Develop Reference

Async IO switch coroutine on multiprocessing.queue.get() not ready

I am writing some code where I have 3 processes (spawned from the main). The first one is a process that uses Async IO to create 3 coroutines and switch between them. The last two processes run independently and generate two outputs that are used in one of the coroutines of the first process.
The communication has been managed using multiprocessing.queue(), the main puts the input data inside queue_source_position_hrir_calculator and queue_source_position_cutoff_calculator, then these two queues are emptied by p2_hrir_computation_process and p3_cutoff_computation_process. These two processes outputs their computation results in two output queues queue_computed_hrirs and queue_computed_cutoff
Finally these two queues are consumed by the Async IO process, in particular inside the input_parameters_coroutine function.
The full code is the following (I will highlight the key parts in following snippets):
import asyncio
import multiprocessing
import numpy as np
import time
from classes.HRIR_interpreter_min_phase_linear_interpolation import HRIR_interpreter_min_phase_linear_interpolation
from classes.object_renderer import ObjectRenderer
#Useful resources: https://bbc.github.io/cloudfit-public-docs/asyncio/asyncio-part-2
#https://realpython.com/async-io-python/
Fs = 44100
# region Async_IO functions
async def audio_input_coroutine(overlay):
for i in range(0,100):
print('Executing audio input coroutine')
print(overlay)
await asyncio.sleep(1/(Fs*4))
async def input_parameters_coroutine(overlay, queue_computed_hrirs,queue_computed_cutoff):
for i in range(0,10):
print('Executing audio input_parameters coroutine')
#print(overlay)
current_hrir = queue_computed_hrirs.get()
print('got current hrir')
current_cutoff = queue_computed_cutoff.get()
print('got current cutoff')
await asyncio.sleep(0.5)
async def audio_output_coroutine(overlay):
for i in range(0,10):
print('Executing audio_output coroutine')
#print(overlay)
await asyncio.sleep(0.5)
async def main_coroutine(overlay, queue_computed_hrirs,queue_computed_cutoff):
await asyncio.gather(audio_input_coroutine(overlay), input_parameters_coroutine(overlay, queue_computed_hrirs,queue_computed_cutoff), audio_output_coroutine(overlay))
def async_IO_main_process(queue_computed_hrirs,queue_computed_cutoff):
overlay = 10
asyncio.run(main_coroutine(overlay, queue_computed_hrirs,queue_computed_cutoff))
# endregion
# region HRIR_computation_process
def compute_hrir(queue_source_position, queue_computed_hrirs):
print('computing hrir')
SOFA_filename = '../HRTF_data/HUTUBS_min_phase.sofa'
# loading the simulated dataset using the support class HRIRInterpreter
HRIRInterpreter = HRIR_interpreter_min_phase_linear_interpolation(SOFA_filename=SOFA_filename)
# variable to check if I have other positions in my input queue
eof_source_position = False
# Un-comment following line to return when no more messages
while not eof_source_position:
#while True:
# print('inside while loop')
time.sleep(1)
# print('state of the queue', queue_source_position.empty())
if not eof_source_position:
position = queue_source_position.get()
if position is None:
eof_source_position = True # end of messages indicator
else:
required_IR = HRIRInterpreter.get_interpolated_IR(position[0], position[1], 1)
queue_computed_hrirs.put(required_IR)
# print('printing computed HRIR:', required_IR)
print('completed hrir computation, adding none to queue')
queue_computed_hrirs.put(None) # end of messages indicator
print('completed hrir process')
# endregion
# region cutoff_computation_process
def compute_cutoff(queue_source_position, queue_computed_cutoff):
print('computing cutoff')
cutoff = 20000
object_renderer = ObjectRenderer()
object_positions = np.array([(20, 0), (40, 0), (100, 0), (225, 0)])
eof_source_position = False
# Un-comment following line to return when no more messages
while not eof_source_position:
#while True:
time.sleep(1)
object_renderer.update_object_position(object_positions)
if not eof_source_position:
print('inside source position update')
source_position = queue_source_position.get()
if source_position is None: # end of messages indicator
eof_source_position = True
else:
cutoff = object_renderer.get_cutoff(azimuth=source_position[0], elevation=source_position[1])
queue_computed_cutoff.put(cutoff)
queue_computed_cutoff.put(None) # end of messages indicator
# endregion
if __name__ == "__main__":
import time
queue_source_position_hrir_calculator = multiprocessing.Queue()
queue_source_position_cutoff_calculator = multiprocessing.Queue()
queue_computed_hrirs = multiprocessing.Queue()
queue_computed_cutoff = multiprocessing.Queue()
i = 0.0
#Basically here I am writing a sequence of positions into the queue
#then I add a None value to detect when I am done with the simulation so the process can end
for _ in range(10):
# print('into main while-> source_position:', source_position[0])
source_position = np.array([i, 0.0])
queue_source_position_hrir_calculator.put(source_position)
queue_source_position_cutoff_calculator.put(source_position)
i += 10
queue_source_position_hrir_calculator.put(None) # "end of messages" indicator
queue_source_position_cutoff_calculator.put(None) # "end of messages" indicator
p1_async_IO_process = multiprocessing.Process(target=async_IO_main_process, args=(queue_computed_hrirs,queue_computed_cutoff)) #process that manages the ASYNC_IO coroutines between DMAs
p2_hrir_computation_process = multiprocessing.Process(target=compute_hrir, args=(queue_source_position_hrir_calculator, queue_computed_hrirs))
p3_cutoff_computation_process = multiprocessing.Process(target=compute_hrir, args=(queue_source_position_cutoff_calculator, queue_computed_cutoff))
p1_async_IO_process.start()
p2_hrir_computation_process.start()
p3_cutoff_computation_process.start()
#temp cycle to join processes
#for _ in range(2):
# current_hrir = queue_computed_hrirs.get()
# current_cutoff = queue_computed_cutoff.get()
print('joining async_IO process')
p1_async_IO_process.join()
print('joined async_IO process')
#NB: to join a process, its qeues must be empty. So before calling the join on p2, I should get the values from the queue_computed_hrirs queue
print('joining hrir computation process')
p2_hrir_computation_process.join()
print('joined hrir computation process')
print('joining hrir computation process')
p2_hrir_computation_process.join()
print('joined hrir computation process')
print('joining cutoff computation process')
p3_cutoff_computation_process.join()
print('joined cutoff computation process')
print("completed main")
The important part of the code is:
async def input_parameters_coroutine(overlay, queue_computed_hrirs,queue_computed_cutoff):
for i in range(0,10):
print('Executing audio input_parameters coroutine')
#print(overlay)
current_hrir = queue_computed_hrirs.get()
print('got current hrir')
current_cutoff = queue_computed_cutoff.get()
print('got current cutoff')
await asyncio.sleep(0.5)
This coroutine receives as input 3 variables overlay (which is a dummy variable I am using for future developments) and the two multiprocessing.Queue() classes, queue_computed_hrirs and queue_computed_cutoff.
At the moment my input_parameters_coroutine gets "stuck" while executing current_hrir = queue_computed_hrirs.get() and current_cutoff = queue_computed_cutoff.get(). I said "stuck" because the code works fine and complete its execution, the problem is that those two commands are blocking, thus my coroutine stops until it has something to get from the queue.
What I would like to achieve is: try to execute current_hrir = queue_computed_hrirs.get(), if it is not possible at that moment, switch to another coroutine and let it execute what it wants, then go back and check if it possible to execute current_hrir = queue_computed_hrirs.get(), if yes go on, if not switch again to another coroutine and let it do its job.
I saw that there are some problems in making async IO and multiprocessing communicate ( What kind of problems (if any) would there be combining asyncio with multiprocessing? , Can I somehow share an asynchronous queue with a subprocess? ) but I wasn't able to find a smart solution to my problem.

What does "The process has forked and you cannot use this CoreFoundation functionality safely. You MUST exec()." error mean?

What does this error mean? Seems this happens when I use some C++ wrapped code under process (no errors when I use it in a plain for loop).
The process has forked and you cannot use this CoreFoundation functionality safely. You MUST exec().
Break on __THE_PROCESS_HAS_FORKED_AND_YOU_CANNOT_USE_THIS_COREFOUNDATION_FUNCTIONALITY___YOU_MUST_EXEC__() to debug.
Example code:
def split_to_chunks(items, k):
n = len(items) // k
chunks = []
for i in range(0, len(items), n):
chunks.append(items[i:i + n])
return chunks
def process_chunk(process_id, img_filepaths, output_dir):
print('Enter process_id:', process_id)
for img_filepath in img_filepaths:
try:
//Code section
except Exception as e:
print(f'Error: process_id: {process_id} : {img_filepath}', e)
print('End process_id:', process_id)
img_filepaths = get_img_filepaths(images_dir)
#n_processes = multiprocessing.cpu_count() // 2
n_processes = 1
img_filepaths_chunks = split_to_chunks(img_filepaths, n_processes)
processes = []
for i in range(n_processes):
process = multiprocessing.Process(
target=process_chunk,
args=(i, img_filepaths_chunks[i], output_dir)
)
processes.append(process)
for process in processes:
process.start()
for i, process in enumerate(processes):
process.join()
if process.exitcode != 0:
print(f'Error: process_id: {i} failed.')
#sys.exit(1)

Seems using 'spawn' like suggested here https://turtlemonvh.github.io/python-multiprocessing-and-corefoundation-libraries.html is solving the problem.
import multiprocessing as mp
mp.set_start_method('spawn')

How to log to single file with multiprocessing.Pool.apply_async

I can't get logging to a single file working with multprocess.Pool.apply_async.
I'm trying to adapt this example from the Logging Cookbook, but it only works for multiprocessing.Process. Passing the logging queue into apply_async doesn't seem to have effect.
I would like to use a Pool so that I can easily manage the number of simultaneous threads.
The following adapted example with multiprocessing.Process works ok for me, except I am not getting log messages from the main process, and I don't think it will work well when I have 100 large jobs.
import logging
import logging.handlers
import numpy as np
import time
import multiprocessing
import pandas as pd
log_file = 'PATH_TO_FILE/log_file.log'
def listener_configurer():
root = logging.getLogger()
h = logging.FileHandler(log_file)
f = logging.Formatter('%(asctime)s %(processName)-10s %(name)s %(levelname)-8s %(message)s')
h.setFormatter(f)
root.addHandler(h)
# This is the listener process top-level loop: wait for logging events
# (LogRecords)on the queue and handle them, quit when you get a None for a
# LogRecord.
def listener_process(queue, configurer):
configurer()
while True:
try:
record = queue.get()
if record is None: # We send this as a sentinel to tell the listener to quit.
break
logger = logging.getLogger(record.name)
logger.handle(record) # No level or filter logic applied - just do it!
except Exception:
import sys, traceback
print('Whoops! Problem:', file=sys.stderr)
traceback.print_exc(file=sys.stderr)
def worker_configurer(queue):
h = logging.handlers.QueueHandler(queue) # Just the one handler needed
root = logging.getLogger()
root.addHandler(h)
# send all messages, for demo; no other level or filter logic applied.
root.setLevel(logging.DEBUG)
# This is the worker process top-level loop, which just logs ten events with
# random intervening delays before terminating.
# The print messages are just so you know it's doing something!
def worker_function(sleep_time, name, queue, configurer):
configurer(queue)
start_message = 'Worker {} started and will now sleep for {}s'.format(name, sleep_time)
logging.info(start_message)
time.sleep(sleep_time)
success_message = 'Worker {} has finished sleeping for {}s'.format(name, sleep_time)
logging.info(success_message)
def main_with_process():
start_time = time.time()
single_thread_time = 0.
queue = multiprocessing.Queue(-1)
listener = multiprocessing.Process(target=listener_process,
args=(queue, listener_configurer))
listener.start()
workers = []
for i in range(10):
name = str(i)
sleep_time = np.random.randint(10) / 2
single_thread_time += sleep_time
worker = multiprocessing.Process(target=worker_function,
args=(sleep_time, name, queue, worker_configurer))
workers.append(worker)
worker.start()
for w in workers:
w.join()
queue.put_nowait(None)
listener.join()
end_time = time.time()
final_message = "Script execution time was {}s, but single-thread time was {}s".format(
(end_time - start_time),
single_thread_time
)
print(final_message)
if __name__ == "__main__":
main_with_process()
But I can't get the following adaptation to work:
def main_with_pool():
start_time = time.time()
queue = multiprocessing.Queue(-1)
listener = multiprocessing.Process(target=listener_process,
args=(queue, listener_configurer))
listener.start()
pool = multiprocessing.Pool(processes=3)
job_list = [np.random.randint(10) / 2 for i in range(10)]
single_thread_time = np.sum(job_list)
for i, sleep_time in enumerate(job_list):
name = str(i)
pool.apply_async(worker_function,
args=(sleep_time, name, queue, worker_configurer))
queue.put_nowait(None)
listener.join()
end_time = time.time()
print("Script execution time was {}s, but single-thread time was {}s".format(
(end_time - start_time),
single_thread_time
))
if __name__ == "__main__":
main_with_pool()
I've tried many slight variations, using multiprocessing.Manager, multiprocessing.Queue, multiprocessing.get_logger, apply_async.get(), but haven't gotten any to work.
I would think there would be an off-the-shelf solution for this. Should I try Celery instead?
thanks

There are actually two separate problems here, which are intertwined:
You cannot pass a multiprocessing.Queue() object as an argument to a Pool-based function (you can pass it to the worker you start directly, but not any "further in" as it were).
You must wait for all the asynchronous workers to complete before you send the None through to your listener process.
To fix the first one, replace:
queue = multiprocessing.Queue(-1)
with:
queue = multiprocessing.Manager().Queue(-1)
as a manager-managed Queue() instance can be passed through.
To fix the second, either collect each result from each asynchronous call, or close the pool and wait for it, e.g.:
pool.close()
pool.join()
queue.put_nowait(None)
or the more complex:
getters = []
for i, sleep_time in enumerate(job_list):
name = str(i)
getters.append(
pool.apply_async(worker_function,
args=(sleep_time, name, queue, worker_configurer))
)
while len(getters):
getters.pop().get()
# optionally, close and join pool here (generally a good idea anyway)
queue.put_nowait(None)
(You should also consider replacing your put_nowait with a waiting version of put and not using unlimited length queues.)

Consider using two queues. The first queue is where you put the data for the workers. Each worker after job completion pushes the results to the second queue. Now consume this second queue to write the log to the file.

[ADDENDUM] Regarding maxtasksperchild=1
you don't really need it. The reason for repeated messages were due to:
you were repeatedly adding queuehandlers to the root logger of a child process. The following code checks if any handlers exist before adding another:
def worker_configurer(queue):
root = logging.getLogger()
# print(f'{root.handlers=}')
if len(root.handlers) == 0:
h = logging.handlers.QueueHandler(queue)
root.addHandler(h)
root.setLevel(logging.DEBUG)

asyncio run multiple tasks while preserving order and handling errors

my function run_tasks(all_tasks, window_size) that takes a generator of asyncio tasks and returns their value while:
run each window (of size window_size) from the all_tasks concurrently
preserve the order of returned results (all_tasks[i] result is results[i])
handle exceptions for each run
My current implementation:
import asyncio
from itertools import islice
# run all tasks and return their results in the same order
# window is the max number of tasks that will run in parallel
def run_tasks(all_tasks, window_size=4):
loop = asyncio.get_event_loop()
while True:
window_tasks = list(islice(all_tasks, window_size))
if not window_tasks:
break
futures = asyncio.wait(window_tasks, loop=loop)
finished, unfinished = loop.run_until_complete(futures)
# sort finished tasks by their launch order.
# removing this line makes returned tasks unordered
finished = sorted(finished, key=lambda f: window_tasks.index(f._coro))
for finished_task in finished:
try:
yield finished_task.result()
except Exception as e:
yield repr(e)
# Example Usage:
# a coroutine that sometime raises exception
async def sleepy(i):
print(f'{i} started')
await asyncio.sleep(10 - i)
print(f'{i} finished')
if i == 5:
raise ValueError('5 is the worst')
return i
# a generator of tasks
all_tasks = (sleepy(i) for i in range(10))
for result in list(run_tasks(all_tasks)):
print(result)
The Problem
The problem with my implementation is that I cannot find a way to sort the tasks in without accessing f._coro which is internal property of asyncio.Task object.
# removing this line makes returned tasks unordered
finished = sorted(finished, key=lambda f: window_tasks.index(f._coro))
I can use asyncio.gather(*tasks) but this will not handle errors.
I am open to suggestions on how to implement this three properties for run_tasks() without access to f._coro.

asyncio.gather can return you errors if you specify it's keyword param return_exceptions. To distinguish real exceptions from exception objects return as result of coroutine, you can wrap your window_tasks with tasks using ensure_future:
futures = [asyncio.ensure_future(t, loop=loop) for t in window_tasks]
gathered = asyncio.gather(*futures, loop=loop, return_exceptions=True)
loop.run_until_complete(gathered)
for fut in futures:
try:
yield fut.result()
except Exception as e:
yield repr(e)

process stop working while queue is not empty

I try to write a script in python to convert url into its corresponding ip. Since the url file is huge (nearly 10GB), so I'm trying to use multiprocessing lib.
I create one process to write output to file and a set of processes to convert url.
Here is my code:
import multiprocessing as mp
import socket
import time
num_processes = mp.cpu_count()
sentinel = None
def url2ip(inqueue, output):
v_url = inqueue.get()
print 'v_url '+v_url
try:
v_ip = socket.gethostbyname(v_url)
output_string = v_url+'|||'+v_ip+'\n'
except:
output_string = v_url+'|||-1'+'\n'
print 'output_string '+output_string
output.put(output_string)
print output.full()
def handle_output(output):
f_ip = open("outputfile", "a")
while True:
output_v = output.get()
if output_v:
print 'output_v '+output_v
f_ip.write(output_v)
else:
break
f_ip.close()
if __name__ == '__main__':
output = mp.Queue()
inqueue = mp.Queue()
jobs = []
proc = mp.Process(target=handle_output, args=(output, ))
proc.start()
print 'run in %d processes' % num_processes
for i in range(num_processes):
p = mp.Process(target=url2ip, args=(inqueue, output))
jobs.append(p)
p.start()
for line in open('inputfile','r'):
print 'ori '+line.strip()
inqueue.put(line.strip())
for i in range(num_processes):
# Send the sentinal to tell Simulation to end
inqueue.put(sentinel)
for p in jobs:
p.join()
output.put(None)
proc.join()
However, it did not work. It did produce several outputs (4 out of 10 urls in the test file) but it just suddenly stops while queues are not empty (I did check queue.empty())
Could anyone suggest what's wrong?Thanks

You're workers exit after processing a single url each, they need to loop internally until they get the sentinel. However, you should probably just look at multiprocessing.pool instead, as that does the bookkeeping for you.