Pick a real performance problem you owned through the whole arc — detect, diagnose, fix, verify — with numbers you can recall. Lead with the symptom and its blast radius: "p99 on the export endpoint walked from five seconds to five minutes as traffic grew", not "it got slow". The strong answer is built on measurement, so narrate the instrument at each step: what metric or profile told you where the time actually went, rather than where you assumed it went. The common miss is fixing the thing you suspected (add a cache) when the real cost was elsewhere (a lock, an N-plus-one, full table scans) — show that you let the profiler, not your prior, pick the target.

The drill that separates this round: "how do you know it was your change?" A deploy rarely lands alone — traffic dips, a cache warms, another team ships, the load test isn't the same shape as production. A strong answer controls for the confound: you held the input fixed (replayed the same query set, the same load profile), changed one thing, compared before-and-after on the same metric, and can say the latency drop tracked the rollout and reverted in the canary that didn't get the change. Then name the next bottleneck — the constraint that becomes tightest once yours is gone — because real systems just move the queue. Keep it first-person and quantified: "I cut p99 from 5 min to 1.2 min, database load down ~80%".