This document describes an older version of Celery (2.2). For the latest stable version please go here.


This module contains the component responsible for consuming messages from the broker, processing the messages and keeping the broker connections up and running.

  • start() is an infinite loop, which only iterates again if the connection is lost. For each iteration (at start, or if the connection is lost) it calls reset_connection(), and starts the consumer by calling consume_messages().

  • reset_connection(), clears the internal queues, establishes a new connection to the broker, sets up the task consumer (+ QoS), and the broadcast remote control command consumer.

    Also if events are enabled it configures the event dispatcher and starts up the hartbeat thread.

  • Finally it can consume messages. consume_messages() is simply an infinite loop waiting for events on the AMQP channels.

    Both the task consumer and the broadcast consumer uses the same callback: receive_message().

  • So for each message received the receive_message() method is called, this checks the payload of the message for either a task key or a control key.

    If the message is a task, it verifies the validity of the message converts it to a celery.worker.job.TaskRequest, and sends it to on_task().

    If the message is a control command the message is passed to on_control(), which in turn dispatches the control command using the control dispatcher.

    It also tries to handle malformed or invalid messages properly, so the worker doesn’t choke on them and die. Any invalid messages are acknowledged immediately and logged, so the message is not resent again, and again.

  • If the task has an ETA/countdown, the task is moved to the eta_schedule so the timer2.Timer can schedule it at its deadline. Tasks without an eta are moved immediately to the ready_queue, so they can be picked up by the Mediator to be sent to the pool.

  • When a task with an ETA is received the QoS prefetch count is also incremented, so another message can be reserved. When the ETA is met the prefetch count is decremented again, though this cannot happen immediately because amqplib doesn’t support doing broker requests across threads. Instead the current prefetch count is kept as a shared counter, so as soon as consume_messages() detects that the value has changed it will send out the actual QoS event to the broker.

  • Notice that when the connection is lost all internal queues are cleared because we can no longer ack the messages reserved in memory. Hoever, this is not dangerous as the broker will resend them to another worker when the channel is closed.

  • WARNING: stop() does not close the connection! This is because some pre-acked messages may be in processing, and they need to be finished before the channel is closed. For celeryd this means the pool must finish the tasks it has acked early, then close the connection.

class celery.worker.consumer.Consumer(ready_queue, eta_schedule, logger, init_callback=<function noop at 0x2aed398>, send_events=False, hostname=None, initial_prefetch_count=2, pool=None, app=None, priority_timer=None)

Listen for messages received from the broker and move them the the ready queue for task processing.


The queue that holds tasks ready for immediate processing.


Scheduler for paused tasks. Reasons for being paused include a countdown/eta or that it’s waiting for retry.


Is events enabled?


Callback to be called the first time the connection is active.


Current hostname. Defaults to the system hostname.


Initial QoS prefetch count for the task channel.


Control command dispatcher. See celery.worker.control.ControlDispatch.




Heart sending out heart beats if events enabled.


The logger used.


Consume messages forever (or until an exception is raised).

on_control(body, message)
on_decode_error(message, exc)

Callback called if the message had decoding errors.

  • message – The message with errors.
  • exc – The original exception instance.

Handle received task.

If the task has an eta we enter it into the ETA schedule, otherwise we move it the ready queue for immediate processing.

receive_message(body, message)

The callback called when a new message is received.


Re-establish connection and set up consumers.


Start the consumer.

If the connection is lost, it tries to re-establish the connection and restarts consuming messages.


Stop consuming.

Does not close connection.


Stop consuming.

celery.worker.consumer.PREFETCH_COUNT_MAX = 65535

Prefetch count can’t exceed short.

class celery.worker.consumer.QoS(consumer, initial_value, logger)

Quality of Service for Channel.

For thread-safe increment/decrement of a channels prefetch count value.

  • consumer – A kombu.messaging.Consumer instance.
  • initial_value – Initial prefetch count value.
  • logger – Logger used to log debug messages.

Decrement the current prefetch count value by one.


Decrement the value, but do not update the qos.

The MainThread will be responsible for calling update() when necessary.


Increment the current prefetch count value by one.

prev = None

Set channel prefetch_count setting.


Update prefetch count with current value.

Previous topic


Next topic


This Page