Bug

BaseReload.restart() and BaseReload.shutdown() in uvicorn/supervisors/basereload.py call self.process.join() without a timeout:

self.process.terminate()
self.process.join()  # blocks forever if worker has open connections

If any client holds a long-lived connection (SSE, WebSocket, chunked streaming), the worker will not exit after receiving SIGTERM — it waits for those connections to close first. process.join() then blocks indefinitely, the reload path never reaches get_subprocess(...), and no new worker is ever started. From the outside, hot-reload silently hangs.

Affected versions

0.30.6 through 0.47.0 (current main as of May 2026).

Reproduction

1. Create a minimal FastAPI/Starlette app with an SSE or WebSocket endpoint that holds a connection open.
2. Connect a client so the connection stays open.
3. Edit any watched file to trigger a reload.
4. Observe: WatchFiles detected changes is logged, the worker receives SIGTERM and logs its shutdown, but no new worker ever appears.

Expected behaviour

The old worker is given a grace period to finish. If it has not exited by then it is force-killed, and the new worker is started immediately afterwards.

Actual behaviour

process.join() never returns. The supervisor thread blocks forever. No new worker is started.

Proposed fix

Reuse timeout_graceful_shutdown (already a config option) as the join timeout, falling back to 5s, then force-kill with SIGKILL if the process is still alive:

def restart(self) -> None:
    if sys.platform == "win32":
        self.is_restarting = True
        assert self.process.pid is not None
        os.kill(self.process.pid, signal.CTRL_C_EVENT)
        sys.stdout.flush()
    else:
        self.process.terminate()
        timeout = self.config.timeout_graceful_shutdown or 5
        self.process.join(timeout=timeout)
        if self.process.is_alive():
            logger.warning("Worker did not exit after %ss, sending SIGKILL", timeout)
            self.process.kill()
            self.process.join()
    self.process = get_subprocess(config=self.config, target=self.target, sockets=self.sockets)
    self.process.start()

Apply the same pattern to shutdown().

timeout_graceful_shutdown is already the configured grace period for lifespan shutdown, so reusing it here is semantically consistent and requires no new config surface.

Workaround

Add a threading.Timer in your app's lifespan shutdown handler that calls os._exit(0) after N seconds. This should not be necessary.