Skip to content

Concurrent-users gauge + job-latency alerting after #38

The two observability items of #2529 ยง3. Both are closed without code in this repo.

Item A โ€” active-users gauge: abandoned

A per-pod count of distinct authenticated users in the last 5 minutes, built in #2957 and removed before merge. It was meant to drive manual capacity tiers counted in users (raise memory, add a pod, need HPA and pgbouncer). The HPA (3โ€“6 backend pods) and pgbouncer now do that, and capacity is measured in CPU per request. People online is known by other means. Its reading also depended on the router's sticky cookie: sum() over pods is exact with it, max() without it.

Item B โ€” job-latency alerting: no change

openshift-app-config#38 (2026-09-07) removed every job-class latency alert. What remained:

  • The 42 s prefill and 184 s upload figures are locust flow times, a trigger then polls. Each request is fast, and a request-duration histogram whose last bucket is 10 s cannot measure a job (1402).
  • Plan prefill's routes enqueue a job or read one job row; the prefill runs outside the request. They belong in route_class="api", so openshift-app-config#74 is closed.
  • Dev's job_poll / job_trigger split has no consumer since #38. It is removed in the ops repo so dev matches stage and prod.
  • Job durations are measured per job_type by #2854. Upload-to-ingested is a pipeline and needs 2049-C4, which stays behind two-maintainer review.