OpenTelemetry Distributed Tracing

OTel tracing across two services.

Step 1: SDK in service A

Service A is the entry point of the trace. Wire the SDK in, name the service, and turn on auto-instrumentation for the HTTP server.

• Install. The language-specific OTel SDK plus the auto-instrumentation package for your HTTP framework.
• Tracer. Initialise once at startup with a stable service.name resource attribute.
• Auto-instrument. The HTTP server middleware emits a span per request without app code changes.
• Exporter. Point at the local Collector via OTEL_EXPORTER_OTLP_ENDPOINT; do not export to backends directly.

Step 2: SDK in service B

• Same as A but different serviceName.
• Instrument HTTP client (auto-injects W3C Traceparent).

Step 3: Collector + Jaeger

The Collector sits between your apps and the trace backend. Always export through it so you can change backends without redeploying services.

• Compose. Run otel-collector and Jaeger via Docker Compose; the standard demo file in the OTel repo works as-is.
• Pipeline. Receivers ingest OTLP from your apps; processors batch and tag; exporters forward to Jaeger.
• Why Collector. Sampling, tagging, and backend swaps live here, not in app code; one config to update.
• Verify. http://localhost:16686 shows Jaeger; service names you set in Step 1 appear in the dropdown.

Step 4: Trace propagation

Distributed tracing only works when the trace context survives the network boundary between A and B. W3C Traceparent is the standard.

• Outbound. Service A's HTTP client auto-injects traceparent on every outgoing request.
• Inbound. Service B's HTTP server middleware reads traceparent and starts a child span under it.
• Result. Jaeger UI shows the cross-service trace as one timeline with both service spans nested.
• Async caveat. Background jobs and message queues need explicit context propagation; auto-instrumentation does not always cover them.

Antipatterns

• Manual instrumentation everywhere. Use auto-instrumentation libraries.
• No sampling. Cost.
• Skipping context propagation in async. Trace breaks.

What to do this week

Three moves. (1) Run the tutorial end-to-end on your own laptop / sandbox. (2) Apply the pattern to one production workload. (3) Document the variations you needed; share with the team.