Would it be possible to create a python Pool that is non-daemonic? I want a pool to be able to call a function that has another pool inside.

I want this because deamon processes cannot create process. Specifically, it will cause the error:

AssertionError: daemonic processes are not allowed to have children

For example, consider the scenario where function_a has a pool which runs function_b which has a pool which runs function_c. This function chain will fail, because function_b is being run in a daemon process, and daemon processes cannot create processes.

The multiprocessing.pool.Pool class creates the worker processes in its __init__ method, makes them daemonic and starts them, and it is not possible to re-set their daemon attribute to False before they are started (and afterwards it’s not allowed anymore). But you can create your own sub-class of multiprocesing.pool.Pool (multiprocessing.Pool is just a wrapper function) and substitute your own multiprocessing.Process sub-class, which is always non-daemonic, to be used for the worker processes.

Here’s a full example of how to do this. The important parts are the two classes NoDaemonProcess and MyPool at the top and to call pool.close() and pool.join() on your MyPool instance at the end.

#!/usr/bin/env python
# -*- coding: UTF-8 -*-

import multiprocessing
# We must import this explicitly, it is not imported by the top-level
# multiprocessing module.
import multiprocessing.pool
import time

from random import randint


class NoDaemonProcess(multiprocessing.Process):
    # make 'daemon' attribute always return False
    def _get_daemon(self):
        return False
    def _set_daemon(self, value):
        pass
    daemon = property(_get_daemon, _set_daemon)

# We sub-class multiprocessing.pool.Pool instead of multiprocessing.Pool
# because the latter is only a wrapper function, not a proper class.
class MyPool(multiprocessing.pool.Pool):
    Process = NoDaemonProcess

def sleepawhile(t):
    print("Sleeping %i seconds..." % t)
    time.sleep(t)
    return t

def work(num_procs):
    print("Creating %i (daemon) workers and jobs in child." % num_procs)
    pool = multiprocessing.Pool(num_procs)

    result = pool.map(sleepawhile,
        [randint(1, 5) for x in range(num_procs)])

    # The following is not really needed, since the (daemon) workers of the
    # child's pool are killed when the child is terminated, but it's good
    # practice to cleanup after ourselves anyway.
    pool.close()
    pool.join()
    return result

def test():
    print("Creating 5 (non-daemon) workers and jobs in main process.")
    pool = MyPool(5)

    result = pool.map(work, [randint(1, 5) for x in range(5)])

    pool.close()
    pool.join()
    print(result)

if __name__ == '__main__':
    test()

I had the necessity to employ a non-daemonic pool in Python 3.7 and ended up adapting the code posted in the accepted answer. Below there’s the snippet that creates the non-daemonic pool:

import multiprocessing.pool

class NoDaemonProcess(multiprocessing.Process):
    @property
    def daemon(self):
        return False

    @daemon.setter
    def daemon(self, value):
        pass


class NoDaemonContext(type(multiprocessing.get_context())):
    Process = NoDaemonProcess

# We sub-class multiprocessing.pool.Pool instead of multiprocessing.Pool
# because the latter is only a wrapper function, not a proper class.
class NestablePool(multiprocessing.pool.Pool):
    def __init__(self, *args, **kwargs):
        kwargs['context'] = NoDaemonContext()
        super(NestablePool, self).__init__(*args, **kwargs)

As the current implementation of multiprocessing has been extensively refactored to be based on contexts, we need to provide a NoDaemonContext class that has our NoDaemonProcess as attribute. NestablePool will then use that context instead of the default one.

That said, I should warn that there are at least two caveats to this approach:

1. It still depends on implementation details of the multiprocessing package, and could therefore break at any time.
2. There are valid reasons why multiprocessing made it so hard to use non-daemonic processes, many of which are explained here. The most compelling in my opinion is:

As for allowing children threads to spawn off children of its own using subprocess runs the risk of creating a little army of zombie ‘grandchildren’ if either the parent or child threads terminate before the subprocess completes and returns.

The multiprocessing module has a nice interface to use pools with processes or threads. Depending on your current use case, you might consider using multiprocessing.pool.ThreadPool for your outer Pool, which will result in threads (that allow to spawn processes from within) as opposed to processes.

It might be limited by the GIL, but in my particular case (I tested both), the startup time for the processes from the outer Pool as created here far outweighed the solution with ThreadPool.

It’s really easy to swap Processes for Threads. Read more about how to use a ThreadPool solution here or here.

On some Python versions replacing standard Pool to custom can raise error: AssertionError: group argument must be None for now.

Here I found a solution that can help:

class NoDaemonProcess(multiprocessing.Process):
    # make 'daemon' attribute always return False
    @property
    def daemon(self):
        return False

    @daemon.setter
    def daemon(self, val):
        pass


class NoDaemonProcessPool(multiprocessing.pool.Pool):

    def Process(self, *args, **kwds):
        proc = super(NoDaemonProcessPool, self).Process(*args, **kwds)
        proc.__class__ = NoDaemonProcess

        return proc

concurrent.futures.ProcessPoolExecutor doesn’t have this limitation. It can have a nested process pool with no problem at all:

from concurrent.futures import ProcessPoolExecutor as Pool
from itertools import repeat
from multiprocessing import current_process
import time

def pid():
    return current_process().pid

def _square(i):  # Runs in inner_pool
    square = i ** 2
    time.sleep(i / 10)
    print(f'{pid()=} {i=} {square=}')
    return square

def _sum_squares(i, j):  # Runs in outer_pool
    with Pool(max_workers=2) as inner_pool:
        squares = inner_pool.map(_square, (i, j))
    sum_squares = sum(squares)
    time.sleep(sum_squares ** .5)
    print(f'{pid()=}, {i=}, {j=} {sum_squares=}')
    return sum_squares

def main():
    with Pool(max_workers=3) as outer_pool:
        for sum_squares in outer_pool.map(_sum_squares, range(5), repeat(3)):
            print(f'{pid()=} {sum_squares=}')

if __name__ == "__main__":
    main()

The above demonstration code was tested with Python 3.8.

A limitation of ProcessPoolExecutor, however, is that it doesn’t have maxtasksperchild. If you need this, consider the answer by Massimiliano instead.

Credit: answer by jfs

The issue I encountered was in trying to import globals between modules, causing the ProcessPool() line to get evaluated multiple times.

globals.py

from processing             import Manager, Lock
from pathos.multiprocessing import ProcessPool
from pathos.threading       import ThreadPool

class SingletonMeta(type):
    def __new__(cls, name, bases, dict):
        dict['__deepcopy__'] = dict['__copy__'] = lambda self, *args: self
        return super(SingletonMeta, cls).__new__(cls, name, bases, dict)

    def __init__(cls, name, bases, dict):
        super(SingletonMeta, cls).__init__(name, bases, dict)
        cls.instance = None

    def __call__(cls,*args,**kw):
        if cls.instance is None:
            cls.instance = super(SingletonMeta, cls).__call__(*args, **kw)
        return cls.instance

    def __deepcopy__(self, item):
        return item.__class__.instance

class Globals(object):
    __metaclass__ = SingletonMeta
    """     
    This class is a workaround to the bug: AssertionError: daemonic processes are not allowed to have children
     
    The root cause is that importing this file from different modules causes this file to be reevalutated each time, 
    thus ProcessPool() gets reexecuted inside that child thread, thus causing the daemonic processes bug    
    """
    def __init__(self):
        print "%s::__init__()" % (self.__class__.__name__)
        self.shared_manager      = Manager()
        self.shared_process_pool = ProcessPool()
        self.shared_thread_pool  = ThreadPool()
        self.shared_lock         = Lock()        # BUG: Windows: global name 'lock' is not defined | doesn't affect cygwin

Then import safely from elsewhere in your code

from globals import Globals
Globals().shared_manager      
Globals().shared_process_pool
Globals().shared_thread_pool  
Globals().shared_lock         

I have written a more expanded wrapper class around pathos.multiprocessing here:

• https://github.com/JamesMcGuigan/python2-timeseries-datapipeline/blob/master/src/util/MultiProcessing.py

As a side note, if your usecase just requires async multiprocess map as a performance optimization, then joblib will manage all your process pools behind the scenes and allow this very simple syntax:

squares = Parallel(-1)( delayed(lambda num: num**2)(x) for x in range(100) )

• https://joblib.readthedocs.io/

I have seen people dealing with this issue by using celery‘s fork of multiprocessing called billiard (multiprocessing pool extensions), which allows daemonic processes to spawn children. The walkaround is to simply replace the multiprocessing module by:

import billiard as multiprocessing