Celery Heimdall is a set of common utilities useful for the Celery background worker framework, built on top of Redis. It's not trying to handle every use case, but to be an easy, modern, and maintainable drop-in solution for 90% of projects.
- Globally unique tasks, allowing only 1 copy of a task to execute at a time, or within a time period (ex: "Don't allow queuing until an hour has passed")
- Global rate limiting. Celery has built-in rate limiting, but it's a rate limit per worker, making it unsuitable for purposes such as limiting requests to an API.
pip install celery-heimdall
Imagine you have a task that starts when a user presses a button. This task takes a long time and a lot of resources to generate a report. You don't want the user to press the button 10 times and start 10 tasks. In this case, you want what Heimdall calls a unique task:
from celery import shared_task
from celery_heimdall import HeimdallTask
@shared_task(base=HeimdallTask, heimdall={'unique': True})
def generate_report(customer_id):
pass
All we've done here is change the base Task class that Celery will use to run the task, and passed in some options for Heimdall to use. This task is now unique - for the given arguments, only 1 will ever run at the same time.
What happens if our task dies, or something goes wrong? We might end up in a situation where our lock never gets cleared, called deadlock. To work around this, we add a maximum time before the task is allowed to be queued again:
from celery import shared_task
from celery_heimdall import HeimdallTask
@shared_task(
base=HeimdallTask,
heimdall={
'unique': True,
'unique_timeout': 60 * 60
}
)
def generate_report(customer_id):
pass
Now, generate_report
will be allowed to run again in an hour even if the
task got stuck, the worker ran out of memory, the machine burst into flames,
etc...
By default, a hash of the task name and its arguments is used as the lock key. But this often might not be what you want. What if you only want one report at a time, even for different customers? Ex:
from celery import shared_task
from celery_heimdall import HeimdallTask
@shared_task(
base=HeimdallTask,
heimdall={
'unique': True,
'key': lambda args, kwargs: 'generate_report'
}
)
def generate_report(customer_id):
pass
By specifying our own key function, we can completely customize how we determine if a task is unique.
By default, if you try to queue up a unique task that is already running,
Heimdall will return the existing task's AsyncResult
. This lets you write
simple code that doesn't need to care if a task is unique or not. Imagine a
simple API endpoint that starts a report when it's hit, but we only want it
to run one at a time. The below is all you need:
import time
from celery import shared_task
from celery_heimdall import HeimdallTask
@shared_task(base=HeimdallTask, heimdall={'unique': True})
def generate_report(customer_id):
time.sleep(10)
def my_api_call(customer_id: int):
return {
'status': 'RUNNING',
'task_id': generate_report.delay(customer_id).id
}
Everytime my_api_call
is called with the same customer_id
, the same
task_id
will be returned by generate_report.delay()
until the original task
has completed.
Sometimes you'll want to catch that the task was already running when you tried to queue it again. We can tell Heimdall to raise an exception in this case:
import time
from celery import shared_task
from celery_heimdall import HeimdallTask, AlreadyQueuedError
@shared_task(
base=HeimdallTask,
heimdall={
'unique': True,
'unique_raises': True
}
)
def generate_report(customer_id):
time.sleep(10)
def my_api_call(customer_id: int):
try:
task = generate_report.delay(customer_id)
return {'status': 'STARTED', 'task_id': task.id}
except AlreadyQueuedError as exc:
return {'status': 'ALREADY_RUNNING', 'task_id': exc.likely_culprit}
By setting unique_raises
to True
when we define our task, an
AlreadyQueuedError
will be raised when you try to queue up a unique task
twice. The AlreadyQueuedError
has two properties:
likely_culprit
, which contains the task ID of the already-running task,expires_in
, which is the time remaining (in seconds) before the already-running task is considered to be expired.
What if we want the task to only run once in an hour, even if it's finished? In those cases, we want it to run, but not clear the lock when it's finished:
from celery import shared_task
from celery_heimdall import HeimdallTask
@shared_task(
base=HeimdallTask,
heimdall={
'unique': True,
'unique_timeout': 60 * 60,
'unique_wait_for_expiry': True
}
)
def generate_report(customer_id):
pass
By setting unique_wait_for_expiry
to True
, the task will finish, and won't
allow another generate_report()
to be queued until unique_timeout
has
passed.
Celery offers rate limiting out of the box. However, this rate limiting applies on a per-worker basis. There's no reliable way to rate limit a task across all your workers. Heimdall makes this easy:
from celery import shared_task
from celery_heimdall import HeimdallTask, RateLimit
@shared_task(
base=HeimdallTask,
heimdall={
'rate_limit': RateLimit((2, 60))
}
)
def download_report_from_amazon(customer_id):
pass
This says "every 60 seconds, only allow this task to run 2 times". If a task can't be run because it would violate the rate limit, it'll be rescheduled.
It's important to note this does not guarantee that your task will run exactly twice a second, just that it won't run more than twice a second. Tasks are rescheduled with a random jitter to prevent the thundering herd problem.
Just like you can dynamically provide a key for a task, you can also dynamically provide a rate limit based off that key.
from celery import shared_task
from celery_heimdall import HeimdallTask, RateLimit
@shared_task(
base=HeimdallTask,
heimdall={
# Provide a lower rate limit for the customer with the ID 10, for everyone
# else provide a higher rate limit.
'rate_limit': RateLimit(lambda args: (1, 30) if args[0] == 10 else (2, 30)),
'key': lambda args, kwargs: f'customer_{args[0]}'
}
)
def download_report_from_amazon(customer_id):
pass
These are more mature projects which inspired this library, and which may support older versions of Celery & Python then this project.
- celery_once, which is unfortunately abandoned and the reason this project exists.
- celery_singleton
- This snippet by Vigrond, and subsequent improvements by various contributors.