[howethomas]

Problem

#154 added vcon.uuid to the metric attributes of 17 links. Every distinct uuid is its own time series. The counters and histograms are cumulative, so each worker re-exports every series it has created on every export interval until the process restarts, and metric volume grows without bound. #224 fixed the jq link after it filled a metrics store (over 1M series and about 460M samples an hour from one counter). The other links have the same flaw at lower volume today; the histograms are the worst case, since each series carries a full set of buckets.

Change

• Remove vcon.uuid from metric attributes in analyze, analyze_and_label, analyze_vcon, check_and_tag, datatrails, deepgram_link, detect_engagement, diet, groq_whisper, hugging_face_whisper, hugging_llm_link, openai_transcribe, post_analysis_to_slack, sampler, scitt, tag_router and webhook (27 call sites). Each attrs dict was checked to feed only metric calls. The uuid stays in log lines and span attributes, where per-event detail belongs.
• common/lib/metrics.py: increment_counter, add_updown_counter and record_histogram drop per-vCon keys (vcon.uuid, vcon_uuid, vcon_id, vcon.id) and log a warning once per metric. This is a backstop so a new or out-of-tree link cannot bring the problem back.

Tests

• The CON-6 tests in conserver/tests/test_link_metrics.py asserted vcon.uuid was present. assert_metric_has_attrs now fails if any data point carries a per-vCon key.
• New TestPerVconAttributeGuard covers counter, up/down counter and histogram: three uuids in, one series out.
• Mutation checks: with the guard disabled, the 3 guard tests fail. With the guard disabled and the links reverted, 27 tests fail. With both in place, all 31 pass.

Notes for operators

Dashboards or alerts that group or filter link metrics by vcon.uuid will lose that dimension. Per-vCon lookups belong in traces or logs.

🤖 Generated with Claude Code