Capacity, performance and cost

Measure the user’s delay and the cost of completed work

Speeding up a 40 ms helper inside a 900 ms request cannot make the page four times faster. Similarly, halving CPU use may leave a fixed-capacity bill unchanged. You will separate elapsed time, resource use, and billed units.

Identify a measured bottleneck, make one relevant change, and compare equivalent workloads. Use the deployment-headroom case to account for capacity temporarily removed during a rollout. Keep teaching numbers, local measurements, and cloud invoices clearly labeled.

The one-box lesson is a worked postmortem of a real experiment: the stack somebody actually used, the ten problems that came up in order, and the fix for each. Two of the results contradict the usual guess — a server that fails on latency with a zero error rate, and a load generator that flatters its own tail.