python - Why can I pass an instance method to multiprocessing.Process, but not a multiprocessing.Pool?

Question

I am trying to write an application that applies a function concurrently with a multiprocessing.Pool. I would like this function to be an instance method (so I can define it differently in different subclasses). This doesn't seem to be possible; as I have learned elsewhere, apparently bound methods can't be pickled. So why does starting a multiprocessing.Process with a bound method as a target work? The following code:

import multiprocessing

def test1():
    print "Hello, world 1"

def increment(x):
    return x + 1

class testClass():
    def process(self):
        process1 = multiprocessing.Process(target=test1)
        process1.start()
        process1.join()
        process2 = multiprocessing.Process(target=self.test2)
        process2.start()
        process2.join()

    def pool(self):
        pool = multiprocessing.Pool(1)
        for answer in pool.imap(increment, range(10)):
            print answer
        print
        for answer in pool.imap(self.square, range(10)):
            print answer

    def test2(self):
        print "Hello, world 2"

    def square(self, x):
        return x * x

def main():
    c = testClass()
    c.process()
    c.pool()

if __name__ == "__main__":
    main()

Produces this output:

Hello, world 1
Hello, world 2
1
2
3
4
5
6
7
8
9
10

Exception in thread Thread-2:
Traceback (most recent call last):
  File "C:Python27Libhreading.py", line 551, in __bootstrap_inner
    self.run()
  File "C:Python27Libhreading.py", line 504, in run
    self.__target(*self.__args, **self.__kwargs)
  File "C:Python27Libmultiprocessingpool.py", line 319, in _handle_tasks
    put(task)
PicklingError: Can't pickle <type 'instancemethod'>: attribute lookup __builtin__.instancemethod failed

Why can Processes handle bound methods, but not Pools?

Answer 1

The pickle module normally can't pickle instance methods:

>>> import pickle
>>> class A(object):
...  def z(self): print "hi"
... 
>>> a = A()
>>> pickle.dumps(a.z)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python2.7/pickle.py", line 1374, in dumps
    Pickler(file, protocol).dump(obj)
  File "/usr/local/lib/python2.7/pickle.py", line 224, in dump
    self.save(obj)
  File "/usr/local/lib/python2.7/pickle.py", line 306, in save
    rv = reduce(self.proto)
  File "/usr/local/lib/python2.7/copy_reg.py", line 70, in _reduce_ex
    raise TypeError, "can't pickle %s objects" % base.__name__
TypeError: can't pickle instancemethod objects

However, the multiprocessing module has a custom Pickler that adds some code to enable this feature:

#
# Try making some callable types picklable
#

from pickle import Pickler
class ForkingPickler(Pickler):
    dispatch = Pickler.dispatch.copy()

    @classmethod
    def register(cls, type, reduce):
        def dispatcher(self, obj):
            rv = reduce(obj)
            self.save_reduce(obj=obj, *rv)
        cls.dispatch[type] = dispatcher

def _reduce_method(m):
    if m.im_self is None:
        return getattr, (m.im_class, m.im_func.func_name)
    else:
        return getattr, (m.im_self, m.im_func.func_name)
ForkingPickler.register(type(ForkingPickler.save), _reduce_method)

You can replicate this using the copy_reg module to see it work for yourself:

>>> import copy_reg
>>> def _reduce_method(m):
...     if m.im_self is None:
...         return getattr, (m.im_class, m.im_func.func_name)
...     else:
...         return getattr, (m.im_self, m.im_func.func_name)
... 
>>> copy_reg.pickle(type(a.z), _reduce_method)
>>> pickle.dumps(a.z)
"c__builtin__
getattr
p0
(ccopy_reg
_reconstructor
p1
(c__main__
A
p2
c__builtin__
object
p3
Ntp4
Rp5
S'z'
p6
tp7
Rp8
."

When you use Process.start to spawn a new process on Windows, it pickles all the parameters you passed to the child process using this custom ForkingPickler:

#
# Windows
#

else:
    # snip...
    from pickle import load, HIGHEST_PROTOCOL

    def dump(obj, file, protocol=None):
        ForkingPickler(file, protocol).dump(obj)

    #
    # We define a Popen class similar to the one from subprocess, but
    # whose constructor takes a process object as its argument.
    #

    class Popen(object):
        '''
        Start a subprocess to run the code of a process object
        '''
        _tls = thread._local()

        def __init__(self, process_obj):
            # create pipe for communication with child
            rfd, wfd = os.pipe()

            # get handle for read end of the pipe and make it inheritable
            ...
            # start process
            ...

            # set attributes of self
            ...

            # send information to child
            prep_data = get_preparation_data(process_obj._name)
            to_child = os.fdopen(wfd, 'wb')
            Popen._tls.process_handle = int(hp)
            try:
                dump(prep_data, to_child, HIGHEST_PROTOCOL)
                dump(process_obj, to_child, HIGHEST_PROTOCOL)
            finally:
                del Popen._tls.process_handle
                to_child.close()

Note the "send information to the child" section. It's using the dump function, which uses ForkingPickler to pickle the data, which means your instance method can be pickled.

Now, when you use methods on multiprocessing.Pool to send a method to a child process, it's using a multiprocessing.Pipe to pickle the data. In Python 2.7, multiprocessing.Pipe is implemented in C, and calls pickle_dumps directly, so it doesn't take advantage of the ForkingPickler. That means pickling the instance method doesn't work.

However, if you use copy_reg to register the instancemethod type, rather than a custom Pickler, all attempts at pickling will be affected. So you can use that to enable pickling instance methods, even via Pool:

import multiprocessing
import copy_reg
import types

def _reduce_method(m):
    if m.im_self is None:
        return getattr, (m.im_class, m.im_func.func_name)
    else:
        return getattr, (m.im_self, m.im_func.func_name)
copy_reg.pickle(types.MethodType, _reduce_method)

def test1():
    print("Hello, world 1")

def increment(x):
    return x + 1

class testClass():
    def process(self):
        process1 = multiprocessing.Process(target=test1)
        process1.start()
        process1.join()
        process2 = multiprocessing.Process(target=self.test2)
        process2.start()
        process2.join()

    def pool(self):
        pool = multiprocessing.Pool(1)
        for answer in pool.imap(increment, range(10)):
            print(answer)
        print
        for answer in pool.imap(self.square, range(10)):
            print(answer)

    def test2(self):
        print("Hello, world 2")

    def square(self, x):
        return x * x

def main():
    c = testClass()
    c.process()
    c.pool()

if __name__ == "__main__":
    main()

Output:

Hello, world 1
Hello, world 2
GOT (0, 0, (True, 1))
GOT (0, 1, (True, 2))
GOT (0, 2, (True, 3))
GOT (0, 3, (True, 4))
GOT (0, 4, (True, 5))
 1GOT (0, 5, (True, 6))

GOT (0, 6, (True, 7))
2
GOT (0, 7, (True, 8))
3
 GOT (0, 8, (True, 9))
GOT (0, 9, (True, 10))
4
5
6
7
8
9
10

GOT (1, 0, (True, 0))
0
GOT (1, 1, (True, 1))
1
GOT (1, 2, (True, 4))
4
GOT (1, 3, (True, 9))
9
 GOT (1, 4, (True, 16))
16
GOT (1, 5, (True, 25))
25
 GOT (1, 6, (True, 36))
36
 GOT (1, 7, (True, 49))
49
 GOT (1, 8, (True, 64))
64
GOT (1, 9, (True, 81))
81
GOT None

Also note that in Python 3.x, pickle can pickle instance method types natively, so none of this stuff matters any more. :)