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 560 -- Core support for typing module and generic types

PEP:560
Title:Core support for typing module and generic types
Author:Ivan Levkivskyi <levkivskyi at gmail.com>
Status:Accepted
Type:Standards Track
Created:03-Sep-2017
Python-Version:3.7
Post-History:09-Sep-2017, 14-Nov-2017
Resolution:https://mail.python.org/pipermail/python-dev/2017-December/151038.html

Abstract

Initially PEP 484 was designed in such way that it would not introduce any changes to the core CPython interpreter. Now type hints and the typing module are extensively used by the community, e.g. PEP 526 and PEP 557 extend the usage of type hints, and the backport of typing on PyPI has 1M downloads/month. Therefore, this restriction can be removed. It is proposed to add two special methods __class_getitem__ and __mro_entries__ to the core CPython for better support of generic types.

Rationale

The restriction to not modify the core CPython interpreter led to some design decisions that became questionable when the typing module started to be widely used. There are three main points of concern: performance of the typing module, metaclass conflicts, and the large number of hacks currently used in typing.

Performance

The typing module is one of the heaviest and slowest modules in the standard library even with all the optimizations made. Mainly this is because subscripted generic types (see PEP 484 for definition of terms used in this PEP) are class objects (see also [1]). There are three main ways how the performance can be improved with the help of the proposed special methods:

  • Creation of generic classes is slow since the GenericMeta.__new__ is very slow; we will not need it anymore.
  • Very long method resolution orders (MROs) for generic classes will be half as long; they are present because we duplicate the collections.abc inheritance chain in typing.
  • Instantiation of generic classes will be faster (this is minor however).

Metaclass conflicts

All generic types are instances of GenericMeta, so if a user uses a custom metaclass, then it is hard to make a corresponding class generic. This is particularly hard for library classes that a user doesn't control. A workaround is to always mix-in GenericMeta:

class AdHocMeta(GenericMeta, LibraryMeta):
    pass

class UserClass(LibraryBase, Generic[T], metaclass=AdHocMeta):
    ...

but this is not always practical or even possible. With the help of the proposed special attributes the GenericMeta metaclass will not be needed.

Hacks and bugs that will be removed by this proposal

  • _generic_new hack that exists because __init__ is not called on instances with a type differing form the type whose __new__ was called, C[int]().__class__ is C.
  • _next_in_mro speed hack will be not necessary since subscription will not create new classes.
  • Ugly sys._getframe hack. This one is particularly nasty since it looks like we can't remove it without changes outside typing.
  • Currently generics do dangerous things with private ABC caches to fix large memory consumption that grows at least as O(N2), see [2]. This point is also important because it was recently proposed to re-implement ABCMeta in C.
  • Problems with sharing attributes between subscripted generics, see [3]. The current solution already uses __getattr__ and __setattr__, but it is still incomplete, and solving this without the current proposal will be hard and will need __getattribute__.
  • _no_slots_copy hack, where we clean up the class dictionary on every subscription thus allowing generics with __slots__.
  • General complexity of the typing module. The new proposal will not only allow to remove the above mentioned hacks/bugs, but also simplify the implementation, so that it will be easier to maintain.

Specification

__class_getitem__

The idea of __class_getitem__ is simple: it is an exact analog of __getitem__ with an exception that it is called on a class that defines it, not on its instances. This allows us to avoid GenericMeta.__getitem__ for things like Iterable[int]. The __class_getitem__ is automatically a class method and does not require @classmethod decorator (similar to __init_subclass__) and is inherited like normal attributes. For example:

class MyList:
    def __getitem__(self, index):
        return index + 1
    def __class_getitem__(cls, item):
        return f"{cls.__name__}[{item.__name__}]"

class MyOtherList(MyList):
    pass

assert MyList()[0] == 1
assert MyList[int] == "MyList[int]"

assert MyOtherList()[0] == 1
assert MyOtherList[int] == "MyOtherList[int]"

Note that this method is used as a fallback, so if a metaclass defines __getitem__, then that will have the priority.

__mro_entries__

If an object that is not a class object appears in the tuple of bases of a class definition, then method __mro_entries__ is searched on it. If found, it is called with the original tuple of bases as an argument. The result of the call must be a tuple, that is unpacked in the base classes in place of this object. (If the tuple is empty, this means that the original bases is simply discarded.) If there are more than one object with __mro_entries__, then all of them are called with the same original tuple of bases. This step happens first in the process of creation of a class, all other steps, including checks for duplicate bases and MRO calculation, happen normally with the updated bases.

Using the method API instead of just an attribute is necessary to avoid inconsistent MRO errors, and perform other manipulations that are currently done by GenericMeta.__new__. The original bases are stored as __orig_bases__ in the class namespace (currently this is also done by the metaclass). For example:

class GenericAlias:
    def __init__(self, origin, item):
        self.origin = origin
        self.item = item
    def __mro_entries__(self, bases):
        return (self.origin,)

class NewList:
    def __class_getitem__(cls, item):
        return GenericAlias(cls, item)

class Tokens(NewList[int]):
    ...

assert Tokens.__bases__ == (NewList,)
assert Tokens.__orig_bases__ == (NewList[int],)
assert Tokens.__mro__ == (Tokens, NewList, object)

Resolution using __mro_entries__ happens only in bases of a class definition statement. In all other situations where a class object is expected, no such resolution will happen, this includes isinstance and issubclass built-in functions.

NOTE: These two method names are reserved for use by the typing module and the generic types machinery, and any other use is discouraged. The reference implementation (with tests) can be found in [4], and the proposal was originally posted and discussed on the typing tracker, see [5].

Dynamic class creation and types.resolve_bases

type.__new__ will not perform any MRO entry resolution. So that a direct call type('Tokens', (List[int],), {}) will fail. This is done for performance reasons and to minimize the number of implicit transformations. Instead, a helper function resolve_bases will be added to the types module to allow an explicit __mro_entries__ resolution in the context of dynamic class creation. Correspondingly, types.new_class will be updated to reflect the new class creation steps while maintaining the backwards compatibility:

def new_class(name, bases=(), kwds=None, exec_body=None):
    resolved_bases = resolve_bases(bases)  # This step is added
    meta, ns, kwds = prepare_class(name, resolved_bases, kwds)
    if exec_body is not None:
        exec_body(ns)
    ns['__orig_bases__'] = bases  # This step is added
    return meta(name, resolved_bases, ns, **kwds)

Backwards compatibility and impact on users who don't use typing

This proposal may break code that currently uses the names __class_getitem__ and __mro_entries__. (But the language reference explicitly reserves all undocumented dunder names, and allows "breakage without warning"; see [6].)

This proposal will support almost complete backwards compatibility with the current public generic types API; moreover the typing module is still provisional. The only two exceptions are that currently issubclass(List[int], List) returns True, while with this proposal it will raise TypeError, and repr() of unsubscripted user-defined generics cannot be tweaked and will coincide with repr() of normal (non-generic) classes.

With the reference implementation I measured negligible performance effects (under 1% on a micro-benchmark) for regular (non-generic) classes. At the same time performance of generics is significantly improved:

  • importlib.reload(typing) is up to 7x faster
  • Creation of user defined generic classes is up to 4x faster (on a micro- benchmark with an empty body)
  • Instantiation of generic classes is up to 5x faster (on a micro-benchmark with an empty __init__)
  • Other operations with generic types and instances (like method lookup and isinstance() checks) are improved by around 10-20%
  • The only aspect that gets slower with the current proof of concept implementation is the subscripted generics cache look-up. However it was already very efficient, so this aspect gives negligible overall impact.

References

[1]Discussion following Mark Shannon's presentation at Language Summit (https://github.com/python/typing/issues/432)
[2]Pull Request to implement shared generic ABC caches (merged) (https://github.com/python/typing/pull/383)
[3]An old bug with setting/accessing attributes on generic types (https://github.com/python/typing/issues/392)
[4]The reference implementation (https://github.com/ilevkivskyi/cpython/pull/2/files, https://github.com/ilevkivskyi/cpython/tree/new-typing)
[5]Original proposal (https://github.com/python/typing/issues/468)
[6]Reserved classes of identifiers (https://docs.python.org/3/reference/lexical_analysis.html#reserved-classes-of-identifiers)
Source: https://github.com/python/peps/blob/master/pep-0560.rst