Karpenter Metrics Dashboard for Kubernetes Autoscaling

Use this dashboard to monitor Karpenter across fleet size and utilization, provisioning and scheduling, disruption and consolidation, controller and cloud provider health, Spot interruption, and NodePool capacity and cost.

Karpenter exposes native karpenter_* metrics on the controller metrics endpoint, and the OpenTelemetry Collector scrapes them for this dashboard.

Dashboard Coverage

Use these panels to:

• Confirm fleet state: See how many nodes Karpenter tracks, how many it owns against the rest of the cluster, whether its view of the cluster is synced, and which build is running.
• Watch provisioning: Track node and NodeClaim creation by NodePool and zone, scheduler queue depth, and how long pods wait to bind and start.
• Find unschedulable work: Monitor pods Karpenter cannot place, and pending pods grouped by their effective zone constraint.
• Follow disruption: Track disruption decisions by reason and NodePool, nodes eligible for disruption, budget consumption, and drain and termination activity.
• Check controller health: Watch cloud provider call latency, call rate, error rate by method, and how long cluster state stays unsynced.
• Handle Spot interruption: Track interruption messages by type and how long they sit on the queue before Karpenter processes them.
• Review capacity and cost: Compare NodePool usage against configured limits, pod requests against allocatable, system overhead, and Karpenter's hourly cost estimate.

Metrics Included

Fleet Overview

• Nodes in Cluster State: Nodes Karpenter tracks.
• Cluster State Synced: 1 when Karpenter's view matches the API server.
• Karpenter-Managed Nodes: Nodes Karpenter provisioned.
• Unschedulable Pods: Pods Karpenter cannot place.
• Cluster Utilization by Resource: Pod requests as a percentage of allocatable.
• Allocatable CPU: Karpenter-Managed vs Unmanaged: Capacity split by the managed label.
• Karpenter Build Info: Running version, Go version, and architecture.

Provisioning & Scheduling

• Nodes Created and NodeClaims Created: Provisioning activity by NodePool, zone, and reason.
• Scheduler Queue Depth: Pods waiting for a scheduling decision.
• Unschedulable Pods Over Time: Unschedulable count by controller.
• Pending Pods by Effective Zone: Pending pods by zone constraint.
• Pod Bound Duration: p50, p95, and p99 time from pod creation to binding.
• Pod Provisioning Startup Duration (p95): Time from Karpenter deciding a pod can schedule until it runs.
• Longest Pod Unbound Wait by Namespace: Age of the oldest unbound pod per namespace.
• Pods by Phase: Pod counts by phase.

Disruption & Consolidation

• Disruption Decisions by Decision & Reason and by NodePool: Voluntary disruption actions taken.
• Nodes Eligible for Disruption by Reason: Nodes currently eligible, by reason.
• Allowed Disruptions by NodePool & Reason and Nodes Consuming Disruption Budget: Budget headroom and usage.
• Nodes Drained and Nodes Terminated: Drain and termination counts by NodePool and zone.
• NodeClaims Disrupted by Reason and NodeClaims Terminated: NodeClaim-level disruption and removal.
• Disruption Decision Evaluation Duration: p50 and p95 time to evaluate a disruption decision.
• Consolidation Timeouts: Times the consolidation algorithm hit its timeout.
• NodeClaim Termination Duration (p95): Time to fully terminate a NodeClaim.

Controller & Cloud Provider Health

• CloudProvider Error Rate by Method & Error: Failed cloud API calls.
• CloudProvider Call Latency by Method: p50 and p95 latency per cloud API method.
• CloudProvider Call Rate by Method: Call throughput per method.
• Cluster State Unsynced Time: How long Karpenter's state has been out of sync.
• Scheduling Simulation Duration by Controller: p50 and p95 scheduling simulation time.

Spot Interruption

• Interruption Messages Received by Type: Spot interruption and instance health notices by message type.
• Interruption Message Queue Duration (p95): Time a message waits on the SQS queue before Karpenter handles it.

Capacity & Cost

• NodePool Usage as % of Limit: Provisioned resources against configured NodePool limits.
• NodePool Hourly Cost Estimate: Karpenter's current hourly cost estimate for a NodePool. This is not a billing source.
• CPU Requests as % of Allocatable: Pod requests against allocatable, split by managed.
• Node System Overhead: Resources reserved for system overhead.
• Oldest Karpenter Node Age by NodePool: Age of the longest-lived Karpenter node.

Dashboard Variables

Use these filter variables:

• nodepool: Filter by Karpenter NodePool name.
• capacity_type: Filter by on-demand, spot, or reserved.
• zone: Filter by availability zone.

Notes

• The Spot Interruption panels need an interruption queue. They stay empty unless you set settings.interruptionQueue to an SQS queue name when installing Karpenter.
• Only the leader replica emits most metrics. Karpenter runs two replicas with leader election. No panel groups by pod, so each one stays correct whichever replica holds the lease.
• Some panels are deliberately unfiltered. karpenter_nodes_* and karpenter_pods_state cover every node and pod in the cluster, not only Karpenter's. Unmanaged nodes carry an empty nodepool and capacity_type, so filtering those panels would silently drop them.
• Counter panels use increase, which needs two datapoints. Karpenter's labeled counters do not exist until their first increment, so a first-ever event on a new cluster does not register. Later events do.
• The instance type catalog metrics are excluded. karpenter_cloudprovider_instance_type_offering_available and _price_estimate emit one series per instance type, zone, and capacity type in the region, and do not scale with cluster size. See the cardinality guidance in the setup guide.

Related Dashboards

• Cert-Manager
• KEDA
• Hostmetrics and Kubernetes