PEP: 557

Title: Data Classes

Author: Eric V. Smith <eric@trueblade.com>

Status: Final

Type: Standards Track

Created: 02-Jun-2017

Python-Version: 3.7

Post-History: 08-Sep-2017, 25-Nov-2017, 30-Nov-2017, 01-Dec-2017, 02-Dec-2017, 06-Jan-2018, 04-Mar-2018

Resolution: https://mail.python.org/pipermail/python-dev/2017-December/151034.html

Notice for Reviewers

This PEP and the initial implementation were drafted in a separate

repo: https://github.com/ericvsmith/dataclasses. Before commenting in

a public forum please at least read the discussion_ listed at the

end of this PEP.

Abstract

This PEP describes an addition to the standard library called Data

Classes. Although they use a very different mechanism, Data Classes

can be thought of as "mutable namedtuples with defaults". Because

Data Classes use normal class definition syntax, you are free to use

inheritance, metaclasses, docstrings, user-defined methods, class

factories, and other Python class features.

A class decorator is provided which inspects a class definition for

variables with type annotations as defined in :pep:`526`, "Syntax for

Variable Annotations". In this document, such variables are called

fields. Using these fields, the decorator adds generated method

definitions to the class to support instance initialization, a repr,

comparison methods, and optionally other methods as described in the

Specification_ section. Such a class is called a Data Class, but

there's really nothing special about the class: the decorator adds

generated methods to the class and returns the same class it was

given.

As an example::

The ``@dataclass`` decorator will add the equivalent of these methods

to the InventoryItem class::

Data Classes save you from writing and maintaining these methods.

Rationale

There have been numerous attempts to define classes which exist

primarily to store values which are accessible by attribute lookup.

Some examples include:

- collections.namedtuple in the standard library.

- typing.NamedTuple in the standard library.

- The popular attrs [#]_ project.

- George Sakkis' recordType recipe [#]_, a mutable data type inspired

by collections.namedtuple.

- Many example online recipes [#]_, packages [#]_, and questions [#]_.

David Beazley used a form of data classes as the motivating example

in a PyCon 2013 metaclass talk [#]_.

So, why is this PEP needed?

With the addition of :pep:`526`, Python has a concise way to specify the

type of class members. This PEP leverages that syntax to provide a

simple, unobtrusive way to describe Data Classes. With two exceptions,

the specified attribute type annotation is completely ignored by Data

Classes.

No base classes or metaclasses are used by Data Classes. Users of

these classes are free to use inheritance and metaclasses without any

interference from Data Classes. The decorated classes are truly

"normal" Python classes. The Data Class decorator should not

interfere with any usage of the class.

One main design goal of Data Classes is to support static type

checkers. The use of :pep:`526` syntax is one example of this, but so is

the design of the ``fields()`` function and the ``@dataclass``

decorator. Due to their very dynamic nature, some of the libraries

mentioned above are difficult to use with static type checkers.

Data Classes are not, and are not intended to be, a replacement

mechanism for all of the above libraries. But being in the standard

library will allow many of the simpler use cases to instead leverage

Data Classes. Many of the libraries listed have different feature

sets, and will of course continue to exist and prosper.

Where is it not appropriate to use Data Classes?

- API compatibility with tuples or dicts is required.

- Type validation beyond that provided by PEPs 484 and 526 is

required, or value validation or conversion is required.

Specification

All of the functions described in this PEP will live in a module named

``dataclasses``.

A function ``dataclass`` which is typically used as a class decorator

is provided to post-process classes and add generated methods,

described below.

The ``dataclass`` decorator examines the class to find ``field``\s. A

``field`` is defined as any variable identified in

``__annotations__``. That is, a variable that has a type annotation.

With two exceptions described below, none of the Data Class machinery

examines the type specified in the annotation.

Note that ``__annotations__`` is guaranteed to be an ordered mapping,

in class declaration order. The order of the fields in all of the

generated methods is the order in which they appear in the class.

The ``dataclass`` decorator will add various "dunder" methods to the

class, described below. If any of the added methods already exist on the

class, a ``TypeError`` will be raised. The decorator returns the same

class that is called on: no new class is created.

The ``dataclass`` decorator is typically used with no parameters and

no parentheses. However, it also supports the following logical

signature::

If ``dataclass`` is used just as a simple decorator with no

parameters, it acts as if it has the default values documented in this

signature. That is, these three uses of ``@dataclass`` are equivalent::

The parameters to ``dataclass`` are:

- ``init``: If true (the default), a ``__init__`` method will be

generated.

- ``repr``: If true (the default), a ``__repr__`` method will be

generated. The generated repr string will have the class name and

the name and repr of each field, in the order they are defined in

the class. Fields that are marked as being excluded from the repr

are not included. For example:

``InventoryItem(name='widget', unit_price=3.0, quantity_on_hand=10)``.

If the class already defines ``__repr__``, this parameter is

ignored.

- ``eq``: If true (the default), an ``__eq__`` method will be

generated. This method compares the class as if it were a tuple of its

fields, in order. Both instances in the comparison must be of the

identical type.

If the class already defines ``__eq__``, this parameter is ignored.

- ``order``: If true (the default is False), ``__lt__``, ``__le__``,

``__gt__``, and ``__ge__`` methods will be generated. These compare

the class as if it were a tuple of its fields, in order. Both

instances in the comparison must be of the identical type. If

``order`` is true and ``eq`` is false, a ``ValueError`` is raised.

If the class already defines any of ``__lt__``, ``__le__``,

``__gt__``, or ``__ge__``, then ``ValueError`` is raised.

- ``unsafe_hash``: If ``False`` (the default), the ``__hash__`` method

is generated according to how ``eq`` and ``frozen`` are set.

If ``eq`` and ``frozen`` are both true, Data Classes will generate a

``__hash__`` method for you. If ``eq`` is true and ``frozen`` is

false, ``__hash__`` will be set to ``None``, marking it unhashable

(which it is). If ``eq`` is false, ``__hash__`` will be left

untouched meaning the ``__hash__`` method of the superclass will be

used (if the superclass is ``object``, this means it will fall back

to id-based hashing).

Although not recommended, you can force Data Classes to create a

``__hash__`` method with ``unsafe_hash=True``. This might be the

case if your class is logically immutable but can nonetheless be

mutated. This is a specialized use case and should be considered

carefully.

If a class already has an explicitly defined ``__hash__`` the

behavior when adding ``__hash__`` is modified. An explicitly

defined ``__hash__`` is defined when:

- ``__eq__`` is defined in the class and ``__hash__`` is defined

with any value other than ``None``.

- ``__eq__`` is defined in the class and any non-``None``

``__hash__`` is defined.

- ``__eq__`` is not defined on the class, and any ``__hash__`` is

defined.

If ``unsafe_hash`` is true and an explicitly defined ``__hash__``

is present, then ``ValueError`` is raised.

If ``unsafe_hash`` is false and an explicitly defined ``__hash__``

is present, then no ``__hash__`` is added.

See the Python documentation [#]_ for more information.

- ``frozen``: If true (the default is False), assigning to fields will

generate an exception. This emulates read-only frozen instances.

If either ``__getattr__`` or ``__setattr__`` is defined in the

class, then ``ValueError`` is raised. See the discussion below.

``field``\s may optionally specify a default value, using normal

Python syntax::

In this example, both ``a`` and ``b`` will be included in the added

``__init__`` method, which will be defined as::

``TypeError`` will be raised if a field without a default value

follows a field with a default value. This is true either when this

occurs in a single class, or as a result of class inheritance.

For common and simple use cases, no other functionality is required.

There are, however, some Data Class features that require additional

per-field information. To satisfy this need for additional

information, you can replace the default field value with a call to

the provided ``field()`` function. The signature of ``field()`` is::

The ``MISSING`` value is a sentinel object used to detect if the

``default`` and ``default_factory`` parameters are provided. This

sentinel is used because ``None`` is a valid value for ``default``.

The parameters to ``field()`` are:

- ``default``: If provided, this will be the default value for this

field. This is needed because the ``field`` call itself replaces

the normal position of the default value.

- ``default_factory``: If provided, it must be a zero-argument

callable that will be called when a default value is needed for this

field. Among other purposes, this can be used to specify fields

with mutable default values, as discussed below. It is an error to

specify both ``default`` and ``default_factory``.

- ``init``: If true (the default), this field is included as a

parameter to the generated ``__init__`` method.

- ``repr``: If true (the default), this field is included in the

string returned by the generated ``__repr__`` method.

- ``compare``: If True (the default), this field is included in the

generated equality and comparison methods (``__eq__``, ``__gt__``,

et al.).

- ``hash``: This can be a bool or ``None``. If True, this field is

included in the generated ``__hash__`` method. If ``None`` (the

default), use the value of ``compare``: this would normally be the

expected behavior. A field should be considered in the hash if

it's used for comparisons. Setting this value to anything other

than ``None`` is discouraged.

One possible reason to set ``hash=False`` but ``compare=True`` would

be if a field is expensive to compute a hash value for, that field

is needed for equality testing, and there are other fields that

contribute to the type's hash value. Even if a field is excluded

from the hash, it will still be used for comparisons.

- ``metadata``: This can be a mapping or None. None is treated as an

empty dict. This value is wrapped in ``types.MappingProxyType`` to

make it read-only, and exposed on the Field object. It is not used

at all by Data Classes, and is provided as a third-party extension

mechanism. Multiple third-parties can each have their own key, to

use as a namespace in the metadata.

If the default value of a field is specified by a call to ``field()``,

then the class attribute for this field will be replaced by the

specified ``default`` value. If no ``default`` is provided, then the

class attribute will be deleted. The intent is that after the

``dataclass`` decorator runs, the class attributes will all contain

the default values for the fields, just as if the default value itself

were specified. For example, after::

The class attribute ``C.z`` will be ``10``, the class attribute

``C.t`` will be ``20``, and the class attributes ``C.x`` and ``C.y``

will not be set.

``Field`` objects

``Field`` objects describe each defined field. These objects are

created internally, and are returned by the ``fields()`` module-level

method (see below). Users should never instantiate a ``Field``

object directly. Its documented attributes are:

- ``name``: The name of the field.

- ``type``: The type of the field.

- ``default``, ``default_factory``, ``init``, ``repr``, ``hash``,

``compare``, and ``metadata`` have the identical meaning and values

as they do in the ``field()`` declaration.

Other attributes may exist, but they are private and must not be

inspected or relied on.

post-init processing

The generated ``__init__`` code will call a method named

``__post_init__``, if it is defined on the class. It will be called

as ``self.__post_init__()``. If no ``__init__`` method is generated,

then ``__post_init__`` will not automatically be called.

Among other uses, this allows for initializing field values that

depend on one or more other fields. For example::

See the section below on init-only variables for ways to pass

parameters to ``__post_init__()``. Also see the warning about how

``replace()`` handles ``init=False`` fields.

Class variables

One place where ``dataclass`` actually inspects the type of a field is

to determine if a field is a class variable as defined in :pep:`526`. It

does this by checking if the type of the field is ``typing.ClassVar``.

If a field is a ``ClassVar``, it is excluded from consideration as a

field and is ignored by the Data Class mechanisms. For more

discussion, see [#]_. Such ``ClassVar`` pseudo-fields are not

returned by the module-level ``fields()`` function.

Init-only variables

The other place where ``dataclass`` inspects a type annotation is to

determine if a field is an init-only variable. It does this by seeing

if the type of a field is of type ``dataclasses.InitVar``. If a field

is an ``InitVar``, it is considered a pseudo-field called an init-only

field. As it is not a true field, it is not returned by the

module-level ``fields()`` function. Init-only fields are added as

parameters to the generated ``__init__`` method, and are passed to

the optional ``__post_init__`` method. They are not otherwise used

by Data Classes.

For example, suppose a field will be initialized from a database, if a

value is not provided when creating the class::

In this case, ``fields()`` will return ``Field`` objects for ``i`` and

``j``, but not for ``database``.

Frozen instances

It is not possible to create truly immutable Python objects. However,

by passing ``frozen=True`` to the ``@dataclass`` decorator you can

emulate immutability. In that case, Data Classes will add

``__setattr__`` and ``__delattr__`` methods to the class. These

methods will raise a ``FrozenInstanceError`` when invoked.

There is a tiny performance penalty when using ``frozen=True``:

``__init__`` cannot use simple assignment to initialize fields, and

must use ``object.__setattr__``.

Inheritance

When the Data Class is being created by the ``@dataclass`` decorator,

it looks through all of the class's base classes in reverse MRO (that

is, starting at ``object``) and, for each Data Class that it finds,

adds the fields from that base class to an ordered mapping of fields.

After all of the base class fields are added, it adds its own fields

to the ordered mapping. All of the generated methods will use this

combined, calculated ordered mapping of fields. Because the fields

are in insertion order, derived classes override base classes. An

example::

The final list of fields is, in order, ``x``, ``y``, ``z``. The final

type of ``x`` is ``int``, as specified in class ``C``.

The generated ``__init__`` method for ``C`` will look like::

Default factory functions

If a field specifies a ``default_factory``, it is called with zero

arguments when a default value for the field is needed. For example,

to create a new instance of a list, use::

If a field is excluded from ``__init__`` (using ``init=False``) and

the field also specifies ``default_factory``, then the default factory

function will always be called from the generated ``__init__``

function. This happens because there is no other way to give the

field an initial value.

Mutable default values

Python stores default member variable values in class attributes.

Consider this example, not using Data Classes::

Note that the two instances of class ``C`` share the same class

variable ``x``, as expected.

Using Data Classes, *if* this code was valid::

it would generate code similar to::

This has the same issue as the original example using class ``C``.

That is, two instances of class ``D`` that do not specify a value for

``x`` when creating a class instance will share the same copy of

``x``. Because Data Classes just use normal Python class creation

they also share this problem. There is no general way for Data

Classes to detect this condition. Instead, Data Classes will raise a

``TypeError`` if it detects a default parameter of type ``list``,

``dict``, or ``set``. This is a partial solution, but it does protect

against many common errors. See `Automatically support mutable

default values`_ in the Rejected Ideas section for more details.

Using default factory functions is a way to create new instances of

mutable types as default values for fields::

Module level helper functions

- ``fields(class_or_instance)``: Returns a tuple of ``Field`` objects

that define the fields for this Data Class. Accepts either a Data

Class, or an instance of a Data Class. Raises ``ValueError`` if not

passed a Data Class or instance of one. Does not return

pseudo-fields which are ``ClassVar`` or ``InitVar``.

- ``asdict(instance, *, dict_factory=dict)``: Converts the Data Class

``instance`` to a dict (by using the factory function

``dict_factory``). Each Data Class is converted to a dict of its

fields, as name:value pairs. Data Classes, dicts, lists, and tuples

are recursed into. For example::

Raises ``TypeError`` if ``instance`` is not a Data Class instance.

- ``astuple(*, tuple_factory=tuple)``: Converts the Data Class

``instance`` to a tuple (by using the factory function

``tuple_factory``). Each Data Class is converted to a tuple of its

field values. Data Classes, dicts, lists, and tuples are recursed

into.

Continuing from the previous example::

Raises ``TypeError`` if ``instance`` is not a Data Class instance.

- ``make_dataclass(cls_name, fields, *, bases=(), namespace=None)``:

Creates a new Data Class with name ``cls_name``, fields as defined

in ``fields``, base classes as given in ``bases``, and initialized

with a namespace as given in ``namespace``. ``fields`` is an

iterable whose elements are either ``name``, ``(name, type)``, or

``(name, type, Field)``. If just ``name`` is supplied,

``typing.Any`` is used for ``type``. This function is not strictly

required, because any Python mechanism for creating a new class with

``__annotations__`` can then apply the ``dataclass`` function to

convert that class to a Data Class. This function is provided as a

convenience. For example::

Is equivalent to::

- ``replace(instance, **changes)``: Creates a new object of the same

type of ``instance``, replacing fields with values from ``changes``.

If ``instance`` is not a Data Class, raises ``TypeError``. If

values in ``changes`` do not specify fields, raises ``TypeError``.

The newly returned object is created by calling the ``__init__``

method of the Data Class. This ensures that

``__post_init__``, if present, is also called.

Init-only variables without default values, if any exist, must be

specified on the call to ``replace`` so that they can be passed to

``__init__`` and ``__post_init__``.

It is an error for ``changes`` to contain any fields that are

defined as having ``init=False``. A ``ValueError`` will be raised

in this case.

Be forewarned about how ``init=False`` fields work during a call to

``replace()``. They are not copied from the source object, but

rather are initialized in ``__post_init__()``, if they're

initialized at all. It is expected that ``init=False`` fields will

be rarely and judiciously used. If they are used, it might be wise

to have alternate class constructors, or perhaps a custom

``replace()`` (or similarly named) method which handles instance

copying.

- ``is_dataclass(class_or_instance)``: Returns True if its parameter

is a dataclass or an instance of one, otherwise returns False.

If you need to know if a class is an instance of a dataclass (and

not a dataclass itself), then add a further check for ``not

isinstance(obj, type)``::

def is_dataclass_instance(obj):

return is_dataclass(obj) and not isinstance(obj, type)

Discussion

python-ideas discussion

This discussion started on python-ideas [#]_ and was moved to a GitHub

repo [#]_ for further discussion. As part of this discussion, we made

the decision to use :pep:`526` syntax to drive the discovery of fields.

Support for automatically setting ``__slots__``?

At least for the initial release, ``__slots__`` will not be supported.

``__slots__`` needs to be added at class creation time. The Data

Class decorator is called after the class is created, so in order to

add ``__slots__`` the decorator would have to create a new class, set

``__slots__``, and return it. Because this behavior is somewhat

surprising, the initial version of Data Classes will not support

automatically setting ``__slots__``. There are a number of

workarounds:

- Manually add ``__slots__`` in the class definition.

- Write a function (which could be used as a decorator) that inspects

the class using ``fields()`` and creates a new class with

``__slots__`` set.

For more discussion, see [#]_.

Why not just use namedtuple?

- Any namedtuple can be accidentally compared to any other with the

same number of fields. For example: ``Point3D(2017, 6, 2) ==

Date(2017, 6, 2)``. With Data Classes, this would return False.

- A namedtuple can be accidentally compared to a tuple. For example,

``Point2D(1, 10) == (1, 10)``. With Data Classes, this would return

False.

- Instances are always iterable, which can make it difficult to add

fields. If a library defines::

Then if a user uses this code as::

then it would not be possible to add a ``second`` field to ``Time``

without breaking the user's code.

- No option for mutable instances.

- Cannot specify default values.

- Cannot control which fields are used for ``__init__``, ``__repr__``,

etc.

- Cannot support combining fields by inheritance.

Why not just use typing.NamedTuple?

For classes with statically defined fields, it does support similar

syntax to Data Classes, using type annotations. This produces a

namedtuple, so it shares ``namedtuple``\s benefits and some of its

downsides. Data Classes, unlike ``typing.NamedTuple``, support

combining fields via inheritance.

Why not just use attrs?

- attrs moves faster than could be accommodated if it were moved in to

the standard library.

- attrs supports additional features not being proposed here:

validators, converters, metadata, etc. Data Classes makes a

tradeoff to achieve simplicity by not implementing these

features.

For more discussion, see [#]_.

post-init parameters

In an earlier version of this PEP before ``InitVar`` was added, the

post-init function ``__post_init__`` never took any parameters.

The normal way of doing parameterized initialization (and not just

with Data Classes) is to provide an alternate classmethod constructor.

For example::

Because the ``__post_init__`` function is the last thing called in the

generated ``__init__``, having a classmethod constructor (which can

also execute code immediately after constructing the object) is

functionally equivalent to being able to pass parameters to a

``__post_init__`` function.

With ``InitVar``\s, ``__post_init__`` functions can now take

parameters. They are passed first to ``__init__`` which passes them

to ``__post_init__`` where user code can use them as needed.

The only real difference between alternate classmethod constructors

and ``InitVar`` pseudo-fields is in regards to required non-field

parameters during object creation. With ``InitVar``\s, using

``__init__`` and the module-level ``replace()`` function ``InitVar``\s

must always be specified. Consider the case where a ``context``

object is needed to create an instance, but isn't stored as a field.

With alternate classmethod constructors the ``context`` parameter is

always optional, because you could still create the object by going

through ``__init__`` (unless you suppress its creation). Which

approach is more appropriate will be application-specific, but both

approaches are supported.

Another reason for using ``InitVar`` fields is that the class author

can control the order of ``__init__`` parameters. This is especially

important with regular fields and ``InitVar`` fields that have default

values, as all fields with defaults must come after all fields without

defaults. A previous design had all init-only fields coming after

regular fields. This meant that if any field had a default value,

then all init-only fields would have to have defaults values, too.

asdict and astuple function names

The names of the module-level helper functions ``asdict()`` and

``astuple()`` are arguably not :pep:`8` compliant, and should be

``as_dict()`` and ``as_tuple()``, respectively. However, after

discussion [#]_ it was decided to keep consistency with

``namedtuple._asdict()`` and ``attr.asdict()``.

Rejected ideas

Copying ``init=False`` fields after new object creation in replace()

Fields that are ``init=False`` are by definition not passed to

``__init__``, but instead are initialized with a default value, or by

calling a default factory function in ``__init__``, or by code in

``__post_init__``.

A previous version of this PEP specified that ``init=False`` fields

would be copied from the source object to the newly created object

after ``__init__`` returned, but that was deemed to be inconsistent

with using ``__init__`` and ``__post_init__`` to initialize the new

object. For example, consider this case::

If ``init=False`` fields were copied from the source to the

destination object after ``__post_init__`` is run, then s2 would end

up begin ``Square(length=2.0, area=1.0)``, instead of the correct

``Square(length=2.0, area=4.0)``.

Automatically support mutable default values

One proposal was to automatically copy defaults, so that if a literal

list ``[]`` was a default value, each instance would get a new list.

There were undesirable side effects of this decision, so the final

decision is to disallow the 3 known built-in mutable types: list,

dict, and set. For a complete discussion of this and other options,

see [#]_.

Examples

Custom __init__ method

Sometimes the generated ``__init__`` method does not suffice. For

example, suppose you wanted to have an object to store ``*args`` and

``**kwargs``::

A complicated example

This code exists in a closed source project::

This can be replaced by::

The Data Class version is more declarative, has less code, supports

``typing``, and includes the other generated functions.

Acknowledgements

The following people provided invaluable input during the development

of this PEP and code: Ivan Levkivskyi, Guido van Rossum, Hynek

Schlawack, Raymond Hettinger, and Lisa Roach. I thank them for their

time and expertise.

A special mention must be made about the ``attrs`` project. It was a

true inspiration for this PEP, and I respect the design decisions they

made.

References

.. [#] attrs project on github

.. [#] George Sakkis' recordType recipe

.. [#] DictDotLookup recipe

.. [#] attrdict package

.. [#] StackOverflow question about data container classes

.. [#] David Beazley metaclass talk featuring data classes

.. [#] Python documentation for __hash__

.. [#] :pep:`ClassVar discussion in PEP 526 <526#class-and-instance-variable-annotations>`

.. [#] Start of python-ideas discussion

.. [#] GitHub repo where discussions and initial development took place

.. [#] Support __slots__?

.. [#] why not just attrs?

.. [#] :pep:`8` names for asdict and astuple

.. [#] Copying mutable defaults

Copyright

This document has been placed in the public domain.