What metrics, logs, and alerts matter most for operating Apache Airflow?

Monitoring Apache Airflow should not stop at process uptime. Build dashboards around service-level impact and the internal mechanics that drive that impact. For Apache Airflow, that means turning DAG design, operators and sensors, executor types (Celery/Kubernetes/Local), scheduler and worker scaling, task retries and backfills, XComs, and connection/variable management into concrete telemetry. Useful alerting separates symptoms from causes: symptom alerts page humans when users or delivery pipelines are affected; cause alerts help responders identify resource exhaustion, bad configuration, certificate expiry, or dependency failure.

For SRE-quality operations, connect alerts to runbooks. Every alert should answer what changed, what is impacted, where to look first, and what action is safe. Avoid noisy alerts that page on harmless transients, and prefer burn-rate or sustained-threshold logic when the signal represents an SLO.