What’s new in Celery 3.1 (Cipater)

Author:Ask Solem (ask at celeryproject.org)

Change history

What’s new documents describe the changes in major versions, we also have a Change history that lists the changes in bugfix releases (0.0.x), while older series are archived under the History section.

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 2.6, 2.7 and 3.3, and also supported on PyPy.

Table of Contents

Make sure you read the important notes before upgrading to this version.

Preface

Deadlocks have long plagued our workers, and while uncommon they are not acceptable. They are also infamous for being extremely hard to diagnose and reproduce, so to make this job easier I wrote a stress test suite that bombards the worker with different tasks in an attempt to break it.

What happens if thousands of worker child processes are killed every second? what if we also kill the broker connection every 10 seconds? These are examples of what the stress test suite will do to the worker, and it reruns these tests using different configuration combinations to find edge case bugs.

The end result was that I had to rewrite the prefork pool to avoid the use of the POSIX semaphore. This was extremely challenging, but after months of hard work the worker now finally passes the stress test suite.

There’s probably more bugs to find, but the good news is that we now have a tool to reproduce them, so should you be so unlucky to experience a bug then we’ll write a test for it and squash it!

Note that I have also moved many broker transports into experimental status: the only transports recommended for production use today is RabbitMQ and Redis.

I don’t have the resources to maintain all of them, so bugs are left unresolved. I wish that someone will step up and take responsibility for these transports or donate resources to improve them, but as the situation is now I don’t think the quality is up to date with the rest of the code-base so I cannot recommend them for production use.

The next version of Celery 3.2 will focus on performance and removing rarely used parts of the library. Work has also started on a new message protocol, supporting multiple languages and more. The initial draft can be found here.

This has probably been the hardest release I’ve worked on, so no introduction to this changelog would be complete without a massive thank you to everyone who contributed and helped me test it!

Thank you for your support!

— Ask Solem

Important Notes

Dropped support for Python 2.5

Celery now requires Python 2.6 or later.

The new dual code base runs on both Python 2 and 3, without requiring the 2to3 porting tool.

注解

This is also the last version to support Python 2.6! From Celery 3.2 and onwards Python 2.7 or later will be required.

Last version to enable Pickle by default

Starting from Celery 3.2 the default serializer will be json.

If you depend on pickle being accepted you should be prepared for this change by explicitly allowing your worker to consume pickled messages using the CELERY_ACCEPT_CONTENT setting:

  1. CELERY_ACCEPT_CONTENT = ['pickle', 'json', 'msgpack', 'yaml']

Make sure you only select the serialization formats you’ll actually be using, and make sure you have properly secured your broker from unwanted access (see the Security Guide).

The worker will emit a deprecation warning if you don’t define this setting.

for Kombu users

Kombu 3.0 no longer accepts pickled messages by default, so if you use Kombu directly then you have to configure your consumers: see the Kombu 3.0 Changelog for more information.

Old command-line programs removed and deprecated

Everyone should move to the new celery umbrella command, so we are incrementally deprecating the old command names.

In this version we’ve removed all commands that are not used in init scripts. The rest will be removed in 3.2.

ProgramNew StatusReplacement
celerydDEPRECATEDcelery worker
celerybeatDEPRECATEDcelery beat
celeryd-multiDEPRECATEDcelery multi
celeryctlREMOVEDcelery inspect|control
celeryevREMOVEDcelery events
camqadmREMOVEDcelery amqp

If this is not a new installation then you may want to remove the old commands:

  1. $ pip uninstall celery
  2. $ # repeat until it fails
  3. # ...
  4. $ pip uninstall celery
  5. $ pip install celery

Please run celery –help for help using the umbrella command.

News

Prefork Pool Improvements

These improvements are only active if you use an async capable transport. This means only RabbitMQ (AMQP) and Redis are supported at this point and other transports will still use the thread-based fallback implementation.

  • Pool is now using one IPC queue per child process.

    Previously the pool shared one queue between all child processes, using a POSIX semaphore as a mutex to achieve exclusive read and write access.

    The POSIX semaphore has now been removed and each child process gets a dedicated queue. This means that the worker will require more file descriptors (two descriptors per process), but it also means that performance is improved and we can send work to individual child processes.

    POSIX semaphores are not released when a process is killed, so killing processes could lead to a deadlock if it happened while the semaphore was acquired. There is no good solution to fix this, so the best option was to remove the semaphore.

  • Asynchronous write operations

    The pool now uses async I/O to send work to the child processes.

  • Lost process detection is now immediate.

    If a child process is killed or exits mysteriously the pool previously had to wait for 30 seconds before marking the task with a WorkerLostError. It had to do this because the outqueue was shared between all processes, and the pool could not be certain whether the process completed the task or not. So an arbitrary timeout of 30 seconds was chosen, as it was believed that the outqueue would have been drained by this point.

    This timeout is no longer necessary, and so the task can be marked as failed as soon as the pool gets the notification that the process exited.

  • Rare race conditions fixed

    Most of these bugs were never reported to us, but was discovered while running the new stress test suite.

Caveats

Long running tasks

The new pool will send tasks to a child process as long as the process inqueue is writable, and since the socket is buffered this means that the processes are, in effect, prefetching tasks.

This benefits performance but it also means that other tasks may be stuck waiting for a long running task to complete:

  1. -> send T1 to Process A
  2. # A executes T1
  3. -> send T2 to Process B
  4. # B executes T2
  5. <- T2 complete
  7. -> send T3 to Process A
  8. # A still executing T1, T3 stuck in local buffer and
  9. # will not start until T1 returns

The buffer size varies based on the operating system: some may have a buffer as small as 64kb but on recent Linux versions the buffer size is 1MB (can only be changed system wide).

You can disable this prefetching behavior by enabling the -Ofair worker option:

  1. $ celery -A proj worker -l info -Ofair

With this option enabled the worker will only write to workers that are available for work, disabling the prefetch behavior.

Max tasks per child

If a process exits and pool prefetch is enabled the worker may have already written many tasks to the process inqueue, and these tasks must then be moved back and rewritten to a new process.

This is very expensive if you have --maxtasksperchild set to a low value (e.g. less than 10), so if you need to enable this option you should also enable -Ofair to turn off the prefetching behavior.

Django supported out of the box

Celery 3.0 introduced a shiny new API, but sadly did not have a solution for Django users.

The situation changes with this version as Django is now supported in core and new Django users coming to Celery are now expected to use the new API directly.

The Django community has a convention where there’s a separate django-x package for every library, acting like a bridge between Django and the library.

Having a separate project for Django users has been a pain for Celery, with multiple issue trackers and multiple documentation sources, and then lastly since 3.0 we even had different APIs.

With this version we challenge that convention and Django users will use the same library, the same API and the same documentation as everyone else.

There is no rush to port your existing code to use the new API, but if you would like to experiment with it you should know that:

  • You need to use a Celery application instance.

    The new Celery API introduced in 3.0 requires users to instantiate the library by creating an application:

    1. from celery import Celery
    3. app = Celery()

  • You need to explicitly integrate Celery with Django

    Celery will not automatically use the Django settings, so you can either configure Celery separately or you can tell it to use the Django settings with:

    1. app.config_from_object('django.conf:settings')

    Neither will it automatically traverse your installed apps to find task modules, but this still available as an option you must enable:

    1. from django.conf import settings
    2. app.autodiscover_tasks(settings.INSTALLED_APPS)

  • You no longer use manage.py

    Instead you use the celery command directly:

    1. celery -A proj worker -l info

    For this to work your app module must store the DJANGO_SETTINGS_MODULE environment variable, see the example in the Django guide.

To get started with the new API you should first read the First Steps with Celery tutorial, and then you should read the Django specific instructions in First steps with Django.

The fixes and improvements applied by the django-celery library are now automatically applied by core Celery when it detects that the DJANGO_SETTINGS_MODULE environment variable is set.

The distribution ships with a new example project using Django in examples/django:

http://github.com/celery/celery/tree/3.1/examples/django

Some features still require the django-celery library:

  • Celery does not implement the Django database or cache result backends.

  • Celery does not ship with the database-based periodic task

    scheduler.

注解

If you’re still using the old API when you upgrade to Celery 3.1 then you must make sure that your settings module contains the djcelery.setup_loader() line, since this will no longer happen as a side-effect of importing the djcelery module.

New users (or if you have ported to the new API) don’t need the setup_loader line anymore, and must make sure to remove it.

Events are now ordered using logical time

Keeping physical clocks in perfect sync is impossible, so using timestamps to order events in a distributed system is not reliable.

Celery event messages have included a logical clock value for some time, but starting with this version that field is also used to order them.

Also, events now record timezone information by including a new utcoffset field in the event message. This is a signed integer telling the difference from UTC time in hours, so e.g. an even sent from the Europe/London timezone in daylight savings time will have an offset of 1.

celery.events.Receiver will automatically convert the timestamps to the local timezone.

注解

The logical clock is synchronized with other nodes in the same cluster (neighbors), so this means that the logical epoch will start at the point when the first worker in the cluster starts.

If all of the workers are shutdown the clock value will be lost and reset to 0, to protect against this you should specify a --statedb so that the worker can persist the clock value at shutdown.

You may notice that the logical clock is an integer value and increases very rapidly. Do not worry about the value overflowing though, as even in the most busy clusters it may take several millennia before the clock exceeds a 64 bits value.

New worker node name format (name@host)

Node names are now constructed by two elements: name and hostname separated by ‘@’.

This change was made to more easily identify multiple instances running on the same machine.

If a custom name is not specified then the worker will use the name ‘celery’ by default, resulting in a fully qualified node name of ‘celery@hostname‘:

  1. $ celery worker -n example.com
  2. [email protected]

To also set the name you must include the @:

  1. $ celery worker -n [email protected]
  2. [email protected]

The worker will identify itself using the fully qualified node name in events and broadcast messages, so where before a worker would identify itself as ‘worker1.example.com’, it will now use ‘celery@worker1.example.com‘.

Remember that the -n argument also supports simple variable substitutions, so if the current hostname is george.example.com then the %h macro will expand into that:

  1. $ celery worker -n worker1@%h
  2. [email protected]

The available substitutions are as follows:

VariableSubstitution
%hFull hostname (including domain name)
%dDomain name only
%nHostname only (without domain name)
%%The character %

Bound tasks

The task decorator can now create “bound tasks”, which means that the task will receive the self argument.

  1. @app.task(bind=True)
  2. def send_twitter_status(self, oauth, tweet):
  3.     try:
  4.         twitter = Twitter(oauth)
  5.         twitter.update_status(tweet)
  6.     except (Twitter.FailWhaleError, Twitter.LoginError) as exc:
  7.         raise self.retry(exc=exc)

Using bound tasks is now the recommended approach whenever you need access to the task instance or request context. Previously one would have to refer to the name of the task instead (send_twitter_status.retry), but this could lead to problems in some configurations.

Mingle: Worker synchronization

The worker will now attempt to synchronize with other workers in the same cluster.

Synchronized data currently includes revoked tasks and logical clock.

This only happens at startup and causes a one second startup delay to collect broadcast responses from other workers.

You can disable this bootstep using the --without-mingle argument.

Gossip: Worker <-> Worker communication

Workers are now passively subscribing to worker related events like heartbeats.

This means that a worker knows what other workers are doing and can detect if they go offline. Currently this is only used for clock synchronization, but there are many possibilities for future additions and you can write extensions that take advantage of this already.

Some ideas include consensus protocols, reroute task to best worker (based on resource usage or data locality) or restarting workers when they crash.

We believe that this is a small addition but one that really opens up for amazing possibilities.

You can disable this bootstep using the --without-gossip argument.

Bootsteps: Extending the worker

By writing bootsteps you can now easily extend the consumer part of the worker to add additional features, like custom message consumers.

The worker has been using bootsteps for some time, but these were never documented. In this version the consumer part of the worker has also been rewritten to use bootsteps and the new Extensions and Bootsteps guide documents examples extending the worker, including adding custom message consumers.

See the Extensions and Bootsteps guide for more information.

注解

Bootsteps written for older versions will not be compatible with this version, as the API has changed significantly.

The old API was experimental and internal but should you be so unlucky to use it then please contact the mailing-list and we will help you port the bootstep to the new API.

New RPC result backend

This new experimental version of the amqp result backend is a good alternative to use in classical RPC scenarios, where the process that initiates the task is always the process to retrieve the result.

It uses Kombu to send and retrieve results, and each client uses a unique queue for replies to be sent to. This avoids the significant overhead of the original amqp backend which creates one queue per task.

By default results sent using this backend will not persist, so they won’t survive a broker restart. You can enable the CELERY_RESULT_PERSISTENT setting to change that.

  1. CELERY_RESULT_BACKEND = 'rpc'
  2. CELERY_RESULT_PERSISTENT = True

Note that chords are currently not supported by the RPC backend.

Time limits can now be set by the client

Two new options have been added to the Calling API: time_limit and soft_time_limit:

  1. >>> res = add.apply_async((2, 2), time_limit=10, soft_time_limit=8)
  3. >>> res = add.subtask((2, 2), time_limit=10, soft_time_limit=8).delay()
  5. >>> res = add.s(2, 2).set(time_limit=10, soft_time_limit=8).delay()

Contributed by Mher Movsisyan.

Redis: Broadcast messages and virtual hosts

Broadcast messages are currently seen by all virtual hosts when using the Redis transport. You can now fix this by enabling a prefix to all channels so that the messages are separated:

  1. BROKER_TRANSPORT_OPTIONS = {'fanout_prefix': True}

Note that you’ll not be able to communicate with workers running older versions or workers that does not have this setting enabled.

This setting will be the default in a future version.

Related to Issue #1490.

pytz replaces python-dateutil dependency

Celery no longer depends on the python-dateutil library, but instead a new dependency on the pytz library was added.

The pytz library was already recommended for accurate timezone support.

This also means that dependencies are the same for both Python 2 and Python 3, and that the requirements/default-py3k.txt file has been removed.

Support for Setuptools extra requirements

Pip now supports the setuptools extra requirements format, so we have removed the old bundles concept, and instead specify setuptools extras.

You install extras by specifying them inside brackets:

  1. $ pip install celery[redis,mongodb]

The above will install the dependencies for Redis and MongoDB. You can list as many extras as you want.

警告

You can’t use the celery-with-* packages anymore, as these will not be updated to use Celery 3.1.

ExtensionRequirement entryType
Rediscelery[redis]transport, result backend
MongoDBcelery[mongodb]transport, result backend
CouchDBcelery[couchdb]transport
Beanstalkcelery[beanstalk]transport
ZeroMQcelery[zeromq]transport
Zookeepercelery[zookeeper]transport
SQLAlchemycelery[sqlalchemy]transport, result backend
librabbitmqcelery[librabbitmq]transport (C amqp client)

The complete list with examples is found in the 捆绑 section.

subtask.__call__() now executes the task directly

A misunderstanding led to Signature.__call__ being an alias of .delay but this does not conform to the calling API of Task which calls the underlying task method.

This means that:

  1. @app.task
  2. def add(x, y):
  3.     return x + y
  5. add.s(2, 2)()

now does the same as calling the task directly:

  1. add(2, 2)

In Other News

  • Now depends on Kombu 3.0.

  • Now depends on billiard version 3.3.

  • Worker will now crash if running as the root user with pickle enabled.

  • Canvas: group.apply_async and chain.apply_async no longer starts separate task.

    That the group and chord primitives supported the “calling API” like other subtasks was a nice idea, but it was useless in practice and often confused users. If you still want this behavior you can define a task to do it for you.

  • New method Signature.freeze() can be used to “finalize” signatures/subtask.

    Regular signature:

    1. >>> s = add.s(2, 2)
    2. >>> result = s.freeze()
    3. >>> result
    4. <AsyncResult: ffacf44b-f8a1-44e9-80a3-703150151ef2>
    5. >>> s.delay()
    6. <AsyncResult: ffacf44b-f8a1-44e9-80a3-703150151ef2>

    Group:

    1. >>> g = group(add.s(2, 2), add.s(4, 4))
    2. >>> result = g.freeze()
    3. <GroupResult: e1094b1d-08fc-4e14-838e-6d601b99da6d [
    4.     70c0fb3d-b60e-4b22-8df7-aa25b9abc86d,
    5.     58fcd260-2e32-4308-a2ea-f5be4a24f7f4]>
    6. >>> g()
    7. <GroupResult: e1094b1d-08fc-4e14-838e-6d601b99da6d [70c0fb3d-b60e-4b22-8df7-aa25b9abc86d, 58fcd260-2e32-4308-a2ea-f5be4a24f7f4]>

  • New ability to specify additional command line options to the worker and beat programs.

    The Celery.user_options attribute can be used to add additional command-line arguments, and expects optparse-style options:

    1. from celery import Celery
    2. from celery.bin import Option
    4. app = Celery()
    5. app.user_options['worker'].add(
    6.     Option('--my-argument'),
    7. )

    See the Extensions and Bootsteps guide for more information.

  • All events now include a pid field, which is the process id of the process that sent the event.

  • Event heartbeats are now calculated based on the time when the event was received by the monitor, and not the time reported by the worker.

    This means that a worker with an out-of-sync clock will no longer show as ‘Offline’ in monitors.

    A warning is now emitted if the difference between the senders time and the internal time is greater than 15 seconds, suggesting that the clocks are out of sync.

  • Monotonic clock support.

    A monotonic clock is now used for timeouts and scheduling.

    The monotonic clock function is built-in starting from Python 3.4, but we also have fallback implementations for Linux and OS X.

  • celery worker now supports a --detach argument to start the worker as a daemon in the background.

  • celery.events.Receiver now sets a local_received field for incoming events, which is set to the time of when the event was received.

  • celery.events.Dispatcher now accepts a groups argument which decides a white-list of event groups that will be sent.

    The type of an event is a string separated by ‘-‘, where the part before the first ‘-‘ is the group. Currently there are only two groups: worker and task.

    A dispatcher instantiated as follows:

    1. app.events.Dispatcher(connection, groups=['worker'])

    will only send worker related events and silently drop any attempts to send events related to any other group.

  • New BROKER_FAILOVER_STRATEGY setting.

    This setting can be used to change the transport failover strategy, can either be a callable returning an iterable or the name of a Kombu built-in failover strategy. Default is “round-robin”.

    Contributed by Matt Wise.

  • Result.revoke will no longer wait for replies.

    You can add the reply=True argument if you really want to wait for responses from the workers.

  • Better support for link and link_error tasks for chords.

    Contributed by Steeve Morin.

  • Worker: Now emits warning if the CELERYD_POOL setting is set to enable the eventlet/gevent pools.

    The -P option should always be used to select the eventlet/gevent pool to ensure that the patches are applied as early as possible.

    If you start the worker in a wrapper (like Django’s manage.py) then you must apply the patches manually, e.g. by creating an alternative wrapper that monkey patches at the start of the program before importing any other modules.

  • There’s a now an ‘inspect clock’ command which will collect the current logical clock value from workers.

  • celery inspect stats now contains the process id of the worker’s main process.

    Contributed by Mher Movsisyan.

  • New remote control command to dump a workers configuration.

    Example:

    1. $ celery inspect conf

    Configuration values will be converted to values supported by JSON where possible.

    Contributed by Mher Movisyan.

  • New settings CELERY_EVENT_QUEUE_TTL and CELERY_EVENT_QUEUE_EXPIRES.

    These control when a monitors event queue is deleted, and for how long events published to that queue will be visible. Only supported on RabbitMQ.

  • New Couchbase result backend.

    This result backend enables you to store and retrieve task results using Couchbase.

    See Couchbase backend settings for more information about configuring this result backend.

    Contributed by Alain Masiero.

  • CentOS init script now supports starting multiple worker instances.

    See the script header for details.

    Contributed by Jonathan Jordan.

  • AsyncResult.iter_native now sets default interval parameter to 0.5

    Fix contributed by Idan Kamara

  • New setting BROKER_LOGIN_METHOD.

    This setting can be used to specify an alternate login method for the AMQP transports.

    Contributed by Adrien Guinet

  • The dump_conf remote control command will now give the string representation for types that are not JSON compatible.

  • Function celery.security.setup_security is now setup_security().

  • Task retry now propagates the message expiry value (Issue #980).

    The value is forwarded at is, so the expiry time will not change. To update the expiry time you would have to pass a new expires argument to retry().

  • Worker now crashes if a channel error occurs.

    Channel errors are transport specific and is the list of exceptions returned by Connection.channel_errors. For RabbitMQ this means that Celery will crash if the equivalence checks for one of the queues in CELERY_QUEUES mismatches, which makes sense since this is a scenario where manual intervention is required.

  • Calling AsyncResult.get() on a chain now propagates errors for previous tasks (Issue #1014).

  • The parent attribute of AsyncResult is now reconstructed when using JSON serialization (Issue #1014).

  • Worker disconnection logs are now logged with severity warning instead of error.

    Contributed by Chris Adams.

  • events.State no longer crashes when it receives unknown event types.

  • SQLAlchemy Result Backend: New CELERY_RESULT_DB_TABLENAMES setting can be used to change the name of the database tables used.

    Contributed by Ryan Petrello.

  • SQLAlchemy Result Backend: Now calls enginge.dispose after fork

    (Issue #1564).

    If you create your own sqlalchemy engines then you must also make sure that these are closed after fork in the worker:

    1. from multiprocessing.util import register_after_fork
    3. engine = create_engine(…)
    4. register_after_fork(engine, engine.dispose)

  • A stress test suite for the Celery worker has been written.

    This is located in the funtests/stress directory in the git repository. There’s a README file there to get you started.

  • The logger named celery.concurrency has been renamed to celery.pool.

  • New command line utility celery graph.

    This utility creates graphs in GraphViz dot format.

    You can create graphs from the currently installed bootsteps:

    1. # Create graph of currently installed bootsteps in both the worker
    2. # and consumer namespaces.
    3. $ celery graph bootsteps | dot -T png -o steps.png
    5. # Graph of the consumer namespace only.
    6. $ celery graph bootsteps consumer | dot -T png -o consumer_only.png
    8. # Graph of the worker namespace only.
    9. $ celery graph bootsteps worker | dot -T png -o worker_only.png

    Or graphs of workers in a cluster:

    1. # Create graph from the current cluster
    2. $ celery graph workers | dot -T png -o workers.png
    4. # Create graph from a specified list of workers
    5. $ celery graph workers nodes:w1,w2,w3 | dot -T png workers.png
    7. # also specify the number of threads in each worker
    8. $ celery graph workers nodes:w1,w2,w3 threads:2,4,6
    10. # …also specify the broker and backend URLs shown in the graph
    11. $ celery graph workers broker:amqp:// backend:redis://
    13. # …also specify the max number of workers/threads shown (wmax/tmax),
    14. # enumerating anything that exceeds that number.
    15. $ celery graph workers wmax:10 tmax:3

  • Changed the way that app instances are pickled.

    Apps can now define a __reduce_keys__ method that is used instead of the old AppPickler attribute. E.g. if your app defines a custom ‘foo’ attribute that needs to be preserved when pickling you can define a __reduce_keys__ as such:

    1. import celery
    3. class Celery(celery.Celery):
    5.     def __init__(self, *args, **kwargs):
    6.         super(Celery, self).__init__(*args, **kwargs)
    7.         self.foo = kwargs.get('foo')
    9.     def __reduce_keys__(self):
    10.         return super(Celery, self).__reduce_keys__().update(
    11.             foo=self.foo,
    12.         )

    This is a much more convenient way to add support for pickling custom attributes. The old AppPickler is still supported but its use is discouraged and we would like to remove it in a future version.

  • Ability to trace imports for debugging purposes.

    The C_IMPDEBUG can be set to trace imports as they occur:

    1. $ C_IMDEBUG=1 celery worker -l info
    1. $ C_IMPDEBUG=1 celery shell

  • Message headers now available as part of the task request.

    Example adding and retrieving a header value:

    1. @app.task(bind=True)
    2. def t(self):
    3.     return self.request.headers.get('sender')
    5. >>> t.apply_async(headers={'sender': 'George Costanza'})

  • New before_task_publish signal dispatched before a task message is sent and can be used to modify the final message fields (Issue #1281).

  • New after_task_publish signal replaces the old task_sent signal.

    The task_sent signal is now deprecated and should not be used.

  • New worker_process_shutdown signal is dispatched in the prefork pool child processes as they exit.

    Contributed by Daniel M Taub.

  • celery.platforms.PIDFile renamed to celery.platforms.Pidfile.

  • MongoDB Backend: Can now be configured using an URL:

    See Example configuration.

  • MongoDB Backend: No longer using deprecated pymongo.Connection.

  • MongoDB Backend: Now disables auto_start_request.

  • MongoDB Backend: Now enables use_greenlets when eventlet/gevent is used.

  • subtask() / maybe_subtask() renamed to signature()/maybe_signature().

    Aliases still available for backwards compatibility.

  • The correlation_id message property is now automatically set to the id of the task.

  • The task message eta and expires fields now includes timezone information.

  • All result backends store_result/mark_as_* methods must now accept a request keyword argument.

  • Events now emit warning if the broken yajl library is used.

  • The celeryd_init signal now takes an extra keyword argument: option.

    This is the mapping of parsed command line arguments, and can be used to prepare new preload arguments (app.user_options[‘preload’]).

  • New callback: Celery.on_configure.

    This callback is called when an app is about to be configured (a configuration key is required).

  • Worker: No longer forks on HUP.

    This means that the worker will reuse the same pid for better support with external process supervisors.

    Contributed by Jameel Al-Aziz.

  • Worker: The log message Got task from broker … was changed to Received task ….

  • Worker: The log message Skipping revoked task … was changed to Discarding revoked task ….

  • Optimization: Improved performance of ResultSet.join_native().

    Contributed by Stas Rudakou.

  • The task_revoked signal now accepts new request argument (Issue #1555).

    The revoked signal is dispatched after the task request is removed from the stack, so it must instead use the Request object to get information about the task.

  • Worker: New -X command line argument to exclude queues (Issue #1399).

    The -X argument is the inverse of the -Q argument and accepts a list of queues to exclude (not consume from):

    1. # Consume from all queues in CELERY_QUEUES, but not the 'foo' queue.
    2. $ celery worker -A proj -l info -X foo

  • Adds C_FAKEFORK envvar for simple init script/multi debugging.

    This means that you can now do:

    1. $ C_FAKEFORK=1 celery multi start 10

    or:

    1. $ C_FAKEFORK=1 /etc/init.d/celeryd start

    to avoid the daemonization step to see errors that are not visible due to missing stdout/stderr.

    A dryrun command has been added to the generic init script that enables this option.

  • New public API to push and pop from the current task stack:

    celery.app.push_current_task() and celery.app.pop_current_task`().

  • RetryTaskError has been renamed to Retry.

    The old name is still available for backwards compatibility.

  • New semi-predicate exception Reject.

    This exception can be raised to reject/requeue the task message, see Reject for examples.

  • Semipredicates documented: (Retry/Ignore/Reject).

Scheduled Removals

  • The BROKER_INSIST setting and the insist argument to ~@connection is no longer supported.

  • The CELERY_AMQP_TASK_RESULT_CONNECTION_MAX setting is no longer supported.

    Use BROKER_POOL_LIMIT instead.

  • The CELERY_TASK_ERROR_WHITELIST setting is no longer supported.

    You should set the ErrorMail attribute of the task class instead. You can also do this using CELERY_ANNOTATIONS:

    1. from celery import Celery
    2. from celery.utils.mail import ErrorMail
    4. class MyErrorMail(ErrorMail):
    5.     whitelist = (KeyError, ImportError)
    7.     def should_send(self, context, exc):
    8.         return isinstance(exc, self.whitelist)
    10. app = Celery()
    11. app.conf.CELERY_ANNOTATIONS = {
    12.     '*': {
    13.         'ErrorMail': MyErrorMails,
    14.     }
    15. }

  • Functions that creates a broker connections no longer supports the connect_timeout argument.

    This can now only be set using the BROKER_CONNECTION_TIMEOUT setting. This is because functions no longer create connections directly, but instead get them from the connection pool.

  • The CELERY_AMQP_TASK_RESULT_EXPIRES setting is no longer supported.

    Use CELERY_TASK_RESULT_EXPIRES instead.

Deprecations

See the Celery Deprecation Timeline.

Fixes

  • AMQP Backend: join did not convert exceptions when using the json serializer.

  • Non-abstract task classes are now shared between apps (Issue #1150).

    Note that non-abstract task classes should not be used in the new API. You should only create custom task classes when you use them as a base class in the @task decorator.

    This fix ensure backwards compatibility with older Celery versions so that non-abstract task classes works even if a module is imported multiple times so that the app is also instantiated multiple times.

  • Worker: Workaround for Unicode errors in logs (Issue #427).

  • Task methods: .apply_async now works properly if args list is None (Issue #1459).

  • Eventlet/gevent/solo/threads pools now properly handles BaseException errors raised by tasks.

  • Autoscale and pool_grow/pool_shrink remote control commands will now also automatically increase and decrease the consumer prefetch count.

    Fix contributed by Daniel M. Taub.

  • celery control pool_ commands did not coerce string arguments to int.

  • Redis/Cache chords: Callback result is now set to failure if the group disappeared from the database (Issue #1094).

  • Worker: Now makes sure that the shutdown process is not initiated multiple times.

  • Multi: Now properly handles both -f and --logfile options (Issue #1541).

Internal changes