k8s-prom-hpa

k8s-prom-hpa is a tutorial showing how Kubernetes can add or remove application replicas using measurements collected by Prometheus. An adapter exposes the chosen measurements to the Horizontal Pod Autoscaler, which decides the desired replica count. It is useful for understanding that connection between monitoring and scaling. The repository's example versions are historical, so use the concept while selecting current APIs and testing whether the signal reflects useful application capacity.

Current guidance

The stefanprodan/k8s-prom-hpa repository demonstrates resource and custom-metric autoscaling using a sample application, Metrics Server, Prometheus and an adapter API. Its instructions target early Kubernetes versions and include an autoscaling/v2beta2 example. Preserve the tutorial’s conceptual distinction between node capacity and workload replicas, but update the APIs and components deliberately.

Current Kubernetes documentation describes autoscaling/v2 behavior and the resource, custom and external metrics interfaces. Prometheus scraping alone does not make a metric available to HPA; verify the adapter’s discovery rules, metric naming, labels and the API response for the target workload. Resource-utilization targets also depend on meaningful resource requests.

Test scaling with a controlled load and observe the metric timestamp, desired replicas, readiness and stabilization behavior. Include a missing or stale metric and a backend that cannot support additional replicas. Check that the chosen signal correlates with useful capacity rather than merely growing when the application is already failing. Pin the adapter and monitoring versions separately from the application and avoid running two controllers against the same replica target.