You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
List endpoints collapse under a single 25s poller: 1000-row operational view builds saturate the API (supports #334) #335
Container subactor-planfile-coding-recovery-v2 (serving 127.0.0.1:18769): 204% CPU, 8.3 GiB; GET /tickets?...limit=1000&view=operational every ~25s from one consumer (subactor/platform hr-control, see subactor/platform#690) drives list endpoints into 6s+ timeouts for EVERYONE — fleet campaigns fail planning (planfile_unavailable:TimeoutError), ticket creation returns 500 intermittently.
Restart only partially helped (2-6s responses resume immediately): the load is chronic, the view rebuild is O(thousands) per request.
Ask (supports open PR #334 'serve the last complete index')
Additional evidence (2026-11-02 15:2x): this is a memory accumulation, not just CPU. Container ran 46h: 8.3 GiB → 12.8 GiB RSS with list endpoints at 6s timeouts and 204% CPU. After restart: 3.2 GiB and 0.3-3s responses — degradation resumes within minutes under the same load (hr-control polling unchanged, see subactor/platform#690). Also observed a full ASGI traceback with 500 on POST /tickets during load. Mitigation in place: a host watchdog restarts the container when list probes time out twice (every 10 min) — removed once a real fix lands.
Activity
tom-sapletta-com commentedon Nov 2, 2026
Additional evidence (2026-11-02 15:2x): this is a memory accumulation, not just CPU. Container ran 46h: 8.3 GiB → 12.8 GiB RSS with list endpoints at 6s timeouts and 204% CPU. After restart: 3.2 GiB and 0.3-3s responses — degradation resumes within minutes under the same load (hr-control polling unchanged, see subactor/platform#690). Also observed a full ASGI traceback with 500 on POST /tickets during load. Mitigation in place: a host watchdog restarts the container when list probes time out twice (every 10 min) — removed once a real fix lands.