[python] How can I recover the return value of a function passed to multiprocessing.Process?

In the example code below, I'd like to recover the return value of the function worker. How can I go about doing this? Where is this value stored?

Example Code:

import multiprocessing

def worker(procnum):
    '''worker function'''
    print str(procnum) + ' represent!'
    return procnum


if __name__ == '__main__':
    jobs = []
    for i in range(5):
        p = multiprocessing.Process(target=worker, args=(i,))
        jobs.append(p)
        p.start()

    for proc in jobs:
        proc.join()
    print jobs

Output:

0 represent!
1 represent!
2 represent!
3 represent!
4 represent!
[<Process(Process-1, stopped)>, <Process(Process-2, stopped)>, <Process(Process-3, stopped)>, <Process(Process-4, stopped)>, <Process(Process-5, stopped)>]

I can't seem to find the relevant attribute in the objects stored in jobs.

A simple solution:

import multiprocessing

output=[]
data = range(0,10)

def f(x):
    return x**2

def handler():
    p = multiprocessing.Pool(64)
    r=p.map(f, data)
    return r

if __name__ == '__main__':
    output.append(handler())

print(output[0])

Output:

[0, 1, 4, 9, 16, 25, 36, 49, 64, 81]

For some reason, I couldn't find a general example of how to do this with Queue anywhere (even Python's doc examples don't spawn multiple processes), so here's what I got working after like 10 tries:

def add_helper(queue, arg1, arg2): # the func called in child processes
    ret = arg1 + arg2
    queue.put(ret)

def multi_add(): # spawns child processes
    q = Queue()
    processes = []
    rets = []
    for _ in range(0, 100):
        p = Process(target=add_helper, args=(q, 1, 2))
        processes.append(p)
        p.start()
    for p in processes:
        ret = q.get() # will block
        rets.append(ret)
    for p in processes:
        p.join()
    return rets

Queue is a blocking, thread-safe queue that you can use to store the return values from the child processes. So you have to pass the queue to each process. Something less obvious here is that you have to get() from the queue before you join the Processes or else the queue fills up and blocks everything.

Update for those who are object-oriented (tested in Python 3.4):

from multiprocessing import Process, Queue

class Multiprocessor():

    def __init__(self):
        self.processes = []
        self.queue = Queue()

    @staticmethod
    def _wrapper(func, queue, args, kwargs):
        ret = func(*args, **kwargs)
        queue.put(ret)

    def run(self, func, *args, **kwargs):
        args2 = [func, self.queue, args, kwargs]
        p = Process(target=self._wrapper, args=args2)
        self.processes.append(p)
        p.start()

    def wait(self):
        rets = []
        for p in self.processes:
            ret = self.queue.get()
            rets.append(ret)
        for p in self.processes:
            p.join()
        return rets

# tester
if __name__ == "__main__":
    mp = Multiprocessor()
    num_proc = 64
    for _ in range(num_proc): # queue up multiple tasks running `sum`
        mp.run(sum, [1, 2, 3, 4, 5])
    ret = mp.wait() # get all results
    print(ret)
    assert len(ret) == num_proc and all(r == 15 for r in ret)

This example shows how to use a list of multiprocessing.Pipe instances to return strings from an arbitrary number of processes:

import multiprocessing

def worker(procnum, send_end):
    '''worker function'''
    result = str(procnum) + ' represent!'
    print result
    send_end.send(result)

def main():
    jobs = []
    pipe_list = []
    for i in range(5):
        recv_end, send_end = multiprocessing.Pipe(False)
        p = multiprocessing.Process(target=worker, args=(i, send_end))
        jobs.append(p)
        pipe_list.append(recv_end)
        p.start()

    for proc in jobs:
        proc.join()
    result_list = [x.recv() for x in pipe_list]
    print result_list

if __name__ == '__main__':
    main()

Output:

0 represent!
1 represent!
2 represent!
3 represent!
4 represent!
['0 represent!', '1 represent!', '2 represent!', '3 represent!', '4 represent!']

This solution uses fewer resources than a multiprocessing.Queue which uses

• a Pipe
• at least one Lock
• a buffer
• a thread

or a multiprocessing.SimpleQueue which uses

• a Pipe
• at least one Lock

It is very instructive to look at the source for each of these types.

For anyone else who is seeking how to get a value from a Process using Queue:

import multiprocessing

ret = {'foo': False}

def worker(queue):
    ret = queue.get()
    ret['foo'] = True
    queue.put(ret)

if __name__ == '__main__':
    queue = multiprocessing.Queue()
    queue.put(ret)
    p = multiprocessing.Process(target=worker, args=(queue,))
    p.start()
    p.join()
    print(queue.get())  # Prints {"foo": True}

Note that in Windows or Jupyter Notebook, with multithreading you have to save this as a file and execute the file. If you do it in a command prompt you will see an error like this:

 AttributeError: Can't get attribute 'worker' on <module '__main__' (built-in)>

The pebble package has a nice abstraction leveraging multiprocessing.Pipe which makes this quite straightforward:

from pebble import concurrent

@concurrent.process
def function(arg, kwarg=0):
    return arg + kwarg

future = function(1, kwarg=1)

print(future.result())

Example from: https://pythonhosted.org/Pebble/#concurrent-decorators

It seems that you should use the multiprocessing.Pool class instead and use the methods .apply() .apply_async(), map()

http://docs.python.org/library/multiprocessing.html?highlight=pool#multiprocessing.pool.AsyncResult

If you are using Python 3, you can use concurrent.futures.ProcessPoolExecutor as a convenient abstraction:

from concurrent.futures import ProcessPoolExecutor

def worker(procnum):
    '''worker function'''
    print(str(procnum) + ' represent!')
    return procnum


if __name__ == '__main__':
    with ProcessPoolExecutor() as executor:
        print(list(executor.map(worker, range(5))))

Output:

0 represent!
1 represent!
2 represent!
3 represent!
4 represent!
[0, 1, 2, 3, 4]

You can use the exit built-in to set the exit code of a process. It can be obtained from the exitcode attribute of the process:

import multiprocessing

def worker(procnum):
    print str(procnum) + ' represent!'
    exit(procnum)

if __name__ == '__main__':
    jobs = []
    for i in range(5):
        p = multiprocessing.Process(target=worker, args=(i,))
        jobs.append(p)
        p.start()

    result = []
    for proc in jobs:
        proc.join()
        result.append(proc.exitcode)
    print result

Output:

0 represent!
1 represent!
2 represent!
3 represent!
4 represent!
[0, 1, 2, 3, 4]

I think the approach suggested by @sega_sai is the better one. But it really needs a code example, so here goes:

import multiprocessing
from os import getpid

def worker(procnum):
    print('I am number %d in process %d' % (procnum, getpid()))
    return getpid()

if __name__ == '__main__':
    pool = multiprocessing.Pool(processes = 3)
    print(pool.map(worker, range(5)))

Which will print the return values:

I am number 0 in process 19139
I am number 1 in process 19138
I am number 2 in process 19140
I am number 3 in process 19139
I am number 4 in process 19140
[19139, 19138, 19140, 19139, 19140]

If you are familiar with map (the Python 2 built-in) this should not be too challenging. Otherwise have a look at sega_Sai's link.

Note how little code is needed. (Also note how processes are re-used).

Thought I'd simplify the simplest examples copied from above, working for me on Py3.6. Simplest is multiprocessing.Pool:

import multiprocessing
import time

def worker(x):
    time.sleep(1)
    return x

pool = multiprocessing.Pool()
print(pool.map(worker, range(10)))

You can set the number of processes in the pool with, e.g., Pool(processes=5). However it defaults to CPU count, so leave it blank for CPU-bound tasks. (I/O-bound tasks often suit threads anyway, as the threads are mostly waiting so can share a CPU core.) Pool also applies chunking optimization.

(Note that the worker method cannot be nested within a method. I initially defined my worker method inside the method that makes the call to pool.map, to keep it all self-contained, but then the processes couldn't import it, and threw "AttributeError: Can't pickle local object outer_method..inner_method". More here. It can be inside a class.)

(Appreciate the original question specified printing 'represent!' rather than time.sleep(), but without it I thought some code was running concurrently when it wasn't.)

Py3's ProcessPoolExecutor is also two lines (.map returns a generator so you need the list()):

from concurrent.futures import ProcessPoolExecutor
with ProcessPoolExecutor() as executor:
    print(list(executor.map(worker, range(10))))

With plain Processes:

import multiprocessing
import time

def worker(x, queue):
    time.sleep(1)
    queue.put(x)

queue = multiprocessing.SimpleQueue()
tasks = range(10)

for task in tasks:
    multiprocessing.Process(target=worker, args=(task, queue,)).start()

for _ in tasks:
    print(queue.get())

Use SimpleQueue if all you need is put and get. The first loop starts all the processes, before the second makes the blocking queue.get calls. I don't think there's any reason to call p.join() too.

I modified vartec's answer a bit since I needed to get the error codes from the function. (Thanks vertec!!! its an awesome trick)

This can also be done with a manager.list but I think is better to have it in a dict and store a list within it. That way, way we keep the function and the results since we can't be sure of the order in which the list will be populated.

from multiprocessing import Process
import time
import datetime
import multiprocessing


def func1(fn, m_list):
    print 'func1: starting'
    time.sleep(1)
    m_list[fn] = "this is the first function"
    print 'func1: finishing'
    # return "func1"  # no need for return since Multiprocess doesnt return it =(

def func2(fn, m_list):
    print 'func2: starting'
    time.sleep(3)
    m_list[fn] = "this is function 2"
    print 'func2: finishing'
    # return "func2"

def func3(fn, m_list):
    print 'func3: starting'
    time.sleep(9)
    # if fail wont join the rest because it never populate the dict
    # or do a try/except to get something in return.
    raise ValueError("failed here")
    # if we want to get the error in the manager dict we can catch the error
    try:
        raise ValueError("failed here")
        m_list[fn] = "this is third"
    except:
        m_list[fn] = "this is third and it fail horrible"
        # print 'func3: finishing'
        # return "func3"


def runInParallel(*fns):  # * is to accept any input in list
    start_time = datetime.datetime.now()
    proc = []
    manager = multiprocessing.Manager()
    m_list = manager.dict()
    for fn in fns:
        # print fn
        # print dir(fn)
        p = Process(target=fn, name=fn.func_name, args=(fn, m_list))
        p.start()
        proc.append(p)
    for p in proc:
        p.join()  # 5 is the time out

    print datetime.datetime.now() - start_time
    return m_list, proc

if __name__ == '__main__':
    manager, proc = runInParallel(func1, func2, func3)
    # print dir(proc[0])
    # print proc[0]._name
    # print proc[0].name
    # print proc[0].exitcode

    # here you can check what did fail
    for i in proc:
        print i.name, i.exitcode  # name was set up in the Process line 53

    # here will only show the function that worked and where able to populate the 
    # manager dict
    for i, j in manager.items():
        print dir(i)  # things you can do to the function
        print i, j