Cost, performance, and sustainability

Cost is a workload property created by requests, replicas, nodes, storage, traffic, observability, and management choices.

Autopilot economics

Autopilot usually charges from requested workload resources and manages bin packing. Accurate requests are still essential. Standard charges for node capacity, so utilization, machine type, node pools, autoscaling, reservations, committed use, and Spot VMs matter.

Practical controls

• Label namespaces and workloads for cost allocation.
• Set ResourceQuota and LimitRange.
• Right-size requests from measurements and VPA recommendations.
• Use HPA and node scaling with tested boundaries.
• Schedule non-production down where possible.
• Use Spot Pods/VMs only for interruption-tolerant work.
• Control logging volume and retention.
• Review cross-zone, internet, and load-balancer data transfer.
• Track cost per request, customer, job, or training run.

Performance

Measure p50/p95/p99 latency, throughput, saturation, startup, scaling, image pull, DNS, service mesh overhead, storage, and downstream limits. Tune the application before only buying more CPU.

Sustainability

Efficient requests, autoscaling, managed services, modern machine families, fewer idle clusters, shorter jobs, sensible data retention, and efficient code reduce both spend and resource use.

Trade-off

The cheapest single-zone cluster may violate availability. Maximum redundancy may waste budget. State the business outcome purchased by each reliability cost.