The Instrumentation Layer: Connecting Traces to Metrics

The Missing Link

Exemplars don't appear automatically just because you have both metrics and tracing instrumented. The metrics library and the tracing library are independent — they don't know about each other by default. You must explicitly wire them together so that when the metrics library records an observation, it can ask the tracing library: "What is the current Trace ID?"

This connection is called a SpanContext supplier or Exemplar sampler, depending on the library.

How It Works

When a request arrives, the tracing library creates a Span and stores the current Trace ID in thread-local storage (or a context propagation mechanism). When the metrics library records a histogram observation for that same request, it calls the SpanContext supplier to retrieve the active Trace ID and attaches it as an Exemplar label.

The key requirement: the metric observation must happen within the same execution context as the active trace span. If you record the metric outside the span's scope, the SpanContext supplier returns nothing, and no Exemplar is attached.

Language Examples

The wiring pattern is consistent across ecosystems, differing only in API shape. In Go (prometheus/client_golang with OpenTelemetry), you read the active span from the request context and, if it's valid, call the histogram's exemplar-aware observe method with the trace and span IDs attached as exemplar labels — a standard .Observe() call has no exemplar support, so the SDK must expose this dedicated method. In Java (Micrometer with OpenTelemetry), the wiring happens once at registry construction — a DefaultExemplarSampler backed by an OpenTelemetrySpanContextSupplier is passed into the PrometheusMeterRegistry, after which every timer observation automatically captures the active trace ID with no per-call changes to business logic. In Python (prometheus_client with OpenTelemetry), you read the current span's context and, if valid, pass an exemplar dict with the formatted trace ID into .observe() alongside the measured duration.

The Full Requirements Checklist

For Exemplars to work end-to-end, every layer of the stack must be configured correctly:

If any layer is missing, Exemplars silently fail — the metrics still work, but the trace links don't appear.