Averages hide tail latency: a few very slow requests can be invisible in the mean but painful for users. p95/p99 show how the slowest 5%/1% behave and help catch queueing and saturation issues.
Advanced answer
Deep dive
Expanding on the short answer — what usually matters in practice:
Explain the "why", not just the "what" (intuition + consequences).
Trade-offs: what you gain/lose (time, memory, complexity, risk).
Edge cases: empty inputs, large inputs, invalid inputs, concurrency.
Examples
A tiny example (an explanation template):
// Example: discuss trade-offs for "why-do-teams-watch-p95/p99-latency,-not-just-ave"
function explain() {
// Start from the core idea:
// Averages hide tail latency: a few very slow requests can be invisible in the mean but pain
}
Common pitfalls
Too generic: no concrete trade-offs or examples.
Mixing average-case and worst-case (e.g., complexity).