How do you implement network observability in a distributed system?

Three pillars: Metrics (request rate, latency percentiles, error rate - Prometheus + Grafana), Logs (structured, correlation IDs, centralized - ELK/Loki), Traces (distributed tracing - Jaeger/Zipkin). Network-specific: RED method (Rate, Errors, Duration), TCP metrics (retransmits, RTT), DNS resolution time. Implementation: OpenTelemetry for instrumentation, sidecar proxies for service mesh metrics, eBPF for kernel-level visibility. Alerting on SLOs, runbooks for common issues. Golden signals per Google SRE.