- Shipped
- October 28, 2026 at 6:34 AM UTC
- Author
- Kamo
- Commit
- c873407
livekit-server v1.13.7 kept 202 closed participants of 176 closed rooms alive (their publisher transport's StreamAllocator polled a closed ICE agent every 15 s, 1.5 CPU) until new SIP legs got no media (2026-10-28). Every Python-SDK (single peer connection) participant that closed from 10-25 on leaked, whatever its close reason; Go (livekit-sip) and JS never did. - livekit-server v1.13.10: carries the warp WebRTC deadlock fixes (webrtc-pion #3/#5, ice #4, dtls #1); not v1.13.11 (single-PC data channel changes). prometheus.port, debug_handler on 6790 so a future leak leaves a goroutine profile. - livekit-reaper: every 60 s deletes rooms no live call can be in (empty, phone leg alone 3 min, orphan ear room, older than 6 h, harness rooms after 30 min, no published track 15 min) and removes stuck participants; at most 10 actions a pass, one JSON log line each. Counts leaked participants from livekit-server's log (Loki) and restarts the server only at >= 10 leaked, 0 live rooms, not twice in 6 h, logging the goroutine profile first. - ServiceMonitor for livekit-server, livekit-sip and the reaper; PrometheusRule (leaked participants, rooms/participants high, busy while idle, reaper busy/deferring/down, livekit-sip down); AlertmanagerConfig to SysAdmin@KamoCRM.com. - apply.sh: a deploy that would roll livekit-server or livekit-sip (Recreate, host ports) waits for 0 live rooms (reaper count, 20 min) or fails without touching them.