"python3 flask asyncio subprocess in route hangs" Answer’s


There are several concerns here:

  • You are creating a new event loop on import, once, but close the event loop in your view. There is no need to close the loop, at all, because a second request will now fail because the loop is closed.

  • The asyncio event loop is not thread safe, and should not be shared between threads. The vast majority of Flask deployments will use threads to handle incoming requests. Your code carries echoes of how this should be handled instead but unfortunately it is not the correct approach. E.g. asyncio.get_child_watcher().attach_loop(eventLoop) is mostly redundant because eventLoop = asyncio.new_event_loop(), if run on the main thread, already does exactly that.

    This is the main candidate for the issues you are seeing.

  • Your code assumes that the executable is in fact present and executable. You should be handling OSError exceptions (and subclasses), because an unqualified s.py would only work if it is made executable, starts with a #! shebang line and is found on the PATH. It won't work just because it is in the same directory, nor would you want to rely on the current working directory anyway.

  • Your code assumes that the process closes stdout at some point. If the subprocess never closes stdout (something that happens automatically when the process exits) then your async for line in process.stdout: loop will wait forever too. Consider adding timeouts to the code to avoid getting blocked on a faulty subprocess.

There are two sections in the Python asyncio documentation that you really would want to read when using asyncio subprocesses in a multi-threaded application:

  • The Concurrency and Multithreading section, explaining that Almost all asyncio objects are not thread safe. You don't want to add tasks to the loop from other threads directly; you want to either use an event loop per thread, or use the asyncio.run_coroutine_threadsafe() function to run a coroutine on a loop in a specific thread.

  • For Python versions up to 3.7, you also need to read the Subprocess and Threads section, because up until that version asyncio uses a non-blocking os.waitpid(-1, os.WNOHANG) call to track child state and relies on using signal handling (which can only be done on the main thread). Python 3.8 removed this restriction (by adding a new child watcher implementation that uses a blocking per-process os.waitpid() call in a separate thread, at the expense of extra memory.

    You don't have to stick to the default child watcher strategy, however. You can use EventLoopPolicy.set_child_watcher() and passing in a different process watcher instance. In practice that means backporting the 3.8 ThreadedChildWatcher implementation.

For your use case, there really no need to need to run a new event loop per thread. Run a single loop, in a separate thread as needed. If you use a loop in a separate thread, depending on your Python version, you may need to have a running loop on the main thread as well or use a different process watcher. Generally speaking, running an asyncio loop on the main thread in a WSGI server is not going to be easy or even possible.

So you need to run a loop, permanently, in a separate thread, and you need to use a child process watcher that works without a main thread loop. Here is an implementation for just that, and this should work for Python versions 3.6 and newer:

import asyncio
import itertools
import logging
import time
import threading

    # Python 3.8 or newer has a suitable process watcher
except AttributeError:
    # backport the Python 3.8 threaded child watcher
    import os
    import warnings

    # Python 3.7 preferred API
    _get_running_loop = getattr(asyncio, "get_running_loop", asyncio.get_event_loop)

    class _Py38ThreadedChildWatcher(asyncio.AbstractChildWatcher):
        def __init__(self):
            self._pid_counter = itertools.count(0)
            self._threads = {}

        def is_active(self):
            return True

        def close(self):

        def __enter__(self):
            return self

        def __exit__(self, exc_type, exc_val, exc_tb):

        def __del__(self, _warn=warnings.warn):
            threads = [t for t in list(self._threads.values()) if t.is_alive()]
            if threads:
                    f"{self.__class__} has registered but not finished child processes",

        def add_child_handler(self, pid, callback, *args):
            loop = _get_running_loop()
            thread = threading.Thread(
                args=(loop, pid, callback, args),
            self._threads[pid] = thread

        def remove_child_handler(self, pid):
            # asyncio never calls remove_child_handler() !!!
            # The method is no-op but is implemented because
            # abstract base class requires it
            return True

        def attach_loop(self, loop):

        def _do_waitpid(self, loop, expected_pid, callback, args):
            assert expected_pid > 0

                pid, status = os.waitpid(expected_pid, 0)
            except ChildProcessError:
                # The child process is already reaped
                # (may happen if waitpid() is called elsewhere).
                pid = expected_pid
                returncode = 255
                    "Unknown child process pid %d, will report returncode 255", pid
                if os.WIFSIGNALED(status):
                    returncode = -os.WTERMSIG(status)
                elif os.WIFEXITED(status):
                    returncode = os.WEXITSTATUS(status)
                    returncode = status

                if loop.get_debug():
                        "process %s exited with returncode %s", expected_pid, returncode

            if loop.is_closed():
                logger.warning("Loop %r that handles pid %r is closed", loop, pid)
                loop.call_soon_threadsafe(callback, pid, returncode, *args)


    # add the watcher to the loop policy

__all__ = ["EventLoopThread", "get_event_loop", "stop_event_loop", "run_coroutine"]

logger = logging.getLogger(__name__)

class EventLoopThread(threading.Thread):
    loop = None
    _count = itertools.count(0)

    def __init__(self):
        name = f"{type(self).__name__}-{next(self._count)}"
        super().__init__(name=name, daemon=True)

    def __repr__(self):
        loop, r, c, d = self.loop, False, True, False
        if loop is not None:
            r, c, d = loop.is_running(), loop.is_closed(), loop.get_debug()
        return (
            f"<{type(self).__name__} {self.name} id={self.ident} "
            f"running={r} closed={c} debug={d}>"

    def run(self):
        self.loop = loop = asyncio.new_event_loop()

                shutdown_asyncgens = loop.shutdown_asyncgens()
            except AttributeError:

    def stop(self):
        loop, self.loop = self.loop, None
        if loop is None:

_lock = threading.Lock()
_loop_thread = None

def get_event_loop():
    global _loop_thread
    if _loop_thread is None:
        with _lock:
            if _loop_thread is None:
                _loop_thread = EventLoopThread()
                # give the thread up to a second to produce a loop
                deadline = time.time() + 1
                while not _loop_thread.loop and time.time() < deadline:

    return _loop_thread.loop

def stop_event_loop():
    global _loop_thread
    with _lock:
        if _loop_thread is not None:
            _loop_thread = None

def run_coroutine(coro):
    return asyncio.run_coroutine_threadsafe(coro, get_event_loop())

The above is the same general 'run async with Flask' solution as I posted for Make a Python asyncio call from a Flask route, but with the addition of the ThreadedChildWatcher backport.

You can then use the loop returned from get_event_loop() to run child processes, by calling run_coroutine_threadsafe():

import asyncio
import locale
import logging

logger = logging.getLogger(__name__)

def get_command_output(cmd, timeout=None):
    encoding = locale.getpreferredencoding(False)

    async def run_async():
            process = await asyncio.create_subprocess_exec(
                cmd, stdout=asyncio.subprocess.PIPE)
        except OSError:
            logging.exception("Process %s could not be started", cmd)
        async for line in process.stdout:
            line = line.decode(encoding)
            # TODO: actually do something with the data.
            print(line, flush=True)

        logging.debug("Process for %s exiting with %i", cmd, process.returncode)

        return await process.wait()

    future = run_coroutine(run_async())
    result = None
        result = future.result(timeout)
    except asyncio.TimeoutError:
        logger.warn('The child process took too long, cancelling the task...')
    except Exception as exc:
        logger.exception(f'The child process raised an exception')
    return result

Note that the above function can take a timeout, in seconds, the maximum amount of time you'll wait for the subprocess to complete.

Friday, August 27, 2021
answered 5 Months ago
Only authorized users can answer the question. Please sign in first, or register a free account.
Not the answer you're looking for? Browse other questions tagged :