This document describes the current stable version of Celery (5.3). For development docs, go here.
Change history for Celery 2.2¶
2011-11-25 04:00 p.m. GMT
[Security: CELERYSA-0001] Daemons would set effective id’s rather than real id’s when the
--gidarguments to celery multi, celeryd_detach, celery beat and celery events were used.
This means privileges weren’t properly dropped, and that it would be possible to regain supervisor privileges later.
2011-06-13 04:00 p.m. BST
These signals can be used to augment logging configuration after Celery has set up logging.
Redis result backend now works with Redis 2.4.4.
--gidoption now works correctly.
worker: Retry wrongfully used the repr of the traceback instead of the string representation.
App.config_from_object: Now loads module, not attribute of module.
Fixed issue where logging of objects would give “<Unrepresentable: …>”
2011-04-15 04:00 p.m. CEST
Now depends on Kombu 1.1.2.
Dependency lists now explicitly specifies that we don’t want python-dateutil 2.x, as this version only supports Python 3.
If you have installed dateutil 2.0 by accident you should downgrade to the 1.5.0 version:
$ pip install -U python-dateutil==1.5.0
$ easy_install -U python-dateutil==1.5.0
WatchedFileHandlerbroke Python 2.5 support (Issue #367).
Task: Don’t use
app.mainif the task name is set explicitly.
Sending emails didn’t work on Python 2.5, due to a bug in the version detection code (Issue #378).
Beat: Adds method
This method can be overridden to change the default value of
An error occurring in process cleanup could mask task errors.
We no longer propagate errors happening at process cleanup, but log them instead. This way they won’t interfere with publishing the task result (Issue #365).
Defining tasks didn’t work properly when using the Django
shell_plusutility (Issue #366).
AsyncResult.getdidn’t accept the
- worker: Fixed a bug where the worker wouldn’t shutdown if a
2011-03-28 06:00 p.m. CEST
Now depends on Kombu 1.0.7
Our documentation is now hosted by Read The Docs (https://docs.celeryq.dev), and all links have been changed to point to the new URL.
Logging: Now supports log rotation using external tools like logrotate.d (Issue #321)
This is accomplished by using the
WatchedFileHandler, which re-opens the file if it’s renamed or deleted.
otherqueuestutorial now documents how to configure Redis/Database result
gevent: Now supports ETA tasks.
But gevent still needs
TaskSet User Guide: now contains TaskSet callback recipes.
Eventlet: New signals:
celery.signalsfor more information.
BROKER_TRANSPORT_OPTIONSsetting can be used to pass additional arguments to a particular broker transport.
worker_pidis now part of the request info as returned by broadcast commands.
TaskSet.apply/Taskset.apply_async now accepts an optional
The taskset_id (if any) is now available in the Task request context.
SQLAlchemy result backend: taskset_id and taskset_id columns now have a unique constraint (tables need to recreated for this to take affect).
Task user guide: Added section about choosing a result backend.
Removed unused attribute
multiprocessing.Pool: Fixes race condition when marking job with
The process may have published a result before it was terminated, but we have no reliable way to detect that this is the case.
So we have to wait for 10 seconds before marking the result with WorkerLostError. This gives the result handler a chance to retrieve the result.
multiprocessing.Pool: Shutdown could hang if rate limits disabled.
There was a race condition when the MainThread was waiting for the pool semaphore to be released. The ResultHandler now terminates after 5 seconds if there are unacked jobs, but no worker processes left to start them (it needs to timeout because there could still be an ack+result that we haven’t consumed from the result queue. It is unlikely we’ll receive any after 5 seconds with no worker processes).
celerybeat: Now creates pidfile even if the
--detachoption isn’t set.
eventlet/gevent: The broadcast command consumer is now running in a separate green-thread.
This ensures broadcast commands will take priority even if there are many active tasks.
worker: Threads now terminates the program by calling
os._exit, as it is the only way to ensure exit in the case of syntax errors, or other unrecoverable errors.
Fixed typo in
worker: Broadcast commands now logs with loglevel debug instead of warning.
AMQP Result Backend: Now resets cached channel if the connection is lost.
Polling results with the AMQP result backend wasn’t working properly.
Rate limits: No longer sleeps if there are no tasks, but rather waits for the task received condition (Performance improvement).
iter(dict)should return keys, not items (Issue #362).
celerybeat: PersistentScheduler now automatically removes a corrupted schedule file (Issue #346).
Programs that doesn’t support positional command-line arguments now provides a user friendly error message.
Programs no longer tries to load the configuration file when showing
Autoscaler: The “all processes busy” log message is now severity debug instead of error.
worker: If the message body can’t be decoded, it’s now passed through
This to ensure we don’t get additional decoding errors when trying to log the failure.
app.config_from_envvarnow works for all loaders.
Now emits a user-friendly error message if the result backend name is unknown (Issue #349).
celery.contrib.batches: Now sets loglevel and logfile in the task request so
task.get_loggerworks with batch tasks (Issue #357).
worker: An exception was raised if using the amqp transport and the prefetch count value exceeded 65535 (Issue #359).
The prefetch count is incremented for every received task with an ETA/countdown defined. The prefetch count is a short, so can only support a maximum value of 65535. If the value exceeds the maximum value we now disable the prefetch count, it’s re-enabled as soon as the value is below the limit again.
cursesmon: Fixed unbound local error (Issue #303).
eventlet/gevent is now imported on demand so autodoc can import the modules without having eventlet/gevent installed.
worker: Ack callback now properly handles
Task.after_returnis now always called after the result has been written.
Cassandra Result Backend: Should now work with the latest
multiprocessing.Pool: No longer cares if the
putlocksemaphore is released too many times (this can happen if one or more worker processes are killed).
SQLAlchemy Result Backend: Now returns accidentally removed
date_doneagain (Issue #325).
Task.request context is now always initialized to ensure calling the task function directly works even if it actively uses the request context.
Exception occurring when iterating over the result from
eventlet: Now properly schedules tasks with an ETA in the past.
2011-02-19 00:00 AM CET
worker: 2.2.3 broke error logging, resulting in tracebacks not being logged.
AMQP result backend: Polling task states didn’t work properly if there were more than one result message in the queue.
TaskSet.apply()now supports an optional
taskset_idkeyword argument (Issue #331).
The current taskset id (if any) is now available in the task context as
SQLAlchemy result backend: date_done was no longer part of the results as it had been accidentally removed. It’s now available again (Issue #325).
SQLAlchemy result backend: Added unique constraint on Task.id and TaskSet.taskset_id. Tables needs to be recreated for this to take effect.
Fixed exception raised when iterating on the result of
Tasks user guide: Added section on choosing a result backend.
2011-02-12 04:00 p.m. CET
Now depends on Kombu 1.0.3
Task.retry now supports a
max_retriesargument, used to change the default value.
multiprocessing.cpu_count may raise
NotImplementedErroron platforms where this isn’t supported (Issue #320).
Coloring of log messages broke if the logged object wasn’t a string.
Fixed several typos in the init-script documentation.
A regression caused Task.exchange and Task.routing_key to no longer have any effect. This is now fixed.
Routing user guide: Fixes typo, routers in
CELERY_ROUTESmust be instances, not classes.
celeryev didn’t create pidfile even though the
--pidfileargument was set.
Task logger format was no longer used (Issue #317).
The id and name of the task is now part of the log message again.
A safe version of
repr()is now used in strategic places to ensure objects with a broken
__repr__doesn’t crash the worker, or otherwise make errors hard to understand (Issue #298).
Remote control command
active_queues: didn’t account for queues added at runtime.
In addition the dictionary replied by this command now has a different structure: the exchange key is now a dictionary containing the exchange declaration in full.
celery worker -Qoption removed unused queue declarations, so routing of tasks could fail.
Queues are no longer removed, but rather app.amqp.queues.consume_from() is used as the list of queues to consume from.
This ensures all queues are available for routing purposes.
celeryctl: Now supports the inspect active_queues command.
2011-02-03 04:00 p.m. CET
celerybeatcouldn’t read the schedule properly, so entries in
CELERYBEAT_SCHEDULEwouldn’t be scheduled.
Task error log message now includes exc_info again.
The eta argument can now be used with task.retry.
Previously it was overwritten by the countdown argument.
celeryd_detach: Now logs errors occurring when executing the celery worker command.
daemonizing tutorial: Fixed typo
Colors in logging broke non-string objects in log messages.
setup_task_loggerno longer makes assumptions about magic task kwargs.
2011-02-02 04:00 p.m. CET
Eventlet pool was leaking memory (Issue #308).
celery.execute.delay_taskwas accidentally removed, now available again.
BasePool.on_terminatestub didn’t exist
celeryd_detach: Adds readable error messages if user/group name doesn’t exist.
Smarter handling of unicode decode errors when logging errors.
2011-02-01 10:00 AM CET
Carrot has been replaced with Kombu
Kombu is the next generation messaging library for Python, fixing several flaws present in Carrot that was hard to fix without breaking backwards compatibility.
Also it adds:
First-class support for virtual transports; Redis, Django ORM, SQLAlchemy, Beanstalk, MongoDB, CouchDB and in-memory.
Consistent error handling with introspection,
The ability to ensure that an operation is performed by gracefully handling connection and channel errors,
This means that ghettoq is no longer needed as the functionality it provided is already available in Celery by default. The virtual transports are also more feature complete with support for exchanges (direct and topic). The Redis transport even supports fanout exchanges so it’s able to perform worker remote control commands.
Magic keyword arguments pending deprecation.
The magic keyword arguments were responsible for many problems and quirks: notably issues with tasks and decorators, and name collisions in keyword arguments for the unaware.
It wasn’t easy to find a way to deprecate the magic keyword arguments, but we think this is a solution that makes sense and it won’t have any adverse effects for existing code.
The path to a magic keyword argument free world is:
the celery.decorators module is deprecated and the decorators can now be found in celery.task.
The decorators in celery.task disables keyword arguments by default
All examples in the documentation have been changed to use celery.task.
This means that the following will have magic keyword arguments enabled (old style):
from celery.decorators import task @task() def add(x, y, **kwargs): print('In task %s' % kwargs['task_id']) return x + y
And this won’t use magic keyword arguments (new style):
from celery.task import task @task() def add(x, y): print('In task %s' % add.request.id) return x + y
In addition, tasks can choose not to accept magic keyword arguments by setting the task.accept_magic_kwargs attribute.
Using the decorators in
PendingDeprecationWarningwith a helpful message urging you to change your code, in version 2.4 this will be replaced with a
DeprecationWarning, and in version 4.0 the
celery.decoratorsmodule will be removed and no longer exist.
Similarly, the task.accept_magic_kwargs attribute will no longer have any effect starting from version 4.0.
The magic keyword arguments are now available as task.request
This is called the context. Using thread-local storage the context contains state that’s related to the current request.
It’s mutable and you can add custom attributes that’ll only be seen by the current task request.
The following context attributes are always available:
Magic Keyword Argument
In addition, the following methods now automatically uses the current context, so you don’t have to pass kwargs manually anymore:
This is great news for I/O-bound tasks!
To change pool implementations you use the
celery worker --poolargument, or globally using the
CELERYD_POOLsetting. This can be the full name of a class, or one of the following aliases: processes, eventlet, gevent.
For more information please see the Concurrency with Eventlet section in the User Guide.
Why not gevent?
For our first alternative concurrency implementation we’ve focused on Eventlet, but there’s also an experimental gevent pool available. This is missing some features, notably the ability to schedule ETA tasks.
Hopefully the gevent support will be feature complete by version 2.3, but this depends on user demand (and contributions).
Python 2.4 support deprecated!
We’re happy^H^H^H^H^Hsad to announce that this is the last version to support Python 2.4.
You’re urged to make some noise if you’re currently stuck with Python 2.4. Complain to your package maintainers, sysadmins and bosses: tell them it’s time to move on!
Apart from wanting to take advantage of
withstatements, coroutines, conditional expressions and enhanced
tryblocks, the code base now contains so many 2.4 related hacks and workarounds it’s no longer just a compromise, but a sacrifice.
If it really isn’t your choice, and you don’t have the option to upgrade to a newer version of Python, you can just continue to use Celery 2.2. Important fixes can be back ported for as long as there’s interest.
worker: Now supports Autoscaling of child worker processes.
--autoscaleoption can be used to configure the minimum and maximum number of child worker processes:
--autoscale=AUTOSCALE Enable autoscaling by providing max_concurrency,min_concurrency. Example: --autoscale=10,3 (always keep 3 processes, but grow to 10 if necessary).
Remote Debugging of Tasks
celery.contrib.rdbis an extended version of
pdbthat enables remote debugging of processes that doesn’t have terminal access.
from celery.contrib import rdb from celery.task import task @task() def add(x, y): result = x + y # set breakpoint rdb.set_trace() return result :func:`~celery.contrib.rdb.set_trace` sets a breakpoint at the current location and creates a socket you can telnet into to remotely debug your task. The debugger may be started by multiple processes at the same time, so rather than using a fixed port the debugger will search for an available port, starting from the base port (6900 by default). The base port can be changed using the environment variable :envvar:`CELERY_RDB_PORT`. By default the debugger will only be available from the local host, to enable access from the outside you have to set the environment variable :envvar:`CELERY_RDB_HOST`. When the worker encounters your breakpoint it will log the following information:: [INFO/MainProcess] Received task: tasks.add[d7261c71-4962-47e5-b342-2448bedd20e8] [WARNING/PoolWorker-1] Remote Debugger:6900: Please telnet 127.0.0.1 6900. Type `exit` in session to continue. [2011-01-18 14:25:44,119: WARNING/PoolWorker-1] Remote Debugger:6900: Waiting for client... If you telnet the port specified you'll be presented with a ``pdb`` shell: .. code-block:: console $ telnet localhost 6900 Connected to localhost. Escape character is '^]'. > /opt/devel/demoapp/tasks.py(128)add() -> return result (Pdb) Enter ``help`` to get a list of available commands, It may be a good idea to read the `Python Debugger Manual`_ if you have never used `pdb` before.
Events are now transient and is using a topic exchange (instead of direct).
The CELERYD_EVENT_EXCHANGE, CELERYD_EVENT_ROUTING_KEY, CELERYD_EVENT_EXCHANGE_TYPE settings are no longer in use.
This means events won’t be stored until there’s a consumer, and the events will be gone as soon as the consumer stops. Also it means there can be multiple monitors running at the same time.
The routing key of an event is the type of event (e.g., worker.started, worker.heartbeat, task.succeeded, etc. This means a consumer can filter on specific types, to only be alerted of the events it cares about.
Each consumer will create a unique queue, meaning it’s in effect a broadcast exchange.
This opens up a lot of possibilities, for example the workers could listen for worker events to know what workers are in the neighborhood, and even restart workers when they go down (or use this information to optimize tasks/autoscaling).
The event exchange has been renamed from
"celeryev"so it doesn’t collide with older versions.
If you’d like to remove the old exchange you can do so by executing the following command:
$ camqadm exchange.delete celeryevent
The worker now starts without configuration, and configuration can be specified directly on the command-line.
Configuration options must appear after the last argument, separated by two dashes:
$ celery worker -l info -I tasks -- broker.host=localhost broker.vhost=/app
Configuration is now an alias to the original configuration, so changes to the original will reflect Celery at runtime.
celery.conf has been deprecated, and modifying celery.conf.ALWAYS_EAGER will no longer have any effect.
The default configuration is now available in the
celery.app.defaultsmodule. The available configuration options and their types can now be introspected.
Remote control commands are now provided by kombu.pidbox, the generic process mailbox.
Internal module celery.worker.listener has been renamed to celery.worker.consumer, and .CarrotListener is now .Consumer.
Previously deprecated modules celery.models and celery.management.commands have now been removed as per the deprecation time-line.
- [Security: Low severity] Removed celery.task.RemoteExecuteTask and
accompanying functions: dmap, dmap_async, and execute_remote.
Executing arbitrary code using pickle is a potential security issue if someone gains unrestricted access to the message broker.
If you really need this functionality, then you’d’ve to add this to your own project.
[Security: Low severity] The stats command no longer transmits the broker password.
One would’ve needed an authenticated broker connection to receive this password in the first place, but sniffing the password at the wire level would’ve been possible if using unencrypted communication.
The internal module celery.task.builtins has been removed.
The module celery.task.schedules is deprecated, and celery.schedules should be used instead.
For example if you have:
from celery.task.schedules import crontab
You should replace that with:
from celery.schedules import crontab
The module needs to be renamed because it must be possible to import schedules without importing the celery.task module.
The following functions have been deprecated and is scheduled for removal in version 2.3:
Use task.apply_async() instead.
Use task.apply() instead.
Use registry.tasks[name].delay() instead.
Importing TaskSet from celery.task.base is now deprecated.
You should use:
>>> from celery.task import TaskSet
New remote control commands:
Returns the queue declarations a worker is currently consuming from.
Added the ability to retry publishing the task message in the event of connection loss or failure.
This is disabled by default but can be enabled using the
CELERY_TASK_PUBLISH_RETRYsetting, and tweaked by the
In addition retry, and retry_policy keyword arguments have been added to Task.apply_async.
Using the retry argument to apply_async requires you to handle the publisher/connection manually.
Periodic Task classes (@periodic_task/PeriodicTask) will not be deprecated as previously indicated in the source code.
But you’re encouraged to use the more flexible
Built-in daemonization support of the worker using celery multi is no longer experimental and is considered production quality.
See Generic init-scripts if you want to use the new generic init scripts.
Added support for message compression using the
CELERY_MESSAGE_COMPRESSIONsetting, or the compression argument to apply_async. This can also be set using routers.
- worker: Now logs stack-trace of all threads when receiving the
SIGUSR1 signal (doesn’t work on CPython 2.4, Windows or Jython).
Inspired by https://gist.github.com/737056
Can now remotely terminate/kill the worker process currently processing a task.
The revoke remote control command now supports a terminate argument Default signal is TERM, but can be specified using the signal argument. Signal can be the uppercase name of any signal defined in the
signalmodule in the Python Standard Library.
Terminating a task also revokes it.
>>> from celery.task.control import revoke >>> revoke(task_id, terminate=True) >>> revoke(task_id, terminate=True, signal='KILL') >>> revoke(task_id, terminate=True, signal='SIGKILL')
TaskSetResult.join_native: Backend-optimized version of join().
If available, this version uses the backends ability to retrieve multiple results at once, unlike join() which fetches the results one by one.
So far only supported by the AMQP result backend. Support for Memcached and Redis may be added later.
Improved implementations of TaskSetResult.join and AsyncResult.wait.
An interval keyword argument have been added to both so the polling interval can be specified (default interval is 0.5 seconds).
A propagate keyword argument have been added to result.wait(), errors will be returned instead of raised if this is set to False.
You should decrease the polling interval when using the database result backend, as frequent polling can result in high database load.
The PID of the child worker process accepting a task is now sent as a field with the
The following fields have been added to all events in the worker class:
sw_ident: Name of worker software (e.g.,
sw_ver: Software version (e.g., 2.2.0).
sw_sys: Operating System (e.g., Linux, Windows, Darwin).
For better accuracy the start time reported by the multiprocessing worker process is used when calculating task duration.
Previously the time reported by the accept callback was used.
- celerybeat: New built-in daemonization support using the –detach
- celeryev: New built-in daemonization support using the –detach
TaskSet.apply_async: Now supports custom publishers by using the publisher argument.
If enabled an event will be sent with every task, so monitors can track tasks before the workers receive them.
- celerybeat: Now reuses the broker connection when calling
The configuration module and loader to use can now be specified on the command-line.
$ celery worker --config=celeryconfig.py --loader=myloader.Loader
Added signals: beat_init and beat_embedded_init
Dispatched when celerybeat starts (either standalone or embedded). Sender is the
Redis result backend: Removed deprecated settings REDIS_TIMEOUT and REDIS_CONNECT_RETRY.
CentOS init-script for celery worker now available in extra/centos.
Now depends on pyparsing version 1.5.0 or higher.
There have been reported issues using Celery with pyparsing 1.4.x, so please upgrade to the latest version.
Lots of new unit tests written, now with a total coverage of 95%.
celeryev Curses Monitor: Improved resize handling and UI layout (Issue #274 + Issue #276)
AMQP Backend: Exceptions occurring while sending task results are now propagated instead of silenced.
the worker will then show the full traceback of these errors in the log.
AMQP Backend: No longer deletes the result queue after successful poll, as this should be handled by the
AMQP Backend: Now ensures queues are declared before polling results.
Windows: worker: Show error if running with -B option.
celerybeatembedded is known not to work on Windows, so users are encouraged to run
celerybeatas a separate service instead.
Windows: Utilities no longer output ANSI color codes on Windows
camqadm: Now properly handles Control-c by simply exiting instead of showing confusing traceback.
Windows: All tests are now passing on Windows.
Remove bin/ directory, and scripts section from
This means we now rely completely on setuptools entry-points.
Jython: worker now runs on Jython using the threaded pool.
All tests pass, but there may still be bugs lurking around the corners.
PyPy: worker now runs on PyPy.
It runs without any pool, so to get parallel execution you must start multiple instances (e.g., using multi).
Sadly an initial benchmark seems to show a 30% performance decrease on
pypy-1.4.1+ JIT. We would like to find out why this is, so stay tuned.
PublisherPool: Experimental pool of task publishers and connections to be used with the retry argument to apply_async.
The example code below will re-use connections and channels, and retry sending of the task message if the connection is lost.
from celery import current_app # Global pool pool = current_app().amqp.PublisherPool(limit=10) def my_view(request): with pool.acquire() as publisher: add.apply_async((2, 2), publisher=publisher, retry=True)