Description
Migrated from GitHub issue #165 (open at the 2026-08-14 tracker migration; full original body in archive/github-issues-2026-08-14.json).
service.version is set on the OTLP metrics resource (the MeterProvider’s Resource), so it is promoted to a service_version label on every emitted metric series. Per the OpenTelemetry → Prometheus compatibility spec that is a deviation: only service.name (+ service.namespace) → job and service.instance.id → instance are meant to become labels. Every other resource attribute — service.version included — belongs on an info metric (target_info, or a *_build_info gauge), per the OpenMetrics 1.0 convention. Promoting a resource attribute to per-series labels is a documented, non-default opt-in.
Impact: each new build mints a whole new series set. After a redeploy the old- and new-version series coexist for the query-lookback window, so any sum-style panel adds both (a transient multiplier), and active-series cardinality grows with the number of versions ever seen. Repos on stable release tags hit the cardinality growth but rarely see the doubling.
Shape of the fix: keep service.version on the OTel resource (semconv-correct, flows to target_info); stop it becoming a per-series label; rely on the existing *_build_info{version=...} gauge joined via group_left where a panel needs it.
Cross-repo consistency pass — the detailed spec citations are in rknightion/graph2otel#104. Sibling issues: tailscale2otel#187, opnsense-exporter#270.
Acceptance Criteria
- #1 service_version no longer appears as a label on ordinary metric series (only on target_info / *_build_info)
- #2 Version stays queryable via the build-info gauge, with a documented group_left example
- #3 Dashboards and alerts under deploy/grafana/ audited; any query aggregating over or filtering on service_version is repointed at the info metric
- #4 The OTLP-to-Prometheus resource-attribute convention is noted in the telemetry setup docs
Definition of Done
- #1 make gate
- #2 go test -tags acceptance ./internal/app/ (only if a §9 acceptance seam changed)
Implementation Notes
GitHub issue #165 was deleted on 2026-08-14 once this task took over the work — gh issue view 165 now 404s. Do not go looking for it. The full original body (2,267 bytes, no comments) is in the archive:
jq ".[] | select(.number == 165)" archive/github-issues-2026-08-14.jsonIt is not in the closed-issues index doc either, because that table indexes the closed set and this one was open. The archive plus this task are the whole record.