Start here. This is the direct spoken answer to practice first.
Overview
Uses representative traffic and tail latency to find the real bottleneck and a safe operating range.
I model the critical user flows with realistic payloads, data distribution, authentication, request mix, and think time. The test ramps through expected and peak concurrency while measuring throughput, errors, and p50, p95, and p99 latency. At the same time I capture per-worker CPU and memory, event-loop delay, connection-pool pressure, database time, and downstream limits. The goal is to find where latency or errors bend upward, not advertise one maximum requests-per-second number.