Track request rate, error rate, latency, saturation, and dependency health.

Latency should be measured using percentiles rather than only averages. Averages can hide a severe tail-latency problem.

Useful measurements include p50, p95, and p99 latency, HTTP status distribution, requests per second, database latency, queue depth, and external dependency failures.

A useful service-level objective might specify that 99% of successful API requests should complete within a target latency under defined conditions.