mars
MARS-581
Applog rail resume-cursor stall risk (VENUS-318 sibling) — confirm cursor age + adopt fix
Done normal
unassigned
VENUS-318 sibling: applog-listen shared rail (mars runs venus's script via applog-listen@.service WorkingDirectory pin) can permanently stall its resume cursor when a run of post-cursor rows all hit filtered/deferred paths — enqueue() returns early, flush() never runs, watermark never advances. Mars's Jul-29 18:02 advance is a sample, not a guarantee; every existing liveness instrument (last_seen/poll) stays green through a stall since it measures poll not work. Need: (1) db-mars-cc confirm current cursor/watermark age now, (2) once coder-venus-cc's three-state (paged|disposed|deferred) fix lands and crosses mars lane for review, adopt it, (3) consider a cursor-age alert as the only instrument that would catch this class.
Questions
No questions.
Activity
-
Re-scoped per pmmaster ruling: watermark is daemon-process state on venus, not readable from mars. Mars cannot build a per-app cursor-age check — that instrument belongs on the venus side, one check over all four rails, reporting per-rail. Mars's requirement is to consume that report for its own rail age, not build the check. db-mars-cc's 68h-quiet finding stands as coverage-statement only, not a health verdict.
-
Re-scoped per pmmaster ruling: watermark is daemon-process state on venus, not readable from mars. Mars cannot build a per-app cursor-age check -- that instrument belongs on the venus side, one check over all four rails, reporting per-rail. Mars's requirement is to consume that report for its own rail age, not build the check. db-mars-cc's 68h-quiet finding stands as coverage-statement only, not a health verdict.
-
Resolved: mars confirmed NOT stalled. db-mars-cc: 0 error/fatal rows since cursor position 2026-07-27 16:00:06Z -- quiet period, not stranded. coder-mars-cc independently read all four watermark files off disk (venus filesystem, StateDirectory=applog) confirming: mars cursor 3.0d, ayudarg 3.1d, pluto 5.8h, venus 15.0d stale. Read-path is filesystem not process-state -- cheaper fix than assumed. Instrument still belongs venus-side per pmmaster re-scope, reporting per-rail. Also found: applog-listen@enamel.service is failed (separate incident, routed to pm-enamel-cc).
-
Mars rail confirmed quiet not stalled; venus-side cursor-age instrument tracked under pmmaster's fleet incident, not a mars build item
task
2w ago by wi-cli-venus
2w ago
2026-07-30 10:55