Kubernetes Cost Optimization
Cost analysis and optimization using kubectl-mcp-server's cost tools.
Quick Cost Analysis
Get Cost Summary
get_namespace_cost(namespace) # Namespace cost breakdown
get_cluster_cost() # Cluster-wide cost summary
Find Unused Resources
find_unused_resources(namespace) # PVCs, ConfigMaps, Secrets
find_orphaned_pvcs(namespace) # Unbound PVCs
Resource Right-Sizing
get_resource_recommendations(namespace) # CPU/memory suggestions
get_pod_metrics(name, namespace) # Current usage
Cost Optimization Workflow
1. Identify Overprovisioned Resources
# Get recommendations
get_resource_recommendations(namespace="production")
# Compare requests vs actual usage
get_pod_metrics(name, namespace)
get_resource_usage(namespace)
2. Find Idle Resources
# Unused PVCs (not mounted)
find_orphaned_pvcs(namespace)
# ConfigMaps/Secrets not referenced
find_unused_resources(namespace)
3. Analyze Node Utilization
get_nodes()
get_node_metrics() # If metrics-server installed
Right-Sizing Guidelines
| Current State | Recommendation |
|---|---|
| CPU usage < 10% of request | Reduce request by 50% |
| CPU usage > 80% of request | Increase request by 25% |
| Memory < 50% of request | Reduce request |
| Memory near limit | Increase limit, monitor OOM |
Cost by Resource Type
Compute (Pods/Deployments)
get_resource_usage(namespace)
get_pod_metrics(name, namespace)
Storage (PVCs)
get_pvc(namespace)
find_orphaned_pvcs(namespace)
Network (LoadBalancers)
get_services(namespace) # Filter type=LoadBalancer
# Each LB has fixed cloud cost
Multi-Cluster Cost Analysis
Compare costs across clusters:
get_cluster_cost(context="production")
get_cluster_cost(context="staging")
get_cluster_cost(context="development")
Cost Reduction Actions
Immediate Wins
- Delete unused PVCs:
find_orphaned_pvcs()then delete - Right-size pods: Apply
get_resource_recommendations() - Scale down dev/staging: Off-hours scaling
Medium-term Optimizations
- Use Spot/Preemptible nodes: For fault-tolerant workloads
- Implement HPA: Auto-scale based on demand
- Use KEDA: Scale to zero for event-driven workloads
Long-term Strategy
- Reserved instances: For stable production workloads
- Multi-tenant clusters: Consolidate small clusters
- Right-size node pools: Match workload requirements
Automated Analysis Script
For comprehensive cost analysis, see scripts/find-overprovisioned.py.
KEDA for Cost Savings
Scale to zero with KEDA:
keda_scaledobjects_list_tool(namespace)
keda_scaledobject_get_tool(name, namespace)
KEDA reduces costs by:
- Scaling pods to 0 when idle
- Event-driven scaling (queue depth, etc.)
- Cron-based scaling for predictable patterns
Related Skills
- k8s-autoscaling - HPA, VPA, KEDA
- k8s-troubleshoot - Resource debugging