Celery Executor

Note

As of Airflow 2.7.0, you need to install the celery provider package to use this executor. This can be done by installing apache-airflow-providers-celery>=3.3.0 or by installing Airflow with the celery extra: pip install 'apache-airflow[celery]'.

CeleryExecutor is one of the ways you can scale out the number of workers. For this to work, you need to setup a Celery backend (RabbitMQ, Redis, Redis Sentinel …) and change your airflow.cfg to point the executor parameter to CeleryExecutor and provide the related Celery settings.

For more information about setting up a Celery broker, refer to the exhaustive Celery documentation on the topic.

The configuration parameters of the Celery Executor can be found in Configuration Reference.

Here are a few imperative requirements for your workers:

  • airflow needs to be installed, and the CLI needs to be in the path

  • Airflow configuration settings should be homogeneous across the cluster

  • Operators that are executed on the worker need to have their dependencies met in that context. For example, if you use the HiveOperator, the Hive CLI needs to be installed on that box, or if you use the MySqlOperator, the required Python library needs to be available in the PYTHONPATH somehow

  • The worker needs to have access to its DAGS_FOLDER, and you need to synchronize the filesystems by your own means. A common setup would be to store your DAGS_FOLDER in a Git repository and sync it across machines using Chef, Puppet, Ansible, or whatever you use to configure machines in your environment. If all your boxes have a common mount point, having your pipelines files shared there should work as well

To kick off a worker, you need to setup Airflow and kick off the worker subcommand

airflow celery worker

Your worker should start picking up tasks as soon as they get fired in its direction. To stop a worker running on a machine you can use:

airflow celery stop

It will try to stop the worker gracefully by sending SIGTERM signal to main Celery process as recommended by Celery documentation.

Note that you can also run Celery Flower, a web UI built on top of Celery, to monitor your workers. You can use the shortcut command to start a Flower web server:

airflow celery flower

Please note that you must have the flower python library already installed on your system. The recommended way is to install the airflow celery bundle.

pip install 'apache-airflow[celery]'

Some caveats:

  • Make sure to use a database backed result backend

  • Make sure to set a visibility timeout in [celery_broker_transport_options] that exceeds the ETA of your longest running task

  • Make sure to specify the password for Redis Server in the [celery_broker_transport_options] section if you are using Redis Sentinel as your broker and the Redis servers are password-protected

  • Make sure to set umask in [worker_umask] to set permissions for newly created files by workers.

  • Tasks can consume resources. Make sure your worker has enough resources to run worker_concurrency tasks

  • Queue names are limited to 256 characters, but each broker backend might have its own restrictions

See Modules Management for details on how Python and Airflow manage modules.

Architecture

digraph A{ rankdir="TB" node[shape="rectangle", style="rounded"] subgraph cluster { label="Cluster"; {rank = same; dag; database} {rank = same; workers; scheduler; web} workers[label="Workers"] scheduler[label="Scheduler"] web[label="Web server"] database[label="Database"] dag[label="DAG files"] subgraph cluster_queue { label="Celery"; {rank = same; queue_broker; queue_result_backend} queue_broker[label="Queue broker"] queue_result_backend[label="Result backend"] } web->workers[label="1"] web->dag[label="2"] web->database[label="3"] workers->dag[label="4"] workers->database[label="5"] workers->queue_result_backend[label="6"] workers->queue_broker[label="7"] scheduler->dag[label="8"] scheduler->database[label="9"] scheduler->queue_result_backend[label="10"] scheduler->queue_broker[label="11"] } }

Airflow consist of several components:

  • Workers - Execute the assigned tasks

  • Scheduler - Responsible for adding the necessary tasks to the queue

  • Web server - HTTP Server provides access to DAG/task status information

  • Database - Contains information about the status of tasks, DAGs, Variables, connections, etc.

  • Celery - Queue mechanism

Please note that the queue at Celery consists of two components:

  • Broker - Stores commands for execution

  • Result backend - Stores status of completed commands

The components communicate with each other in many places

  • [1] Web server –> Workers - Fetches task execution logs

  • [2] Web server –> DAG files - Reveal the DAG structure

  • [3] Web server –> Database - Fetch the status of the tasks

  • [4] Workers –> DAG files - Reveal the DAG structure and execute the tasks

  • [5] Workers –> Database - Gets and stores information about connection configuration, variables and XCOM.

  • [6] Workers –> Celery’s result backend - Saves the status of tasks

  • [7] Workers –> Celery’s broker - Stores commands for execution

  • [8] Scheduler –> DAG files - Reveal the DAG structure and execute the tasks

  • [9] Scheduler –> Database - Store a DAG run and related tasks

  • [10] Scheduler –> Celery’s result backend - Gets information about the status of completed tasks

  • [11] Scheduler –> Celery’s broker - Put the commands to be executed

Task execution process

../../_images/run_task_on_celery_executor.png

Sequence diagram - task execution process

Initially, two processes are running:

  • SchedulerProcess - process the tasks and run using CeleryExecutor

  • WorkerProcess - observes the queue waiting for new tasks to appear

  • WorkerChildProcess - waits for new tasks

Two databases are also available:

  • QueueBroker

  • ResultBackend

During this process, two 2 process are created:

  • LocalTaskJobProcess - It logic is described by LocalTaskJob. It is monitoring RawTaskProcess. New processes are started using TaskRunner.

  • RawTaskProcess - It is process with the user code e.g. execute().

[1] SchedulerProcess processes the tasks and when it finds a task that needs to be done, sends it to the QueueBroker.
[2] SchedulerProcess also begins to periodically query ResultBackend for the status of the task.
[3] QueueBroker, when it becomes aware of the task, sends information about it to one WorkerProcess.
[4] WorkerProcess assigns a single task to a one WorkerChildProcess.
[5] WorkerChildProcess performs the proper task handling functions - execute_command(). It creates a new process - LocalTaskJobProcess.
[6] LocalTaskJobProcess logic is described by LocalTaskJob class. It starts new process using TaskRunner.
[7][8] Process RawTaskProcess and LocalTaskJobProcess is stopped when they have finished their work.
[10][12] WorkerChildProcess notifies the main process - WorkerProcess about the end of the task and the availability of subsequent tasks.
[11] WorkerProcess saves status information in ResultBackend.
[13] When SchedulerProcess asks ResultBackend again about the status, it will get information about the status of the task.

Queues

When using the CeleryExecutor, the Celery queues that tasks are sent to can be specified. queue is an attribute of BaseOperator, so any task can be assigned to any queue. The default queue for the environment is defined in the airflow.cfg’s operators -> default_queue. This defines the queue that tasks get assigned to when not specified, as well as which queue Airflow workers listen to when started.

Workers can listen to one or multiple queues of tasks. When a worker is started (using command airflow celery worker), a set of comma-delimited queue names (with no whitespace) can be given (e.g. airflow celery worker -q spark,quark). This worker will then only pick up tasks wired to the specified queue(s).

This can be useful if you need specialized workers, either from a resource perspective (for say very lightweight tasks where one worker could take thousands of tasks without a problem), or from an environment perspective (you want a worker running from within the Spark cluster itself because it needs a very specific environment and security rights).

Was this entry helpful?