Autoscaling on Kubernetes

On Kubernetes, you can automatically adjust the number of authentik server and worker replicas independently:

• Scale servers on CPU usage with the Helm chart's Horizontal Pod Autoscaler (HPA). This requires Kubernetes resource metrics, but does not require KEDA or Prometheus.
• Scale workers on queued tasks with KEDA. Queue length can reveal a backlog even when workers use little CPU, such as when they wait for external services. The metric includes delayed tasks and excludes tasks already running.

Keep at least one worker running to handle scheduled tasks and file changes. For the number of tasks each replica can run concurrently, see Worker scaling.