Alert on postgres_logs 'checkpoint complete' with write= over a few seconds - would have fired BEFORE the 2026-09-04 student-visible 500s
RESCUED FROM MARS-691, which db-mars-cc retired on 2026-09-13 (GO from pmmaster-evolutiva-cc). Filed so it does not die with the reminder that carried it - the same evaporation MARS-691's own retirement note warned about. FILED-NOT-STARTED; fleet stop of 2026-09-12 02:56Z stands, and my whole queue is parked pending Elazar's ruling on the bin-venus-ca handoff. WANT: an alert on Mars postgres_logs lines 'checkpoint complete' whose write= field exceeds a few seconds. db-mars-cc's evidence: on 2026-09-04 this signature was present BEFORE any student saw a 500, so it is a leading indicator, not a post-mortem one. WHY IT IS AN ALERT AND NOT A REMINDER, which is the reasoning that retired MARS-691 and should not be re-derived: the conditions worth acting on here are INCIDENT-triggered, not time-triggered. A daily nudge to re-measure a healthy system returns clean every day forever at a DM a day across three agents - a recurring per-invocation cost with no possible yield. Fire on the event; never ask a human to go and look. SHAPE: same unit pattern as the MARS-703 WAL-archiver monitor - mars-first, end to end (unit, credential, alert, heartbeat, forced red) before any replication to other refs. Do not build four instances at once. THRESHOLD IS NOT SPECIFIED AND MUST NOT BE COPIED. 'a few seconds' is db-mars-cc's wording, not a derived line. Derive it from an OBSERVED baseline of the checkpoint write= distribution on mars, and print the baseline alongside the threshold on every run; an underivable threshold is UNKNOWN, never a fallback. See reference-a-threshold-copied-from-a-cadence-goes-stale and reference-a-guc-is-not-a-baseline. VERIFICATION WHEN STARTED: force the red against a synthetic slow-checkpoint line before accepting it, and assert on the ALERT FIRING, not on a clean run - a monitor that has never been seen to fail is a green line, not a control. Also confirm the empty path writes evidence it ran (the VENUS-398 lesson: a timer that never fires must not be indistinguishable from one that fires and finds nothing). LANE: Mars DB rows moved to db03-mars-ca under Elazar's runtime switch (db-mars-cc at 24% weekly vs codex 99%), pmmaster ruling, gated on pre-apply review by db02-mars-cc. Confirm the current holder before starting.
Activity
-
Parked: fleet stop stands and my queue is parked pending Elazar's ruling on the handoff. Rescued from MARS-691 at retirement so it does not evaporate.
-
bin-venus-cc