be omitted (in the case of a full=False TypeDict), #151
7. datetime, date, time & timedelta validation is improved, see the speedate rust library I built specifically for this purpose for more details
8. Powerful "priority" system for optionally merging or overriding config in sub-models

Removed Features 😐

1. __root__ custom root models are no longer necessary since validation on any support data type is supported without a model
2. parse_file and parse_raw, partially replaced with .model_validate_json()
3. TypeError are no longer considered as validation errors, but rather as internal errors, this is to better catch errors in argument names in function validators.
4. Subclasses of builtin types like str, bytes and int are coerced to their parent builtin type, this is a limitation of how pydantic-core converts these types to rust types during validation, if you have a specific need to keep the type, you can use wrap validators (see above) or custom type validation

Conversion Table

The table below provisionally defines what times are allowed to which field types.

An updated and complete version of this table will be included in the docs for V2.

Field Type	Input	Mode	Input Source	Conditions
str	str	both	python, JSON	-
str	bytes	lax	python	assumes UTF-8, error on unicode decoding error
str	bytearray	lax	python	assumes UTF-8, error on unicode decoding error
bytes	bytes	both	python	-
bytes	str	both	JSON	-
bytes	str	lax	python	-
bytes	bytearray	lax	python	-
int	int	strict	python, JSON	max abs value 2^64 - i64 is used internally, bool explicitly forbidden
int	int	lax	python, JSON	i64
int	float	lax	python, JSON	i64, must be exact int, e.g. f % 1 == 0, nan, inf raise errors
int	Decimal	lax	python, JSON	i64, must be exact int, e.g. f % 1 == 0
int	bool	lax	python, JSON	-
int	str	lax	python, JSON	i64, must be numeric only, e.g. [0-9]+
float	float	strict	python, JSON	bool explicitly forbidden
float	float	lax	python, JSON	-
float	int	lax	python, JSON	-
float	str	lax	python, JSON	must match [0-9]+(\.[0-9]+)?
float	Decimal	lax	python	-
float	bool	lax	python, JSON	-
bool	bool	both	python, JSON	-
bool	int	lax	python, JSON	allowed: 0, 1
bool	float	lax	python, JSON	allowed: 0, 1
bool	Decimal	lax	python, JSON	allowed: 0, 1
bool	str	lax	python, JSON	allowed: 'f', 'n', 'no', 'off', 'false', 't', 'y', 'on', 'yes', 'true'
None	None	both	python, JSON	-
date	date	both	python	-
date	datetime	lax	python	must be exact date, eg. no H, M, S, f
date	str	both	JSON	format YYYY-MM-DD
date	str	lax	python	format YYYY-MM-DD
date	bytes	lax	python	format YYYY-MM-DD (UTF-8)
date	int	lax	python, JSON	interpreted as seconds or ms from epoch, see speedate, must be exact date
date	float	lax	python, JSON	interpreted as seconds or ms from epoch, see speedate, must be exact date
datetime	datetime	both	python	-
datetime	date	lax	python	-
datetime	str	both	JSON	format YYYY-MM-DDTHH:MM:SS.f etc. see speedate
datetime	str	lax	python	format YYYY-MM-DDTHH:MM:SS.f etc. see speedate
datetime	bytes	lax	python	format YYYY-MM-DDTHH:MM:SS.f etc. see speedate, (UTF-8)
datetime	int	lax	python, JSON	interpreted as seconds or ms from epoch, see speedate
datetime	float	lax	python, JSON	interpreted as seconds or ms from epoch, see speedate
time	time	both	python	-
time	str	both	JSON	format HH:MM:SS.FFFFFF etc. see speedate
time	str	lax	python	format HH:MM:SS.FFFFFF etc. see speedate
time	bytes	lax	python	format HH:MM:SS.FFFFFF etc. see speedate, (UTF-8)
time	int	lax	python, JSON	interpreted as seconds, range 0 - 86399
time	float	lax	python, JSON	interpreted as seconds, range 0 - 86399.9*
time	Decimal	lax	python, JSON	interpreted as seconds, range 0 - 86399.9*
timedelta	timedelta	both	python	-
timedelta	str	both	JSON	format ISO8601 etc. see speedate
timedelta	str	lax	python	format ISO8601 etc. see speedate
timedelta	bytes	lax	python	format ISO8601 etc. see speedate, (UTF-8)
timedelta	int	lax	python, JSON	interpreted as seconds
timedelta	float	lax	python, JSON	interpreted as seconds
timedelta	Decimal	lax	python, JSON	interpreted as seconds
dict	dict	both	python	-
dict	Object	both	JSON	-
dict	mapping	lax	python	must implement the mapping interface and have an items() method
TypeDict	dict	both	python	-
TypeDict	Object	both	JSON	-
TypeDict	Any	both	python	builtins not allowed, uses getattr, requires from_attributes=True
TypeDict	mapping	lax	python	must implement the mapping interface and have an items() method
list	list	both	python	-
list	Array	both	JSON	-
list	tuple	lax	python	-
list	set	lax	python	-
list	frozenset	lax	python	-
list	dict_keys	lax	python	-
tuple	tuple	both	python	-
tuple	Array	both	JSON	-
tuple	list	lax	python	-
tuple	set	lax	python	-
tuple	frozenset	lax	python	-
tuple	dict_keys	lax	python	-
set	set	both	python	-
set	Array	both	JSON	-
set	list	lax	python	-
set	tuple	lax	python	-
set	frozenset	lax	python	-
set	dict_keys	lax	python	-
frozenset	frozenset	both	python	-
frozenset	Array	both	JSON	-
frozenset	list	lax	python	-
frozenset	tuple	lax	python	-
frozenset	set	lax	python	-
frozenset	dict_keys	lax	python	-
is_instance	Any	both	python	isinstance() check returns True
is_instance	-	both	JSON	never valid
callable	Any	both	python	callable() check returns True
callable	-	both	JSON	never valid

!!!note The ModelClass validator (use to create instances of a class) uses the TypeDict validator, then creates an instance with __dict__ and __fields_set__ set, so same rules apply as TypeDict.