The OpenMetrics Standard

Why a New Format Was Needed

The original Prometheus text exposition format was designed purely for metrics. It had no concept of attaching trace context to a data point. OpenMetrics is the standardized evolution of the Prometheus format — now an official CNCF standard — and it adds first-class support for Exemplars.

When Prometheus scrapes a target that exposes OpenMetrics format, it can read and store the Exemplar data alongside the metric values. This requires Prometheus to be configured with --enable-feature=exemplar-storage (Prometheus 2.26+).

The Syntax

A standard Prometheus histogram bucket looks like this:

http_request_duration_seconds_bucket{le="2.0"} 847

This tells you: 847 requests completed in under 2 seconds. That's the aggregate.

With an Exemplar in OpenMetrics format, the same line becomes:

http_request_duration_seconds_bucket{le="2.0"} 847 # {trace_id="abc123",span_id="def456"} 1.7 1686000000.000

Breaking down the Exemplar portion (everything after #):

• {trace_id="abc123",span_id="def456"} — the labels linking to the trace
• 1.7 — the exact value of the sampled request (1.7 seconds)
• 1686000000.000 — the Unix timestamp of the observation

The # character is not a comment here (as it would be in standard Prometheus format). In OpenMetrics, # followed by {...} is the Exemplar syntax. This is a deliberate design choice to maintain backward compatibility — older parsers that don't understand OpenMetrics will treat the Exemplar as a comment and ignore it safely.

Querying Exemplars in Prometheus

Prometheus stores Exemplars in a separate in-memory ring buffer (not in the TSDB). You can query them directly via the /api/v1/query_exemplars endpoint, passing the metric query and a time range. The response includes the Exemplar labels, value, and timestamp for each metric series that has Exemplar data in that window.