This document describes the current stable version of Celery (5.2). For development docs, go here.

Frequently Asked Questions¶

General ¶

What kinds of things should I use Celery for?¶

Answer: Queue everything and delight everyone is a good article describing why you’d use a queue in a web context.

These are some common use cases:

Running something in the background. For example, to finish the web request as soon as possible, then update the users page incrementally. This gives the user the impression of good performance and “snappiness”, even though the real work might actually take some time.
Running something after the web request has finished.
Making sure something is done, by executing it asynchronously and using retries.
Scheduling periodic work.

And to some degree:

Distributed computing.
Parallel execution.

Misconceptions ¶

Does Celery really consist of 50.000 lines of code?¶

Answer: No, this and similarly large numbers have been reported at various locations.

The numbers as of this writing are:

core: 7,141 lines of code.

tests: 14,209 lines.

backends, contrib, compat utilities: 9,032 lines.

Lines of code isn’t a useful metric, so even if Celery did consist of 50k lines of code you wouldn’t be able to draw any conclusions from such a number.

Does Celery have many dependencies?¶

A common criticism is that Celery uses too many dependencies. The rationale behind such a fear is hard to imagine, especially considering code reuse as the established way to combat complexity in modern software development, and that the cost of adding dependencies is very low now that package managers like pip and PyPI makes the hassle of installing and maintaining dependencies a thing of the past.

Celery has replaced several dependencies along the way, and the current list of dependencies are:

celery ¶

kombu

Kombu is part of the Celery ecosystem and is the library used to send and receive messages. It’s also the library that enables us to support many different message brokers. It’s also used by the OpenStack project, and many others, validating the choice to separate it from the Celery code-base.

billiard

Billiard is a fork of the Python multiprocessing module containing many performance and stability improvements. It’s an eventual goal that these improvements will be merged back into Python one day.

It’s also used for compatibility with older Python versions that don’t come with the multiprocessing module.

pytz

The pytz module provides timezone definitions and related tools.

kombu ¶

Kombu depends on the following packages:

amqp

The underlying pure-Python amqp client implementation. AMQP being the default broker this is a natural dependency.

Note

To handle the dependencies for popular configuration choices Celery defines a number of “bundle” packages, see Bundles.

Is Celery heavy-weight?¶

Celery poses very little overhead both in memory footprint and performance.

But please note that the default configuration isn’t optimized for time nor space, see the Optimizing guide for more information.

Is Celery dependent on pickle?¶

Answer: No, Celery can support any serialization scheme.

We have built-in support for JSON, YAML, Pickle, and msgpack. Every task is associated with a content type, so you can even send one task using pickle, another using JSON.

The default serialization support used to be pickle, but since 4.0 the default is now JSON. If you require sending complex Python objects as task arguments, you can use pickle as the serialization format, but see notes in Serializers.

If you need to communicate with other languages you should use a serialization format suited to that task, which pretty much means any serializer that’s not pickle.

You can set a global default serializer, the default serializer for a particular Task, or even what serializer to use when sending a single task instance.

Is Celery for Django only?¶

Answer: No, you can use Celery with any framework, web or otherwise.

Do I have to use AMQP/RabbitMQ?¶

Answer: No, although using RabbitMQ is recommended you can also use Redis, SQS, or Qpid.

See Backends and Brokers for more information.

Redis as a broker won’t perform as well as an AMQP broker, but the combination RabbitMQ as broker and Redis as a result store is commonly used. If you have strict reliability requirements you’re encouraged to use RabbitMQ or another AMQP broker. Some transports also use polling, so they’re likely to consume more resources. However, if you for some reason aren’t able to use AMQP, feel free to use these alternatives. They will probably work fine for most use cases, and note that the above points are not specific to Celery; If using Redis/database as a queue worked fine for you before, it probably will now. You can always upgrade later if you need to.

Is Celery multilingual?¶

Answer: Yes.

worker is an implementation of Celery in Python. If the language has an AMQP client, there shouldn’t be much work to create a worker in your language. A Celery worker is just a program connecting to the broker to process messages.

Also, there’s another way to be language-independent, and that’s to use REST tasks, instead of your tasks being functions, they’re URLs. With this information you can even create simple web servers that enable preloading of code. Simply expose an endpoint that performs an operation, and create a task that just performs an HTTP request to that endpoint.

You can also use Flower’s REST API to invoke tasks.

Troubleshooting ¶

MySQL is throwing deadlock errors, what can I do?¶

Answer: MySQL has default isolation level set to REPEATABLE-READ, if you don’t really need that, set it to READ-COMMITTED. You can do that by adding the following to your my.cnf:

[mysqld]
transaction-isolation = READ-COMMITTED

For more information about InnoDB`s transaction model see MySQL - The InnoDB Transaction Model and Locking in the MySQL user manual.

(Thanks to Honza Kral and Anton Tsigularov for this solution)

The worker isn’t doing anything, just hanging ¶

Answer: See MySQL is throwing deadlock errors, what can I do?, or Why is Task.delay/apply*/the worker just hanging?.

Task results aren’t reliably returning ¶

Answer: If you’re using the database backend for results, and in particular using MySQL, see MySQL is throwing deadlock errors, what can I do?.

Why is Task.delay/apply*/the worker just hanging?¶

Answer: There’s a bug in some AMQP clients that’ll make it hang if it’s not able to authenticate the current user, the password doesn’t match or the user doesn’t have access to the virtual host specified. Be sure to check your broker logs (for RabbitMQ that’s /var/log/rabbitmq/rabbit.log on most systems), it usually contains a message describing the reason.

Does it work on FreeBSD?¶

Answer: Depends;

When using the RabbitMQ (AMQP) and Redis transports it should work out of the box.

For other transports the compatibility prefork pool is used and requires a working POSIX semaphore implementation, this is enabled in FreeBSD by default since FreeBSD 8.x. For older version of FreeBSD, you have to enable POSIX semaphores in the kernel and manually recompile billiard.

Luckily, Viktor Petersson has written a tutorial to get you started with Celery on FreeBSD here: http://www.playingwithwire.com/2009/10/how-to-get-celeryd-to-work-on-freebsd/

I’m having IntegrityError: Duplicate Key errors. Why?¶

Answer: See MySQL is throwing deadlock errors, what can I do?. Thanks to @@howsthedotcom.

Why aren’t my tasks processed?¶

Answer: With RabbitMQ you can see how many consumers are currently receiving tasks by running the following command:

$ rabbitmqctl list_queues -p <myvhost> name messages consumers
Listing queues ...
celery     2891    2

This shows that there’s 2891 messages waiting to be processed in the task queue, and there are two consumers processing them.

One reason that the queue is never emptied could be that you have a stale worker process taking the messages hostage. This could happen if the worker wasn’t properly shut down.

When a message is received by a worker the broker waits for it to be acknowledged before marking the message as processed. The broker won’t re-send that message to another consumer until the consumer is shut down properly.

If you hit this problem you have to kill all workers manually and restart them:

$ pkill 'celery worker'

$ # - If you don't have pkill use:
$ # ps auxww | awk '/celery worker/ {print $2}' | xargs kill

You may have to wait a while until all workers have finished executing tasks. If it’s still hanging after a long time you can kill them by force with:

$ pkill -9 'celery worker'

$ # - If you don't have pkill use:
$ # ps auxww | awk '/celery worker/ {print $2}' | xargs kill -9

Why won’t my Task run?¶

Answer: There might be syntax errors preventing the tasks module being imported.

You can find out if Celery is able to run the task by executing the task manually:

>>> from myapp.tasks import MyPeriodicTask
>>> MyPeriodicTask.delay()

Watch the workers log file to see if it’s able to find the task, or if some other error is happening.

Why won’t my periodic task run?¶

Answer: See Why won’t my Task run?.

How do I purge all waiting tasks?¶

Answer: You can use the celery purge command to purge all configured task queues:

$ celery -A proj purge

or programmatically:

>>> from proj.celery import app
>>> app.control.purge()
1753

If you only want to purge messages from a specific queue you have to use the AMQP API or the celery amqp utility:

$ celery -A proj amqp queue.purge <queue name>

The number 1753 is the number of messages deleted.

You can also start the worker with the --purge option enabled to purge messages when the worker starts.

I’ve purged messages, but there are still messages left in the queue?¶

Answer: Tasks are acknowledged (removed from the queue) as soon as they’re actually executed. After the worker has received a task, it will take some time until it’s actually executed, especially if there are a lot of tasks already waiting for execution. Messages that aren’t acknowledged are held on to by the worker until it closes the connection to the broker (AMQP server). When that connection is closed (e.g., because the worker was stopped) the tasks will be re-sent by the broker to the next available worker (or the same worker when it has been restarted), so to properly purge the queue of waiting tasks you have to stop all the workers, and then purge the tasks using celery.control.purge().

Results ¶

How do I get the result of a task if I have the ID that points there?¶

Answer: Use task.AsyncResult:

>>> result = my_task.AsyncResult(task_id)
>>> result.get()

This will give you a AsyncResult instance using the tasks current result backend.

If you need to specify a custom result backend, or you want to use the current application’s default backend you can use app.AsyncResult:

>>> result = app.AsyncResult(task_id)
>>> result.get()

Security ¶

Isn’t using pickle a security concern?¶

Answer: Indeed, since Celery 4.0 the default serializer is now JSON to make sure people are choosing serializers consciously and aware of this concern.

It’s essential that you protect against unauthorized access to your broker, databases and other services transmitting pickled data.

Note that this isn’t just something you should be aware of with Celery, for example also Django uses pickle for its cache client.

For the task messages you can set the task_serializer setting to “json” or “yaml” instead of pickle.

Similarly for task results you can set result_serializer.

For more details of the formats used and the lookup order when checking what format to use for a task see Serializers

Can messages be encrypted?¶

Answer: Some AMQP brokers supports using SSL (including RabbitMQ). You can enable this using the broker_use_ssl setting.

It’s also possible to add additional encryption and security to messages, if you have a need for this then you should contact the Mailing list.

Is it safe to run celery worker as root?¶

Answer: No!

We’re not currently aware of any security issues, but it would be incredibly naive to assume that they don’t exist, so running the Celery services (celery worker, celery beat, celeryev, etc) as an unprivileged user is recommended.

Brokers ¶

Why is RabbitMQ crashing?¶

Answer: RabbitMQ will crash if it runs out of memory. This will be fixed in a future release of RabbitMQ. please refer to the RabbitMQ FAQ: https://www.rabbitmq.com/faq.html#node-runs-out-of-memory

Note

This is no longer the case, RabbitMQ versions 2.0 and above includes a new persister, that’s tolerant to out of memory errors. RabbitMQ 2.1 or higher is recommended for Celery.

If you’re still running an older version of RabbitMQ and experience crashes, then please upgrade!

Misconfiguration of Celery can eventually lead to a crash on older version of RabbitMQ. Even if it doesn’t crash, this can still consume a lot of resources, so it’s important that you’re aware of the common pitfalls.

Events.

Running worker with the -E option will send messages for events happening inside of the worker.

Events should only be enabled if you have an active monitor consuming them, or if you purge the event queue periodically.

AMQP backend results.

When running with the AMQP result backend, every task result will be sent as a message. If you don’t collect these results, they will build up and RabbitMQ will eventually run out of memory.

This result backend is now deprecated so you shouldn’t be using it. Use either the RPC backend for rpc-style calls, or a persistent backend if you need multi-consumer access to results.

Results expire after 1 day by default. It may be a good idea to lower this value by configuring the result_expires setting.

If you don’t use the results for a task, make sure you set the ignore_result option:

@app.task(ignore_result=True)
def mytask():
    pass

class MyTask(Task):
    ignore_result = True

Can I use Celery with ActiveMQ/STOMP?¶

Answer: No. It used to be supported by Carrot (our old messaging library) but isn’t currently supported in Kombu (our new messaging library).

What features aren’t supported when not using an AMQP broker?¶

This is an incomplete list of features not available when using the virtual transports:

Remote control commands (supported only by Redis).

Monitoring with events may not work in all virtual transports.

The header and fanout exchange types
(fanout is supported by Redis).

Tasks ¶

How can I reuse the same connection when calling tasks?¶

Answer: See the broker_pool_limit setting. The connection pool is enabled by default since version 2.5.

sudo in a `subprocess` returns `None`¶

There’s a sudo configuration option that makes it illegal for process without a tty to run sudo:

Defaults requiretty

If you have this configuration in your /etc/sudoers file then tasks won’t be able to call sudo when the worker is running as a daemon. If you want to enable that, then you need to remove the line from /etc/sudoers.

See: http://timelordz.com/wiki/Apache_Sudo_Commands

Why do workers delete tasks from the queue if they’re unable to process them?¶

Answer:

The worker rejects unknown tasks, messages with encoding errors and messages that don’t contain the proper fields (as per the task message protocol).

If it didn’t reject them they could be redelivered again and again, causing a loop.

Recent versions of RabbitMQ has the ability to configure a dead-letter queue for exchange, so that rejected messages is moved there.

Can I call a task by name?¶

Answer: Yes, use app.send_task().

You can also call a task by name, from any language, using an AMQP client:

>>> app.send_task('tasks.add', args=[2, 2], kwargs={})
<AsyncResult: 373550e8-b9a0-4666-bc61-ace01fa4f91d>

To use chain, chord or group with tasks called by name, use the Celery.signature() method:

>>> chain(
...     app.signature('tasks.add', args=[2, 2], kwargs={}),
...     app.signature('tasks.add', args=[1, 1], kwargs={})
... ).apply_async()
<AsyncResult: e9d52312-c161-46f0-9013-2713e6df812d>

Can I get the task id of the current task?¶

Answer: Yes, the current id and more is available in the task request:

@app.task(bind=True)
def mytask(self):
    cache.set(self.request.id, "Running")

For more information see Task Request.

If you don’t have a reference to the task instance you can use app.current_task:

>>> app.current_task.request.id

But note that this will be any task, be it one executed by the worker, or a task called directly by that task, or a task called eagerly.

To get the current task being worked on specifically, use current_worker_task:

>>> app.current_worker_task.request.id

Note

Both current_task, and current_worker_task can be None.

Can I specify a custom task_id?¶

Answer: Yes, use the task_id argument to Task.apply_async():

>>> task.apply_async(args, kwargs, task_id='…')

Can I use decorators with tasks?¶

Answer: Yes, but please see note in the sidebar at Basics.

Can I use natural task ids?¶

Answer: Yes, but make sure it’s unique, as the behavior for two tasks existing with the same id is undefined.

The world will probably not explode, but they can definitely overwrite each others results.

Can I run a task once another task has finished?¶

Answer: Yes, you can safely launch a task inside a task.

A common pattern is to add callbacks to tasks:

from celery.utils.log import get_task_logger

logger = get_task_logger(__name__)

@app.task
def add(x, y):
    return x + y

@app.task(ignore_result=True)
def log_result(result):
    logger.info("log_result got: %r", result)

Invocation:

>>> (add.s(2, 2) | log_result.s()).delay()

See Canvas: Designing Work-flows for more information.

Can I cancel the execution of a task?¶

Answer: Yes, Use result.revoke():

>>> result = add.apply_async(args=[2, 2], countdown=120)
>>> result.revoke()

or if you only have the task id:

>>> from proj.celery import app
>>> app.control.revoke(task_id)

The latter also support passing a list of task-ids as argument.

Why aren’t my remote control commands received by all workers?¶

Answer: To receive broadcast remote control commands, every worker node creates a unique queue name, based on the nodename of the worker.

If you have more than one worker with the same host name, the control commands will be received in round-robin between them.

To work around this you can explicitly set the nodename for every worker using the -n argument to worker:

$ celery -A proj worker -n worker1@%h
$ celery -A proj worker -n worker2@%h

where %h expands into the current hostname.

Can I send some tasks to only some servers?¶

Answer: Yes, you can route tasks to one or more workers, using different message routing topologies, and a worker instance can bind to multiple queues.

See Routing Tasks for more information.

Can I disable prefetching of tasks?¶

Answer: Maybe! The AMQP term “prefetch” is confusing, as it’s only used to describe the task prefetching limit. There’s no actual prefetching involved.

Disabling the prefetch limits is possible, but that means the worker will consume as many tasks as it can, as fast as possible.

A discussion on prefetch limits, and configuration settings for a worker that only reserves one task at a time is found here: Prefetch Limits.

Can I change the interval of a periodic task at runtime?¶

Answer: Yes, you can use the Django database scheduler, or you can create a new schedule subclass and override is_due():

from celery.schedules import schedule

class my_schedule(schedule):

    def is_due(self, last_run_at):
        return run_now, next_time_to_check

Does Celery support task priorities?¶

Answer: Yes, RabbitMQ supports priorities since version 3.5.0, and the Redis transport emulates priority support.

You can also prioritize work by routing high priority tasks to different workers. In the real world this usually works better than per message priorities. You can use this in combination with rate limiting, and per message priorities to achieve a responsive system.

Should I use retry or acks_late?¶

Answer: Depends. It’s not necessarily one or the other, you may want to use both.

Task.retry is used to retry tasks, notably for expected errors that is catch-able with the try block. The AMQP transaction isn’t used for these errors: if the task raises an exception it’s still acknowledged!

The acks_late setting would be used when you need the task to be executed again if the worker (for some reason) crashes mid-execution. It’s important to note that the worker isn’t known to crash, and if it does it’s usually an unrecoverable error that requires human intervention (bug in the worker, or task code).

In an ideal world you could safely retry any task that’s failed, but this is rarely the case. Imagine the following task:

@app.task
def process_upload(filename, tmpfile):
    # Increment a file count stored in a database
    increment_file_counter()
    add_file_metadata_to_db(filename, tmpfile)
    copy_file_to_destination(filename, tmpfile)

If this crashed in the middle of copying the file to its destination the world would contain incomplete state. This isn’t a critical scenario of course, but you can probably imagine something far more sinister. So for ease of programming we have less reliability; It’s a good default, users who require it and know what they are doing can still enable acks_late (and in the future hopefully use manual acknowledgment).

In addition Task.retry has features not available in AMQP transactions: delay between retries, max retries, etc.

So use retry for Python errors, and if your task is idempotent combine that with acks_late if that level of reliability is required.

Can I schedule tasks to execute at a specific time?¶

Answer: Yes. You can use the eta argument of Task.apply_async().

Can I safely shut down the worker?¶

Answer: Yes, use the TERM signal.

This will tell the worker to finish all currently executing jobs and shut down as soon as possible. No tasks should be lost even with experimental transports as long as the shutdown completes.

You should never stop worker with the KILL signal (kill -9), unless you’ve tried TERM a few times and waited a few minutes to let it get a chance to shut down.

Also make sure you kill the main worker process only, not any of its child processes. You can direct a kill signal to a specific child process if you know the process is currently executing a task the worker shutdown is depending on, but this also means that a WorkerLostError state will be set for the task so the task won’t run again.

Identifying the type of process is easier if you have installed the setproctitle module:

$ pip install setproctitle

With this library installed you’ll be able to see the type of process in ps listings, but the worker must be restarted for this to take effect.

But it may still work and we are happy to accept patches.

Frequently Asked Questions¶

sudo in a subprocess returns None¶

sudo in a `subprocess` returns `None`¶