Opscraft
Amazon EKS · Lesson 9 of 11 · 3 min read

Cost optimisation and multi-tenancy

Where EKS money goes

Nodes (usually the bulk), NAT gateway data processing, cross-AZ data transfer, load balancers, EBS volumes and snapshots, CloudWatch logs/metrics ingestion, and the per-cluster fee (×3–4 in extended support).

Levers

  • Right-size requests — the scheduler packs by requests, so inflated requests waste nodes. Compare requests vs actual usage (Kubecost/OpenCost, Container Insights, VPA recommendations).
  • Karpenter consolidation and diverse instance types; Spot for stateless and batch workloads; Graviton for ~20% savings; Savings Plans for the on-demand baseline.
  • Scale dev/test to zero outside working hours (KEDA cron, Karpenter limits, scheduled scaling) and delete idle clusters.
  • One ALB for many services via IngressGroups.
  • VPC endpoints (S3 gateway, ECR) to cut NAT charges; topology-aware routing to reduce cross-AZ traffic.
  • Log volume control — drop debug logs, set retention.
  • Stay on standard support versions.

"Cost spiked after moving to EKS — where do you look?" — node utilisation (requests vs usage), NAT and cross-AZ data in Cost Explorer, number of load balancers, idle dev clusters, on-demand vs Spot mix, logging volume.

Enlarge diagramRight-size requestspacking uses requestsKarpenter + Spotconsolidation, diverse typesGravitonabout 20% cheaperScale dev to zeronights and weekendsOne ALBIngressGroupsVPC endpointscut NAT data charges
Figure 1. The cost levers that usually matter most on EKS.
Swipe sideways to see all of it, or tap to open it full screen.

Multi-tenant cluster design

For several teams on one cluster:

  • Namespace per team/app, with labels for ownership and cost allocation.
  • Access entries/RBAC scoped to namespaces.
  • ResourceQuotas and LimitRanges per namespace.
  • NetworkPolicies — default deny between namespaces.
  • Policy engine — Kyverno/Gatekeeper guardrails; PSA restricted.
  • Separate node pools (taints/tolerations or Karpenter NodePools) for teams with special or noisy workloads.
  • Cost allocation by namespace/labels (Kubecost, split cost allocation data for EKS in AWS CUR).
  • For hard isolation (compliance, untrusted tenants), separate clusters or accounts instead.
Enlarge diagramNAMESPACE TEAM-ORDERSNAMESPACE TEAM-PAYMENTSRBACaccess entry scoped hereResourceQuotaCPU, memory, podsNetworkPolicydefault deny across namespacesRBACaccess entry scoped hereResourceQuotaCPU, memory, podsNetworkPolicydefault deny across namespaces
Figure 2. Several teams on one cluster: each namespace gets its own fences.
Swipe sideways to see all of it, or tap to open it full screen.