How to use multiprocessing queue in Python?

Question

I'm having much trouble trying to understand just how the multiprocessing queue works on python and how to implement it. Lets say I have two python modules that access data from a shared file, let's call these two modules a writer and a reader. My plan is to have both the reader and writer put requests into two separate multiprocessing queues, and then have a third process pop these requests in a loop and execute as such.

My main problem is that I really don't know how to implement multiprocessing.queue correctly, you cannot really instantiate the object for each process since they will be separate queues, how do you make sure that all processes relate to a shared queue (or in this case, queues)

Answer 1

Short Summary

As of CY2023, the technique described in this answer is quite out of date. These days, use concurrent.futures.ProcessPoolExecutor() instead of multiprocessing, below...

This answer describes the benefits and shortcomings of using concurrent.futures.ProcessPoolExecutor(). FYI, multiple python processes are sometimes used instead of threading to get the most benefit from concurrency. That said, python threading works pretty well as long as there is sufficient CPU activity to avoid the GIL (activity such as sending / receiving network traffic).

Original Answer

My main problem is that I really don't know how to implement multiprocessing.queue correctly, you cannot really instantiate the object for each process since they will be separate queues, how do you make sure that all processes relate to a shared queue (or in this case, queues)

This is a simple example of a reader and writer sharing a single queue... The writer sends a bunch of integers to the reader; when the writer runs out of numbers, it sends 'DONE', which lets the reader know to break out of the read loop.

You can spawn as many reader processes as you like...

from multiprocessing import Process, Queue
import time
import sys


def reader_proc(queue):
    """Read from the queue; this spawns as a separate Process"""
    while True:
        msg = queue.get()  # Read from the queue and do nothing
        if msg == "DONE":
            break


def writer(count, num_of_reader_procs, queue):
    """Write integers into the queue.  A reader_proc() will read them from the queue"""
    for ii in range(0, count):
        queue.put(ii)  # Put 'count' numbers into queue

    ### Tell all readers to stop...
    for ii in range(0, num_of_reader_procs):
        queue.put("DONE")


def start_reader_procs(qq, num_of_reader_procs):
    """Start the reader processes and return all in a list to the caller"""
    all_reader_procs = list()
    for ii in range(0, num_of_reader_procs):
        ### reader_p() reads from qq as a separate process...
        ###    you can spawn as many reader_p() as you like
        ###    however, there is usually a point of diminishing returns
        reader_p = Process(target=reader_proc, args=((qq),))
        reader_p.daemon = True
        reader_p.start()  # Launch reader_p() as another proc

        all_reader_procs.append(reader_p)

    return all_reader_procs


if __name__ == "__main__":
    num_of_reader_procs = 2
    qq = Queue()  # writer() writes to qq from _this_ process
    for count in [10**4, 10**5, 10**6]:
        assert 0 < num_of_reader_procs < 4
        all_reader_procs = start_reader_procs(qq, num_of_reader_procs)

        writer(count, len(all_reader_procs), qq)  # Queue stuff to all reader_p()
        print("All reader processes are pulling numbers from the queue...")

        _start = time.time()
        for idx, a_reader_proc in enumerate(all_reader_procs):
            print("    Waiting for reader_p.join() index %s" % idx)
            a_reader_proc.join()  # Wait for a_reader_proc() to finish

            print("        reader_p() idx:%s is done" % idx)

        print(
            "Sending {0} integers through Queue() took {1} seconds".format(
                count, (time.time() - _start)
            )
        )
        print("")

Answer 2

Here's a dead simple usage of multiprocessing.Queue and multiprocessing.Process that allows callers to send an "event" plus arguments to a separate process that dispatches the event to a "do_" method on the process. (Python 3.4+)

import multiprocessing as mp
import collections

Msg = collections.namedtuple('Msg', ['event', 'args'])

class BaseProcess(mp.Process):
    """A process backed by an internal queue for simple one-way message passing.
    """
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.queue = mp.Queue()

    def send(self, event, *args):
        """Puts the event and args as a `Msg` on the queue
        """
        msg = Msg(event, args)
        self.queue.put(msg)

    def dispatch(self, msg):
        event, args = msg

        handler = getattr(self, "do_%s" % event, None)
        if not handler:
            raise NotImplementedError("Process has no handler for [%s]" % event)

        handler(*args)

    def run(self):
        while True:
            msg = self.queue.get()
            self.dispatch(msg)

Usage:

class MyProcess(BaseProcess):
    def do_helloworld(self, arg1, arg2):
        print(arg1, arg2)

if __name__ == "__main__":
    process = MyProcess()
    process.start()
    process.send('helloworld', 'hello', 'world')

The send happens in the parent process, the do_* happens in the child process.

I left out any exception handling that would obviously interrupt the run loop and exit the child process. You can also customize it by overriding run to control blocking or whatever else.

This is really only useful in situations where you have a single worker process, but I think it's a relevant answer to this question to demonstrate a common scenario with a little more object-orientation.

Answer 3

Just made a simple and general example for demonstrating passing a message over a Queue between 2 standalone programs. It doesn't directly answer the OP's question but should be clear enough indicating the concept.

Server:

multiprocessing-queue-manager-server.py

import asyncio
import concurrent.futures
import multiprocessing
import multiprocessing.managers
import queue
import sys
import threading
from typing import Any, AnyStr, Dict, Union


class QueueManager(multiprocessing.managers.BaseManager):

    def get_queue(self, ident: Union[AnyStr, int, type(None)] = None) -> multiprocessing.Queue:
        pass


def get_queue(ident: Union[AnyStr, int, type(None)] = None) -> multiprocessing.Queue:
    global q

    if not ident in q:
        q[ident] = multiprocessing.Queue()

    return q[ident]


q: Dict[Union[AnyStr, int, type(None)], multiprocessing.Queue] = dict()
delattr(QueueManager, 'get_queue')


def init_queue_manager_server():
    if not hasattr(QueueManager, 'get_queue'):
        QueueManager.register('get_queue', get_queue)


def serve(no: int, term_ev: threading.Event):
    manager: QueueManager
    with QueueManager(authkey=QueueManager.__name__.encode()) as manager:
        print(f"Server address {no}: {manager.address}")

        while not term_ev.is_set():
            try:
                item: Any = manager.get_queue().get(timeout=0.1)
                print(f"Client {no}: {item} from {manager.address}")
            except queue.Empty:
                continue


async def main(n: int):
    init_queue_manager_server()
    term_ev: threading.Event = threading.Event()
    executor: concurrent.futures.ThreadPoolExecutor = concurrent.futures.ThreadPoolExecutor()

    i: int
    for i in range(n):
        asyncio.ensure_future(asyncio.get_running_loop().run_in_executor(executor, serve, i, term_ev))

    # Gracefully shut down
    try:
        await asyncio.get_running_loop().create_future()
    except asyncio.CancelledError:
        term_ev.set()
        executor.shutdown()
        raise


if __name__ == '__main__':
    asyncio.run(main(int(sys.argv[1])))

Client:

multiprocessing-queue-manager-client.py

import multiprocessing
import multiprocessing.managers
import os
import sys
from typing import AnyStr, Union


class QueueManager(multiprocessing.managers.BaseManager):

    def get_queue(self, ident: Union[AnyStr, int, type(None)] = None) -> multiprocessing.Queue:
        pass


delattr(QueueManager, 'get_queue')


def init_queue_manager_client():
    if not hasattr(QueueManager, 'get_queue'):
        QueueManager.register('get_queue')


def main():
    init_queue_manager_client()

    manager: QueueManager = QueueManager(sys.argv[1], authkey=QueueManager.__name__.encode())
    manager.connect()

    message = f"A message from {os.getpid()}"
    print(f"Message to send: {message}")
    manager.get_queue().put(message)


if __name__ == '__main__':
    main()

Usage

Server:

$ python3 multiprocessing-queue-manager-server.py N

N is a integer indicating how many servers should be created. Copy one of the <server-address-N> output by the server and make it the first argument of each multiprocessing-queue-manager-client.py.

Client:

python3 multiprocessing-queue-manager-client.py <server-address-1>

Result

Server:

Client 1: <item> from <server-address-1>

Gist: https://gist.github.com/89062d639e40110c61c2f88018a8b0e5

---

UPD: Created a package here.

Server:

import ipcq


with ipcq.QueueManagerServer(address=ipcq.Address.AUTO, authkey=ipcq.AuthKey.AUTO) as server:
    server.get_queue().get()

Client:

import ipcq


client = ipcq.QueueManagerClient(address=ipcq.Address.AUTO, authkey=ipcq.AuthKey.AUTO)
client.get_queue().put('a message')

Answer 4

A multi-producers and multi-consumers example, verified. It should be easy to modify it to cover other cases, single/multi producers, single/multi consumers.

from multiprocessing import Process, JoinableQueue
import time
import os

q = JoinableQueue()

def producer():
    for item in range(30):
        time.sleep(2)
        q.put(item)
    pid = os.getpid()
    print(f'producer {pid} done')


def worker():
    while True:
        item = q.get()
        pid = os.getpid()
        print(f'pid {pid} Working on {item}')
        print(f'pid {pid} Finished {item}')
        q.task_done()

for i in range(5):
    p = Process(target=worker, daemon=True).start()

# send thirty task requests to the worker
producers = []
for i in range(2):
    p = Process(target=producer)
    producers.append(p)
    p.start()

# make sure producers done
for p in producers:
    p.join()

# block until all workers are done
q.join()
print('All work completed')

Explanation:

1. Two producers and five consumers in this example.
2. JoinableQueue is used to make sure all elements stored in queue will be processed. 'task_done' is for worker to notify an element is done. 'q.join()' will wait for all elements marked as done.
3. With #2, there is no need to join wait for every worker.
4. But it is important to join wait for every producer to store element into queue. Otherwise, program exit immediately.

Answer 5

I had a look at multiple answers across stack overflow and the web while trying to set-up a way of doing multiprocessing using queues for passing around large pandas dataframes. It seemed to me that every answer was re-iterating the same kind of solutions without any consideration of the multitude of edge cases one will definitely come across when setting up calculations like these. The problem is that there is many things at play at the same time. The number of tasks, the number of workers, the duration of each task and possible exceptions during task execution. All of these make synchronization tricky and most answers do not address how you can go about it. So this is my take after fiddling around for a few hours, hopefully this will be generic enough for most people to find it useful.

Some thoughts before any coding examples. Since queue.Empty or queue.qsize() or any other similar method is unreliable for flow control, any code of the like

while True:
    try:
        task = pending_queue.get_nowait()
    except queue.Empty:
        break

is bogus. This will kill the worker even if milliseconds later another task turns up in the queue. The worker will not recover and after a while ALL the workers will disappear as they randomly find the queue momentarily empty. The end result will be that the main multiprocessing function (the one with the join() on the processes) will return without all the tasks having completed. Nice. Good luck debugging through that if you have thousands of tasks and a few are missing.

The other issue is the use of sentinel values. Many people have suggested adding a sentinel value in the queue to flag the end of the queue. But to flag it to whom exactly? If there is N workers, assuming N is the number of cores available give or take, then a single sentinel value will only flag the end of the queue to one worker. All the other workers will sit waiting for more work when there is none left. Typical examples I've seen are

while True:
    task = pending_queue.get()
    if task == SOME_SENTINEL_VALUE:
        break

One worker will get the sentinel value while the rest will wait indefinitely. No post I came across mentioned that you need to submit the sentinel value to the queue AT LEAST as many times as you have workers so that ALL of them get it.

The other issue is the handling of exceptions during task execution. Again these should be caught and managed. Moreover, if you have a completed_tasks queue you should independently count in a deterministic way how many items are in the queue before you decide that the job is done. Again relying on queue sizes is bound to fail and returns unexpected results.

In the example below, the par_proc() function will receive a list of tasks including the functions with which these tasks should be executed alongside any named arguments and values.

import multiprocessing as mp
import dill as pickle
import queue
import time
import psutil

SENTINEL = None


def do_work(tasks_pending, tasks_completed):
    # Get the current worker's name
    worker_name = mp.current_process().name

    while True:
        try:
            task = tasks_pending.get_nowait()
        except queue.Empty:
            print(worker_name + ' found an empty queue. Sleeping for a while before checking again...')
            time.sleep(0.01)
        else:
            try:
                if task == SENTINEL:
                    print(worker_name + ' no more work left to be done. Exiting...')
                    break

                print(worker_name + ' received some work... ')
                time_start = time.perf_counter()
                work_func = pickle.loads(task['func'])
                result = work_func(**task['task'])
                tasks_completed.put({work_func.__name__: result})
                time_end = time.perf_counter() - time_start
                print(worker_name + ' done in {} seconds'.format(round(time_end, 5)))
            except Exception as e:
                print(worker_name + ' task failed. ' + str(e))
                tasks_completed.put({work_func.__name__: None})


def par_proc(job_list, num_cpus=None):

    # Get the number of cores
    if not num_cpus:
        num_cpus = psutil.cpu_count(logical=False)

    print('* Parallel processing')
    print('* Running on {} cores'.format(num_cpus))

    # Set-up the queues for sending and receiving data to/from the workers
    tasks_pending = mp.Queue()
    tasks_completed = mp.Queue()

    # Gather processes and results here
    processes = []
    results = []

    # Count tasks
    num_tasks = 0

    # Add the tasks to the queue
    for job in job_list:
        for task in job['tasks']:
            expanded_job = {}
            num_tasks = num_tasks + 1
            expanded_job.update({'func': pickle.dumps(job['func'])})
            expanded_job.update({'task': task})
            tasks_pending.put(expanded_job)

    # Use as many workers as there are cores (usually chokes the system so better use less)
    num_workers = num_cpus

    # We need as many sentinels as there are worker processes so that ALL processes exit when there is no more
    # work left to be done.
    for c in range(num_workers):
        tasks_pending.put(SENTINEL)

    print('* Number of tasks: {}'.format(num_tasks))

    # Set-up and start the workers
    for c in range(num_workers):
        p = mp.Process(target=do_work, args=(tasks_pending, tasks_completed))
        p.name = 'worker' + str(c)
        processes.append(p)
        p.start()

    # Gather the results
    completed_tasks_counter = 0
    while completed_tasks_counter < num_tasks:
        results.append(tasks_completed.get())
        completed_tasks_counter = completed_tasks_counter + 1

    for p in processes:
        p.join()

    return results

And here is a test to run the above code against

def test_parallel_processing():
    def heavy_duty1(arg1, arg2, arg3):
        return arg1 + arg2 + arg3

    def heavy_duty2(arg1, arg2, arg3):
        return arg1 * arg2 * arg3

    task_list = [
        {'func': heavy_duty1, 'tasks': [{'arg1': 1, 'arg2': 2, 'arg3': 3}, {'arg1': 1, 'arg2': 3, 'arg3': 5}]},
        {'func': heavy_duty2, 'tasks': [{'arg1': 1, 'arg2': 2, 'arg3': 3}, {'arg1': 1, 'arg2': 3, 'arg3': 5}]},
    ]

    results = par_proc(task_list)

    job1 = sum([y for x in results if 'heavy_duty1' in x.keys() for y in list(x.values())])
    job2 = sum([y for x in results if 'heavy_duty2' in x.keys() for y in list(x.values())])

    assert job1 == 15
    assert job2 == 21

plus another one with some exceptions

def test_parallel_processing_exceptions():
    def heavy_duty1_raises(arg1, arg2, arg3):
        raise ValueError('Exception raised')
        return arg1 + arg2 + arg3

    def heavy_duty2(arg1, arg2, arg3):
        return arg1 * arg2 * arg3

    task_list = [
        {'func': heavy_duty1_raises, 'tasks': [{'arg1': 1, 'arg2': 2, 'arg3': 3}, {'arg1': 1, 'arg2': 3, 'arg3': 5}]},
        {'func': heavy_duty2, 'tasks': [{'arg1': 1, 'arg2': 2, 'arg3': 3}, {'arg1': 1, 'arg2': 3, 'arg3': 5}]},
    ]

    results = par_proc(task_list)

    job1 = sum([y for x in results if 'heavy_duty1' in x.keys() for y in list(x.values())])
    job2 = sum([y for x in results if 'heavy_duty2' in x.keys() for y in list(x.values())])

    assert not job1
    assert job2 == 21

Hope that is helpful.

Answer 6

We implemented two versions of this, one a simple multi thread pool that can execute many types of callables, making our lives much easier and the second version that uses processes, which is less flexible in terms of callables and requires and extra call to dill.

Setting frozen_pool to true will freeze execution until finish_pool_queue is called in either class.

Thread Version:

'''
Created on Nov 4, 2019

@author: Kevin
'''
from threading import Lock, Thread
from Queue import Queue
import traceback
from helium.loaders.loader_retailers import print_info
from time import sleep
import signal
import os

class ThreadPool(object):
    def __init__(self, queue_threads, *args, **kwargs):
        self.frozen_pool = kwargs.get('frozen_pool', False)
        self.print_queue = kwargs.get('print_queue', True)
        self.pool_results = []
        self.lock = Lock()
        self.queue_threads = queue_threads
        self.queue = Queue()
        self.threads = []

        for i in range(self.queue_threads):
            t = Thread(target=self.make_pool_call)
            t.daemon = True
            t.start()
            self.threads.append(t)

    def make_pool_call(self):
        while True:
            if self.frozen_pool:
                #print '--> Queue is frozen'
                sleep(1)
                continue

            item = self.queue.get()
            if item is None:
                break

            call = item.get('call', None)
            args = item.get('args', [])
            kwargs = item.get('kwargs', {})
            keep_results = item.get('keep_results', False)

            try:
                result = call(*args, **kwargs)

                if keep_results:
                    self.lock.acquire()
                    self.pool_results.append((item, result))
                    self.lock.release()

            except Exception as e:
                self.lock.acquire()
                print e
                traceback.print_exc()
                self.lock.release()
                os.kill(os.getpid(), signal.SIGUSR1)

            self.queue.task_done()

    def finish_pool_queue(self):
        self.frozen_pool = False

        while self.queue.unfinished_tasks > 0:
            if self.print_queue:
                print_info('--> Thread pool... %s' % self.queue.unfinished_tasks)
            sleep(5)

        self.queue.join()

        for i in range(self.queue_threads):
            self.queue.put(None)

        for t in self.threads:
            t.join()

        del self.threads[:]

    def get_pool_results(self):
        return self.pool_results

    def clear_pool_results(self):
        del self.pool_results[:]

Process Version:

  '''
Created on Nov 4, 2019

@author: Kevin
'''
import traceback
from helium.loaders.loader_retailers import print_info
from time import sleep
import signal
import os
from multiprocessing import Queue, Process, Value, Array, JoinableQueue, Lock,\
    RawArray, Manager
from dill import dill
import ctypes
from helium.misc.utils import ignore_exception
from mem_top import mem_top
import gc

class ProcessPool(object):
    def __init__(self, queue_processes, *args, **kwargs):
        self.frozen_pool = Value(ctypes.c_bool, kwargs.get('frozen_pool', False))
        self.print_queue = kwargs.get('print_queue', True)
        self.manager = Manager()
        self.pool_results = self.manager.list()
        self.queue_processes = queue_processes
        self.queue = JoinableQueue()
        self.processes = []

        for i in range(self.queue_processes):
            p = Process(target=self.make_pool_call)
            p.start()
            self.processes.append(p)

        print 'Processes', self.queue_processes

    def make_pool_call(self):
        while True:
            if self.frozen_pool.value:
                sleep(1)
                continue

            item_pickled = self.queue.get()

            if item_pickled is None:
                #print '--> Ending'
                self.queue.task_done()
                break

            item = dill.loads(item_pickled)

            call = item.get('call', None)
            args = item.get('args', [])
            kwargs = item.get('kwargs', {})
            keep_results = item.get('keep_results', False)

            try:
                result = call(*args, **kwargs)

                if keep_results:
                    self.pool_results.append(dill.dumps((item, result)))
                else:
                    del call, args, kwargs, keep_results, item, result

            except Exception as e:
                print e
                traceback.print_exc()
                os.kill(os.getpid(), signal.SIGUSR1)

            self.queue.task_done()

    def finish_pool_queue(self, callable=None):
        self.frozen_pool.value = False

        while self.queue._unfinished_tasks.get_value() > 0:
            if self.print_queue:
                print_info('--> Process pool... %s' % (self.queue._unfinished_tasks.get_value()))

            if callable:
                callable()

            sleep(5)

        for i in range(self.queue_processes):
            self.queue.put(None)

        self.queue.join()
        self.queue.close()

        for p in self.processes:
            with ignore_exception: p.join(10)
            with ignore_exception: p.terminate()

        with ignore_exception: del self.processes[:]

    def get_pool_results(self):
        return self.pool_results

    def clear_pool_results(self):
        del self.pool_results[:]

def test(eg):
        print 'EG', eg

Call with either:

tp = ThreadPool(queue_threads=2)
tp.queue.put({'call': test, 'args': [random.randint(0, 100)]})
tp.finish_pool_queue()

or

pp = ProcessPool(queue_processes=2)
pp.queue.put(dill.dumps({'call': test, 'args': [random.randint(0, 100)]}))
pp.queue.put(dill.dumps({'call': test, 'args': [random.randint(0, 100)]}))
pp.finish_pool_queue()

Answer 7

in "from queue import Queue" there is no module called queue, instead multiprocessing should be used. Therefore, it should look like "from multiprocessing import Queue"