Skip to content

Cluster Metrics

Use cluster metrics to judge whether an AKS issue is rooted in the control plane, cluster-wide capacity, or an individual workload.

Topic Groups

Control-plane and cluster-wide metrics

Metric Source What it means Common use Denominator / cardinality notes
apiserver_request_duration_seconds / ApiRequestDurationSeconds Managed Prometheus for Prometheus-native views; Azure Monitor platform metric for Azure Metrics charts API server request latency. Rising values mean slower control-plane responses or overloaded clients. Confirm whether deployments, scaling, or upgrades are slowed by API latency rather than by node or pod issues. Split by verb or route carefully. Cardinality grows quickly when you keep Method and Route dimensions or Prometheus labels together.
apiserver_request_total / ApiRequestCount Managed Prometheus for Prometheus-native views; Azure Monitor platform metric for Azure Metrics charts Request volume against the Kubernetes API. Error-code or route spikes often explain failed automation. Correlate deployment storms, noisy controllers, or repeated client retries with control-plane pressure. Rates are usually more useful than raw counts. Preserve only the labels needed for verb, resource, or status-code analysis.
etcd_db_total_size_in_bytes Managed Prometheus Size trend for the etcd backing store. Steady growth is expected; sudden jumps can indicate heavy object churn. Watch for abnormal control-plane growth before upgrades, mass rollouts, or runaway custom-resource usage. Treat it as a single-cluster trend metric, not a per-workload KPI. Compare slope over time rather than single-sample values.
Node Count Azure Monitor platform metric Current number of nodes in the managed cluster. Verify whether expected scale-out or scale-in actually occurred during incidents or maintenance. Interpret together with node pool intent. A cluster-wide total can hide imbalance across pools.

Interpretation priorities

  • Start with API latency and request volume when kubectl, controllers, or upgrade workflows feel slow.
  • Use Node Count to distinguish “cluster did not add capacity” from “cluster added nodes but pods still failed.”
  • Treat etcd_db_total_size_in_bytes as an early-warning trend signal, not an alert threshold by itself.

Usage Notes

  • Prefer Azure Monitor platform metrics when you want fast, low-friction cluster-wide alerting and long-lived Azure-native dashboards.
  • Prefer Managed Prometheus when you need Prometheus labels, rate functions, or Grafana panels that correlate control-plane signals with workload metrics.
  • For incident triage, correlate cluster metrics with node and autoscaler metrics before assuming the control plane is the only problem.

See Also

Sources