Implement Distributed Tracing with Jaeger and Tempo

Distributed Tracing implements request-flow visibility across microservices using Jaeger and Grafana Tempo. It defines the core concepts - a Trace as the end-to-end request journey, a Span as a single operation within it, Context as metadata propagated between services, Tags as filterable key-value pairs, and Logs as timestamped events within a span - and documents Jaeger deployment via Kubernetes (Jaeger Operator with Elasticsearch storage) or Docker Compose (jaegertracing/all-in-one with UI, collector, gRPC, and Zipkin-compatible ports). Application instrumentation is shown via OpenTelemetry across three languages: Python/Flask (TracerProvider with a JaegerExporter and BatchSpanProcessor, FlaskInstrumentor, and manual spans with set_attribute calls for things like db.system and db.statement), Node.js/Express (NodeTracerProvider, HttpInstrumentation/ExpressInstrumentation auto-instrumentation, and manual span start/end), and Go (a JaegerExporter-backed TracerProvider registered globally, with span.SetAttributes and span.RecordError). Context propagation across service boundaries uses W3C traceparent/tracestate HTTP headers, injected via OpenTelemetry's propagation.inject() in both Python and Node.js.

Use this skill to debug latency issues, understand service dependencies, identify bottlenecks, trace error propagation, or analyze request paths across a distributed system. It is not for tasks unrelated to distributed tracing or for work in a different domain or tool outside this scope.

Given a tracing task, the skill outputs deployable Jaeger and Tempo configuration, including a Tempo Kubernetes ConfigMap with S3 trace storage and a Deployment manifest, and three sampling strategies with configuration examples: probabilistic, rate-limiting, and adaptive/TraceIdRatioBased.

### Sample 1% of traces
sampler:
  type: probabilistic
  param: 0.01

It also documents Jaeger query syntax for finding slow requests (duration > 1s) or errors (error=true, tags.http.status_code >= 500), log-correlation code that attaches the current span's trace_id to structured log entries, and troubleshooting steps for two common failure modes: no traces appearing (check collector endpoint, network connectivity, sampling config) and high latency overhead (reduce sampling rate, use a batch span processor, check exporter config).

Built on Jaeger, Grafana Tempo, and OpenTelemetry SDKs/exporters across Python, Node.js, and Go, with Jaeger automatically generating service dependency graphs showing relationships, request rates, error rates, and average latencies. Related skills are prometheus-configuration for metrics and grafana-dashboards for visualization, plus slo-implementation for latency SLOs. Ten best practices are stated: sample appropriately (1-10% in production), add meaningful tags like user_id and request_id, propagate context across all service boundaries, log exceptions in spans, use consistent operation naming, monitor tracing overhead (under 1% CPU impact), set up alerts for trace errors, implement distributed context/baggage, use span events for milestones, and document instrumentation standards.

Engineers operating microservices who need to trace a request across service boundaries to diagnose latency, failures, or dependency issues - setting up Jaeger/Tempo infrastructure, instrumenting applications with OpenTelemetry, and tuning sampling so tracing overhead stays acceptable in production.