Rightsizing Automation

Automate the rightsizing.

Inputs to rightsizing

Per-resource utilisation: CPU, memory, IO. Multiple weeks of data to capture peak and steady-state.

Per-resource cost: instance class price, EBS storage cost, RDS class price. Pulled from billing.

Workload characteristics: bursty, steady, batch. Different shapes need different rightsizing approaches.

Recommendation engines

AWS Compute Optimizer: free; per-resource recommendations. Conservative; good baseline.

Vendor tools (Vantage, Cloudability, Kubecost): more aggressive; multi-cloud; richer reporting.

Custom: scripts on top of metrics and billing APIs. Tailored to specific workload patterns.

Apply with safety

Recommendations reviewed by engineers. Auto-apply for low-risk changes (gp2 to gp3); manual approval for instance class changes.

Staged rollout: apply to non-prod first; observe; promote.

Rollback ready. Document the original size before changing; reverse if performance degrades.

Recurring rightsizing

Quarterly: full rightsizing audit. Workloads grow and shrink; periodic check catches drift.

Monthly: review of recent changes. Did the rightsized resources still meet performance targets?

Annual: deeper review including instance class evolution. New cloud instance types come out; old ones retire.

Outcome metrics

Per-quarter cost reduction from rightsizing. Trend over time.

Performance metric stability. Are we maintaining latency and throughput SLOs? Rightsizing should not regress these.

Engineering time invested per dollar saved. The math should justify continued automation investment.