Currently multiprocessing uses the pickle module for its serialization of objects to be communicated between processes.  Specifically v2 of the pickle protocols is now exclusively used to provide maximum compatibility, motivated by the desire for multiple versions of Python to be used simultaneously with multiprocessing.

Per conversations in issue26507, issue23403, and others, multiprocessing should offer the option to specify what serialization is to be used for the transport of data between processes.  Besides supporting requests to use a different version of the pickle protocol or using 3rd party tools like dill, a hook to specify the means for reducing objects to a transmittable form opens a door for other creative or higher performance strategies.

Ultimately, this is not an enhancement to add functionality but rather to reorganize the existing internals of multiprocessing to permit better control over its use of serialization.

Attaching patch containing refactorizations but missing update to docs for the purposes of review.

Introduces three new things:
* a function to get the current serializer/reducer
* a function to set the serializer/reducer
* an abstract base class to assist others in rolling their own reducers

LGTM

New changeset 7381b1b50e00 by Davin Potts in branch 'default':
Issue #28053: Applying refactorings, docs and other cleanup to follow.
https://hg.python.org/cpython/rev/7381b1b50e00

Docs need updating still.

Documentation would be appreciated. I have a project that uses BaseManager, Client, and Listener to create some servers and clients. I would like to update the project to work with Python 3 and would prefer to update the clients and the servers separately (i.e. switch the client to Python 3 while the server is run with Python 2.7). However, BaseManager uses connection.Client which uses connection._ConnectionBase which uses reduction.ForkingPickler without a protocol argument. It seems the default protocol is 3 on Python 3.6 and 2 on Python 2.7 (contrary to the comment above about v2 being used). I just want to set the protocol version to 2 in Python 3.6. Can I do that with the changes added by this patch?

I tried creating pickle2reducer.py like this:

from multiprocessing.reduction import ForkingPickler, AbstractReducer

class ForkingPickler2(ForkingPickler):
    def __init__(self, *args):
        if len(args) > 1:
            args[1] = 2
        else:
            args.append(2)
        super().__init__(*args)

    @classmethod
    def dumps(cls, obj, protocol=2):
        return ForkingPickler.dumps(obj, protocol)


def dump(obj, file, protocol=2):
    ForkingPickler2(file, protocol).dump(obj)


class Pickle2Reducer(AbstractReducer):
    ForkingPickler = ForkingPickler2
    register = ForkingPickler2.register
    dump = dump

and then putting

import pickle2reducer
multiprocessing.reducer = pickle2reducer.Pickle2Reducer()

at the top of my module before

import multiprocessing.connection

but I still see "ValueError: unsupported pickle protocol: 3" on the server when I connect with a Python 3.6 client.

Just to follow up in case anyone comes across my last message later:

I just had to change the last line from

multiprocessing.reducer = pickle2reducer.Pickle2Reducer()

to

multiprocessing.context._default_context.reducer = pickle2reducer.Pickle2Reducer()

I'd like to know if the work here will be completed soon :-) This currently lacks documentation but also tests.

In particular, looking at the code (and that is supported by Will S' comment above), I feel the API isn't working as intended (i.e. setting the "reducer" property won't actually change the underlying parameters since the "reduction" module is imported eagerly).