performance-engineer

Start by establishing a baseline with metrics and profiling, then use distributed tracing tools like OpenTelemetry or Jaeger to isolate slow spans. Combine CPU, memory, and I/O profiles to identify the highest-impact hotspot before optimizing.

k6, JMeter, Gatling, Locust, and Artillery are common choices for API load testing. Select based on scripting language preference and CI/CD integration needs, and always test with production-like data and traffic patterns.

Instrument services with OpenTelemetry SDKs to emit traces and metrics, export them to collectors, and visualize in Prometheus with Grafana dashboards. Define SLIs and SLOs to track latency, error rate, and saturation over time.

Load testing production is risky and should be avoided without explicit approvals and safeguards. Prefer staging environments that mirror production, or use staged rollouts with rollback plans and rate-limited synthetic traffic.

Skip it when the task is pure feature development with no performance goals, when no metrics or profiling data is available, or when only a non-technical summary is needed. Optimization without measurement risks wasted effort and regressions.