Cost, performance, and sustainability

The goal is not the smallest bill. It is the required outcome with the least waste and acceptable risk.

Main cost drivers

Count EKS cluster fees, extended-support fees, EC2 or Fargate compute, EBS/EFS, load balancers, NAT processing, cross-AZ and internet transfer, logs, metrics, security tools, support, and engineer time.

Reduce waste

• Right-size requests with measurements.
• Use HPA plus Karpenter, Auto Mode, or node scaling intentionally.
• Keep stable baseline on appropriate On-Demand or Savings Plans.
• Use diversified Spot for interruption-tolerant capacity.
• Evaluate Graviton and efficient images.
• Remove idle clusters, load balancers, volumes, snapshots, and ECR images.
• Use VPC endpoints and locality when they reduce unnecessary transfer/NAT cost.
• Control log retention, metric cardinality, and trace sampling.

Performance

Measure latency percentiles, throughput, errors, saturation, scheduling delay, image pull time, DNS, database connections, and dependency latency. Load-test autoscaling and recovery, not only steady state.

Unit economics

Tag and allocate by owner, environment, application, and cost center. Track cost per order, user, tenant, build, or request. A growing bill may be healthy when unit cost falls and business volume grows.

Feynman check

A half-empty bus may look cheap, but cost per passenger is high. Good platform economics packs work safely without making one crowded bus a single failure.