performance-testing

Define stages in the options object to ramp virtual users, then add thresholds like 'http_req_duration': ['p(95)<500'] and 'http_req_failed': ['rate<0.01']. k6 exits with a failure code if any threshold is breached, which CI pipelines can use to block merges.

k6 uses JavaScript and suits scripted scenarios with built-in thresholds and cloud execution. Locust uses Python, supports weighted task distributions, and offers a web UI plus headless CSV output for CI. Choose based on team language preference and reporting needs.

Derive thresholds from production SLOs rather than arbitrary numbers, for example p99 under 500ms and error rate under 0.1%. First establish a baseline on staging, then fail builds if p99 degrades more than 20% from that baseline.

Rising latency with VU count usually indicates connection pool exhaustion. Confirm by checking pool wait metrics, then fix by increasing pool size or adding a connection proxy like PgBouncer.

Run go test -bench=. -count=10 -benchmem and save output before and after the change, then use benchstat to compare the two files. Include b.ReportAllocs() and b.ResetTimer() in benchmarks for accurate allocation and timing data.

A single machine saturates its network interface before the target at high VU counts, producing misleading results. Above roughly 500 virtual users, use distributed execution such as k6 Cloud or multiple Locust workers.