health_check raises ValueError: process object is closed after inference subprocess exits (missing same guard as _send_kill_signal, see #4500/#4694)

[jeser-aj]

Bug Description

worker.py's health_check HTTP handler calls self._inference_executor.is_alive() without guarding against the underlying multiprocessing.Process already being closed. Once the shared inference subprocess exits, SupervisedProc._supervise_task() calls self._proc.close() (supervised_proc.py:300). Any health_check request that arrives after that point raises an unhandled ValueError: process object is closed, which propagates out of the handler as a 500 instead of the intended 503.

This is the same race already fixed in _send_kill_signal() (see #4500, #4694) — that fix wrapped the identical self._proc.is_alive() call in try/except ValueError, but health_check() in worker.py was never updated with the same guard.

Stack Trace

Traceback (most recent call last):
  File "/usr/local/lib/python3.11/multiprocessing/process.py", line 101, in _check_closed
    raise ValueError("process object is closed")
ValueError: process object is closed

  File "livekit/agents/ipc/inference_proc_executor.py", line 104, in is_alive
    return self._proc.is_alive()
  File "livekit/agents/worker.py", line 363, in health_check
    if self._inference_executor and not self._inference_executor.is_alive():
  File "aiohttp/web_protocol.py", line 510, in _handle_request
    resp = await request_handler(request)

Environment

• livekit-agents 1.2.18 (confirmed still present, unfixed, in latest 1.6.7)
• Python 3.11
• Linux (k8s)

Steps to Reproduce

1. Run a worker with a plugin that registers an InferenceRunner (e.g. livekit.plugins.turn_detector.english.EnglishModel), which spawns a shared inference subprocess for the worker.
2. Cause the inference subprocess to exit (crash, OOM, unresponsive ping/pong timeout, etc.) without the whole worker process restarting.
3. Once _supervise_task() runs self._proc.close(), every subsequent hit to the worker's / health-check HTTP endpoint raises ValueError: process object is closed instead of returning a 503.
4. This repeats on every health-check poll indefinitely (e.g. every ~1 min from a k8s probe), since nothing recreates the inference subprocess for that worker instance.

Expected Behavior

health_check() should treat a closed inference process the same as a dead one and return 503 inference process not running, not raise an unhandled exception.

Suggested Fix

Same pattern already used in _send_kill_signal():

async def health_check(_: Any) -> web.Response:
    try:
        inference_alive = not self._inference_executor or self._inference_executor.is_alive()
    except ValueError:
        inference_alive = False

    if not inference_alive:
        return web.Response(status=503, text="inference process not running")

    return web.Response(text="OK")

Impact

• Every health-check poll after the inference subprocess dies throws an unhandled exception and returns 500 instead of 503, which can confuse liveness/readiness probes and floods error monitoring (we saw ~20k occurrences in Sentry over 2 days from a single worker pod).
• No indication in the code that a dead shared inference subprocess is ever respawned within a running worker instance, so this is a persistent failure mode until the pod is restarted.

[tulanisahil1-beep]

Confirmed on current main — same race as #4500/#4694.

After the inference subprocess exits, _supervise_task calls self._proc.close(), then health_check → InferenceProcExecutor.is_alive() → ValueError: process object is closed (500 instead of 503).

Fixed by guarding is_alive() the same way _send_kill_signal already does:

def is_alive(self) -> bool:
    try:
        return self._proc.is_alive()
    except ValueError:  # process object is closed
        return False

[tulanisahil1-beep]

PR: #6673