This document describes the current stable version of Celery (5.4). For development docs, go here.

Change history

This document contains change notes for bugfix releases in the 4.3.x series, please see What’s new in Celery 4.3 (rhubarb) for an overview of what’s new in Celery 4.3.



2020-09-10 1:00 P.M UTC+3:00


Omer Katz

  • Limit vine version to be below 5.0.0.

    Contributed by Omer Katz



2019-03-31 7:00 P.M UTC+3:00


Omer Katz

  • Added support for broadcasting using a regular expression pattern or a glob pattern to multiple Pidboxes.

    This allows you to inspect or ping multiple workers at once.

    Contributed by Dmitry Malinovsky & Jason Held

  • Added support for PEP 420 namespace packages.

    This allows you to load tasks from namespace packages.

    Contributed by Colin Watson

  • Added acks_on_failure_or_timeout as a setting instead of a task only option.

    This was missing from the original PR but now added for completeness.

    Contributed by Omer Katz

  • Added the task_received signal.

    Contributed by Omer Katz

  • Fixed a crash of our CLI that occurred for everyone using Python < 3.6.

    The crash was introduced in acd6025 by using the ModuleNotFoundError exception which was introduced in Python 3.6.

    Contributed by Omer Katz

  • Fixed a crash that occurred when using the Redis result backend while the result_expires is set to None.

    Contributed by Toni Ruža & Omer Katz

  • Added support the DNS seedlist connection format for the MongoDB result backend.

    This requires the dnspython package which will be installed by default when installing the dependencies for the MongoDB result backend.

    Contributed by George Psarakis

  • Bump the minimum eventlet version to 0.24.1.

    Contributed by George Psarakis

  • Replace the msgpack-python package with msgpack.

    We’re no longer using the deprecated package. See our important notes for this release for further details on how to upgrade.

    Contributed by Daniel Hahler

  • Allow scheduling error handlers which are not registered tasks in the current worker.

    These kind of error handlers are now possible:

    from celery import Signature
      'bar', args=['foo'],
      link_error=Signature('msg.err', queue='msg')
  • Additional fixes and enhancements to the SSL support of the Redis broker and result backend.

    Contributed by Jeremy Cohen

Code Cleanups, Test Coverage & CI Improvements by:

  • Omer Katz

  • Florian Chardin

Documentation Fixes by:

  • Omer Katz

  • Samuel Huang

  • Amir Hossein Saeid Mehr

  • Dmytro Litvinov

4.3.0 RC2


2019-03-03 9:30 P.M UTC+2:00


Omer Katz

  • Filesystem Backend: Added meaningful error messages for filesystem backend.

    Contributed by Lars Rinn

  • New Result Backend: Added the ArangoDB backend.

    Contributed by Dilip Vamsi Moturi

  • Django: Prepend current working directory instead of appending so that the project directory will have precedence over system modules as expected.

    Contributed by Antonin Delpeuch

  • Bump minimum py-redis version to 3.2.0.

    Due to multiple bugs in earlier versions of py-redis that were causing issues for Celery, we were forced to bump the minimum required version to 3.2.0.

    Contributed by Omer Katz

  • Dependencies: Bump minimum required version of Kombu to 4.4

    Contributed by Omer Katz

4.3.0 RC1


2019-02-20 5:00 PM IST


Omer Katz

  • Canvas: celery.chain.apply() does not ignore keyword arguments anymore when applying the chain.

    Contributed by Korijn van Golen

  • Result Set: Don’t attempt to cache results in a celery.result.ResultSet.

    During a join, the results cache was populated using celery.result.ResultSet.get(), if one of the results contains an exception, joining unexpectedly failed.

    The results cache is now removed.

    Contributed by Derek Harland

  • Application: celery.Celery.autodiscover_tasks() now attempts to import the package itself when the related_name keyword argument is None.

    Contributed by Alex Ioannidis

  • Windows Support: On Windows 10, stale PID files prevented celery beat to run. We now remove them when a SystemExit is raised.

    Contributed by :github_user:`na387`

  • Task: Added the new task_acks_on_failure_or_timeout setting.

    Acknowledging SQS messages on failure or timing out makes it impossible to use dead letter queues.

    We introduce the new option acks_on_failure_or_timeout, to ensure we can totally fallback on native SQS message lifecycle, using redeliveries for retries (in case of slow processing or failure) and transitions to dead letter queue after defined number of times.

    Contributed by Mario Kostelac

  • RabbitMQ Broker: Adjust HA headers to work on RabbitMQ 3.x.

    This change also means we’re ending official support for RabbitMQ 2.x.

    Contributed by Asif Saif Uddin

  • Command Line: Improve celery update error handling.

    Contributed by Federico Bond

  • Canvas: Support chords with task_always_eager set to True.

    Contributed by Axel Haustant

  • Result Backend: Optionally store task properties in result backend.

    Setting the result_extended configuration option to True enables storing additional task properties in the result backend.

    Contributed by John Arnold

  • Couchbase Result Backend: Allow the Couchbase result backend to automatically detect the serialization format.

    Contributed by Douglas Rohde

  • New Result Backend: Added the Azure Block Blob Storage result backend.

    The backend is implemented on top of the azure-storage library which uses Azure Blob Storage for a scalable low-cost PaaS backend.

    The backend was load tested via a simple nginx/gunicorn/sanic app hosted on a DS4 virtual machine (4 vCores, 16 GB RAM) and was able to handle 600+ concurrent users at ~170 RPS.

    The commit also contains a live end-to-end test to facilitate verification of the backend functionality. The test is activated by setting the AZUREBLOCKBLOB_URL environment variable to azureblockblob://{ConnectionString} where the value for ConnectionString can be found in the Access Keys pane of a Storage Account resources in the Azure Portal.

    Contributed by Clemens Wolff

  • Task: now accepts keyword arguments.

    This allows passing extra fields to the result backend. These fields are unused by default but custom result backends can use them to determine how to store results.

    Contributed by Christopher Dignam

  • Gracefully handle consumer kombu.exceptions.DecodeError.

    When using the v2 protocol the worker no longer crashes when the consumer encounters an error while decoding a message.

    Contributed by Steven Sklar

  • Deployment: Fix init.d service stop.

    Contributed by Marcus McHale

  • Django: Drop support for Django < 1.11.

    Contributed by Asif Saif Uddin

  • Django: Remove old djcelery loader.

    Contributed by Asif Saif Uddin

  • Result Backend: celery.worker.request.Request now passes to the backend’s store_result functions.

    Since the class currently passes self to these functions, revoking a task resulted in corrupted task result data when django-celery-results was used.

    Contributed by Kiyohiro Yamaguchi

  • Worker: Retry if the heartbeat connection dies.

    Previously, we keep trying to write to the broken connection. This results in a memory leak because the event dispatcher will keep appending the message to the outbound buffer.

    Contributed by Raf Geens

  • Celery Beat: Handle microseconds when scheduling.

    Contributed by K Davis

  • Asynpool: Fixed deadlock when closing socket.

    Upon attempting to close a socket, celery.concurrency.asynpool.AsynPool only removed the queue writer from the hub but did not remove the reader. This led to a deadlock on the file descriptor and eventually the worker stopped accepting new tasks.

    We now close both the reader and the writer file descriptors in a single loop iteration which prevents the deadlock.

    Contributed by Joshua Engelman

  • Celery Beat: Correctly consider timezone when calculating timestamp.

    Contributed by :github_user:`yywing`

  • Celery Beat: celery.beat.Scheduler.schedules_equal() can now handle either arguments being a None value.

    Contributed by :github_user:` ratson`

  • Documentation/Sphinx: Fixed Sphinx support for shared_task decorated functions.

    Contributed by Jon Banafato

  • New Result Backend: Added the CosmosDB result backend.

    This change adds a new results backend. The backend is implemented on top of the pydocumentdb library which uses Azure CosmosDB for a scalable, globally replicated, high-performance, low-latency and high-throughput PaaS backend.

    Contributed by Clemens Wolff

  • Application: Added configuration options to allow separate multiple apps to run on a single RabbitMQ vhost.

    The newly added event_exchange and control_exchange configuration options allow users to use separate Pidbox exchange and a separate events exchange.

    This allow different Celery applications to run separately on the same vhost.

    Contributed by Artem Vasilyev

  • Result Backend: Forget parent result metadata when forgetting a result.

    Contributed by :github_user:`tothegump`

  • Task Store task arguments inside celery.exceptions.MaxRetriesExceededError.

    Contributed by Anthony Ruhier

  • Result Backend: Added the result_accept_content setting.

    This feature allows to configure different accepted content for the result backend.

    A special serializer (auth) is used for signed messaging, however the result_serializer remains in json, because we don’t want encrypted content in our result backend.

    To accept unsigned content from the result backend, we introduced this new configuration option to specify the accepted content from the backend.

    Contributed by Benjamin Pereto

  • Canvas: Fixed error callback processing for class based tasks.

    Contributed by Victor Mireyev

  • New Result Backend: Added the S3 result backend.

    Contributed by Florian Chardin

  • Task: Added support for Cythonized Celery tasks.

    Contributed by Andrey Skabelin

  • Riak Result Backend: Warn Riak backend users for possible Python 3.7 incompatibilities.

    Contributed by George Psarakis

  • Python Runtime: Added Python 3.7 support.

    Contributed by Omer Katz & Asif Saif Uddin

  • Auth Serializer: Revamped the auth serializer.

    The auth serializer received a complete overhaul. It was previously horribly broken.

    We now depend on cryptography instead of pyOpenSSL for this serializer.

    Contributed by Benjamin Pereto

  • Command Line: celery report now reports kernel version along with other platform details.

    Contributed by Omer Katz

  • Canvas: Fixed chords with chains which include sub chords in a group.

    Celery now correctly executes the last task in these types of canvases:

    c = chord(

    Contributed by Maximilien Cuony

  • Canvas: Complex canvases with error callbacks no longer raises an AttributeError.

    Very complex canvases such as this no longer raise an AttributeError which prevents constructing them.

    We do not know why this bug occurs yet.

    Contributed by Manuel Vázquez Acosta

  • Command Line: Added proper error messages in cases where app cannot be loaded.

    Previously, celery crashed with an exception.

    We now print a proper error message.

    Contributed by Omer Katz

  • Task: Added the task_default_priority setting.

    You can now set the default priority of a task using the task_default_priority setting. The setting’s value will be used if no priority is provided for a specific task.

    Contributed by :github_user:`madprogrammer`

  • Dependencies: Bump minimum required version of Kombu to 4.3 and Billiard to 3.6.

    Contributed by Asif Saif Uddin

  • Result Backend: Fix memory leak.

    We reintroduced weak references to bound methods for AsyncResult callback promises, after adding full weakref support for Python 2 in vine. More details can be found in celery/celery#4839.

    Contributed by George Psarakis and :github_user:`monsterxx03`.

  • Task Execution: Fixed roundtrip serialization for eager tasks.

    When doing the roundtrip serialization for eager tasks, the task serializer will always be JSON unless the serializer argument is present in the call to If the serializer argument is present but is ‘pickle’, an exception will be raised as pickle-serialized objects cannot be deserialized without specifying to serialization.loads what content types should be accepted. The Producer’s serializer seems to be set to None, causing the default to JSON serialization.

    We now continue to use (in order) the serializer argument to, if present, or the Producer’s serializer if not None. If the Producer’s serializer is None, it will use the Celery app’s task_serializer configuration entry as the serializer.

    Contributed by Brett Jackson

  • Redis Result Backend: The celery.backends.redis.ResultConsumer class no longer assumes celery.backends.redis.ResultConsumer.start() to be called before celery.backends.redis.ResultConsumer.drain_events().

    This fixes a race condition when using the Gevent workers pool.

    Contributed by Noam Kush

  • Task: Added the task_inherit_parent_priority setting.

    Setting the task_inherit_parent_priority configuration option to True will make Celery tasks inherit the priority of the previous task linked to it.


    c = celery.chain(
      add.s(2), # priority=None
      add.s(3).set(priority=5), # priority=5
      add.s(4), # priority=5
      add.s(5).set(priority=3), # priority=3
      add.s(6), # priority=3
    def child_task(self):
    def parent_task(self):
    # child_task will also have priority=5
    parent_task.apply_async(args=[], priority=5)

    Contributed by :github_user:`madprogrammer`

  • Canvas: Added the result_chord_join_timeout setting.

    Previously, celery.result.GroupResult.join() had a fixed timeout of 3 seconds.

    The result_chord_join_timeout setting now allows you to change it.

    Contributed by :github_user:`srafehi`

Code Cleanups, Test Coverage & CI Improvements by:

  • Jon Dufresne

  • Asif Saif Uddin

  • Omer Katz

  • Brett Jackson

  • Bruno Alla

  • :github_user:`tothegump`

  • Bojan Jovanovic

  • Florian Chardin

  • :github_user:`walterqian`

  • Fabian Becker

  • Lars Rinn

  • :github_user:`madprogrammer`

  • Ciaran Courtney

Documentation Fixes by:

  • Lewis M. Kabui

  • Dash Winterson

  • Shanavas M

  • Brett Randall

  • Przemysław Suliga

  • Joshua Schmid

  • Asif Saif Uddin

  • Xiaodong

  • Vikas Prasad

  • Jamie Alessio

  • Lars Kruse

  • Guilherme Caminha

  • Andrea Rabbaglietti

  • Itay Bittan

  • Noah Hall

  • Peng Weikang

  • Mariatta Wijaya

  • Ed Morley

  • Paweł Adamczak

  • :github_user:`CoffeeExpress`

  • :github_user:`aviadatsnyk`

  • Brian Schrader

  • Josue Balandrano Coronel

  • Tom Clancy

  • Sebastian Wojciechowski

  • Meysam Azad

  • Willem Thiart

  • Charles Chan

  • Omer Katz

  • Milind Shakya