load-testing

Run a stress test that steps the arrival rate and records throughput, p50/p95/p99, error rate, and resource saturation at each step. The knee appears where latency climbs sharply, and the resource metric tracking p99 names the bottleneck, such as a connection pool.

A closed model holds a fixed number of virtual users that self-throttle when the server slows, hiding queue collapse. An open model fixes the arrival rate so work piles up under degradation, matching real traffic. Use k6's ramping-arrival-rate executor for open models.

Averages are dominated by fast requests and hide tail latency that users actually experience. A build can show a 3 ms average regression while its p99 jumps from 380 ms to 4 seconds. Always report p50, p95, p99, and max instead.

Soak tests should run 4 to 24 hours at 60-80% of peak load. A leak of 4 MB per hour is invisible in a 20-minute run but takes down a pod in days, so short tests cannot reveal leaks, fd exhaustion, or disk growth.

Common causes include testing against small dev datasets that change query plans and cache hit rates, replaying one hardcoded id so caches absorb all load, and using closed models that self-throttle. Match production row counts, key skew, and use arrival-rate executors.

Verify generator CPU stays below 70%, file descriptors and ephemeral ports have headroom, and k6 dropped_iterations is zero. Nonzero dropped iterations means maxVUs was too low and the intended arrival rate was never actually offered.