python-peps/peps/pep-3154.rst

PEP: 3154
Title: Pickle protocol version 4
Version: $Revision$
Last-Modified: $Date$
Author: Antoine Pitrou <solipsis@pitrou.net>
Status: Final
Type: Standards Track
Content-Type: text/x-rst
Created: 11-Aug-2011
Python-Version: 3.4
Post-History: `12-Aug-2011 <https://mail.python.org/pipermail/python-dev/2011-August/112821.html>`__
Resolution: https://mail.python.org/pipermail/python-dev/2013-November/130439.html


Abstract
========

Data serialized using the pickle module must be portable across Python
versions.  It should also support the latest language features as well
as implementation-specific features.  For this reason, the pickle
module knows about several protocols (currently numbered from 0 to 3),
each of which appeared in a different Python version.  Using a
low-numbered protocol version allows to exchange data with old Python
versions, while using a high-numbered protocol allows access to newer
features and sometimes more efficient resource use (both CPU time
required for (de)serializing, and disk size / network bandwidth
required for data transfer).


Rationale
=========

The latest current protocol, coincidentally named protocol 3, appeared
with Python 3.0 and supports the new incompatible features in the
language (mainly, unicode strings by default and the new bytes
object).  The opportunity was not taken at the time to improve the
protocol in other ways.

This PEP is an attempt to foster a number of incremental improvements
in a new pickle protocol version.  The PEP process is used in order to
gather as many improvements as possible, because the introduction of a
new pickle protocol should be a rare occurrence.


Proposed changes
================

Framing
-------

Traditionally, when unpickling an object from a stream (by calling
``load()`` rather than ``loads()``), many small ``read()``
calls can be issued on the file-like object, with a potentially huge
performance impact.

Protocol 4, by contrast, features binary framing.  The general structure
of a pickle is thus the following::

    +------+------+
    | 0x80 | 0x04 |              protocol header (2 bytes)
    +------+------+
    |  OP  |                     FRAME opcode (1 byte)
    +------+------+-----------+
    | MM MM MM MM MM MM MM MM |  frame size (8 bytes, little-endian)
    +------+------------------+
    | .... |                     first frame contents (M bytes)
    +------+
    |  OP  |                     FRAME opcode (1 byte)
    +------+------+-----------+
    | NN NN NN NN NN NN NN NN |  frame size (8 bytes, little-endian)
    +------+------------------+
    | .... |                     second frame contents (N bytes)
    +------+
      etc.

To keep the implementation simple, it is forbidden for a pickle opcode
to straddle frame boundaries.  The pickler takes care not to produce such
pickles, and the unpickler refuses them.  Also, there is no "last frame"
marker.  The last frame is simply the one which ends with a STOP opcode.

A well-written C implementation doesn't need additional memory copies
for the framing layer, preserving general (un)pickling efficiency.

.. note::

   How the pickler decides to partition the pickle stream into frames is an
   implementation detail.  For example, "closing" a frame as soon as it
   reaches ~64 KiB is a reasonable choice for both performance and pickle
   size overhead.

Binary encoding for all opcodes
-------------------------------

The GLOBAL opcode, which is still used in protocol 3, uses the
so-called "text" mode of the pickle protocol, which involves looking
for newlines in the pickle stream.  It also complicates the implementation
of binary framing.

Protocol 4 forbids use of the GLOBAL opcode and replaces it with
STACK_GLOBAL, a new opcode which takes its operand from the stack.

Serializing more "lookupable" objects
-------------------------------------

By default, pickle is only able to serialize module-global functions and
classes.  Supporting other kinds of objects, such as unbound methods [4]_,
is a common request. Actually, third-party support for some of them, such
as bound methods, is implemented in the multiprocessing module [5]_.

The ``__qualname__`` attribute from :pep:`3155` makes it possible to
lookup many more objects by name.  Making the STACK_GLOBAL opcode accept
dot-separated names would allow the standard pickle implementation to
support all those kinds of objects.

64-bit opcodes for large objects
--------------------------------

Current protocol versions export object sizes for various built-in
types (str, bytes) as 32-bit ints.  This forbids serialization of
large data [1]_.  New opcodes are required to support very large bytes
and str objects.

Native opcodes for sets and frozensets
--------------------------------------

Many common built-in types (such as str, bytes, dict, list, tuple)
have dedicated opcodes to improve resource consumption when
serializing and deserializing them; however, sets and frozensets
don't.  Adding such opcodes would be an obvious improvement.  Also,
dedicated set support could help remove the current impossibility of
pickling self-referential sets [2]_.

Calling __new__ with keyword arguments
--------------------------------------

Currently, classes whose ``__new__`` mandates the use of keyword-only
arguments can not be pickled (or, rather, unpickled) [3]_.  Both a new
special method (``__getnewargs_ex__``) and a new opcode (NEWOBJ_EX)
are needed.  The ``__getnewargs_ex__`` method, if it exists, must
return a two-tuple ``(args, kwargs)`` where the first item is the
tuple of positional arguments and the second item is the dict of
keyword arguments for the class's ``__new__`` method.

Better string encoding
----------------------

Short str objects currently have their length coded as a 4-bytes
integer, which is wasteful.  A specific opcode with a 1-byte length
would make many pickles smaller.

Smaller memoization
-------------------

The PUT opcodes all require an explicit index to select in which entry
of the memo dictionary the top-of-stack is memoized.  However, in practice
those numbers are allocated in sequential order.  A new opcode, MEMOIZE,
will instead store the top-of-stack in at the index equal to the current
size of the memo dictionary.  This allows for shorter pickles, since PUT
opcodes are emitted for all non-atomic datatypes.


Summary of new opcodes
======================

These reflect the state of the proposed implementation (thanks mostly
to Alexandre Vassalotti's work):

* ``FRAME``: introduce a new frame (followed by the 8-byte frame size
  and the frame contents).

* ``SHORT_BINUNICODE``: push a utf8-encoded str object with a one-byte
  size prefix (therefore less than 256 bytes long).

* ``BINUNICODE8``: push a utf8-encoded str object with an eight-byte
  size prefix (for strings longer than 2**32 bytes, which therefore cannot
  be serialized using ``BINUNICODE``).

* ``BINBYTES8``: push a bytes object with an eight-byte size prefix
  (for bytes objects longer than 2**32 bytes, which therefore cannot be
  serialized using ``BINBYTES``).

* ``EMPTY_SET``: push a new empty set object on the stack.

* ``ADDITEMS``: add the topmost stack items to the set (to be used with
  ``EMPTY_SET``).

* ``FROZENSET``: create a frozenset object from the topmost stack items,
  and push it on the stack.

* ``NEWOBJ_EX``: take the three topmost stack items ``cls``, ``args``
  and ``kwargs``, and push the result of calling
  ``cls.__new__(*args, **kwargs)``.

* ``STACK_GLOBAL``: take the two topmost stack items ``module_name`` and
  ``qualname``, and push the result of looking up the dotted ``qualname``
  in the module named ``module_name``.

* ``MEMOIZE``: store the top-of-stack object in the memo dictionary with
  an index equal to the current size of the memo dictionary.


Alternative ideas
=================

Prefetching
-----------

Serhiy Storchaka suggested to replace framing with a special PREFETCH
opcode (with a 2- or 4-bytes argument) to declare known pickle chunks
explicitly. Large data may be pickled outside such chunks.  A naïve
unpickler should be able to skip the PREFETCH opcode and still decode
pickles properly, but good error handling would require checking that
the PREFETCH length falls on an opcode boundary.


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

In alphabetic order:

* Alexandre Vassalotti, for starting the second :pep:`3154` implementation [6]_

* Serhiy Storchaka, for discussing the framing proposal [6]_

* Stefan Mihaila, for starting the first :pep:`3154` implementation as a
  Google Summer of Code project mentored by Alexandre Vassalotti [7]_.


References
==========

.. [1] "pickle not 64-bit ready":
   http://bugs.python.org/issue11564

.. [2] "Cannot pickle self-referencing sets":
   http://bugs.python.org/issue9269

.. [3] "pickle/copyreg doesn't support keyword only arguments in __new__":
   http://bugs.python.org/issue4727

.. [4] "pickle should support methods":
   http://bugs.python.org/issue9276

.. [5] Lib/multiprocessing/forking.py:
   http://hg.python.org/cpython/file/baea9f5f973c/Lib/multiprocessing/forking.py#l54

.. [6] Implement PEP 3154, by Alexandre Vassalotti
   http://bugs.python.org/issue17810

.. [7] Implement PEP 3154, by Stefan Mihaila
   http://bugs.python.org/issue15642


Copyright
=========

This document has been placed in the public domain.



..
   Local Variables:
   mode: indented-text
   indent-tabs-mode: nil
   sentence-end-double-space: t
   fill-column: 70
   coding: utf-8
   End: