Detect and diagnose hung voice workers from outside the event loop

A voice worker can hang without exiting. The container stays running and the deploy looks healthy while health checks time out, no calls start and the remaining runners fill up. An in-process asyncio heartbeat freezes with the loop, and an on-failure restart policy only replaces processes that exit. Run the server under a parent supervisor that probes a static liveness endpoint every few seconds. After several failures it requests a stack dump, sends SIGTERM, then SIGKILL, so the platform restarts it. To learn what blocked the loop, use a watchdog thread or signal-triggered faulthandler dump, since an in-loop lag monitor can only report after the loop is free again. One dump showed garbage collection inside the inbound resample path. Alert on missing telemetry as unhealthy, not OK.

Applies when: a Python realtime voice service (Pipecat, LiveKit agents) shares one event loop across calls. Keywords: Pipecat event loop latency blocking watchdog stack dump faulthandler liveness supervisor zombie worker restart policy LiveKit