Notice: While Javascript is not essential for this website, your interaction with the content will be limited. Please turn Javascript on for the full experience.

PEP 557 -- Data Classes

PEP:557
Title:Data Classes
Author:Eric V. Smith <eric at trueblade.com>
Status:Accepted
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
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:

@dataclass
class InventoryItem:
    '''Class for keeping track of an item in inventory.'''
    name: str
    unit_price: float
    quantity_on_hand: int = 0

    def total_cost(self) -> float:
        return self.unit_price * self.quantity_on_hand

The @dataclass decorator will add the equivalent of these methods to the InventoryItem class:

def __init__(self, name: str, unit_price: float, quantity_on_hand: int = 0) -> None:
    self.name = name
    self.unit_price = unit_price
    self.quantity_on_hand = quantity_on_hand
def __repr__(self):
    return f'InventoryItem(name={self.name!r}, unit_price={self.unit_price!r}, quantity_on_hand={self.quantity_on_hand!r})'
def __eq__(self, other):
    if other.__class__ is self.__class__:
        return (self.name, self.unit_price, self.quantity_on_hand) == (other.name, other.unit_price, other.quantity_on_hand)
    return NotImplemented
def __ne__(self, other):
    if other.__class__ is self.__class__:
        return (self.name, self.unit_price, self.quantity_on_hand) != (other.name, other.unit_price, other.quantity_on_hand)
    return NotImplemented
def __lt__(self, other):
    if other.__class__ is self.__class__:
        return (self.name, self.unit_price, self.quantity_on_hand) < (other.name, other.unit_price, other.quantity_on_hand)
    return NotImplemented
def __le__(self, other):
    if other.__class__ is self.__class__:
        return (self.name, self.unit_price, self.quantity_on_hand) <= (other.name, other.unit_price, other.quantity_on_hand)
    return NotImplemented
def __gt__(self, other):
    if other.__class__ is self.__class__:
        return (self.name, self.unit_price, self.quantity_on_hand) > (other.name, other.unit_price, other.quantity_on_hand)
    return NotImplemented
def __ge__(self, other):
    if other.__class__ is self.__class__:
        return (self.name, self.unit_price, self.quantity_on_hand) >= (other.name, other.unit_price, other.quantity_on_hand)
    return NotImplemented

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:

  • collection.namedtuple in the standard library.
  • typing.NamedTuple in the standard library.
  • The popular attrs [1] project.
  • George Sakkis' recordType recipe [2], a mutable data type inspired by collections.namedtuple.
  • Many example online recipes [3], packages [4], and questions [5]. David Beazley used a form of data classes as the motivating example in a PyCon 2013 metaclass talk [6].

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 fields. 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:

def dataclass(*, init=True, repr=True, eq=True, order=False, hash=None, frozen=False)

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:

@dataclass
class C:
    ...

@dataclass()
class C:
    ...

@dataclass(init=True, repr=True, eq=True, order=False, hash=None, frozen=False)
class C:
    ...

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).

  • eq: If true (the default), __eq__ and __ne__ 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.

  • 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.

  • hash: Either a bool or None. If None (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 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.

    See the Python documentation [7] for more information.

  • frozen: If true (the default is False), assigning to fields will generate an exception. This emulates read-only frozen instances. See the discussion below.

fields may optionally specify a default value, using normal Python syntax:

@dataclass
class C:
    a: int       # 'a' has no default value
    b: int = 0   # assign a default value for 'b'

In this example, both a and b will be included in the added __init__ method, which will be defined as:

def __init__(self, a: int, b: int = 0):

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:

def field(*, default=_MISSING, default_factory=_MISSING, repr=True,
          hash=None, init=True, compare=True, metadata=None)

The _MISSING value is a sentinel object used to detect if the default and default_factory parameters are provided. Users should never use _MISSING or depend on its value. 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:

@dataclass
class C:
    x: int
    y: int = field(repr=False)
    z: int = field(repr=False, default=10)
    t: int = 20

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 not __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:

@dataclass
class C:
    a: float
    b: float
    c: float = field(init=False)

    def __post_init__(self):
        self.c = self.a + self.b

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 [8]. 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 initialzed from a database, if a value is not provided when creating the class:

@dataclass
class C:
    i: int
    j: int = None
    database: InitVar[DatabaseType] = None

    def __post_init__(self, database):
        if self.j is None and database is not None:
            self.j = database.lookup('j')

c = C(10, database=my_database)

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:

@dataclass
class Base:
    x: Any = 15.0
    y: int = 0

@dataclass
class C(Base):
    z: int = 10
    x: int = 15

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:

def __init__(self, x: int = 15, y: int = 0, z: int = 10):

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:

l: list = field(default_factory=list)

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:

class C:
    x = []
    def add(self, element):
        self.x += element

o1 = C()
o2 = C()
o1.add(1)
o2.add(2)
assert o1.x == [1, 2]
assert o1.x is o2.x

Note that the two instances of class C share the same class variable x, as expected.

Using Data Classes, if this code was valid:

@dataclass
class D:
    x: List = []
    def add(self, element):
        self.x += element

it would generate code similar to:

class D:
    x = []
    def __init__(self, x=x):
        self.x = x
    def add(self, element):
        self.x += element

assert D().x is D().x

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:

@dataclass
class D:
    x: list = field(default_factory=list)

assert D().x is not D().x

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:

    @dataclass
    class Point:
         x: int
         y: int
    
    @dataclass
    class C:
         l: List[Point]
    
    p = Point(10, 20)
    assert asdict(p) == {'x': 10, 'y': 20}
    
    c = C([Point(0, 0), Point(10, 4)])
    assert asdict(c) == {'l': [{'x': 0, 'y': 0}, {'x': 10, 'y': 4}]}
    

    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:

    assert astuple(p) == (10, 20)
    assert astuple(c) == ([(0, 0), (10, 4)],)
    

    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. 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:

    C = make_dataclass('C',
                       [('x', int),
                        ('y', int, field(default=5))],
                       namespace={'add_one': lambda self: self.x + 1})
    

    Is equivalent to:

    @dataclass
    class C:
        x: int
        y: int = 5
    
        def add_one(self):
            return self.x + 1
    
  • 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.

Discussion

python-ideas discussion

This discussion started on python-ideas [9] and was moved to a GitHub repo [10] 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 [11].

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:

    Time = namedtuple('Time', ['hour', 'minute'])
    def get_time():
        return Time(12, 0)
    

    Then if a user uses this code as:

    hour, minute = get_time()
    

    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 namedtuples 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 [12].

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:

@dataclass
class C:
    x: int

    @classmethod
    def from_file(cls, filename):
        with open(filename) as fl:
            file_value = int(fl.read())
        return C(file_value)

c = C.from_file('file.txt')

Because the __post_init__ function is the last thing called in the generated __init__, having a classmethod constructor (which can also execute code immmediately after constructing the object) is functionally equivalent to being able to pass parameters to a __post_init__ function.

With InitVars, __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 InitVars, using __init__ and the module-level replace() function InitVars 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 [13] 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:

@dataclass
class Square:
    length: float
    area: float = field(init=False, default=0.0)

    def __post_init__(self):
        self.area = self.length * self.length

s1 = Square(1.0)
s2 = replace(s1, length=2.0)

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 [14].

Providing isdataclass()

An earlier version of this PEP defined an isdataclass(obj) helper function. However, there was no known use case for this, and there was debate on whether it should return True for instances or classes or both. In the end, isdataclass() was removed.

The supported way of writing a function that checks if an object is a dataclass instance or class is:

def isdataclass(obj):
    try:
        dataclasses.fields(obj)
        return True
    except TypeError:
        return False

If needed, a further check for isinstance(obj, type) can be added to discern if obj is a class.

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:

@dataclass(init=False)
class ArgHolder:
    args: List[Any]
    kwargs: Mapping[Any, Any]

    def __init__(self, *args, **kwargs):
        self.args = args
        self.kwargs = kwargs

a = ArgHolder(1, 2, three=3)

A complicated example

This code exists in a closed source project:

class Application:
    def __init__(self, name, requirements, constraints=None, path='', executable_links=None, executables_dir=()):
        self.name = name
        self.requirements = requirements
        self.constraints = {} if constraints is None else constraints
        self.path = path
        self.executable_links = [] if executable_links is None else executable_links
        self.executables_dir = executables_dir
        self.additional_items = []

    def __repr__(self):
        return f'Application({self.name!r},{self.requirements!r},{self.constraints!r},{self.path!r},{self.executable_links!r},{self.executables_dir!r},{self.additional_items!r})'

This can be replaced by:

@dataclass
class Application:
    name: str
    requirements: List[Requirement]
    constraints: Dict[str, str] = field(default_factory=dict)
    path: str = ''
    executable_links: List[str] = field(default_factory=list)
    executable_dir: Tuple[str] = ()
    additional_items: List[str] = field(init=False, default_factory=list)

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.

Source: https://github.com/python/peps/blob/master/pep-0557.rst