-
Notifications
You must be signed in to change notification settings - Fork 2
plat 347
Category: Performance
Severity: P1
Status: Fixed
Observed: 2026-09-21 on RTS
The production release was fully built and verified but activation waited on
/api/health reporting active_sessions: 1, even though
in_flight_requests: 0 and no user work was observable. Retained chats and
stale logical ownership can keep that aggregate non-zero, so it is not a safe
deployment gate.
The old policy could wait for an hour and then abandon a valid release. It also did not prevent new turns from starting during the wait, so it was neither a true drain nor a deterministic zero-downtime handoff.
RTS production activation is explicitly breaking. After the build, tests, and
release-asset checks pass, deployment swaps the release and restarts the three
user services immediately. Clients reconnect to the new runtime. The deploy no
longer uploads or consumes DRAIN_TIMEOUT_SECONDS.
Build isolation remains unchanged: a failed build never changes the active release, and only RTS user services are restarted.
Auto-synced from docs/ on main. Edit there, not here.