Helm Deployment and Executor Choice Quiz
Chart 1.22.0 / Airflow 3.3.1 / KEDA 2.20.
1. Which chart and repository alias does this guide use?
Show answer
The Apache Airflow project's official chart. With the repository registered as apache-airflow, commands use apache-airflow/airflow. Do not mix independent community-chart values.
2. What are chart 1.22.0's actual Airflow/executor defaults?
Show answer
Airflow 3.2.2 and CeleryExecutor. The guide explicitly overrides them for its 3.3.1/KubernetesExecutor profile.
3. Does Chart.yaml enforce every documented minimum version?
Show answer
Do not assume so. Use Helm 3.19.0+ and compatible Kubernetes/Airflow. Chart 1.16's README specified Kubernetes 1.29+; even the reviewed 1.22 templates rendered for 1.29, which is not proof of support.
4. Does executor alone determine Redis and worker resource kinds?
Show answer
It influences them, but Redis/external-broker settings and worker persistence also matter. Workers can be Deployments or StatefulSets, and Celery can use an external broker.
5. Does disabling PostgreSQL configure an external database?
Show answer
No. Prepare the database, migration privileges, URI/TLS and a connection reference such as metadataSecretName. With that Secret set, metadataConnection is not the authoritative source.
6. Should Fernet/API/JWT keys be regenerated on upgrades?
Show answer
Preserve them during ordinary upgrades. Fernet-key loss can break access to encrypted data, and JWT changes can affect active task authentication. Handle backup and rotation deliberately.
7. How does the example avoid default admin/admin?
Show answer
It disables createUserJob and uses interactive airflow users create for the selected FAB auth manager, prompting twice for the password. Other auth managers/SSO need their own procedures.
8. Does KubernetesExecutor always take 1–2 minutes to start?
Show answer
No. Measure image cache/pull, Kubernetes API/scheduler, node availability/provisioning, quotas and runtime initialization. Celery also has cold starts after scale-to-zero.
9. Can KubernetesExecutor use an arbitrary GPU/CLI worker image?
Show answer
The worker needs a compatible Airflow task runtime and DAG/provider dependencies. This differs from KubernetesPodOperator launching a child pod with a workload image.
10. Does having no worker pods make total Airflow idle cost zero?
Show answer
Control-plane, database, broker, node, storage and logging costs can remain. Per-pod execution or scale-to-zero does not guarantee zero total cost or complete security isolation.
11. What are the chart KEDA timing defaults versus the example?
Show answer
Defaults are pollingInterval=5s and cooldownPeriod=30s. The example explicitly chooses 10s/300s. Distinguish scale-to-zero cooldown from HPA scaling/stabilization above zero.
12. What should worker_concurrency become in the SQL?
Show answer
A numeric value rendered by the chart, 4 in this example—not a database column named worker_concurrency. Count the relevant running/queued tasks for the worker's queues/executors, divide and round up.
13. Why can a k8s alias cause mixed-executor scaling errors?
Show answer
TaskInstance stores task.executor. If the chart excludes only the literal KubernetesExecutor, tasks stored as k8s can enter the Celery count. Use filters matching actual stored values.
14. What if the query returns 25 with maxReplicaCount=20?
Show answer
The maximum bounds workers, so not all demand is immediately satisfied. Inspect backlog, concurrency, worker resources, database/broker load and node capacity together.
15. Is Airflow-pod TLS configuration enough for KEDA database access?
Show answer
KEDA is a separate database client needing DNS, networking, permissions, a compatible URI and CA. A URI's sslrootcert path must be readable by the scaler; the chart does not copy it automatically.
16. Is KEDA limited to Deployments?
Show answer
No. The chart targets a Deployment or StatefulSet according to Celery persistence. KubernetesExecutor task pods are not the same replica-pool pattern, but that does not prohibit other KEDA uses.
17. Does successful Helm --wait finish validation?
Show answer
Check successful migrations, Ready long-running components, DAG delivery, worker execution, Execution API, results and logs. Test workload-driven scaling, return to idle and recovery for KEDA profiles.
18. Does Helm uninstall immediately delete all PostgreSQL data?
Show answer
Pod removal and PVC/PV data lifecycles differ. Review PVC retention, StorageClass reclaim policy and external-database deletion/backups separately; do not indiscriminately remove shared namespaces, Secrets and databases.