Karpenter Dashboard
Shows every Karpenter NodePool on one screen in Wave Console: nodes, cost per hour, Spot ratio, disruptions, and thrashing alerts.
A NodePool creates and removes 6 nodes in one hour, and no screen shows why
The 6 nodes are illustrative. karpenter.sh v1.14 docs: Metrics, Disruption, Getting Started with Karpenter.
By default, Karpenter reports its state as Prometheus metrics and Kubernetes events. Its getting-started Grafana stack does not include the persistent storage that monitoring a production deployment needs (karpenter.sh).
Karpenter resets the consolidation timer whenever a Pod is added to or removed from a node (karpenter.sh). When nodes are created and removed again and again, the reason is spread over NodeClaims, events, metrics, and the AWS bill.
The 6 nodes and their lifespans are illustrative. Metrics and events: karpenter.sh v1.14 docs. The Thrashing Pattern alert and its fields: Wave Karpenter Dashboard docs.
How Wave does it
- Karpentercreates and removes NodeClaims
- exportsPrometheusmetrics in Prometheus format
- chartsGrafanaexample dashboards, no persistent storage
- you joinEngineermatches metrics, events, and the AWS bill by hand
- Wave Agentreads nodes, NodePools, and NodeClaims
- gRPCWave Coreprocesses the Karpenter data
- checksThrashing checknodes created and removed in rapid succession
- alertsWave ConsoleKarpenter > Overview: nodes, cost per hour, Spot ratio, alerts
Collect from Karpenter
The Wave Agent sends node, NodePool, and NodeClaim data to Wave Core over gRPC. It needs Amazon EKS, Karpenter v0.32 or later, and one Helm value, karpenter.enabled.
See cost and Spot in one place
Karpenter > Overview in Wave Console shows Karpenter nodes, total cost per hour, Spot ratio, active NodePools, and Spot interruptions. Each NodePool shows its disruptions and thrashing over 30 days.
Catch thrashing
Wave flags nodes that are created and removed in rapid succession. The Thrashing Pattern alert shows the affected nodes, the reason, their average lifespan, and the average Pods affected.
Look into one NodePool
Each NodePool has a node lifecycle timeline, its NodeClaims with their conditions and cost per hour, and its disruption events.