homepage

The multiprocessing module modifies the global copyreg.dispatch_table to
extend pickling to several built-in types (mostly callable types). In my
humble opinion, this is unacceptable for a standard library module.
Here is an example of a behaviour change made by multiprocessing:
 >>> import pickle
 >>> pickle.dumps(str.isalpha)
 Traceback (most recent call last):
 ...
 _pickle.PicklingError: Can't pickle <class 'method_descriptor'>:
attribute lookup builtins.method_descriptor failed
 >>> import multiprocessing.util
 >>> pickle.dumps(str.isalpha)
 
b'\x80\x03cbuiltins\ngetattr\nq\x00cbuiltins\nstr\nq\x01X\x07\x00\x00\x00isalphaq\x02\x86q\x03Rq\x04.'
There was some discussion in issue 558238 about allowing methods to be
pickled. However, no consensus was reached.
In addition, there is a bug in the method pickling support added by
multiprocessing in Python 3.0:
 Python 2.6b1+ (trunk:64899:64900, Jul 13 2008, 13:33:02) 
 [GCC 4.2.3 (Ubuntu 4.2.3-2ubuntu7)] on linux2
 Type "help", "copyright", "credits" or "license" for more information.
 >>> import multiprocessing.util, pickle
 >>> class A(object):
 ... def foo(self): pass
 ... 
 >>> pickle.dumps(A().foo, 2)
 
'\x80\x02c__builtin__\ngetattr\nq\x00c__main__\nA\nq\x01)\x81q\x02}q\x03bU\x03fooq\x04\x86q\x05Rq\x06.'
 >>> pickle.dumps(A.foo, 2)
 
'\x80\x02c__builtin__\ngetattr\nq\x00c__main__\nA\nq\x01U\x03fooq\x02\x86q\x03Rq\x04.'
 
 Python 3.0b1+ (py3k:64876M, Jul 11 2008, 12:20:51) 
 [GCC 4.2.3 (Ubuntu 4.2.3-2ubuntu7)] on linux2
 Type "help", "copyright", "credits" or "license" for more information.
 >>> import multiprocessing.util, pickle
 >>> class A(object):
 ... def foo(self): pass
 ... 
 >>> pickle.dumps(A().foo, 2)
 Traceback (most recent call last):
 ...
 pickle.PicklingError: Can't pickle <class 'method'>: it's not found as
builtins.method
 >>> pickle.dumps(A.foo, 2)
 Traceback (most recent call last):
 ...
 pickle.PicklingError: Can't pickle <function foo at 0xb7b1392c>: it's
not found as __main__.foo
A better solution for the interprocess communication needs of
multiprocessing would be to define custom subclass of Pickler with
extended support for built-in types. If needed, I am willing to extend
the _pickle module in Python 3.0 to support modification to its
"dispatch table", like the undocumented dispatch_table attribute in
pickle.Pickler.