token_bucket_refill tasks starts to grow in exponents about django-celery-token-bucket

[LennyLip]

Is the API that you're trying to throttle really giving one request per second?

Thanks for the response. Yes, I need to access to API, that limit is one request per second. I reset tasks queue (it was 39247 refill tasks in queue) 5 hours ago, and it works correctly for now. Can't determine when it starts to grow for now, will trying to investigate.

[lociii]

Thanks for the report. We've never tried running the refill every second. I can imagine that this will run into some race conditions.
We're running the following schedule without any issues

CELERY_TOKEN_BUCKETS: Dict[str, TokenBucket] = {
    "xxx": TokenBucket(
        name="xxx",
        schedule=schedules.crontab(minute=0),
        amount=3000,
        maximum=3000,
    )
}

Is the API that you're trying to throttle really giving one request per second?

[LennyLip]

I set the refill task (with the same config) to run once per 2 seconds. Nothing changed, it works correctly 5-8 hours and then starts to grow in exponents.

/data # redis-cli llen celery
(integer) 19067

[LennyLip]

Ok, I'm ending up with separate queue for refill task and looks like it working ok now.
One moment will be nice, if we can use "exponential backoff" for countdown argument, or at least a random jitter countdown.
Thank you for your work!

[LennyLip]

race conditions.

Yes, you are right. Task execution time sometimes exceed 1 second, so second task starts before stop previous one. So, it's starts to grow in exponents.

Looks like, I need external script (management command may be) to refill queue.

[LennyLip]

BTW, I need to purge queue before refill it. That's why I have a ton of messages, really. Any reason you didn't do it?

[LennyLip]

Random countdown. Useful feature as for me.

[lociii]

BTW, I need to purge queue before refill it. That's why I have a ton of messages, really. Any reason you didn't do it?

This seems pretty weird. The kombu queue is set to only hold a certain maximum of messages.
See https://github.com/RegioHelden/django-celery-token-bucket/blob/main/django_celery_token_bucket/bucket.py#L34

Random countdown. Useful feature as for me.

I'd implement that a bit differently but could make sense sometimes.