I asked this on Stackoverflow and discovered from the discussions there that it might be a Python bug.

http://stackoverflow.com/questions/9859222/python-multiprocessing-with-maxtasksperchild

HOW TO REPRODUCE
================

This seemingly-simple program isn't working for me unless I remove the maxtasksperchild parameter. What am I doing wrong?

from multiprocessing import Pool
import os
import sys

def f(x):
  print "pid: ", os.getpid(), " got: ", x
  sys.stdout.flush()
  return [x, x+1]

def cb(r):
  print "got result: ", r

if __name__ == '__main__':
  pool = Pool(processes=1, maxtasksperchild=9)
  keys = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
  result = pool.map_async(f, keys, chunksize=1, callback=cb)
  pool.close()
  pool.join()

When I run it, I get:

$ python doit.py
pid:  6409  got:  1
pid:  6409  got:  2
pid:  6409  got:  3
pid:  6409  got:  4
pid:  6409  got:  5
pid:  6409  got:  6
pid:  6409  got:  7
pid:  6409  got:  8
pid:  6409  got:  9

And it hangs. That is, the new worker to process the 10th element didn't get spawned.

In another terminal, I see:

$ ps -C python
  PID TTY          TIME CMD
 6408 pts/11   00:00:00 python
 6409 pts/11   00:00:00 python <defunct>

This is done on Ubuntu 11.10 running python 2.7.2+ (installed from ubuntu packages).

MY HYPOTHESIS
=============

This is based on skimming the code and turning on logging.

The call to pool.close() (which the docs say I should call before calling pool.join()), sets the flag pool._state to CLOSE. The function Pool._handle_workers relies on that flag being 'RUN' to kick off new worker processes. Which doesn't happen, leading to everything getting stuck.

One workaround for the bug is to sleep after the map_async call for about 10 seconds before pool.close() is called. That works as it should (because pool._state doesn't get set to CLOSE until the all the jobs have finished).

Sorry if I missed something, didn't RTFM etc.