We have a flask app that uses a lot of memory for ML models, and I'm trying to reduce the memory footprint by using gunicorn's preload option, but when I add the --preload flag, and deploy that (with -w 4, to a docker container running on GKE), it will handle just a few requests, and then hang until it times out, at which point gunicorn will start another worker to replace it and the same thing will happen. It's not clear yet how many requests each worker will process before hanging (possibly just 1... possibly a few)
The timeout is over 10 minutes, so it seems to be hanging indefinitely.
This does not happen at all if I remove the --preload flag.
What is it about the --preload flag that could be causing the workers to hang indefinitely?
Related
We are running a django web application via Gunicorn + NGNIX (on a linux server). We find that the jobs we are executing via APScheduler are being executed 3 times each (which is undesirable). We solved the problem on our local computers, however, the problem remains when being run via Gunicorn.
This was the most informative thread:
Make sure only one worker launches the apscheduler event in a pyramid web app running multiple workers
Our problem is that we want to execute the following code: env/bin/gunicorn module_containing_app:app -b 0.0.0.0:8080 --workers 3 --preload, we are unsure about two things:
We have a domain "example.com", should this be taken into consideration?
We are trying to find the env/bin/gunicorn folder/path with no luck. We are also trying to understand what to replace module_containing_app:app with because of this.
I operate a Django 1.9 website on Heroku, with Celery 3.1.23. RabbitMQ is used as a broker.
After restarting the beat worker, the memory usage is always around 497Mb. This results in frequent Error R14 (Memory quota exceeded) as it quickly reaches the 512Mb limit.
How can I analyze what is in the memory at startup? I.e. how can I get a detail of what is in the memory when restarted?
Here is a detail of memory consumption obtained with the beta Heroku log-runtime-metrics:
heroku/beat.1:
source=beat.1 dyno=heroku.52346831.1ea92181-ab6d-461c-90fa-61fa8fef2c18
sample#memory_total=497.66MB
sample#memory_rss=443.91MB
sample#memory_cache=20.43MB
sample#memory_swap=33.33MB
sample#memory_pgpgin=282965pages
sample#memory_pgpgout=164606pages
sample#memory_quota=512.00MB
I had the same problem. Searching around, I followed How many CPU cores has a heroku dyno? and Celery immediately exceeds memory on Heroku.
So I typed:
heroku run grep -c processor /proc/cpuinfo -a <app_name>
It returned 8. So I added --concurrency=4 to my Procfile line:
worker: celery -A <app> worker -l info -O fair --without-gossip --without-mingle --without-heartbeat --concurrency=4
And memory usage seemed to be divided by almost 2:
I have a Python/Django Project running on uwsgi/nginx. For asynchronous task we are using rabbitmq/celeryd and supervisord to manage all the daemons
Versions:
python: 2.7
django: 1.9.7
celery: 3.1.23
django-celery: 3.1.17
Celery has 10 queue of type Direct (say queue1, queue2, ...)
Each queue is handled by a separate celeryd process which is manage via supervisord. each supervisord process looks as following
[program:app_queue_worker]
command=/var/www/myproj/venv/bin/celery worker -A myproj -c 2 --queue=queue1 --loglevel=INFO
directory=/var/www/myproj/
user=ubuntu
numprocs=1
autostart=true
autorestart=true
startsecs=10
exitcodes=1
stopwaitsecs = 600
killasgroup=true
priority=1000
Hence Supervisord is running 10 Mainprocess and 20 Worker process
Other Thing I have noticed is uwsgi also spawns some celery workers(Dont understand how and why, YET ) with concurrency=2. So if I have 4 uwsgi process running i will have an addition 10 celery workers running
All these workers are each taking 200-300M memory? Something is wrong here I feel it but I am not able to put my finger on it. Celery shouldn't be running such memory heavy process?
Note: Debug=False, there is no memory leakage due to debug
Can someone please comment on the architecture if it is correct or wrong?
Would it be better to run 2-3 celery MainProcesses which listen all queues at once and increase its concurrency?
Update : celery.py Config
from celery import Celery
# set the default Django settings module for the 'celery' program.
os.environ.setdefault('DJANGO_SETTINGS_MODULE', 'MyProject.settings')
from django.conf import settings # noqa
from chatterbox import celery_settings
app = Celery('MyProject')
# Using a string here means the worker will not have to
# pickle the object when using Windows.
app.config_from_object('django.conf:settings')
app.conf.update(
CELERY_RESULT_BACKEND='djcelery.backends.database:DatabaseBackend',
CELERYD_CONCURRENCY=1,
)
app.autodiscover_tasks(lambda: settings.INSTALLED_APPS)
There is no simple answer to this.
To me, the fact that uwsgi spawns celery workers is wrong.
Creating only worker processes that consume all queues might lead to the situation where long running tasks make some queues overflow whereas separate workers that consume specific queues with short running tasks could make the situation better. Everything depends on your use case.
The 300mb residual memory is quite a lot. If the tasks are i/o bound go multi-thread/gevent. However, if the tasks are CPU bound, you have no other option than to scale process wise.
If you start a celery worker with concurrency of n, it will spawn n + 1 process by default. Since you are spawning 10 workers with a concurrency of 2, celery will start 30 processes.
Each worker consumes ~60MB(~30MB for main process & 2*~15MB for subprocesses) of memory when not consuming queues. It might vary depending what your worker is doing. If you start 10 workers, it will consume ~600MB of memory.
I am not sure how you came to know that uwsgi also spawns some celery workers. Only supervisor should spawn the process.
You can run just 1 celery worker which listens to all queues with a concurrency of 20. This will reduce your memory usage at the cost of flexibility. With this setup, you can't start/stop consuming from selected queues. Also, there is no guarantee that all queues will be consumed equally.
We use Celery with our Django webapp to manage offline tasks; some of these tasks can run up to 120 seconds.
Whenever we make any code modifications, we need to restart Celery to have it reload the new Python code. Our current solution is to send a SIGTERM to the main Celery process (kill -s 15 `cat /var/run/celeryd.pid`), then to wait for it to die and restart it (python manage.py celeryd --pidfile=/var/run/celeryd.pid [...]).
Because of the long-running tasks, this usually means the shutdown will take a minute or two, during which no new tasks are processed, causing a noticeable delay to users currently on the site. I'm looking for a way to tell Celery to shutdown, but then immediately launch a new Celery instance to start running new tasks.
Things that didn't work:
Sending SIGHUP to the main process: this caused Celery to attempt to "restart," by doing a warm shutdown and then relaunching itself. Not only does this take a long time, it doesn't even work, because apparently the new process launches before the old one dies, so the new one complains ERROR: Pidfile (/var/run/celeryd.pid) already exists. Seems we're already running? (PID: 13214) and dies immediately. (This looks like a bug in Celery itself; I've let them know about it.)
Sending SIGTERM to the main process and then immediately launching a new instance: same issue with the Pidfile.
Disabling the Pidfile entirely: without it, we have no way of telling which of the 30 Celery process are the main process that needs to be sent a SIGTERM when we want it to do a warm shutdown. We also have no reliable way to check if the main process is still alive.
celeryd has --autoreload option. If enabled, celery worker (main process) will detect changes in celery modules and restart all worker processes. In contrast to SIGHUP signal, autoreload restarts each process independently when the current executing task finishes. It means while one worker process is restarting the remaining processes can execute tasks.
http://celery.readthedocs.org/en/latest/userguide/workers.html#autoreloading
I've recently fixed the bug with SIGHUP: https://github.com/celery/celery/pull/662
rm *.pyc
This causes the updated tasks to be reloaded. I discovered this trick recently, I just hope there are no nasty side effects.
Well you using SIGHUP (1) for warm shutdown of celery. I am not sure if it actually causes a warm shutdown. But SIGINT (2) would cause a warm shutdown. Try SIGINT in place of SIGHUP and then start celery manually in your script (I guess).
Can you launch it with a custom pid file name. Possibly timestamped, and key off of that to know which PID to kill?
CELERYD_PID_FILE="/var/run/celery/%n_{timestamp}.pid"
^I dont know the timestamp syntax but maybe you do or you can find it?
then use the current system time to kill off any old pids and launch a new one?
A little late, but that can fixed by deleting the file called celerybeat.pid.
Worked for me.
I think you can try this:
kill -s HUP ``cat /var/run/celeryd.pid``
python manage.py celeryd --pidfile=/var/run/celeryd.pid
HUP may recycle every free worker and leave executing workers keep running and HUP will let these workers be trusted. Then you can safely restart a new celery worker main process and workers. Old workers may be killed itself when task has been finished.
I've use this way in our production and it seems safe now. Hope this can help you!
We're having issues with our celery daemon being very flaky. We use a fabric deployment script to restart the daemon whenever we push changes, but for some reason this is causing massive issues.
Whenever the deployment script is run the celery processes are left in some pseudo dead state. They will (unfortunately) still consume tasks from rabbitmq, but they won't actually do anything. Confusingly a brief inspection would indicate everything seems to be "fine" in this state, celeryctl status shows one node online and ps aux | grep celery shows 2 running processes.
However, attempting to run /etc/init.d/celeryd stop manually results in the following error:
start-stop-daemon: warning: failed to kill 30360: No such process
While in this state attempting to run celeryd start appears to work correctly, but in fact does nothing. The only way to fix the issue is to manually kill the running celery processes and then start them again.
Any ideas what's going on here? We also don't have complete confirmation, but we think the problem also develops after a few days (with no activity this is a test server currently) on it's own with no deployment.
I can't say that I know what's ailing your setup, but I've always used supervisord to run celery -- maybe the issue has to do with upstart? Regardless, I've never experienced this with celery running on top of supervisord.
For good measure, here's a sample supervisor config for celery:
[program:celeryd]
directory=/path/to/project/
command=/path/to/project/venv/bin/python manage.py celeryd -l INFO
user=nobody
autostart=true
autorestart=true
startsecs=10
numprocs=1
stdout_logfile=/var/log/sites/foo/celeryd_stdout.log
stderr_logfile=/var/log/sites/foo/celeryd_stderr.log
; Need to wait for currently executing tasks to finish at shutdown.
; Increase this if you have very long running tasks.
stopwaitsecs = 600
Restarting celeryd in my fab script is then as simple as issuing a sudo supervisorctl restart celeryd.