Cast AI for Karpenter brings Cast AI's optimization capabilities to clusters running open-source Karpenter. This page provides an overview of available features and how they integrate with your existing Karpenter setup.

For a conceptual introduction to Cast AI for Karpenter, see Cast AI for Karpenter overview.

Feature availability

The following features are available for Karpenter-managed clusters:

How Cast AI features work with Karpenter

Cast AI features are designed to extend Karpenter rather than replace it. The integration follows these principles:

Karpenter remains the provisioner
Node creation and deletion continue to flow through Karpenter. Cast AI influences decisions by modifying Karpenter CRDs and providing optimization signals, but Karpenter executes the actual infrastructure changes.

CRD-native configuration
Where possible, Cast AI stores configuration in Kubernetes-native formats. Your existing NodePools and EC2NodeClasses remain the source of truth for provisioning constraints.

Incremental enablement
Each feature can be enabled independently. You can start with cost reporting only, then gradually enable optimization features as you build confidence.

Feature details

Evictor

Cast AI's Evictor integrates with Karpenter to improve Node utilization while minimizing workload disruption.

Although Karpenter has native consolidation, Evictor provides significantly better bin-packing through its integration with Workload Autoscaler and Container Live Migration. This coordination allows Evictor to consolidate based on actual resource usage rather than just requested resources, achieving utilization levels that Karpenter's consolidation alone cannot reach.

What it adds to Karpenter:

• Coordination with Workload Autoscaler to consolidate Pods according to optimized resource requests, even if they are yet to be applied to the Pod and are pending
• Container Live Migration support for eligible workloads, preserving Pod state and TCP connections when moving Pods from Node to Node (graceful fallback to traditional eviction included)
• Progressive consolidation that respects Pod Disruption Budgets
• Superior bin-packing efficiency by considering actual workload usage patterns, not just static resource requests

How it differs from standard Cast AI:

Coordinating with Karpenter consolidation

When Evictor is enabled, Cast AI marks Nodes with karpenter.sh/do-not-disrupt to prevent Karpenter from consolidating them. This ensures Evictor and Karpenter's consolidation don't conflict.

Evictor respects your NodePool's consolidateAfter setting. Evictor will not consolidate a Node until both its own grace period and Karpenter's consolidateAfter window have passed. For example, if your NodePool specifies:

disruption:
  consolidateAfter: 30m
  consolidationPolicy: WhenEmptyOrUnderutilized

Evictor will wait at least 30 minutes before considering the Node for consolidation, preventing conflicts with Karpenter's consolidation policy.

Consolidation priority

When both Evictor and Karpenter's native consolidation are enabled, Karpenter's consolidation takes precedence. Evictor will not attempt to consolidate a node until:

1. Karpenter's consolidateAfter window has passed
2. Evictor's own grace period has elapsed

This ensures predictable behavior and prevents conflicts between the two consolidation mechanisms. If you notice Evictor not acting on apparently underutilized nodes, check your NodePool's consolidateAfter setting first—Evictor is likely waiting for that window to pass.

Protecting Nodes from consolidation

To exclude specific Nodes from Evictor consolidation, apply the autoscaling.cast.ai/removal-disabled label. This works similarly to Karpenter's karpenter.sh/do-not-disrupt label and allows you to manually protect critical Nodes.

For Evictor documentation, see Evictor.

Rebalancer

The Rebalancer optimizes your entire cluster by identifying Nodes that could be replaced with more cost-effective alternatives for your workloads.

What it adds to Karpenter:

• Cross-NodePool optimization that Karpenter doesn't perform natively
• Awareness of Reserved Instances and Savings Plans
• Coordinated replacements that maintain workload stability
• Integration with Workload Autoscaler to rebalance based on optimized resource requirements

How it differs from standard Cast AI:

Rebalancing scope

Cast AI for Karpenter only rebalances nodes that were created and are managed by Karpenter. The following nodes are excluded from rebalancing:

• Legacy nodes created before Karpenter was installed
• EKS-managed node group nodes
• Nodes created by other provisioners (Cluster Autoscaler, etc.)

This limitation exists by design: Cast AI for Karpenter operates through Karpenter's CRDs and permissions, and does not have cloud-side control over non-Karpenter nodes. If you need to optimize non-Karpenter nodes, consider migrating them to Karpenter management or using Cast AI's standard Autoscaler.

Rebalancer capabilities not available with Karpenter

When using Rebalancer with Karpenter, the following capabilities from standard Cast AI Rebalancer are not available:

These limitations exist because Karpenter manages the Node lifecycle. Rebalancer coordinates with Karpenter by cordoning Nodes and creating new NodeClaims, while Karpenter handles the actual provisioning, drain, and deletion.

Tracking rebalancing progress

The Cast AI console may not show detailed progress during rebalancing. For full visibility, inspect the Rebalancing CRD:

kubectl get rebalancing -n castai-agent
kubectl describe rebalancing <name> -n castai-agent

The Rebalancing CRD contains the current phase, human-readable status messages, and related events for each step. This is the source of truth for rebalancing status and troubleshooting.

For Rebalancer documentation, see Rebalancer.

Spot intelligence

Cast AI improves Karpenter's Spot Instance handling with predictive capabilities and reliability improvements.

What it adds to Karpenter:

• Interruption prediction — Identifies at-risk Nodes before AWS announces interruptions
• Spot reliability model (coming in future releases) — Steers toward historically stable Spot pools
• Spot fallback recovery (coming in future releases) — Automatically returns to Spot when capacity becomes available again

How interruption prediction works

Cast AI's Spot interruption prediction provides up to 30 minutes of advance warning before interruptions occur—significantly extending AWS's standard 2-minute notice. The prediction mechanism differs based on whether Container Live Migration is enabled:

In both cases, the extended prediction window provides significantly more time for graceful workload migration compared to waiting for AWS's native interruption signal.

How it differs from standard Cast AI:

For Spot handling documentation, see Spot Instances and Spot Handler.

Workload Autoscaler

Workload Autoscaler continuously rightsizes workloads based on actual resource usage.

What it adds to Karpenter:

• Automatic adjustment of CPU and memory requests to match actual usage
• Tighter bin-packing as rightsized workloads require less capacity
• Integration with Evictor and Rebalancer for coordinated optimization

How it differs from standard Cast AI:

Workload Autoscaler behavior is largely identical whether you're using Cast AI for Karpenter or Cast AI's Autoscaler—it operates at the workload level independently of Node provisioning.

For Workload Autoscaler documentation, see Workload Autoscaling.

Pod mutations

Pod mutations automate Pod spec adjustments to simplify workload configuration and reduce manual efforts by teams.

What it adds to Karpenter:

• Automatic application of labels, tolerations, and NodeSelectors
• Simplified onboarding without modifying Deployment manifests
• Consistent Pod configuration across workloads

How it differs from standard Cast AI:

Pod mutations work identically with Karpenter and standard Cast AI. The mutations apply to Pod specs before creation, independent of which autoscaler provisions Nodes.

For Pod mutations documentation, see Pod mutations.

Cost reporting

The savings report and other cost monitoring capabilities provide visibility into your cluster's optimization potential without making any changes.

What it provides:

• Current vs. optimized cost comparison
• Node utilization and bin-packing analysis
• Commitment utilization tracking
• Spot adoption opportunities
• Workload rightsizing recommendations

How it differs from standard Cast AI:

Cost reporting works identically for Karpenter clusters. The analysis examines your current state and models what Cast AI optimization could achieve.

For general cost monitoring, see Cost Monitoring.

Labels reference

Cast AI for Karpenter uses specific labels to coordinate optimization activities with Karpenter. Understanding these labels helps with troubleshooting and protecting critical nodes.

Node labels

Workload labels

Protecting nodes from optimization

To exclude a node from Evictor consolidation only:

kubectl label node <node-name> autoscaling.cast.ai/removal-disabled=true

The node remains available for Karpenter's native consolidation.

To exclude a node from both Evictor and Karpenter consolidation:

kubectl label node <node-name> karpenter.sh/do-not-disrupt=true

Note: When Evictor is enabled, Kentroller automatically applies karpenter.sh/do-not-disrupt to nodes to prevent conflicts between Evictor and Karpenter consolidation. Evictor then manages consolidation exclusively for those nodes.

Features not available with Karpenter

Some Cast AI capabilities require tighter integration with Node scheduling than the Karpenter-layered approach allows:

Additionally, the following Rebalancer capabilities are not available when using Karpenter (see Rebalancing scope for details):

• Aggressive mode
• Graceful eviction controls
• Paused drain configuration
• Progress bar in UI
• Container Live Migration (available with Evictor only)

To benefit from these capabilities, consider migrating to Cast AI Autoscaler.

Related resources

• Cast AI for Karpenter overview — Conceptual introduction
• Getting started — Connect your cluster