Performance is a distribution, not a number. Drive the load up and watch latency percentiles, throughput, and error rate move — then make a budget (thresholds) decide whether the run passes.
The same idea powers tools like k6, where thresholds turn performance budgets into CI gates.
Latency distribution