Scaling Out with 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, …) 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.

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 worker

Your worker should start picking up tasks as soon as they get fired in its direction.

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 airflow flower to start a Flower web server.

Please note that you must have the flower python library already installed on your system. The recommend 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

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