Skip to content
GreenKube

Karpenter & Cluster Autoscaler

Karpenter and the Cluster Autoscaler are not recommendation sources. GreenKube reads them only to estimate how much released capacity really becomes removed nodes: the realization factor. The Karpenter NodePool heuristic (greenkube.nodepool) and its NodePool policy advice were removed in this release.

The controller reads, read only:

  • the Cluster Autoscaler status ConfigMap kube-system/cluster-autoscaler-status;
  • Karpenter NodePools (consolidation policy, disruption budgets) and NodeClaims (claim → pool);
  • Karpenter DisruptionBlocked events (a blocker: the pool’s factor is halved) and Unconsolidatable events (reported, factor unchanged);
  • do-not-disrupt / safe-to-evict annotations and blocking PodDisruptionBudgets;
  • with Prometheus, Karpenter’s consolidation metrics (decisions, eligible nodes, pool cost). Metrics that are not exported change nothing.

These feed the per-pool realization factor. Events that cannot be read (RBAC) are reported as “Karpenter events could not be read”, never as “no blocker”. HPA reads stay too: they feed the rightsizing evidence and the risk model.

get/list on karpenter.sh NodePools and NodeClaims, and list on core events, all in the chart’s ClusterRole.

impact:
autoscalerHint: "" # none, cluster_autoscaler or karpenter, when detection cannot see it
# realizationFactor: 0.5 # pin the factor for every projection

config.recommendations.karpenterEnabled and sources.settings["greenkube.nodepool"] are accepted and ignored since this release; the chart’s install notes name them (Upgrading).