The three signals
Metrics are cheap aggregates good for alerting; logs are high-cardinality detail good for forensics; traces follow one request across services and expose where latency actually goes. OpenTelemetry gives one vendor-neutral instrumentation layer for all three.
SLOs and error budgets
Define service level objectives from user-visible behaviour — availability and latency percentiles, never averages. The gap between the objective and 100% is an error budget: spend it on shipping features, and freeze risky changes when it is exhausted.
Operating
Alert on symptoms users feel, not on every resource threshold, or on-call drowns in noise. Practise runbooks, use blameless postmortems with action items that get scheduled, and design for graceful degradation: timeouts, retries with jitter, circuit breakers and bulkheads.