Production Deployment¶

It is time to deploy your DAG in production. To do this, first, you need to make sure that the Airflow is itself production-ready. Let’s see what precautions you need to take.

Database backend¶

Airflow comes with an SQLite backend by default. This allows the user to run Airflow without any external database. However, such a setup is meant to be used for testing purposes only; running the default setup in production can lead to data loss in multiple scenarios. If you want to run production-grade Airflow, make sure you configure the backend to be an external database such as PostgreSQL or MySQL.

You can change the backend using the following config

[database]
sql_alchemy_conn = my_conn_string

Once you have changed the backend, airflow needs to create all the tables required for operation. Create an empty DB and give Airflow’s user permission to CREATE/ALTER it. Once that is done, you can run -

airflow db migrate

migrate keeps track of migrations already applied, so it’s safe to run as often as you need.

Note

Prior to Airflow version 2.7.0, airflow db upgrade was used to apply migrations, however, it has been deprecated in favor of airflow db migrate.

Multi-Node Cluster¶

Airflow uses SequentialExecutor by default. However, by its nature, the user is limited to executing at most one task at a time. Sequential Executor also pauses the scheduler when it runs a task, hence it is not recommended in a production setup. You should use the LocalExecutor for a single machine. For a multi-node setup, you should use the Kubernetes executor or the Celery executor.

Once you have configured the executor, it is necessary to make sure that every node in the cluster contains the same configuration and DAGs. Airflow sends simple instructions such as “execute task X of DAG Y”, but does not send any DAG files or configuration. You can use a simple cronjob or any other mechanism to sync DAGs and configs across your nodes, e.g., checkout DAGs from git repo every 5 minutes on all nodes.

Logging¶

If you are using disposable nodes in your cluster, configure the log storage to be a distributed file system (DFS) such as S3 and GCS, or external services such as Stackdriver Logging, Elasticsearch or Amazon CloudWatch. This way, the logs are available even after the node goes down or gets replaced. See Logging for Tasks for configurations.

Note

The logs only appear in your DFS after the task has finished. You can view the logs while the task is running in UI itself.

Configuration¶

Airflow comes bundled with a default airflow.cfg configuration file. You should use environment variables for configurations that change across deployments e.g. metadata DB, password, etc. You can accomplish this using the format AIRFLOW__{SECTION}__{KEY}

AIRFLOW__DATABASE__SQL_ALCHEMY_CONN=my_conn_id
AIRFLOW__WEBSERVER__BASE_URL=http://host:port

Some configurations such as the Airflow Backend connection URI can be derived from bash commands as well:

sql_alchemy_conn_cmd = bash_command_to_run

Scheduler Uptime¶

Airflow users occasionally report instances of the scheduler hanging without a trace, for example in these issues:

To mitigate these issues, make sure you have a health check set up that will detect when your scheduler has not heartbeat in a while.

Production Container Images¶

We provide a Docker Image (OCI) for Apache Airflow for use in a containerized environment. Consider using it to guarantee that software will always run the same no matter where it’s deployed.

Helm Chart for Kubernetes¶

Helm provides a simple mechanism to deploy software to a Kubernetes cluster. We maintain an official Helm chart for Airflow that helps you define, install, and upgrade deployment. The Helm Chart uses our official Docker image and Dockerfile that is also maintained and released by the community.

Live-upgrading Airflow¶

Airflow is by-design a distributed system and while the basic Airflow deployment requires usually a complete Airflow restart to upgrade, it is possible to upgrade Airflow without any downtime when you run Airflow in a distributed deployment.

Such a live upgrade is possible when there are no changes in Airflow metadata database schema, so you should aim to do it when you upgrade Airflow patch-level (bugfix) versions of the same minor Airflow version or when upgrading between adjacent minor versions (feature) of Airflow after reviewing the release notes and Reference for Database Migrations and making sure there are no changes in the database schema between them.

In some cases when database migration is not significant, such live migration could also potentially be possible with upgrading Airflow database first and between MINOR versions, however, this is not recommended and you should only do it on your own risk, carefully reviewing the modifications to be applied to the database schema and assessing the risk of such upgrade - it requires deep knowledge of Airflow database ERD Schema of the Database and reviewing the Reference for Database Migrations. You should always thoroughly test such upgrade in a staging environment first. Usually cost connected with such live upgrade preparation will be higher than the cost of a short downtime of Airflow, so we strongly discourage such live upgrades.

Make sure to test such live upgrade procedure in a staging environment before you do it in production, to avoid any surprises and side-effects.

When it comes to live-upgrading the Webserver, Triggerer components, if you run them in separate environments and have more than one instances for each of them, you can rolling-restart them one by one, without any downtime. This should usually be done as the first step in your upgrade procedure.

When you are running a deployment with separate DAG processor, in a Separate DAG processing deployment the DAG processor is not horizontally scaled - even if you have more of them there is usually one DAG processor running at a time per specific folder, so you can just stop it and start the new one - but since the DAG processor is not a critical component, it’s ok for it to experience a short downtime.

When it comes to upgrading the schedulers and workers, you can use the live upgrade capabilities of the executor you use:

For the Local executor your tasks are running as subprocesses of scheduler and you cannot upgrade the Scheduler without killing the tasks run by it. You can either pause all your DAGs and wait for the running tasks to complete or just stop the scheduler and kill all the tasks it runs - then you will need to clear and restart those tasks manually after the upgrade is completed (or rely on retry being set for stopped tasks).
For the Celery executor, you have to first put your workers in offline mode (usually by setting a single TERM signal to the workers), wait until the workers finish all the running tasks, and then upgrade the code (for example by replacing the image the workers run in and restart the workers). You can monitor your workers via flower monitoring tool and see the number of running tasks going down to zero. Once the workers are upgraded, they will be automatically put in online mode and start picking up new tasks. You can then upgrade the Scheduler in a rolling restart mode.
For the Kubernetes executor, you can upgrade the scheduler triggerer, webserver in a rolling restart mode, and generally you should not worry about the workers, as they are managed by the Kubernetes cluster and will be automatically adopted by Schedulers when they are upgraded and restarted.
For the :doc:CeleryKubernetesExecutor <../core-concepts/executor/celery-kubernetes>, you follow the same procedure as for the CeleryExecutor - you put the workers in offline mode, wait for the running tasks to complete, upgrade the workers, and then upgrade the scheduler, triggerer and webserver in a rolling restart mode - which should also adopt tasks run via the KubernetesExecutor part of the executor.

Most of the rolling-restart upgrade scenarios are implemented in the Helm Chart for Apache Airflow, so you can use it to upgrade your Airflow deployment without any downtime - especially in case you do patch-level upgrades of Airflow.

Kerberos-authenticated workers¶

Apache Airflow has a built-in mechanism for authenticating the operation with a KDC (Key Distribution Center). Airflow has a separate command airflow kerberos that acts as token refresher. It uses the pre-configured Kerberos Keytab to authenticate in the KDC to obtain a valid token, and then refreshing valid token at regular intervals within the current token expiry window.

Each request for refresh uses a configured principal, and only keytab valid for the principal specified is capable of retrieving the authentication token.

The best practice to implement proper security mechanism in this case is to make sure that worker workloads have no access to the Keytab but only have access to the periodically refreshed, temporary authentication tokens. This can be achieved in Docker environment by running the airflow kerberos command and the worker command in separate containers - where only the airflow kerberos token has access to the Keytab file (preferably configured as secret resource). Those two containers should share a volume where the temporary token should be written by the airflow kerberos and read by the workers.

In the Kubernetes environment, this can be realized by the concept of sidecar, where both Kerberos token refresher and worker are part of the same Pod. Only the Kerberos sidecar has access to Keytab secret and both containers in the same Pod share the volume, where temporary token is written by the sidecar container and read by the worker container.

This concept is implemented in the Helm Chart for Apache Airflow.

Secured Server and Service Access on Google Cloud¶

This section describes techniques and solutions for securely accessing servers and services when your Airflow environment is deployed on Google Cloud, or you connect to Google services, or you are connecting to the Google API.

IAM and Service Accounts¶

You should not rely on internal network segmentation or firewalling as our primary security mechanisms. To protect your organization’s data, every request you make should contain sender identity. In the case of Google Cloud, the identity is provided by the IAM and Service account. Each Compute Engine instance has an associated service account identity. It provides cryptographic credentials that your workload can use to prove its identity when making calls to Google APIs or third-party services. Each instance has access only to short-lived credentials. If you use Google-managed service account keys, then the private key is always held in escrow and is never directly accessible.

If you are using Kubernetes Engine, you can use Workload Identity to assign an identity to individual pods.

For more information about service accounts in the Airflow, see Google Cloud Connection

Impersonate Service Accounts¶

If you need access to other service accounts, you can impersonate other service accounts to exchange the token with the default identity to another service account. Thus, the account keys are still managed by Google and cannot be read by your workload.

It is not recommended to generate service account keys and store them in the metadata database or the secrets backend. Even with the use of the backend secret, the service account key is available for your workload.

Access to Compute Engine Instance¶

If you want to establish an SSH connection to the Compute Engine instance, you must have the network address of this instance and credentials to access it. To simplify this task, you can use ComputeEngineHook instead of SSHHook

The ComputeEngineHook support authorization with Google OS Login service. It is an extremely robust way to manage Linux access properly as it stores short-lived ssh keys in the metadata service, offers PAM modules for access and sudo privilege checking and offers the nsswitch user lookup into the metadata service as well.

It also solves the discovery problem that arises as your infrastructure grows. You can use the instance name instead of the network address.

Access to Amazon Web Service¶

Thanks to the Web Identity Federation, you can exchange the Google Cloud Platform identity to the Amazon Web Service identity, which effectively means access to Amazon Web Service platform. For more information, see: Google Cloud to AWS authentication using Web Identity Federation