Quick interview answer
Derive it from production analytics: peak concurrent users, transaction mix, think time and peak-hour distribution — never from a guessed round number.
Detailed explanation
Pull 90 days of APM/access-log data, identify the busiest hour, model the transaction mix by percentage, add think times, then define baseline, peak and stress profiles with explicit pass/fail SLAs (p95 latency, error rate, throughput).
Real-world example
Peak-hour checkout mix modelled from Grafana data caught a DB pool exhaustion bug.
Interview tip
Always state SLAs as p95/p99, not averages.