cloud-cost-optimization

Start by right-sizing instances from measured p95 CPU, memory, and IO over a representative week, since over-provisioning is the largest lever. Then cover the steady baseline with savings plans or reservations, automate storage tiering, and kill idle resources like orphaned volumes and 24/7 dev environments.

Pull real CPU, memory, and IO metrics over at least a week covering peaks, then move each workload to the smallest instance family that holds its p95 with headroom. For containers, set requests from observed p95 rather than guesses, and schedule recurring reviews since usage drifts over time.

Use reserved instances or savings plans for the predictable baseline of usage that is always present, and spot instances for interruptible, fault-tolerant work like batch jobs, CI, and stateless workers. Keep the variable top of the load curve on-demand where flexibility justifies the premium.

Egress is metered for internet traffic, cross-region transfer, and even cross-availability-zone traffic, which surprises teams because it is invisible in architecture diagrams. Reduce it by co-locating chatty services, keeping compute next to its data, using CDNs for repeated content, and routing over private endpoints.

Cost optimization should never delete redundancy that SLOs require, since optimizing away reliability trades short-term savings for outage risk. Buy-and-forget reservations also become sunk costs when workloads shrink, so commitments and right-sizing need recurring review rather than one-time fixes.