Production Guide¶

The following are things to consider when using this Helm chart in a production environment.

Database¶

It is advised to set up an external database for the Airflow metastore. The default Helm chart deploys a Postgres database running in a container. For production usage, a database running on a dedicated machine or leveraging a cloud provider’s database service such as AWS RDS, should be used. Embedded Postgres lacks stability, monitoring and persistence features that you need for a production database. It is only there to make it easier to test the Helm Chart in a “standalone” version, but you might experience data loss when you are using it. Supported databases and versions can be found at Set up a Database Backend.

Note

When using the helm chart, you do not need to initialize the db with airflow db migrate as outlined in Set up a Database Backend.

To disable deployment of Postgres pod, set below values in your values.yaml file:

values.yaml¶

postgresql:
  enabled: false

To provide the database credentials to Airflow, you have 2 options - in your values file or in a Kubernetes Secret.

Values file¶

This is the simpler options, as the chart will create a Kubernetes Secret for you. However, keep in mind your credentials will be in your values file.

values.yaml¶

data:
  metadataConnection:
    user: <username>
    pass: <password>
    protocol: postgresql
    host: <hostname>
    port: 5432
    db: <database name>

Warning

Due to security concerns, it is not advised to store Airflow database user credentials directly in the values.yaml file.

Kubernetes Secret¶

You can store the credentials in a Kubernetes Secret (it requires manual creation).

Note

Any special character in the username/password must be URL encoded.

kubectl create secret generic mydatabase --from-literal=connection=postgresql://user:pass@host:5432/db

After secret creation, configure the chart to use the secret:

values.yaml¶

data:
  metadataSecretName: mydatabase

Metadata DB Cleanup¶

It is recommended to periodically clean up the Airflow metadata database to remove old records and keep the database size manageable. A Kubernetes CronJob can be enabled for this purpose:

values.yaml¶

databaseCleanup:
  enabled: true
  retentionDays: 90

For details regarding the airflow db clean command, see db clean usage and for additional options which can be configured via helm chart values, see parameters reference.

PgBouncer¶

If you are using PostgreSQL as your database, you will likely want to enable PgBouncer as well. Due to distributed nature of Airflow, it can open a lot of database connections. Using a connection pooler can significantly reduce the number of open connections on the database.

Database credentials stored Values file¶

values.yaml¶

pgbouncer:
  enabled: true

Database credentials stored Kubernetes Secret¶

The default connection string in this case will not work. You need to modify accordingly the Kubernetes secret:

kubectl create secret generic mydatabase --from-literal=connection=postgresql://user:pass@pgbouncer_svc_name.deployment_namespace:6543/airflow-metadata

Furthermore, two additional Kubernetes Secret are required for PgBouncer to be able to properly work:

airflow-pgbouncer-stats secret:

kubectl create secret generic airflow-pgbouncer-stats --from-literal=connection=postgresql://user:pass@127.0.0.1:6543/pgbouncer?sslmode=disable

airflow-pgbouncer-config secret:

airflow-pgbouncer-config¶

apiVersion: v1
kind: Secret
metadata:
  name: airflow-pgbouncer-config
data:
  pgbouncer.ini: dmFsdWUtMg0KDQo=
  users.txt: dmFsdWUtMg0KDQo=

where:

pgbouncer.ini value is equal to the base64 encoded version of below text:

pgbouncer.ini¶

[databases]
airflow-metadata = host={external_database_host} dbname={external_database_dbname} port=5432 pool_size=10

[pgbouncer]
pool_mode = transaction
listen_port = 6543
listen_addr = *
auth_type = scram-sha-256
auth_file = /etc/pgbouncer/users.txt
stats_users = postgres
ignore_startup_parameters = extra_float_digits
max_client_conn = 100
verbose = 0
log_disconnections = 0
log_connections = 0

server_tls_sslmode = prefer
server_tls_ciphers = normal

users.txt value is equal to the base64 encoded version of below text:
users.txt¶
```
"{ external_database_username }" "{ external_database_pass }"
```

In the values.yaml below secret-related parameters should be adjusted like:

values.yaml¶

pgbouncer:
  enabled: true
  configSecretName: airflow-pgbouncer-config
  metricsExporterSidecar:
    statsSecretName: airflow-pgbouncer-stats

Note

Depending on the size of your Airflow instance, you may want to adjust the following as well (defaults are shown):

values.yaml¶

pgbouncer:
  # The maximum number of connections to PgBouncer
  maxClientConn: 100
  # The maximum number of server connections to the metadata database from PgBouncer
  metadataPoolSize: 10
  # The maximum number of server connections to the result backend database from PgBouncer
  resultBackendPoolSize: 5

API Secret Key¶

You should set a static API secret key when deploying with Airflow chart as it will help ensure your Airflow components only restart when necessary.

Note

This section also applies to the webserver for Airflow 2 (simply replace api with webserver).

Warning

You should use a different secret key for every instance you run, as this key is used to sign session cookies and perform other security related functions.

Follow below steps to create static API secret key:

Generate a strong secret key:

python3 -c 'import secrets; print(secrets.token_hex(16))'

Add the secret to your values file:
values.yaml¶
```
apiSecretKey: <secret_key>
```
or create a Kubernetes Secret and use apiSecretKeySecretName:
values.yaml¶
```
apiSecretKeySecretName: my-api-secret
# Where the random key is under `webserver-secret-key` in the k8s Secret
```
Warning

Due to security concerns, it is advised to use Kubernetes Secret instead of setting API secret key directly in the values file.

Example to create a Kubernetes Secret from kubectl:

kubectl create secret generic my-api-secret --from-literal="api-secret-key=$(python3 -c 'import secrets; print(secrets.token_hex(16))')"

The API secret key is also used to authorize requests to Celery workers when logs are retrieved. The token generated using the secret key has a short expiry time though. Make sure that time on ALL the machines that you run Airflow components on is synchronized (for example using ntpd). You might get “forbidden” errors when the logs are accessed otherwise.

Eviction configuration¶

When running Airflow along with the Kubernetes Cluster Autoscaler, it is important to configure whether pods can be safely evicted. This setting can be configured in the Airflow chart at different levels:

values.yaml¶

workers:
  safeToEvict: true
scheduler:
  safeToEvict: true
apiServer:
  safeToEvict: true

workers.safeToEvict defaults to false, and when using KubernetesExecutor workers.safeToEvict shouldn’t be set to true as the workers may be removed before finishing.

Extending and customizing Airflow Image¶

The Apache Airflow community, releases Docker Images which are reference images for Apache Airflow. However, Airflow has more than 60 community managed providers (installable via extras) and some of the default extras/providers installed are not used by everyone. Sometimes other extras/providers are needed, sometimes (very often actually) you need to add your own custom dependencies, packages or even custom providers, or add custom tools and binaries that are needed in your deployment.

In Kubernetes and Docker terms, this means that you need another image with your specific requirements. This is why you should learn how to build your own Docker (or more properly Container) image.

Typical scenarios where you would like to use your custom image are adding:

apt packages,
PyPI packages,
binary resources necessary for your deployment,
custom tools needed in your deployment.

See Extending Airflow Image and/or Building the image for more details on how you can extend, customize and test the modifications of Airflow image.

Managing Dag Files¶

See Manage Dag files.

knownHosts¶

If you are using dags.gitSync.sshKeySecret, you should also set dags.gitSync.knownHosts. Here we will show the process for GitHub, but the same can be done for any provider:

Grab GitHub’s public key:

ssh-keyscan -t rsa github.com > github_public_key

Print the fingerprint for the public key:
```
ssh-keygen -lf github_public_key
```
Compare that output with GitHub’s SSH key fingerprints.
If values are the same, add the public key to your values. It’ll look something like this:
values.yaml¶
```
dags:
  gitSync:
    knownHosts: |
      github.com ssh-rsa AAAA...1/wsjk=
```

External Scheduler¶

To use an external Scheduler instance:

values.yaml¶

scheduler:
  enabled: false

Ensure that your external scheduler is connected to the same redis host as workers.

Accessing the Airflow UI¶

How you access the Airflow UI will depend on your environment; however, the chart does support various options.

External API Server¶

To use an external API Server:

values.yaml¶

apiServer:
  enabled: false

Ingress¶

You can create and configure Ingress objects. See the Ingress chart parameters. For more information on Ingress, see the Kubernetes Ingress documentation.

LoadBalancer Service¶

You can change the Service type for the API Server to be LoadBalancer, and set any necessary annotations:

values.yaml¶

apiServer:
  service:
    type: LoadBalancer

For more information on LoadBalancer Services, see the Kubernetes LoadBalancer Service Documentation.

Logging¶

Depending on your choice of executor, task logs may not work out of the box. All logging choices can be found at Manage logs.

Metrics¶

The chart supports sending metrics to an existing StatsD instance or provide a Prometheus endpoint.

Prometheus Endpoint¶

The metrics endpoint is available at svc/{{ .Release.Name }}-statsd:9102/metrics.

External StatsD¶

To use an external StatsD instance:

values.yaml¶

statsd:
  enabled: false
config:
  metrics:
    statsd_on: true
    statsd_host: ...
    statsd_port: ...

IPv6 StatsD¶

To use an StatsD instance with IPv6 address. Example with Kubernetes with IPv6 enabled:

values.yaml¶

statsd:
  enabled: true
config:
  metrics:
    statsd_on: 'True'
    statsd_host: ...
    statsd_ipv6: 'True'
    statsd_port: ...
    statsd_prefix: airflow

Datadog¶

If you are using a Datadog agent in your environment, this will enable Airflow to export metrics to the Datadog agent.

values.yaml¶

statsd:
  enabled: false
config:
  metrics:
    statsd_on: true
    statsd_port: 8125
extraEnv: |-
  - name: AIRFLOW__METRICS__STATSD_HOST
    valueFrom:
      fieldRef:
        fieldPath: status.hostIP

Celery Backend¶

If you are using CeleryExecutor or CeleryKubernetesExecutor, you can bring your own Celery backend.

By default, the chart will deploy Redis. However, you can use any supported Celery backend instead:

values.yaml¶

redis:
  enabled: false
data:
  brokerUrl: redis://redis-user:password@redis-host:6379/0

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

Security Context¶

Constraints¶

A Security Context Constraint (SCC) is a OpenShift construct that works as a RBAC rule. However, it targets Pods instead of users. When defining a SCC, one can control actions and resources a POD can perform or access during startup and runtime.

The SCCs are split into different levels or categories with the restricted SCC being the default one assigned to Pods. When deploying Airflow to OpenShift, one can leverage the SCCs and allow the Pods to start containers utilizing the anyuid SCC.

In order to enable the usage of SCCs, one must set the parameter rbac.createSCCRoleBinding to true as shown below:

values.yaml¶

rbac:
  create: true
  createSCCRoleBinding: true

In this chart, SCCs are bound to the Pods via RoleBindings meaning that the option rbac.create must also be set to true in order to fully enable the SCC usage.

For more information about SCCs and what can be achieved with this construct, please refer to Managing security context constraints.

Configuration¶

In Kubernetes a securityContext can be used to define user ids, group ids and capabilities such as running a container in privileged mode.

When deploying an application to Kubernetes, it is recommended to give the least privilege to containers to reduce access and protect the host where the container is running.

In the Airflow Helm chart, the securityContext can be configured in several ways:

uid - configures the global uid or RunAsUser
gid - configures the global gid or fsGroup
securityContexts - same as uid, but allows for setting all Pod securityContext options and Container securityContext options

The same way one can configure the global securityContexts. It is also possible to configure different values for specific workloads by setting their local securityContexts as follows:

values.yaml¶

scheduler:
  securityContexts:
    pod:
      runAsUser: 5000
      fsGroup: 0
    containers:
      allowPrivilegeEscalation: false

In the example above, the scheduler pod securityContext will be set to runAsUser: 5000 and fsGroup: 0. The scheduler container securityContext will be set to allowPrivilegeEscalation: false.

As one can see, the local setting will take precedence over the global setting when defined. The following explains the precedence rule for securityContexts options in this chart:

values.yaml¶

uid: 40000
gid: 0

securityContexts:
  pod:
    runAsUser: 50000
    fsGroup: 0

scheduler:
  securityContexts:
    pod:
      runAsUser: 1001
      fsGroup: 0

This will generate the following scheduler deployment:

airflow-scheduler¶

kind: Deployment
apiVersion: apps/v1
metadata:
  name: airflow-scheduler
spec:
  template:
    spec:
      securityContext:    # As the securityContexts was defined in ``scheduler``, its value will take priority
        runAsUser: 1001
        fsGroup: 0

If we remove both the securityContexts and scheduler.securityContexts from the example above:

values.yaml¶

uid: 40000
gid: 0

securityContexts: {}

scheduler:
  securityContexts: {}

it will generate the following scheduler deployment:

airflow-scheduler¶

kind: Deployment
apiVersion: apps/v1
metadata:
  name: airflow-scheduler
spec:
  template:
    spec:
      securityContext:
        runAsUser: 40000   # As the securityContext was not defined in ``scheduler`` or ``podSecurity``, the value from uid will be used
        fsGroup: 0         # As the securityContext was not defined in ``scheduler`` or ``podSecurity``, the value from gid will be used
      initContainers:
        - name: wait-for-airflow-migrations
      ...
      containers:
        - name: scheduler
      ...

And finally if we set securityContexts, but not scheduler.securityContexts:

values.yaml¶

uid: 40000
gid: 0

securityContexts:
  pod:
    runAsUser: 50000
    fsGroup: 0

scheduler:
  securityContexts: {}

This will generate the following scheduler deployment:

airflow-scheduler¶

kind: Deployment
apiVersion: apps/v1
metadata:
  name: airflow-scheduler
spec:
  template:
    spec:
      securityContext:     # As the securityContexts was not defined in ``scheduler``, the values from securityContexts will take priority
        runAsUser: 50000
        fsGroup: 0
      initContainers:
        - name: wait-for-airflow-migrations
      ...
      containers:
        - name: scheduler
      ...

Built-in secrets and environment variables¶

The Helm Chart by default uses Kubernetes Secrets to store secrets that are needed by Airflow. The contents of those secrets are by default turned into environment variables that are read by Airflow.

Note

Some of the environment variables have several variants to support older versions of Airflow.

By default, the secret names are determined from the Release Name used when the Helm Chart, but you can also use a different secret to set the variables or disable using secrets entirely and rely on environment variables (specifically if you want to use _CMD or __SECRET variant of the environment variable).

However, Airflow supports other variants of setting secret configuration. You can specify a system command to retrieve and automatically rotate the secret (by defining variable with _CMD suffix) or to retrieve a variable from secret backed (by defining the variable with _SECRET suffix).

If the <VARIABLE_NAME> is set, it takes precedence over the _CMD and _SECRET variant, so if you want to set one of the _CMD or _SECRET variants, you must disable the built in variables retrieved from Kubernetes secrets, by setting .Values.enableBuiltInSecretEnvVars.<VARIABLE_NAME> to false.

For example in order to use a command to retrieve the DB connection, you should (in your values.yaml file) specify:

values.yaml¶

extraEnv:
  AIRFLOW_CONN_AIRFLOW_DB_CMD: "/usr/local/bin/retrieve_connection_url"
enableBuiltInSecretEnvVars:
  AIRFLOW_CONN_AIRFLOW_DB: false

Here is the full list of secrets that can be disabled and replaced by _CMD and _SECRET variants:

Default secret name if secret name not specified	Use a different Kubernetes Secret	Airflow Environment Variable
`<RELEASE_NAME>-airflow-metadata`	`.Values.data.metadataSecretName`	`AIRFLOW_CONN_AIRFLOW_DB` `AIRFLOW__DATABASE__SQL_ALCHEMY_CONN`
`<RELEASE_NAME>-fernet-key`	`.Values.fernetKeySecretName`	`AIRFLOW__CORE__FERNET_KEY`
`<RELEASE_NAME>-api-secret-key`	`.Values.apiSecretKeySecretName`	`AIRFLOW__API__SECRET_KEY`
`<RELEASE_NAME>-jwt-secret`	`.Values.jwtSecretName`	`AIRFLOW__API_AUTH__JWT_SECRET`
`<RELEASE_NAME>-webserver-secret-key`	`.Values.webserverSecretKeySecretName`	`AIRFLOW__WEBSERVER__SECRET_KEY`
`<RELEASE_NAME>-airflow-result-backend`	`.Values.data.resultBackendSecretName`	`AIRFLOW__CELERY__RESULT_BACKEND`
`<RELEASE_NAME>-airflow-broker-url`	`.Values.data.brokerUrlSecretName`	`AIRFLOW__CELERY__BROKER_URL`
`<RELEASE_NAME>-elasticsearch`	`.Values.elasticsearch.secretName`	`AIRFLOW__ELASTICSEARCH__HOST`

There are also a number of secrets, which names are also determined from the release name, that do not need to be disabled. This is because either they do not follow the _CMD or _SECRET pattern, are variables which do not start with AIRFLOW__, or they do not have a corresponding variable.

There is also AIRFLOW__CELERY__FLOWER_BASIC_AUTH, that does not need to be disabled, even if you want set the _CMD and _SECRET variant. This variable is not set by default. It is only set when .Values.flower.secretName is set or when .Values.flower.user and .Values.flower.password are set. If you do not set any of the .Values.flower.* variables, you can freely configure flower Basic Auth using the _CMD or _SECRET variant without disabling the basic variant.

Default secret name if secret name not specified	Use a different Kubernetes Secret	Airflow Environment Variable
`<RELEASE_NAME>-redis-password`	`.Values.redis.passwordSecretName`	`REDIS_PASSWORD`
`<RELEASE_NAME>-pgbouncer-config`	`.Values.pgbouncer.configSecretName`
`<RELEASE_NAME>-pgbouncer-certificates`
`<RELEASE_NAME>-kerberos-keytab`
`<RELEASE_NAME>-flower`	`.Values.flower.secretName`	`AIRFLOW__CELERY__FLOWER_BASIC_AUTH`

A secret named <RELEASE_NAME>-registry is also created when .Values.registry.connection is defined and neither .Values.registry.secretName nor .Values.imagePullSecrets is set. However, this behavior is deprecated in favor of explicitly defining .Values.imagePullSecrets.

You can read more about advanced ways of setting configuration variables in the Setting Configuration Options.

Service Account Token Volume Configuration¶

When using pod-launching executors (CeleryExecutor, CeleryKubernetesExecutor, KubernetesExecutor, LocalKubernetesExecutor), you can configure how Kubernetes service account tokens are mounted into pods. This provides enhanced security control and compatibility with security policies like Kyverno.

Background¶

By default, Kubernetes automatically mounts service account tokens into pods via the automountServiceAccountToken setting. However, for security reasons, you might want to disable automatic mounting and manually configure service account token volumes instead.

This feature addresses Bug #59099 where scheduler.serviceAccount.automountServiceAccountToken: false was ignored when using the KubernetesExecutor. The solution implements a defense-in-depth approach with both ServiceAccount-level and Pod-level controls.

Container-Specific Security¶

The Service Account Token Volume is mounted only in containers that require Kubernetes API access, implementing the Principle of Least Privilege:

Scheduler Container: Receives Service Account Token as it needs API access for pod management
Init Container “wait-for-airflow-migrations”: No Service Account Token as it only performs database migrations
Sidecar Container “scheduler-log-groomer”: No Service Account Token as it only performs log cleanup operations

This container-specific approach ensures that:

Database Migration Container: Only accesses the database for schema updates as no Kubernetes API access required
Log Groomer Container: Only performs filesystem operations for log cleanup as no API access required
Scheduler Container: Requires API access for launching and managing pods with pod-launching executors

Security Benefits:

Explicit control: Manual configuration makes token mounting intentional and visible
Policy compliance: Compatible with security policies that restrict automountServiceAccountToken: true
Defense-in-depth: Provides both ServiceAccount-level and Pod-level security controls
Custom expiration: Allows setting shorter token lifetimes for enhanced security
Container isolation: Only scheduler container receives API access, reducing attack surface
Principle of Least Privilege: Each container receives only the minimum required permissions

Configuration Options¶

The service account token volume configuration is available for the scheduler component and includes the following options:

values.yaml¶

scheduler:
  serviceAccount:
    automountServiceAccountToken: false
    serviceAccountTokenVolume:
      enabled: true
      mountPath: /var/run/secrets/kubernetes.io/serviceaccount
      volumeName: kube-api-access
      expirationSeconds: 3600
      audience: ~

Security Implications¶

Manual token volumes should be used when:

Security policies require explicit control over service account token mounting
Using security policy engines like Kyverno that restrict automatic token mounting
Implementing defense-in-depth security strategies
You need custom token expiration times or audiences
Compliance frameworks mandate container-specific privilege assignment

Use Cases and Examples¶

For comprehensive configuration examples, security scenarios, and detailed use cases, see Service Account Token Volume Examples.

Supported Executors¶

The service account token volume configuration is only effective for pod-launching executors:

CeleryExecutor - when launching Celery worker pods
CeleryKubernetesExecutor - for both Celery workers and Kubernetes task pods
KubernetesExecutor - when launching task pods in Kubernetes
LocalKubernetesExecutor - for Kubernetes task pods in local mode

For other executors (LocalExecutor, SequentialExecutor), this configuration has no effect as they don’t launch additional pods.

Migration from Automatic to Manual Token Mounting¶

To migrate from automatic to manual token mounting:

Test the configuration in a non-production environment first

Update your values.yaml:

values.yaml¶

scheduler:
  serviceAccount:
    automountServiceAccountToken: false
    serviceAccountTokenVolume:
      enabled: true

Deploy the changes using Helm upgrade
Verify that the scheduler can still launch pods successfully
Monitor for any authentication issues in the logs

Troubleshooting¶

Common Issues:

Authentication failures: Ensure serviceAccountTokenVolume.enabled is set to true when automountServiceAccountToken is false
Permission denied: Verify that the service account has the necessary RBAC permissions
Token expiration: Check if expirationSeconds is too short for your workload patterns

Debugging:

Check the scheduler logs for authentication-related errors:

kubectl logs deployment/airflow-scheduler -n <namespace>

Verify the projected volume is mounted correctly:

kubectl describe pod <scheduler-pod-name> -n <namespace>

Backward Compatibility¶

This feature maintains full backward compatibility:

Existing deployments with automountServiceAccountToken: true continue to work unchanged
The serviceAccountTokenVolume configuration is only applied when explicitly enabled
Default values ensure no breaking changes for existing installations

For more information about Kubernetes service account tokens and projected volumes, see the Kubernetes documentation on service account tokens.