Skip to content

K8s Right-Sizing

Over-provisioned pods waste money and hide real utilization. Oracle finds consolidation opportunities your kube-scheduler misses.

Prerequisites

  • CloudSlash installed and initialized
  • kubectl configured and pointing to your cluster
  • K8s plugin installed: cs plugin install k8s

Scan Your Cluster

cs scan

The K8s plugin scans all namespaces: pods, services, deployments, daemonsets, statefulsets, and more. Results appear immediately in the table output.

Insights Tab

View Findings in the

cs daemon

Open http://localhost:8080 → Insights tab. Look for:

  • Over-provisioned pods: CPU/memory requests far exceed actual usage
  • Idle services: services with no backing pods or zero traffic
  • Ghost namespaces: namespaces with no active workloads
  • Orphaned PVCs: persistent volume claims not mounted by any pod

Bin Packing

Review the 2D Analysis

The Tetris engine in CloudSlash models your cluster as a 2D bin packing problem. It fits workloads into nodes by CPU and memory simultaneously, using Best Fit Decreasing (BFD) to find the tightest valid packing.

In the Solver Results section of the Insights tab, you see:

  • Current node utilization vs. optimal packing
  • How many nodes could be removed while fitting all workloads
  • Estimated monthly savings from the right-sized cluster

DEVI Recommendations

Ask for

In the TUI, press / to open the AI prompt bar and ask:

What's the most over-provisioned namespace in my cluster?

Which pods have resource requests 10x their actual usage?

DEVI queries the Universal Graph directly and returns precise answers with resource CRNs, costs, and recommended actions.

Track Improvements

After applying right-sizing changes, run the next scan and compare:

cs diff --from=2026-04-01    # Compare to a specific date
cs trend --days=30           # Show savings trend

Common Findings

Finding Heuristic Typical Cause
Idle pod idle_pod Deployment scaled to zero, not removed
Ghost namespace empty_namespace Dev environment abandoned
Orphaned PVC unbound_pvc StatefulSet deleted, PVC left behind
Over-requested deployment over_provisioned_workload Copy-paste resource limits from prod

What's Next