python-peps/peps/pep-0688.rst

PEP: 688
Title: Making the buffer protocol accessible in Python
Author: Jelle Zijlstra <jelle.zijlstra@gmail.com>
Discussions-To: https://discuss.python.org/t/19756
Status: Accepted
Type: Standards Track
Topic: Typing
Content-Type: text/x-rst
Created: 23-Apr-2022
Python-Version: 3.12
Post-History: `23-Apr-2022 <https://mail.python.org/archives/list/typing-sig@python.org/thread/CX7GPSIYQEL23RXMYL66GAKGP4RLUD7P/>`__,
              `25-Apr-2022 <https://discuss.python.org/t/15265>`__,
              `06-Oct-2022 <https://discuss.python.org/t/19756>`__,
              `26-Oct-2022 <https://mail.python.org/archives/list/typing-sig@python.org/thread/XH5ZK2MSZIQLL62PYZ6I5532SQKKVCBL/>`__
Resolution: https://discuss.python.org/t/pep-688-making-the-buffer-protocol-accessible-in-python/15265/35


Abstract
========

This PEP proposes a Python-level API for the buffer protocol,
which is currently accessible only to C code. This allows type
checkers to evaluate whether objects implement the protocol.


Motivation
==========

The CPython C API provides a versatile mechanism for accessing the
underlying memory of an object—the `buffer protocol <https://docs.python.org/3/c-api/buffer.html>`__
introduced in :pep:`3118`.
Functions that accept binary data are usually written to handle any
object implementing the buffer protocol. For example, at the time of writing,
there are around 130 functions in CPython using the Argument Clinic
``Py_buffer`` type, which accepts the buffer protocol.

Currently, there is no way for Python code to inspect whether an object
supports the buffer protocol. Moreover, the static type system
does not provide a type annotation to represent the protocol.
This is a `common problem <https://github.com/python/typing/issues/593>`__
when writing type annotations for code that accepts generic buffers.

Similarly, it is impossible for a class written in Python to support
the buffer protocol. A buffer class in
Python would give users the ability to easily wrap a C buffer object, or to test
the behavior of an API that consumes the buffer protocol. Granted, this is not
a particularly common need. However, there has been a
`CPython feature request <https://github.com/python/cpython/issues/58006>`__
for supporting buffer classes written in Python that has been open since 2012.


Rationale
=========

Current options
---------------

There are two known workarounds for annotating buffer types in
the type system, but neither is adequate.

First, the `current workaround <https://github.com/python/typeshed/blob/2a0fc1b582ef84f7a82c0beb39fa617de2539d3d/stdlib/_typeshed/__init__.pyi#L194>`__
for buffer types in typeshed is a type alias
that lists well-known buffer types in the standard library, such as
``bytes``, ``bytearray``, ``memoryview``, and ``array.array``. This
approach works for the standard library, but it does not extend to
third-party buffer types.

Second, the `documentation <https://docs.python.org/3.10/library/typing.html#typing.ByteString>`__
for ``typing.ByteString`` currently states:

   This type represents the types ``bytes``, ``bytearray``, and
   ``memoryview`` of byte sequences.

   As a shorthand for this type, ``bytes`` can be used to annotate
   arguments of any of the types mentioned above.

Although this sentence has been in the documentation
`since 2015 <https://github.com/python/cpython/commit/2a19d956ab92fc9084a105cc11292cb0438b322f>`__,
the use of ``bytes`` to include these other types is not specified
in any of the typing PEPs. Furthermore, this mechanism has a number of
problems. It does not include all possible buffer types, and it
makes the ``bytes`` type ambiguous in type annotations. After all,
there are many operations that are valid on ``bytes`` objects, but
not on ``memoryview`` objects, and it is perfectly possible for
a function to accept ``bytes`` but not ``memoryview`` objects.
A mypy user
`reports <https://github.com/python/mypy/issues/12643#issuecomment-1105914159>`__
that this shortcut has caused significant problems for the ``psycopg`` project.

Kinds of buffers
----------------

The C buffer protocol supports
`many options <https://docs.python.org/3.10/c-api/buffer.html#buffer-request-types>`__,
affecting strides, contiguity, and support for writing to the buffer. Some of these
options would be useful in the type system. For example, typeshed
currently provides separate type aliases for writable and read-only
buffers.

However, in the C buffer protocol, most of these options cannot be
queried directly on the type object. The only way to figure out
whether an object supports a particular flag is to actually
ask for the buffer. For some types, such as ``memoryview``,
the supported flags depend on the instance. As a result, it would
be difficult to represent support for these flags in the type system.


Specification
=============

Python-level buffer protocol
----------------------------

We propose to add two Python-level special methods, ``__buffer__``
and ``__release_buffer__``. Python
classes that implement these methods are usable as buffers from C
code. Conversely, classes implemented in C that support the
buffer protocol acquire synthesized methods accessible from Python
code.

The ``__buffer__`` method is called to create a buffer from a Python
object, for example by the ``memoryview()`` constructor.
It corresponds to the ``bf_getbuffer`` C slot.
The Python signature for this method is
``def __buffer__(self, flags: int, /) -> memoryview: ...``. The method
must return a ``memoryview`` object. If the ``bf_getbuffer`` slot
is invoked on a Python class with a ``__buffer__`` method,
the interpreter extracts the underlying ``Py_buffer`` from the
``memoryview`` returned by the method
and returns it to the C caller. Similarly, if Python code calls the
``__buffer__`` method on an instance of a C class that
implements ``bf_getbuffer``, the returned buffer is wrapped in a
``memoryview`` for consumption by Python code.

The ``__release_buffer__`` method should be called when a caller no
longer needs the buffer returned by ``__buffer__``. It corresponds to the
``bf_releasebuffer`` C slot. This is an
optional part of the buffer protocol.
The Python signature for this method is
``def __release_buffer__(self, buffer: memoryview, /) -> None: ...``.
The buffer to be released is wrapped in a ``memoryview``. When this
method is invoked through CPython's buffer API (for example, through
calling ``memoryview.release`` on a ``memoryview`` returned by
``__buffer__``), the passed ``memoryview`` is the same object
as was returned by ``__buffer__``. It is
also possible to call ``__release_buffer__`` on a C class that
implements ``bf_releasebuffer``.

If ``__release_buffer__`` exists on an object,
Python code that calls ``__buffer__`` directly on the object must
call ``__release_buffer__`` on the same object when it is done
with the buffer. Otherwise, resources used by the object may
not be reclaimed. Similarly, it is a programming error
to call ``__release_buffer__`` without a previous call to
``__buffer__``, or to call it multiple times for a single call
to ``__buffer__``. For objects that implement the C buffer protocol,
calls to ``__release_buffer__`` where the argument is not a
``memoryview`` wrapping the same object will raise an exception.
After a valid call to ``__release_buffer__``, the ``memoryview``
is invalidated (as if its ``release()`` method had been called),
and any subsequent calls to ``__release_buffer__`` with the same
``memoryview`` will raise an exception.
The interpreter will ensure that misuse
of the Python API will not break invariants at the C level -- for
example, it will not cause memory safety violations.

``inspect.BufferFlags``
-----------------------

To help implementations of ``__buffer__``, we add ``inspect.BufferFlags``,
a subclass of ``enum.IntFlag``. This enum contains all flags defined in the
C buffer protocol. For example, ``inspect.BufferFlags.SIMPLE`` has the same
value as the ``PyBUF_SIMPLE`` constant.

``collections.abc.Buffer``
--------------------------

We add a new abstract base classes, ``collections.abc.Buffer``,
which requires the ``__buffer__`` method.
This class is intended primarily for use in type annotations:

.. code-block:: python

   def need_buffer(b: Buffer) -> memoryview:
       return memoryview(b)

   need_buffer(b"xy")  # ok
   need_buffer("xy")  # rejected by static type checkers


It can also be used in ``isinstance`` and ``issubclass`` checks:

.. code-block:: pycon

   >>> from collections.abc import Buffer
   >>> isinstance(b"xy", Buffer)
   True
   >>> issubclass(bytes, Buffer)
   True
   >>> issubclass(memoryview, Buffer)
   True
   >>> isinstance("xy", Buffer)
   False
   >>> issubclass(str, Buffer)
   False

In the typeshed stub files, the class should be defined as a ``Protocol``,
following the precedent of other simple ABCs in ``collections.abc`` such as
``collections.abc.Iterable`` or ``collections.abc.Sized``.

Example
-------

The following is an example of a Python class that implements the
buffer protocol:

.. code-block:: python

   import contextlib
   import inspect

   class MyBuffer:
       def __init__(self, data: bytes):
           self.data = bytearray(data)
           self.view = None

       def __buffer__(self, flags: int) -> memoryview:
           if flags != inspect.BufferFlags.FULL_RO:
               raise TypeError("Only BufferFlags.FULL_RO supported")
           if self.view is not None:
               raise RuntimeError("Buffer already held")
           self.view = memoryview(self.data)
           return self.view

       def __release_buffer__(self, view: memoryview) -> None:
           assert self.view is view  # guaranteed to be true
           self.view.release()
           self.view = None

       def extend(self, b: bytes) -> None:
           if self.view is not None:
               raise RuntimeError("Cannot extend held buffer")
           self.data.extend(b)

   buffer = MyBuffer(b"capybara")
   with memoryview(buffer) as view:
       view[0] = ord("C")

       with contextlib.suppress(RuntimeError):
           buffer.extend(b"!")  # raises RuntimeError

   buffer.extend(b"!")  # ok, buffer is no longer held

   with memoryview(buffer) as view:
       assert view.tobytes() == b"Capybara!"


Equivalent for older Python versions
------------------------------------

New typing features are usually backported to older Python versions
in the `typing_extensions <https://pypi.org/project/typing-extensions/>`_
package. Because the buffer protocol
is currently accessible only in C, this PEP cannot be fully implemented
in a pure-Python package like ``typing_extensions``. As a temporary
workaround, an abstract base class ``typing_extensions.Buffer``
will be provided for Python versions
that do not have ``collections.abc.Buffer`` available.

After this PEP is implemented, inheriting from ``collections.abc.Buffer`` will
not be necessary to indicate that an object supports the buffer protocol.
However, in older Python versions, it will be necessary to explicitly
inherit from ``typing_extensions.Buffer`` to indicate to type checkers that
a class supports the buffer protocol, since objects supporting the buffer
protocol will not have a ``__buffer__`` method. It is expected that this
will happen primarily in stub files, because buffer classes are necessarily
implemented in C code, which cannot have types defined inline.
For runtime uses, the ``ABC.register`` API can be used to register
buffer classes with ``typing_extensions.Buffer``.


No special meaning for ``bytes``
--------------------------------

The special case stating that ``bytes`` may be used as a shorthand
for other ``ByteString`` types will be removed from the ``typing``
documentation.
With ``collections.abc.Buffer`` available as an alternative, there will be no good
reason to allow ``bytes`` as a shorthand.
Type checkers currently implementing this behavior
should deprecate and eventually remove it.


Backwards Compatibility
=======================

``__buffer__`` and ``__release_buffer__`` attributes
----------------------------------------------------

As the runtime changes in this PEP only add new functionality, there are
few backwards compatibility concerns.

However, code that uses a ``__buffer__`` or ``__release_buffer__`` attribute for
other purposes may be affected. While all dunders are technically reserved for the
language, it is still good practice to ensure that a new dunder does not
interfere with too much existing code, especially widely used packages. A survey
of publicly accessible code found:

- PyPy `supports <https://doc.pypy.org/en/latest/__pypy__-module.html#generally-available-functionality>`__
  a ``__buffer__`` method with compatible semantics to those proposed in this
  PEP. A PyPy core developer `expressed his support <https://discuss.python.org/t/pep-688-making-the-buffer-protocol-accessible-in-python/15265/34>`__
  for this PEP.
- pyzmq `implements <https://github.com/zeromq/pyzmq/blob/fe18dc55516ef50d168fc02f8550a67ff5b5633d/zmq/backend/cffi/message.py#L190>`__
  a PyPy-compatible ``__buffer__`` method.
- mpi4py `defines <https://github.com/mpi4py/mpi4py/blob/453b87d0da37c5914b91afb511b188556dff2a9c/src/mpi4py/typing.py#L66>`__
  a ``SupportsBuffer`` protocol that would be equivalent to this PEP's ``collections.abc.Buffer``.
- NumPy used to have an undocumented behavior where it would access a ``__buffer__`` attribute
  (not method) to get an object's buffer. This was `removed <https://github.com/numpy/numpy/pull/13049>`__
  in 2019 for NumPy 1.17. The behavior would have last worked in NumPy 1.16, which only supported
  Python 3.7 and older. Python 3.7 will have reached its end of life by the time this PEP is expected to
  be implemented.

Thus, this PEP's use of the ``__buffer__`` method will improve interoperability with
PyPy and not interfere with the current versions of any major Python packages.

No publicly accessible code uses the name ``__release_buffer__``.

Removal of the ``bytes`` special case
-------------------------------------

Separately, the recommendation to remove the special behavior for
``bytes`` in type checkers does have a backwards compatibility
impact on their users. An `experiment <https://github.com/python/mypy/pull/12661>`__
with mypy shows that several major open source projects that use it
for type checking will see new errors if the ``bytes`` promotion
is removed. Many of these errors can be fixed by improving
the stubs in typeshed, as has already been done for the
`builtins <https://github.com/python/typeshed/pull/7631>`__,
`binascii <https://github.com/python/typeshed/pull/7677>`__,
`pickle <https://github.com/python/typeshed/pull/7678>`__, and
`re <https://github.com/python/typeshed/pull/7679>`__ modules.
A `review <https://github.com/python/typeshed/issues/9006>`__ of all
usage of ``bytes`` types in typeshed is in progress.
Overall, the change improves type safety and makes the type system
more consistent, so we believe the migration cost is worth it.


How to Teach This
=================

We will add notes pointing to ``collections.abc.Buffer`` in appropriate places in the
documentation, such as `typing.readthedocs.io <https://typing.readthedocs.io/en/latest/>`__
and the `mypy cheat sheet <https://mypy.readthedocs.io/en/stable/cheat_sheet_py3.html>`__.
Type checkers may provide additional pointers in their error messages. For example,
when they encounter a buffer object being passed to a function that
is annotated to only accept ``bytes``, the error message could include a note suggesting
the use of ``collections.abc.Buffer`` instead.


Reference Implementation
========================

An implementation of this PEP is
`available <https://github.com/python/cpython/compare/main...JelleZijlstra:pep688v2?expand=1>`__
in the author's fork.


Rejected Ideas
==============

``types.Buffer``
----------------

An earlier version of this PEP proposed adding a new ``types.Buffer`` type with
an ``__instancecheck__`` implemented in C so that ``isinstance()`` checks can be
used to check whether a type implements the buffer protocol. This avoids the
complexity of exposing the full buffer protocol to Python code, while still
allowing the type system to check for the buffer protocol.

However, that approach
does not compose well with the rest of the type system, because ``types.Buffer``
would be a nominal type, not a structural one. For example, there would be no way
to represent "an object that supports both the buffer protocol and ``__len__``". With
the current proposal, ``__buffer__`` is like any other special method, so a
``Protocol`` can be defined combining it with another method.

More generally, no other part of Python works like the proposed ``types.Buffer``.
The current proposal is more consistent with the rest of the language, where
C-level slots usually have corresponding Python-level special methods.

Keep ``bytearray`` compatible with ``bytes``
--------------------------------------------

It has been suggested to remove the special case where ``memoryview`` is
always compatible with ``bytes``, but keep it for ``bytearray``, because
the two types have very similar interfaces. However, several standard
library functions (e.g., ``re.compile``, ``socket.getaddrinfo``, and most
functions accepting path-like arguments) accept
``bytes`` but not ``bytearray``. In most codebases, ``bytearray`` is also
not a very common type. We prefer to have users spell out accepted types
explicitly (or use ``Protocol`` from :pep:`544` if only a specific set of
methods is required). This aspect of the proposal was `specifically
discussed <https://mail.python.org/archives/list/typing-sig@python.org/thread/XH5ZK2MSZIQLL62PYZ6I5532SQKKVCBL/>`__
on the typing-sig mailing list, without any strong disagreement from the
typing community.

Distinguish between mutable and immutable buffers
-------------------------------------------------

The most frequently used distinction within buffer types is
whether or not the buffer is mutable. Some functions accept only
mutable buffers (e.g., ``bytearray``, some ``memoryview`` objects),
others accept all buffers.

An earlier version of this PEP proposed using the presence of the
``bf_releasebuffer`` slot to determine whether a buffer type is mutable.
This rule holds for most standard library buffer types, but the relationship
between mutability and the presence of this slot is not absolute. For
example, ``numpy`` arrays are mutable but do not have this slot.

The current buffer protocol does not provide any way to reliably
determine whether a buffer type represents a mutable or immutable
buffer. Therefore, this PEP does not add type system support
for this distinction.
The question can be revisited in the future if the buffer protocol
is enhanced to provide static introspection support.
A `sketch <https://discuss.python.org/t/introspection-and-mutable-xor-shared-semantics-for-pybuffer/20314>`_
for such a mechanism exists.


Acknowledgments
===============

Many people have provided useful feedback on drafts of this PEP.
Petr Viktorin has been particularly helpful in improving my understanding
of the subtleties of the buffer protocol.


Copyright
=========

This document is placed in the public domain or under the
CC0-1.0-Universal license, whichever is more permissive.