Looppool 0.4.0

Tornado IO loop process pool with message passing

Looppool

Looppool is a Python 3 package for running worker process pool of Tornado IO loops. It’s useful for a heavy asynchronous application which doesn’t fit into single process due to increasing CPU usage and suffers from IO loop blocking (see set_blocking_log_threshold [1]).

It was developed as a part of performance optimisation of a Tornado data extraction application. The application mixed IO-bound and CPU-bound tasks. Moreover, the CPU-bound tasks were highly coupled with IO loop. Because of such coupling much simpler approach of concurrent.futures [3] wouldn’t have helped.

Design

A picture worth a thousand words.

A few observations and notes:

1. Messages are off-loaded to IO loops immediately but task and result queues are protected by semaphores to not off-load more tasks on IO loops than the queue size.

2. In fact there are n task queue, a queue per worker. Task messages are distributed evenly between workers.

3. add_callback [2] is safe (and only safe) method to pass control from other thread to IO loop’s thread.

4. Because queue message handlers (fn1 and fn2) are called from IO loop they can be coroutines.

5. Pool stops its workers by sending PoisonPill task message per worker.

Worker

There is a base class for a Worker. It represents a process worker that runs its own Tornado IO loop. It processes task messages and puts them in result queue (or just puts without receiving anything). It requires override of _process_message(self, task) in a subclass and mandates that once the task is done, self._task_done() is called (directly in try-finally or with @task_done), like:

def _process_message(self, task):
    try:
        result = 'some processing'
        self._put_nowait(result)
    finally:
        self._task_done()

WorkerSubclass._process_message may be plain function or coroutine. More details are available in the package’s unit test module [5].

Note

For Pool.process_message there is also requirement to call result_done argument once message is done. You can also wrap it with @task_done in simple case when result is considered ingested on return of process_message.

Both requirements are bound to semaphores that limit running tasks and pending results. It also affects how pool is stopped as it waits for running tasks to complete.

Stateful worker

If you want to run stateful workers, for instance, use some periodically calculated lookup table, but don’t want to calculate it in every workers (e.g. burden of maintaining database connection), you can send n tasks and each of n workers is guaranteed to receive it.

You can also send messages to workers individually. Pool.put_nowait has optional argument worker_num.

Note

If your process start method [4] is fork (default on *nix platforms), you can share some static data from parent process.

Installation

pip install Looppool

Usage

Note

Note that the following example has 1-to-1 correspondence of input to output messages. Because worker has IO loop at its disposal, it can, for example, subscribe to something by input message and put results later without task message.

#!/usr/bin/env python3


import looppool
from tornado import gen, ioloop, httpclient


class FetchWorker(looppool.Worker):

    _http_client = None
    '''Tornado asynchronous HTTP client'''


    def _initialise(self):
        self._http_client = httpclient.AsyncHTTPClient()

    @looppool.task_done
    @gen.coroutine
    def _process_message(self, url):
        response = yield self._http_client.fetch(url)
        self._put_nowait((url, response.headers.get('server')))


@looppool.task_done
def process_message(result, result_done):
    print(result)


@gen.coroutine
def main():
    loop = ioloop.IOLoop.instance()
    pool = looppool.Pool(loop, pool_size = 4, worker_class = FetchWorker)
    pool.process_message = process_message
    pool.start()

    urls = [
        'https://python.org/',
        'http://tornadoweb.org/',
        'https://google.com/',
        'https://stackoverflow.com/',
    ]
    list(map(pool.put_nowait, urls))

    pool.stop()


if __name__ == '__main__':
    ioloop.IOLoop.instance().run_sync(main)

Maintenance

Maintaining a process group instead of one process is more tricky thing to do. Initially, you may want to see if your pool instance has actually spawned any processes. Here’s what you can do visualise your process tree, which has main process, one or two (depending on start method) multiprocessing helper processes, and processes of your looppool pools (you can use different pools for different purposes):

htop -p $(pgrep -d"," -g $(pgrep -f "main-process-name-or-its-start-args"))

top will also work but is limited to 20 PIDs. Enforcing the process tree stop is also different. If something goes wrong process tree should be killed like:

kill -9 -- -$(pgrep -f "main-process-name-or-its-start-args")

Killing only main process will leave helper and worker processes running.

Note

Pool workers intentionally ignore SIGINT and SIGTERM because these signals propagate to children from parent process and break normal, message-based shutdown.

You can improve the names of your worker processes by setting them in worker’s initialiser with setproctitle [6] (see example below).

Monitoring

Generally it’s very important to know how well your application behaves. Even more important it is for single-threaded (asynchronous) and multi-process applications. For the former is critical to know that the process doesn’t use 100% CPU except for rare peaks, which would otherwise impair IO loop’s ability to schedule tasks. For the latter CPU usage shows how well current number of workers handle the load. This is being said about application metrics.

looppool comes with build-in loopppol.utility.ResourceReporter which periodically (10 seconds by default) sends metrics as CPU usage, memory usage (RSS) and length of IO loop backlogs (ioloop._handlers, ioloop._callbacks and ioloop._timeouts) to statsd-compatible server [8].

Logging

Multi-process logging is complicated. Most important part of logging is error reporting. Sentry [7] goes a great solution to error reporting problem. It seamlessly integrates with logging and is suggested tool to know what errors occur in your workers.

Instrumentated worker

For the following code you will need to run pip install raven statsd setproctitle.

#!/usr/bin/env python3


import logging

import looppool
from looppool.utility import ResourceReporter
from statsd import StatsClient
from raven import Client
from raven.exceptions import InvalidDsn
from raven.handlers.logging import SentryHandler
from setproctitle import setproctitle


class InstrumentatedWorker(looppool.Worker):

    _resource_reporter = None
    '''CPU, RSS and IO loop stats reporter'''


    def _initialise(self):
        setproctitle('python APP_NAME POOL_NAME pool worker')

        statsd = StatsClient('localhost', 8125, 'APP_PREFIX')
        self._resource_reporter = ResourceReporter(self._ioloop, statsd,
            'worker.instrumentated.process.{}'.format(self._number))

        try:
            handler = SentryHandler(Client('SENTRY_DSN'))
            handler.setLevel(logging.WARNING)
        except InvalidDsn:
            logging.exception('Cannot configure Sentry handler')
        else:
            logging.basicConfig(handlers=[handler], level=logging.WARNING)

    @looppool.task_done
    def _process_message(self, message):
        self._put_nowait((message, self._number))

    def start(self):
        super().start()

        self._resource_reporter.start()

    def join(self):
        self._resource_reporter.stop()

        super().join()

[1]

http://www.tornadoweb.org/en/stable/ioloop.html#tornado.ioloop.IOLoop.set_blocking_log_threshold

[2]

http://www.tornadoweb.org/en/stable/ioloop.html#tornado.ioloop.IOLoop.add_callback

[3]

https://docs.python.org/3/library/concurrent.futures.html

[4]

https://docs.python.org/3/library/multiprocessing.html#multiprocessing.set_start_method

[5]

https://bitbucket.org/saaj/looppool/src/default/looppool/test.py

[6]

https://pypi.python.org/pypi/setproctitle

[7]

https://pypi.python.org/pypi/sentry

[8]

https://github.com/etsy/statsd/wiki#server-implementations