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

What’s new in Celery 4.0 (latentcall)

Author:

Ask Solem (ask at celeryproject.org)

Celery is a simple, flexible, and reliable distributed system to process vast amounts of messages, while providing operations with the tools required to maintain such a system.

It’s a task queue with focus on real-time processing, while also supporting task scheduling.

Celery has a large and diverse community of users and contributors, you should come join us on IRC or our mailing-list.

To read more about Celery you should go read the introduction.

While this version is backward compatible with previous versions it’s important that you read the following section.

This version is officially supported on CPython 2.7, 3.4, and 3.5. and also supported on PyPy.

Preface

Welcome to Celery 4!

This is a massive release with over two years of changes. Not only does it come with many new features, but it also fixes a massive list of bugs, so in many ways you could call it our “Snow Leopard” release.

The next major version of Celery will support Python 3.5 only, where we are planning to take advantage of the new asyncio library.

This release would not have been possible without the support of my employer, Robinhood (we’re hiring!).

  • Ask Solem

Dedicated to Sebastian “Zeb” Bjørnerud (RIP), with special thanks to Ty Wilkins, for designing our new logo, all the contributors who help make this happen, and my colleagues at Robinhood.

Wall of Contributors

Aaron McMillin, Adam Chainz, Adam Renberg, Adriano Martins de Jesus, Adrien Guinet, Ahmet Demir, Aitor Gómez-Goiri, Alan Justino, Albert Wang, Alex Koshelev, Alex Rattray, Alex Williams, Alexander Koshelev, Alexander Lebedev, Alexander Oblovatniy, Alexey Kotlyarov, Ali Bozorgkhan, Alice Zoë Bevan–McGregor, Allard Hoeve, Alman One, Amir Rustamzadeh, Andrea Rabbaglietti, Andrea Rosa, Andrei Fokau, Andrew Rodionoff, Andrew Stewart, Andriy Yurchuk, Aneil Mallavarapu, Areski Belaid, Armenak Baburyan, Arthur Vuillard, Artyom Koval, Asif Saifuddin Auvi, Ask Solem, Balthazar Rouberol, Batiste Bieler, Berker Peksag, Bert Vanderbauwhede, Brendan Smithyman, Brian Bouterse, Bryce Groff, Cameron Will, ChangBo Guo, Chris Clark, Chris Duryee, Chris Erway, Chris Harris, Chris Martin, Chillar Anand, Colin McIntosh, Conrad Kramer, Corey Farwell, Craig Jellick, Cullen Rhodes, Dallas Marlow, Daniel Devine, Daniel Wallace, Danilo Bargen, Davanum Srinivas, Dave Smith, David Baumgold, David Harrigan, David Pravec, Dennis Brakhane, Derek Anderson, Dmitry Dygalo, Dmitry Malinovsky, Dongweiming, Dudás Ádám, Dustin J. Mitchell, Ed Morley, Edward Betts, Éloi Rivard, Emmanuel Cazenave, Fahad Siddiqui, Fatih Sucu, Feanil Patel, Federico Ficarelli, Felix Schwarz, Felix Yan, Fernando Rocha, Flavio Grossi, Frantisek Holop, Gao Jiangmiao, George Whewell, Gerald Manipon, Gilles Dartiguelongue, Gino Ledesma, Greg Wilbur, Guillaume Seguin, Hank John, Hogni Gylfason, Ilya Georgievsky, Ionel Cristian Mărieș, Ivan Larin, James Pulec, Jared Lewis, Jason Veatch, Jasper Bryant-Greene, Jeff Widman, Jeremy Tillman, Jeremy Zafran, Jocelyn Delalande, Joe Jevnik, Joe Sanford, John Anderson, John Barham, John Kirkham, John Whitlock, Jonathan Vanasco, Joshua Harlow, João Ricardo, Juan Carlos Ferrer, Juan Rossi, Justin Patrin, Kai Groner, Kevin Harvey, Kevin Richardson, Komu Wairagu, Konstantinos Koukopoulos, Kouhei Maeda, Kracekumar Ramaraju, Krzysztof Bujniewicz, Latitia M. Haskins, Len Buckens, Lev Berman, lidongming, Lorenzo Mancini, Lucas Wiman, Luke Pomfrey, Luyun Xie, Maciej Obuchowski, Manuel Kaufmann, Marat Sharafutdinov, Marc Sibson, Marcio Ribeiro, Marin Atanasov Nikolov, Mathieu Fenniak, Mark Parncutt, Mauro Rocco, Maxime Beauchemin, Maxime Vdb, Mher Movsisyan, Michael Aquilina, Michael Duane Mooring, Michael Permana, Mickaël Penhard, Mike Attwood, Mitchel Humpherys, Mohamed Abouelsaoud, Morris Tweed, Morton Fox, Môshe van der Sterre, Nat Williams, Nathan Van Gheem, Nicolas Unravel, Nik Nyby, Omer Katz, Omer Korner, Ori Hoch, Paul Pearce, Paulo Bu, Pavlo Kapyshin, Philip Garnero, Pierre Fersing, Piotr Kilczuk, Piotr Maślanka, Quentin Pradet, Radek Czajka, Raghuram Srinivasan, Randy Barlow, Raphael Michel, Rémy Léone, Robert Coup, Robert Kolba, Rockallite Wulf, Rodolfo Carvalho, Roger Hu, Romuald Brunet, Rongze Zhu, Ross Deane, Ryan Luckie, Rémy Greinhofer, Samuel Giffard, Samuel Jaillet, Sergey Azovskov, Sergey Tikhonov, Seungha Kim, Simon Peeters, Spencer E. Olson, Srinivas Garlapati, Stephen Milner, Steve Peak, Steven Sklar, Stuart Axon, Sukrit Khera, Tadej Janež, Taha Jahangir, Takeshi Kanemoto, Tayfun Sen, Tewfik Sadaoui, Thomas French, Thomas Grainger, Tomas Machalek, Tobias Schottdorf, Tocho Tochev, Valentyn Klindukh, Vic Kumar, Vladimir Bolshakov, Vladimir Gorbunov, Wayne Chang, Wieland Hoffmann, Wido den Hollander, Wil Langford, Will Thompson, William King, Yury Selivanov, Vytis Banaitis, Zoran Pavlovic, Xin Li, 許邱翔, @allenling, @alzeih, @bastb, @bee-keeper, @ffeast, @firefly4268, @flyingfoxlee, @gdw2, @gitaarik, @hankjin, @lvh, @m-vdb, @kindule, @mdk:, @michael-k, @mozillazg, @nokrik, @ocean1, @orlo666, @raducc, @wanglei, @worldexception, @xBeAsTx.

Note

This wall was automatically generated from git history, so sadly it doesn’t not include the people who help with more important things like answering mailing-list questions.

Upgrading from Celery 3.1

Step 1: Upgrade to Celery 3.1.25

If you haven’t already, the first step is to upgrade to Celery 3.1.25.

This version adds forward compatibility to the new message protocol, so that you can incrementally upgrade from 3.1 to 4.0.

Deploy the workers first by upgrading to 3.1.25, this means these workers can process messages sent by clients using both 3.1 and 4.0.

After the workers are upgraded you can upgrade the clients (e.g. web servers).

Step 2: Update your configuration with the new setting names

This version radically changes the configuration setting names, to be more consistent.

The changes are fully backwards compatible, so you have the option to wait until the old setting names are deprecated, but to ease the transition we have included a command-line utility that rewrites your settings automatically.

See Lowercase setting names for more information.

Step 3: Read the important notes in this document

Make sure you are not affected by any of the important upgrade notes mentioned in the following section.

An especially important note is that Celery now checks the arguments you send to a task by matching it to the signature (Task argument checking).

Step 4: Upgrade to Celery 4.0

At this point you can upgrade your workers and clients with the new version.

Important Notes

Dropped support for Python 2.6

Celery now requires Python 2.7 or later, and also drops support for Python 3.3 so supported versions are:

  • CPython 2.7

  • CPython 3.4

  • CPython 3.5

  • PyPy 5.4 (pypy2)

  • PyPy 5.5-alpha (pypy3)

Last major version to support Python 2

Starting from Celery 5.0 only Python 3.5+ will be supported.

To make sure you’re not affected by this change you should pin the Celery version in your requirements file, either to a specific version: celery==4.0.0, or a range: celery>=4.0,<5.0.

Dropping support for Python 2 will enable us to remove massive amounts of compatibility code, and going with Python 3.5 allows us to take advantage of typing, async/await, asyncio, and similar concepts there’s no alternative for in older versions.

Celery 4.x will continue to work on Python 2.7, 3.4, 3.5; just as Celery 3.x still works on Python 2.6.

Django support

Celery 4.x requires Django 1.8 or later, but we really recommend using at least Django 1.9 for the new transaction.on_commit feature.

A common problem when calling tasks from Django is when the task is related to a model change, and you wish to cancel the task if the transaction is rolled back, or ensure the task is only executed after the changes have been written to the database.

transaction.atomic enables you to solve this problem by adding the task as a callback to be called only when the transaction is committed.

Example usage:

from functools import partial
from django.db import transaction

from .models import Article, Log
from .tasks import send_article_created_notification

def create_article(request):
    with transaction.atomic():
        article = Article.objects.create(**request.POST)
        # send this task only if the rest of the transaction succeeds.
        transaction.on_commit(partial(
            send_article_created_notification.delay, article_id=article.pk))
        Log.objects.create(type=Log.ARTICLE_CREATED, object_pk=article.pk)

Removed features

  • Microsoft Windows is no longer supported.

    The test suite is passing, and Celery seems to be working with Windows, but we make no guarantees as we are unable to diagnose issues on this platform. If you are a company requiring support on this platform, please get in touch.

  • Jython is no longer supported.

Features removed for simplicity

  • Webhook task machinery (celery.task.http) has been removed.

    Nowadays it’s easy to use the requests module to write webhook tasks manually. We would love to use requests but we are simply unable to as there’s a very vocal ‘anti-dependency’ mob in the Python community

    If you need backwards compatibility you can simply copy + paste the 3.1 version of the module and make sure it’s imported by the worker: https://github.com/celery/celery/blob/3.1/celery/task/http.py

  • Tasks no longer sends error emails.

    This also removes support for app.mail_admins, and any functionality related to sending emails.

  • celery.contrib.batches has been removed.

    This was an experimental feature, so not covered by our deprecation timeline guarantee.

    You can copy and pase the existing batches code for use within your projects: https://github.com/celery/celery/blob/3.1/celery/contrib/batches.py

Features removed for lack of funding

We announced with the 3.1 release that some transports were moved to experimental status, and that there’d be no official support for the transports.

As this subtle hint for the need of funding failed we’ve removed them completely, breaking backwards compatibility.

  • Using the Django ORM as a broker is no longer supported.

    You can still use the Django ORM as a result backend: see django-celery-results - Using the Django ORM/Cache as a result backend section for more information.

  • Using SQLAlchemy as a broker is no longer supported.

    You can still use SQLAlchemy as a result backend.

  • Using CouchDB as a broker is no longer supported.

    You can still use CouchDB as a result backend.

  • Using IronMQ as a broker is no longer supported.

  • Using Beanstalk as a broker is no longer supported.

In addition some features have been removed completely so that attempting to use them will raise an exception:

  • The --autoreload feature has been removed.

    This was an experimental feature, and not covered by our deprecation timeline guarantee. The flag is removed completely so the worker will crash at startup when present. Luckily this flag isn’t used in production systems.

  • The experimental threads pool is no longer supported and has been removed.

  • The force_execv feature is no longer supported.

    The celery worker command now ignores the --no-execv, --force-execv, and the CELERYD_FORCE_EXECV setting.

    This flag will be removed completely in 5.0 and the worker will raise an error.

  • The old legacy “amqp” result backend has been deprecated, and will be removed in Celery 5.0.

    Please use the rpc result backend for RPC-style calls, and a persistent result backend for multi-consumer results.

We think most of these can be fixed without considerable effort, so if you’re interested in getting any of these features back, please get in touch.

Now to the good news…

New Task Message Protocol

This version introduces a brand new task message protocol, the first major change to the protocol since the beginning of the project.

The new protocol is enabled by default in this version and since the new version isn’t backwards compatible you have to be careful when upgrading.

The 3.1.25 version was released to add compatibility with the new protocol so the easiest way to upgrade is to upgrade to that version first, then upgrade to 4.0 in a second deployment.

If you wish to keep using the old protocol you may also configure the protocol version number used:

app = Celery()
app.conf.task_protocol = 1

Read more about the features available in the new protocol in the news section found later in this document.

Lowercase setting names

In the pursuit of beauty all settings are now renamed to be in all lowercase and some setting names have been renamed for consistency.

This change is fully backwards compatible so you can still use the uppercase setting names, but we would like you to upgrade as soon as possible and you can do this automatically using the celery upgrade settings command:

$ celery upgrade settings proj/settings.py

This command will modify your module in-place to use the new lower-case names (if you want uppercase with a “CELERY” prefix see block below), and save a backup in proj/settings.py.orig.

For Django users and others who want to keep uppercase names

If you’re loading Celery configuration from the Django settings module then you’ll want to keep using the uppercase names.

You also want to use a CELERY_ prefix so that no Celery settings collide with Django settings used by other apps.

To do this, you’ll first need to convert your settings file to use the new consistent naming scheme, and add the prefix to all Celery related settings:

$ celery upgrade settings proj/settings.py --django

After upgrading the settings file, you need to set the prefix explicitly in your proj/celery.py module:

app.config_from_object('django.conf:settings', namespace='CELERY')

You can find the most up to date Django Celery integration example here: First steps with Django.

Note

This will also add a prefix to settings that didn’t previously have one, for example BROKER_URL should be written CELERY_BROKER_URL with a namespace of CELERY CELERY_BROKER_URL.

Luckily you don’t have to manually change the files, as the celery upgrade settings --django program should do the right thing.

The loader will try to detect if your configuration is using the new format, and act accordingly, but this also means you’re not allowed to mix and match new and old setting names, that’s unless you provide a value for both alternatives.

The major difference between previous versions, apart from the lower case names, are the renaming of some prefixes, like celerybeat_ to beat_, celeryd_ to worker_.

The celery_ prefix has also been removed, and task related settings from this name-space is now prefixed by task_, worker related settings with worker_.

Apart from this most of the settings will be the same in lowercase, apart from a few special ones:

Setting name

Replace with

CELERY_MAX_CACHED_RESULTS

result_cache_max

CELERY_MESSAGE_COMPRESSION

result_compression/task_compression.

CELERY_TASK_RESULT_EXPIRES

result_expires

CELERY_RESULT_DBURI

result_backend

CELERY_RESULT_ENGINE_OPTIONS

database_engine_options

-*-_DB_SHORT_LIVED_SESSIONS

database_short_lived_sessions

CELERY_RESULT_DB_TABLE_NAMES

database_db_names

CELERY_ACKS_LATE

task_acks_late

CELERY_ALWAYS_EAGER

task_always_eager

CELERY_ANNOTATIONS

task_annotations

CELERY_MESSAGE_COMPRESSION

task_compression

CELERY_CREATE_MISSING_QUEUES

task_create_missing_queues

CELERY_DEFAULT_DELIVERY_MODE

task_default_delivery_mode

CELERY_DEFAULT_EXCHANGE

task_default_exchange

CELERY_DEFAULT_EXCHANGE_TYPE

task_default_exchange_type

CELERY_DEFAULT_QUEUE

task_default_queue

CELERY_DEFAULT_RATE_LIMIT

task_default_rate_limit

CELERY_DEFAULT_ROUTING_KEY

task_default_routing_key

-"-_EAGER_PROPAGATES_EXCEPTIONS

task_eager_propagates

CELERY_IGNORE_RESULT

task_ignore_result

CELERY_TASK_PUBLISH_RETRY

task_publish_retry

CELERY_TASK_PUBLISH_RETRY_POLICY

task_publish_retry_policy

CELERY_QUEUES

task_queues

CELERY_ROUTES

task_routes

CELERY_SEND_TASK_SENT_EVENT

task_send_sent_event

CELERY_TASK_SERIALIZER

task_serializer

CELERYD_TASK_SOFT_TIME_LIMIT

task_soft_time_limit

CELERYD_TASK_TIME_LIMIT

task_time_limit

CELERY_TRACK_STARTED

task_track_started

CELERY_DISABLE_RATE_LIMITS

worker_disable_rate_limits

CELERY_ENABLE_REMOTE_CONTROL

worker_enable_remote_control

CELERYD_SEND_EVENTS

worker_send_task_events

You can see a full table of the changes in New lowercase settings.

Json is now the default serializer

The time has finally come to end the reign of pickle as the default serialization mechanism, and json is the default serializer starting from this version.

This change was announced with the release of Celery 3.1.

If you’re still depending on pickle being the default serializer, then you have to configure your app before upgrading to 4.0:

task_serializer = 'pickle'
result_serializer = 'pickle'
accept_content = {'pickle'}

The Json serializer now also supports some additional types:

  • datetime, time, date

    Converted to json text, in ISO-8601 format.

  • Decimal

    Converted to json text.

  • django.utils.functional.Promise

    Django only: Lazy strings used for translation etc., are evaluated and conversion to a json type is attempted.

  • uuid.UUID

    Converted to json text.

You can also define a __json__ method on your custom classes to support JSON serialization (must return a json compatible type):

class Person:
    first_name = None
    last_name = None
    address = None

    def __json__(self):
        return {
            'first_name': self.first_name,
            'last_name': self.last_name,
            'address': self.address,
        }

The Task base class no longer automatically register tasks

The Task class is no longer using a special meta-class that automatically registers the task in the task registry.

Instead this is now handled by the app.task decorators.

If you’re still using class based tasks, then you need to register these manually:

class CustomTask(Task):
    def run(self):
        print('running')
CustomTask = app.register_task(CustomTask())

The best practice is to use custom task classes only for overriding general behavior, and then using the task decorator to realize the task:

@app.task(bind=True, base=CustomTask)
def custom(self):
    print('running')

This change also means that the abstract attribute of the task no longer has any effect.

Task argument checking

The arguments of the task are now verified when calling the task, even asynchronously:

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

>>> add.delay(8, 8)
<AsyncResult: f59d71ca-1549-43e0-be41-4e8821a83c0c>

>>> add.delay(8)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "celery/app/task.py", line 376, in delay
    return self.apply_async(args, kwargs)
  File "celery/app/task.py", line 485, in apply_async
    check_arguments(*(args or ()), **(kwargs or {}))
TypeError: add() takes exactly 2 arguments (1 given)

You can disable the argument checking for any task by setting its typing attribute to False:

>>> @app.task(typing=False)
... def add(x, y):
...     return x + y

Or if you would like to disable this completely for all tasks you can pass strict_typing=False when creating the app:

app = Celery(..., strict_typing=False)

Redis Events not backward compatible

The Redis fanout_patterns and fanout_prefix transport options are now enabled by default.

Workers/monitors without these flags enabled won’t be able to see workers with this flag disabled. They can still execute tasks, but they cannot receive each others monitoring messages.

You can upgrade in a backward compatible manner by first configuring your 3.1 workers and monitors to enable the settings, before the final upgrade to 4.0:

BROKER_TRANSPORT_OPTIONS = {
    'fanout_patterns': True,
    'fanout_prefix': True,
}

Redis Priorities Reversed

Priority 0 is now lowest, 9 is highest.

This change was made to make priority support consistent with how it works in AMQP.

Contributed by Alex Koshelev.

Django: Auto-discover now supports Django app configurations

The autodiscover_tasks() function can now be called without arguments, and the Django handler will automatically find your installed apps:

app.autodiscover_tasks()

The Django integration example in the documentation has been updated to use the argument-less call.

This also ensures compatibility with the new, ehm, AppConfig stuff introduced in recent Django versions.

Worker direct queues no longer use auto-delete

Workers/clients running 4.0 will no longer be able to send worker direct messages to workers running older versions, and vice versa.

If you’re relying on worker direct messages you should upgrade your 3.x workers and clients to use the new routing settings first, by replacing celery.utils.worker_direct() with this implementation:

from kombu import Exchange, Queue

worker_direct_exchange = Exchange('C.dq2')

def worker_direct(hostname):
    return Queue(
        '{hostname}.dq2'.format(hostname),
        exchange=worker_direct_exchange,
        routing_key=hostname,
    )

This feature closed Issue #2492.

Old command-line programs removed

Installing Celery will no longer install the celeryd, celerybeat and celeryd-multi programs.

This was announced with the release of Celery 3.1, but you may still have scripts pointing to the old names, so make sure you update these to use the new umbrella command:

Program

New Status

Replacement

celeryd

REMOVED

celery worker

celerybeat

REMOVED

celery beat

celeryd-multi

REMOVED

celery multi

News

New protocol highlights

The new protocol fixes many problems with the old one, and enables some long-requested features:

  • Most of the data are now sent as message headers, instead of being serialized with the message body.

    In version 1 of the protocol the worker always had to deserialize the message to be able to read task meta-data like the task id, name, etc. This also meant that the worker was forced to double-decode the data, first deserializing the message on receipt, serializing the message again to send to child process, then finally the child process deserializes the message again.

    Keeping the meta-data fields in the message headers means the worker doesn’t actually have to decode the payload before delivering the task to the child process, and also that it’s now possible for the worker to reroute a task written in a language different from Python to a different worker.

  • A new lang message header can be used to specify the programming language the task is written in.

  • Worker stores results for internal errors like ContentDisallowed, and other deserialization errors.

  • Worker stores results and sends monitoring events for unregistered task errors.

  • Worker calls callbacks/errbacks even when the result is sent by the parent process (e.g., WorkerLostError when a child process terminates, deserialization errors, unregistered tasks).

  • A new origin header contains information about the process sending the task (worker node-name, or PID and host-name information).

  • A new shadow header allows you to modify the task name used in logs.

    This is useful for dispatch like patterns, like a task that calls any function using pickle (don’t do this at home):

    from celery import Task
from celery.utils.imports import qualname

class call_as_task(Task):

    def shadow_name(self, args, kwargs, options):
        return 'call_as_task:{0}'.format(qualname(args[0]))

    def run(self, fun, *args, **kwargs):
        return fun(*args, **kwargs)
call_as_task = app.register_task(call_as_task())

  • New argsrepr and kwargsrepr fields contain textual representations of the task arguments (possibly truncated) for use in logs, monitors, etc.

    This means the worker doesn’t have to deserialize the message payload to display the task arguments for informational purposes.

  • Chains now use a dedicated chain field enabling support for chains of thousands and more tasks.

  • New parent_id and root_id headers adds information about a tasks relationship with other tasks.

    • parent_id is the task id of the task that called this task

    • root_id is the first task in the work-flow.

    These fields can be used to improve monitors like flower to group related messages together (like chains, groups, chords, complete work-flows, etc).

  • app.TaskProducer replaced by app.amqp.create_task_message() and app.amqp.send_task_message().

    Dividing the responsibilities into creating and sending means that people who want to send messages using a Python AMQP client directly, don’t have to implement the protocol.

    The app.amqp.create_task_message() method calls either app.amqp.as_task_v2(), or app.amqp.as_task_v1() depending on the configured task protocol, and returns a special task_message tuple containing the headers, properties and body of the task message.

See also

The new task protocol is documented in full here: Version 2.

Prefork Pool Improvements

Tasks now log from the child process

Logging of task success/failure now happens from the child process executing the task. As a result logging utilities, like Sentry can get full information about tasks, including variables in the traceback stack.

-Ofair is now the default scheduling strategy

To re-enable the default behavior in 3.1 use the -Ofast command-line option.

There’s been lots of confusion about what the -Ofair command-line option does, and using the term “prefetch” in explanations have probably not helped given how confusing this terminology is in AMQP.

When a Celery worker using the prefork pool receives a task, it needs to delegate that task to a child process for execution.

The prefork pool has a configurable number of child processes (--concurrency) that can be used to execute tasks, and each child process uses pipes/sockets to communicate with the parent process:

  • inqueue (pipe/socket): parent sends task to the child process

  • outqueue (pipe/socket): child sends result/return value to the parent.

In Celery 3.1 the default scheduling mechanism was simply to send the task to the first inqueue that was writable, with some heuristics to make sure we round-robin between them to ensure each child process would receive the same amount of tasks.

This means that in the default scheduling strategy, a worker may send tasks to the same child process that is already executing a task. If that task is long running, it may block the waiting task for a long time. Even worse, hundreds of short-running tasks may be stuck behind a long running task even when there are child processes free to do work.

The -Ofair scheduling strategy was added to avoid this situation, and when enabled it adds the rule that no task should be sent to the a child process that is already executing a task.

The fair scheduling strategy may perform slightly worse if you have only short running tasks.

Limit child process resident memory size

You can now limit the maximum amount of memory allocated per prefork pool child process by setting the worker --max-memory-per-child option, or the worker_max_memory_per_child setting.

The limit is for RSS/resident memory size and is specified in kilobytes.

A child process having exceeded the limit will be terminated and replaced with a new process after the currently executing task returns.

See Max memory per child setting for more information.

Contributed by Dave Smith.

One log-file per child process

Init-scrips and celery multi now uses the %I log file format option (e.g., /var/log/celery/%n%I.log).

This change was necessary to ensure each child process has a separate log file after moving task logging to the child process, as multiple processes writing to the same log file can cause corruption.

You’re encouraged to upgrade your init-scripts and celery multi arguments to use this new option.

Transports

RabbitMQ priority queue support

See RabbitMQ Message Priorities for more information.

Contributed by Gerald Manipon.

Configure broker URL for read/write separately

New broker_read_url and broker_write_url settings have been added so that separate broker URLs can be provided for connections used for consuming/publishing.

In addition to the configuration options, two new methods have been added the app API:

  • app.connection_for_read()

  • app.connection_for_write()

These should now be used in place of app.connection() to specify the intent of the required connection.

Note

Two connection pools are available: app.pool (read), and app.producer_pool (write). The latter doesn’t actually give connections but full kombu.Producer instances.

def publish_some_message(app, producer=None):
    with app.producer_or_acquire(producer) as producer:
        ...

def consume_messages(app, connection=None):
    with app.connection_or_acquire(connection) as connection:
        ...

RabbitMQ queue extensions support

Queue declarations can now set a message TTL and queue expiry time directly, by using the message_ttl and expires arguments

New arguments have been added to Queue that lets you directly and conveniently configure RabbitMQ queue extensions in queue declarations:

  • Queue(expires=20.0)

    Set queue expiry time in float seconds.

    See kombu.Queue.expires.

  • Queue(message_ttl=30.0)

    Set queue message time-to-live float seconds.

    See kombu.Queue.message_ttl.

  • Queue(max_length=1000)

    Set queue max length (number of messages) as int.

    See kombu.Queue.max_length.

  • Queue(max_length_bytes=1000)

    Set queue max length (message size total in bytes) as int.

    See kombu.Queue.max_length_bytes.

  • Queue(max_priority=10)

    Declare queue to be a priority queue that routes messages based on the priority field of the message.

    See kombu.Queue.max_priority.

Amazon SQS transport now officially supported

The SQS broker transport has been rewritten to use async I/O and as such joins RabbitMQ, Redis and QPid as officially supported transports.

The new implementation also takes advantage of long polling, and closes several issues related to using SQS as a broker.

This work was sponsored by Nextdoor.

Apache QPid transport now officially supported

Contributed by Brian Bouterse.

Redis: Support for Sentinel

You can point the connection to a list of sentinel URLs like:

sentinel://0.0.0.0:26379;sentinel://0.0.0.0:26380/...

where each sentinel is separated by a ;. Multiple sentinels are handled by kombu.Connection constructor, and placed in the alternative list of servers to connect to in case of connection failure.

Contributed by Sergey Azovskov, and Lorenzo Mancini.

Tasks

Task Auto-retry Decorator

Writing custom retry handling for exception events is so common that we now have built-in support for it.

For this a new autoretry_for argument is now supported by the task decorators, where you can specify a tuple of exceptions to automatically retry for:

from twitter.exceptions import FailWhaleError

@app.task(autoretry_for=(FailWhaleError,))
def refresh_timeline(user):
    return twitter.refresh_timeline(user)

See Automatic retry for known exceptions for more information.

Contributed by Dmitry Malinovsky.

Task.replace Improvements

  • self.replace(signature) can now replace any task, chord or group, and the signature to replace with can be a chord, group or any other type of signature.

  • No longer inherits the callbacks and errbacks of the existing task.

    If you replace a node in a tree, then you wouldn’t expect the new node to inherit the children of the old node.

  • Task.replace_in_chord has been removed, use .replace instead.

  • If the replacement is a group, that group will be automatically converted to a chord, where the callback “accumulates” the results of the group tasks.

    A new built-in task (celery.accumulate was added for this purpose)

Contributed by Steeve Morin, and Ask Solem.

Remote Task Tracebacks

The new task_remote_tracebacks will make task tracebacks more useful by injecting the stack of the remote worker.

This feature requires the additional tblib library.

Contributed by Ionel Cristian Mărieș.

Handling task connection errors

Connection related errors occurring while sending a task is now re-raised as a kombu.exceptions.OperationalError error:

>>> try:
...     add.delay(2, 2)
... except add.OperationalError as exc:
...     print('Could not send task %r: %r' % (add, exc))

See Connection Error Handling for more information.

Gevent/Eventlet: Dedicated thread for consuming results

When using gevent, or eventlet there is now a single thread responsible for consuming events.

This means that if you have many calls retrieving results, there will be a dedicated thread for consuming them:

result = add.delay(2, 2)

# this call will delegate to the result consumer thread:
#   once the consumer thread has received the result this greenlet can
# continue.
value = result.get(timeout=3)

This makes performing RPC calls when using gevent/eventlet perform much better.

AsyncResult.then(on_success, on_error)

The AsyncResult API has been extended to support the promise protocol.

This currently only works with the RPC (amqp) and Redis result backends, but lets you attach callbacks to when tasks finish:

import gevent.monkey
monkey.patch_all()

import time
from celery import Celery

app = Celery(broker='amqp://', backend='rpc')

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

def on_result_ready(result):
    print('Received result for id %r: %r' % (result.id, result.result,))

add.delay(2, 2).then(on_result_ready)

time.sleep(3)  # run gevent event loop for a while.

Demonstrated using gevent here, but really this is an API that’s more useful in callback-based event loops like twisted, or tornado.

New Task Router API

The task_routes setting can now hold functions, and map routes now support glob patterns and regexes.

Instead of using router classes you can now simply define a function:

def route_for_task(name, args, kwargs, options, task=None, **kwargs):
    from proj import tasks

    if name == tasks.add.name:
        return {'queue': 'hipri'}

If you don’t need the arguments you can use start arguments, just make sure you always also accept star arguments so that we have the ability to add more features in the future:

def route_for_task(name, *args, **kwargs):
    from proj import tasks
    if name == tasks.add.name:
        return {'queue': 'hipri', 'priority': 9}

Both the options argument and the new task keyword argument are new to the function-style routers, and will make it easier to write routers based on execution options, or properties of the task.

The optional task keyword argument won’t be set if a task is called by name using app.send_task().

For more examples, including using glob/regexes in routers please see task_routes and Automatic routing.

Canvas Refactor

The canvas/work-flow implementation have been heavily refactored to fix some long outstanding issues.

  • Error callbacks can now take real exception and traceback instances (Issue #2538).

    >>> add.s(2, 2).on_error(log_error.s()).delay()

    Where log_error could be defined as:

    @app.task
def log_error(request, exc, traceback):
    with open(os.path.join('/var/errors', request.id), 'a') as fh:
        print('--\n\n{0} {1} {2}'.format(
            task_id, exc, traceback), file=fh)

    See Canvas: Designing Work-flows for more examples.

  • chain(a, b, c) now works the same as a | b | c.

    This means chain may no longer return an instance of chain, instead it may optimize the workflow so that e.g. two groups chained together becomes one group.

  • Now unrolls groups within groups into a single group (Issue #1509).

  • chunks/map/starmap tasks now routes based on the target task

  • chords and chains can now be immutable.

  • Fixed bug where serialized signatures weren’t converted back into signatures (Issue #2078)

    Fix contributed by Ross Deane.

  • Fixed problem where chains and groups didn’t work when using JSON serialization (Issue #2076).

    Fix contributed by Ross Deane.

  • Creating a chord no longer results in multiple values for keyword argument ‘task_id’ (Issue #2225).

    Fix contributed by Aneil Mallavarapu.

  • Fixed issue where the wrong result is returned when a chain contains a chord as the penultimate task.

    Fix contributed by Aneil Mallavarapu.

  • Special case of group(A.s() | group(B.s() | C.s())) now works.

  • Chain: Fixed bug with incorrect id set when a subtask is also a chain.

  • group | group is now flattened into a single group (Issue #2573).

  • Fixed issue where group | task wasn’t upgrading correctly to chord (Issue #2922).

  • Chords now properly sets result.parent links.

  • chunks/map/starmap are now routed based on the target task.

  • Signature.link now works when argument is scalar (not a list)

    (Issue #2019).

  • group() now properly forwards keyword arguments (Issue #3426).

    Fix contributed by Samuel Giffard.

  • A chord where the header group only consists of a single task is now turned into a simple chain.

  • Passing a link argument to group.apply_async() now raises an error (Issue #3508).

  • chord | sig now attaches to the chord callback (Issue #3356).

Periodic Tasks

New API for configuring periodic tasks

This new API enables you to use signatures when defining periodic tasks, removing the chance of mistyping task names.

An example of the new API is here.

Optimized Beat implementation

The celery beat implementation has been optimized for millions of periodic tasks by using a heap to schedule entries.

Contributed by Ask Solem and Alexander Koshelev.

Schedule tasks based on sunrise, sunset, dawn and dusk

See Solar schedules for more information.

Contributed by Mark Parncutt.

Result Backends

RPC Result Backend matured

Lots of bugs in the previously experimental RPC result backend have been fixed and can now be considered to production use.

Contributed by Ask Solem, Morris Tweed.

Redis: Result backend optimizations

result.get() is now using pub/sub for streaming task results

Calling result.get() when using the Redis result backend used to be extremely expensive as it was using polling to wait for the result to become available. A default polling interval of 0.5 seconds didn’t help performance, but was necessary to avoid a spin loop.

The new implementation is using Redis Pub/Sub mechanisms to publish and retrieve results immediately, greatly improving task round-trip times.

Contributed by Yaroslav Zhavoronkov and Ask Solem.

New optimized chord join implementation

This was an experimental feature introduced in Celery 3.1, that could only be enabled by adding ?new_join=1 to the result backend URL configuration.

We feel that the implementation has been tested thoroughly enough to be considered stable and enabled by default.

The new implementation greatly reduces the overhead of chords, and especially with larger chords the performance benefit can be massive.

New Riak result backend introduced

See conf-riak-result-backend for more information.

Contributed by Gilles Dartiguelongue, Alman One and NoKriK.

New CouchDB result backend introduced

See CouchDB backend settings for more information.

Contributed by Nathan Van Gheem.

New Consul result backend introduced

Add support for Consul as a backend using the Key/Value store of Consul.

Consul has an HTTP API where through you can store keys with their values.

The backend extends KeyValueStoreBackend and implements most of the methods.

Mainly to set, get and remove objects.

This allows Celery to store Task results in the K/V store of Consul.

Consul also allows to set a TTL on keys using the Sessions from Consul. This way the backend supports auto expiry of Task results.

For more information on Consul visit https://consul.io/

The backend uses python-consul for talking to the HTTP API. This package is fully Python 3 compliant just as this backend is:

$ pip install python-consul

That installs the required package to talk to Consul’s HTTP API from Python.

You can also specify consul as an extension in your dependency on Celery:

$ pip install celery[consul]

See Bundles for more information.

Contributed by Wido den Hollander.

Brand new Cassandra result backend

A brand new Cassandra backend utilizing the new cassandra-driver library is replacing the old result backend using the older pycassa library.

See Cassandra backend settings for more information.

To depend on Celery with Cassandra as the result backend use:

$ pip install celery[cassandra]

You can also combine multiple extension requirements, please see Bundles for more information.

New Elasticsearch result backend introduced

See Elasticsearch backend settings for more information.

To depend on Celery with Elasticsearch as the result bakend use:

$ pip install celery[elasticsearch]

You can also combine multiple extension requirements, please see Bundles for more information.

Contributed by Ahmet Demir.

New File-system result backend introduced

See File-system backend settings for more information.

Contributed by Môshe van der Sterre.

Event Batching

Events are now buffered in the worker and sent as a list, reducing the overhead required to send monitoring events.

For authors of custom event monitors there will be no action required as long as you’re using the Python Celery helpers (Receiver) to implement your monitor.

However, if you’re parsing raw event messages you must now account for batched event messages, as they differ from normal event messages in the following way:

  • The routing key for a batch of event messages will be set to <event-group>.multi where the only batched event group is currently task (giving a routing key of task.multi).

  • The message body will be a serialized list-of-dictionaries instead of a dictionary. Each item in the list can be regarded as a normal event message body.

In Other News…

Requirements

  • Now depends on Kombu 4.0.

  • Now depends on billiard version 3.5.

  • No longer depends on anyjson. Good-bye old friend :(

Tasks

  • The “anon-exchange” is now used for simple name-name direct routing.

    This increases performance as it completely bypasses the routing table, in addition it also improves reliability for the Redis broker transport.

  • An empty ResultSet now evaluates to True.

    Fix contributed by Colin McIntosh.

  • The default routing key (task_default_routing_key) and exchange name (task_default_exchange) is now taken from the task_default_queue setting.

    This means that to change the name of the default queue, you now only have to set a single setting.

  • New task_reject_on_worker_lost setting, and reject_on_worker_lost task attribute decides what happens when the child worker process executing a late ack task is terminated.

    Contributed by Michael Permana.

  • Task.subtask renamed to Task.signature with alias.

  • Task.subtask_from_request renamed to Task.signature_from_request with alias.

  • The delivery_mode attribute for kombu.Queue is now respected (Issue #1953).

  • Routes in task-routes can now specify a Queue instance directly.

    Example:

    task_routes = {'proj.tasks.add': {'queue': Queue('add')}}

  • AsyncResult now raises ValueError if task_id is None. (Issue #1996).

  • Retried tasks didn’t forward expires setting (Issue #3297).

  • result.get() now supports an on_message argument to set a callback to be called for every message received.

  • New abstract classes added:

  • Task.replace now properly forwards callbacks (Issue #2722).

    Fix contributed by Nicolas Unravel.

  • Task.replace: Append to chain/chord (Closes #3232)

    Fixed issue #3232, adding the signature to the chain (if there’s any). Fixed the chord suppress if the given signature contains one.

    Fix contributed by @honux.

  • Task retry now also throws in eager mode.

    Fix contributed by Feanil Patel.

Beat

  • Fixed crontab infinite loop with invalid date.

    When occurrence can never be reached (example, April, 31th), trying to reach the next occurrence would trigger an infinite loop.

    Try fixing that by raising a RuntimeError after 2,000 iterations

    (Also added a test for crontab leap years in the process)

    Fix contributed by Romuald Brunet.

  • Now ensures the program exits with a non-zero exit code when an exception terminates the service.

    Fix contributed by Simon Peeters.

App

  • Dates are now always timezone aware even if enable_utc is disabled (Issue #943).

    Fix contributed by Omer Katz.

  • Config: App preconfiguration is now also pickled with the configuration.

    Fix contributed by Jeremy Zafran.

  • The application can now change how task names are generated using

    the gen_task_name() method.

    Contributed by Dmitry Malinovsky.

  • App has new app.current_worker_task property that returns the task that’s currently being worked on (or None). (Issue #2100).

Logging

  • get_task_logger() now raises an exception if trying to use the name “celery” or “celery.task” (Issue #3475).

Execution Pools

  • Eventlet/Gevent: now enables AMQP heartbeat (Issue #3338).

  • Eventlet/Gevent: Fixed race condition leading to “simultaneous read” errors (Issue #2755).

  • Prefork: Prefork pool now uses poll instead of select where available (Issue #2373).

  • Prefork: Fixed bug where the pool would refuse to shut down the worker (Issue #2606).

  • Eventlet: Now returns pool size in celery inspect stats command.

    Contributed by Alexander Oblovatniy.

Testing

  • Celery is now a pytest plugin, including fixtures useful for unit and integration testing.

    See the testing user guide for more information.

Transports

  • amqps:// can now be specified to require SSL.

  • Redis Transport: The Redis transport now supports the broker_use_ssl option.

    Contributed by Robert Kolba.

  • JSON serializer now calls obj.__json__ for unsupported types.

    This means you can now define a __json__ method for custom types that can be reduced down to a built-in json type.

    Example:

    class Person:
    first_name = None
    last_name = None
    address = None

    def __json__(self):
        return {
            'first_name': self.first_name,
            'last_name': self.last_name,
            'address': self.address,
        }

  • JSON serializer now handles datetime’s, Django promise, UUID and Decimal.

  • New Queue.consumer_arguments can be used for the ability to set consumer priority via x-priority.

    See https://www.rabbitmq.com/consumer-priority.html

    Example:

    consumer = Consumer(channel, consumer_arguments={'x-priority': 3})

  • Queue/Exchange: no_declare option added (also enabled for internal amq. exchanges).

Programs

  • Celery is now using argparse, instead of optparse.

  • All programs now disable colors if the controlling terminal is not a TTY.

  • celery worker: The -q argument now disables the startup banner.

  • celery worker: The “worker ready” message is now logged using severity info, instead of warn.

  • celery multi: %n format for is now synonym with %N to be consistent with celery worker.

  • celery inspect/celery control: now supports a new --json option to give output in json format.

  • celery inspect registered: now ignores built-in tasks.

  • celery purge now takes -Q and -X options used to specify what queues to include and exclude from the purge.

  • New celery logtool: Utility for filtering and parsing celery worker log-files

  • celery multi: now passes through %i and %I log file formats.

  • General: %p can now be used to expand to the full worker node-name in log-file/pid-file arguments.

  • A new command line option

    --executable is now available for daemonizing programs (celery worker and celery beat).

    Contributed by Bert Vanderbauwhede.

  • celery worker: supports new --prefetch-multiplier option.

    Contributed by Mickaël Penhard.

  • The --loader argument is now always effective even if an app argument is set (Issue #3405).

  • inspect/control now takes commands from registry

    This means user remote-control commands can also be used from the command-line.

    Note that you need to specify the arguments/and type of arguments for the arguments to be correctly passed on the command-line.

    There are now two decorators, which use depends on the type of command: @inspect_command + @control_command:

    from celery.worker.control import control_command

@control_command(
    args=[('n', int)]
    signature='[N=1]',
)
def something(state, n=1, **kwargs):
    ...

    Here args is a list of args supported by the command. The list must contain tuples of (argument_name, type).

    signature is just the command-line help used in e.g. celery -A proj control --help.

    Commands also support variadic arguments, which means that any arguments left over will be added to a single variable. Here demonstrated by the terminate command which takes a signal argument and a variable number of task_ids:

    from celery.worker.control import control_command

@control_command(
    args=[('signal', str)],
    signature='<signal> [id1, [id2, [..., [idN]]]]',
    variadic='ids',
)
def terminate(state, signal, ids, **kwargs):
    ...

    This command can now be called using:

    $ celery -A proj control terminate SIGKILL id1 id2 id3`

    See Writing your own remote control commands for more information.

Worker

  • Improvements and fixes for LimitedSet.

    Getting rid of leaking memory + adding minlen size of the set: the minimal residual size of the set after operating for some time. minlen items are kept, even if they should’ve been expired.

    Problems with older and even more old code:

    1. Heap would tend to grow in some scenarios (like adding an item multiple times).

    2. Adding many items fast wouldn’t clean them soon enough (if ever).

    3. When talking to other workers, revoked._data was sent, but it was processed on the other side as iterable. That means giving those keys new (current) time-stamp. By doing this workers could recycle items forever. Combined with 1) and 2), this means that in large set of workers, you’re getting out of memory soon.

    All those problems should be fixed now.

    This should fix issues #3095, #3086.

    Contributed by David Pravec.

  • New settings to control remote control command queues.

    • control_queue_expires

      Set queue expiry time for both remote control command queues, and remote control reply queues.

    • control_queue_ttl

      Set message time-to-live for both remote control command queues, and remote control reply queues.

    Contributed by Alan Justino.

  • The worker_shutdown signal is now always called during shutdown.

    Previously it would not be called if the worker instance was collected by gc first.

  • Worker now only starts the remote control command consumer if the broker transport used actually supports them.

  • Gossip now sets x-message-ttl for event queue to heartbeat_interval s. (Issue #2005).

  • Now preserves exit code (Issue #2024).

  • Now rejects messages with an invalid ETA value (instead of ack, which means they will be sent to the dead-letter exchange if one is configured).

  • Fixed crash when the -purge argument was used.

  • Log–level for unrecoverable errors changed from error to critical.

  • Improved rate limiting accuracy.

  • Account for missing timezone information in task expires field.

    Fix contributed by Albert Wang.

  • The worker no longer has a Queues bootsteps, as it is now

    superfluous.

  • Now emits the “Received task” line even for revoked tasks. (Issue #3155).

  • Now respects broker_connection_retry setting.

    Fix contributed by Nat Williams.

  • New control_queue_ttl and control_queue_expires settings now enables you to configure remote control command message TTLs, and queue expiry time.

    Contributed by Alan Justino.

  • New celery.worker.state.requests enables O(1) loookup of active/reserved tasks by id.

  • Auto-scale didn’t always update keep-alive when scaling down.

    Fix contributed by Philip Garnero.

  • Fixed typo options_list -> option_list.

    Fix contributed by Greg Wilbur.

  • Some worker command-line arguments and Worker() class arguments have been renamed for consistency.

    All of these have aliases for backward compatibility.

    • --send-events -> --task-events

    • --schedule -> --schedule-filename

    • --maxtasksperchild -> --max-tasks-per-child

    • Beat(scheduler_cls=) -> Beat(scheduler=)

    • Worker(send_events=True) -> Worker(task_events=True)

    • Worker(task_time_limit=) -> Worker(time_limit=)

    • Worker(task_soft_time_limit=) -> Worker(soft_time_limit=)

    • Worker(state_db=) -> Worker(statedb=)

    • Worker(working_directory=) -> Worker(workdir=)

Debugging Utilities

  • celery.contrib.rdb: Changed remote debugger banner so that you can copy and paste the address easily (no longer has a period in the address).

    Contributed by Jonathan Vanasco.

  • Fixed compatibility with recent psutil versions (Issue #3262).

Signals

  • App: New signals for app configuration/finalization:

  • Task: New task signals for rejected task messages:

    • celery.signals.task_rejected.

    • celery.signals.task_unknown.

  • Worker: New signal for when a heartbeat event is sent.

    • celery.signals.heartbeat_sent

      Contributed by Kevin Richardson.

Events

  • Event messages now uses the RabbitMQ x-message-ttl option to ensure older event messages are discarded.

    The default is 5 seconds, but can be changed using the event_queue_ttl setting.

  • Task.send_event now automatically retries sending the event on connection failure, according to the task publish retry settings.

  • Event monitors now sets the event_queue_expires setting by default.

    The queues will now expire after 60 seconds after the monitor stops consuming from it.

  • Fixed a bug where a None value wasn’t handled properly.

    Fix contributed by Dongweiming.

  • New event_queue_prefix setting can now be used to change the default celeryev queue prefix for event receiver queues.

    Contributed by Takeshi Kanemoto.

  • State.tasks_by_type and State.tasks_by_worker can now be used as a mapping for fast access to this information.

Deployment

  • Generic init-scripts now support CELERY_SU and CELERYD_SU_ARGS environment variables to set the path and arguments for su (su(1)).

  • Generic init-scripts now better support FreeBSD and other BSD systems by searching /usr/local/etc/ for the configuration file.

    Contributed by Taha Jahangir.

  • Generic init-script: Fixed strange bug for celerybeat where restart didn’t always work (Issue #3018).

  • The systemd init script now uses a shell when executing services.

    Contributed by Tomas Machalek.

Result Backends

  • Redis: Now has a default socket timeout of 120 seconds.

    The default can be changed using the new redis_socket_timeout setting.

    Contributed by Raghuram Srinivasan.

  • RPC Backend result queues are now auto delete by default (Issue #2001).

  • RPC Backend: Fixed problem where exception wasn’t deserialized properly with the json serializer (Issue #2518).

    Fix contributed by Allard Hoeve.

  • CouchDB: The backend used to double-json encode results.

    Fix contributed by Andrew Stewart.

  • CouchDB: Fixed typo causing the backend to not be found (Issue #3287).

    Fix contributed by Andrew Stewart.

  • MongoDB: Now supports setting the result_serialzier setting to bson to use the MongoDB libraries own serializer.

    Contributed by Davide Quarta.

  • MongoDB: URI handling has been improved to use

    database name, user and password from the URI if provided.

    Contributed by Samuel Jaillet.

  • SQLAlchemy result backend: Now ignores all result engine options when using NullPool (Issue #1930).

  • SQLAlchemy result backend: Now sets max char size to 155 to deal with brain damaged MySQL Unicode implementation (Issue #1748).

  • General: All Celery exceptions/warnings now inherit from common CeleryError/CeleryWarning. (Issue #2643).

Documentation Improvements

Contributed by:

  • Adam Chainz

  • Amir Rustamzadeh

  • Arthur Vuillard

  • Batiste Bieler

  • Berker Peksag

  • Bryce Groff

  • Daniel Devine

  • Edward Betts

  • Jason Veatch

  • Jeff Widman

  • Maciej Obuchowski

  • Manuel Kaufmann

  • Maxime Beauchemin

  • Mitchel Humpherys

  • Pavlo Kapyshin

  • Pierre Fersing

  • Rik

  • Steven Sklar

  • Tayfun Sen

  • Wieland Hoffmann

Reorganization, Deprecations, and Removals

Incompatible changes

  • Prefork: Calling result.get() or joining any result from within a task now raises RuntimeError.

    In previous versions this would emit a warning.

  • celery.worker.consumer is now a package, not a module.

  • Module celery.worker.job renamed to celery.worker.request.

  • Beat: Scheduler.Publisher/.publisher renamed to .Producer/.producer.

  • Result: The task_name argument/attribute of app.AsyncResult was removed.

    This was historically a field used for pickle compatibility, but is no longer needed.

  • Backends: Arguments named status renamed to state.

  • Backends: backend.get_status() renamed to backend.get_state().

  • Backends: backend.maybe_reraise() renamed to .maybe_throw()

    The promise API uses .throw(), so this change was made to make it more consistent.

    There’s an alias available, so you can still use maybe_reraise until Celery 5.0.

Unscheduled Removals

  • The experimental celery.contrib.methods feature has been removed, as there were far many bugs in the implementation to be useful.

  • The CentOS init-scripts have been removed.

    These didn’t really add any features over the generic init-scripts, so you’re encouraged to use them instead, or something like supervisor.

Reorganization Deprecations

These symbols have been renamed, and while there’s an alias available in this version for backward compatibility, they will be removed in Celery 5.0, so make sure you rename these ASAP to make sure it won’t break for that release.

Chances are that you’ll only use the first in this list, but you never know:

Scheduled Removals

Modules

  • Module celery.worker.job has been renamed to celery.worker.request.

    This was an internal module so shouldn’t have any effect. It’s now part of the public API so must not change again.

  • Module celery.task.trace has been renamed to celery.app.trace as the celery.task package is being phased out. The module will be removed in version 5.0 so please change any import from:

    from celery.task.trace import X

    to:

    from celery.app.trace import X

  • Old compatibility aliases in the celery.loaders module has been removed.

    • Removed celery.loaders.current_loader(), use: current_app.loader

    • Removed celery.loaders.load_settings(), use: current_app.conf

Result

  • AsyncResult.serializable() and celery.result.from_serializable

    has been removed:

    Use instead:

    >>> tup = result.as_tuple()
>>> from celery.result import result_from_tuple
>>> result = result_from_tuple(tup)

  • Removed BaseAsyncResult, use AsyncResult for instance checks instead.

  • Removed TaskSetResult, use GroupResult instead.

    • TaskSetResult.total -> len(GroupResult)

    • TaskSetResult.taskset_id -> GroupResult.id

  • Removed ResultSet.subtasks, use ResultSet.results instead.

TaskSet

TaskSet has been removed, as it was replaced by the group construct in Celery 3.0.

If you have code like this:

>>> from celery.task import TaskSet

>>> TaskSet(add.subtask((i, i)) for i in xrange(10)).apply_async()

You need to replace that with:

>>> from celery import group
>>> group(add.s(i, i) for i in xrange(10))()

Events

  • Removals for class celery.events.state.Worker:

    • Worker._defaults attribute.

      Use {k: getattr(worker, k) for k in worker._fields}.

    • Worker.update_heartbeat

      Use Worker.event(None, timestamp, received)

    • Worker.on_online

      Use Worker.event('online', timestamp, received, fields)

    • Worker.on_offline

      Use Worker.event('offline', timestamp, received, fields)

    • Worker.on_heartbeat

      Use Worker.event('heartbeat', timestamp, received, fields)

  • Removals for class celery.events.state.Task:

    • Task._defaults attribute.

      Use {k: getattr(task, k) for k in task._fields}.

    • Task.on_sent

      Use Worker.event('sent', timestamp, received, fields)

    • Task.on_received

      Use Task.event('received', timestamp, received, fields)

    • Task.on_started

      Use Task.event('started', timestamp, received, fields)

    • Task.on_failed

      Use Task.event('failed', timestamp, received, fields)

    • Task.on_retried

      Use Task.event('retried', timestamp, received, fields)

    • Task.on_succeeded

      Use Task.event('succeeded', timestamp, received, fields)

    • Task.on_revoked

      Use Task.event('revoked', timestamp, received, fields)

    • Task.on_unknown_event

      Use Task.event(short_type, timestamp, received, fields)

    • Task.update

      Use Task.event(short_type, timestamp, received, fields)

    • Task.merge

      Contact us if you need this.

Magic keyword arguments

Support for the very old magic keyword arguments accepted by tasks is finally removed in this version.

If you’re still using these you have to rewrite any task still using the old celery.decorators module and depending on keyword arguments being passed to the task, for example:

from celery.decorators import task

@task()
def add(x, y, task_id=None):
    print('My task id is %r' % (task_id,))

should be rewritten into:

from celery import task

@task(bind=True)
def add(self, x, y):
    print('My task id is {0.request.id}'.format(self))

Removed Settings

The following settings have been removed, and is no longer supported:

Logging Settings

Setting name

Replace with

CELERYD_LOG_LEVEL

celery worker --loglevel

CELERYD_LOG_FILE

celery worker --logfile

CELERYBEAT_LOG_LEVEL

celery beat --loglevel

CELERYBEAT_LOG_FILE

celery beat --logfile

CELERYMON_LOG_LEVEL

celerymon is deprecated, use flower

CELERYMON_LOG_FILE

celerymon is deprecated, use flower

CELERYMON_LOG_FORMAT

celerymon is deprecated, use flower

Task Settings

Setting name

Replace with

CELERY_CHORD_PROPAGATES

N/A

Changes to internal API

Deprecation Time-line Changes

See the Celery Deprecation Time-line.