When will a child process be stuck when running multiple workers?

[papa99do]

Hello community,

I was trying to test the robustness of multi-process management feature of uvicorn. In the document, it mentioned that when a child worker process is accidentally stuck, the parent process will kill it and spawn a new one. I was wondering how I could simulate this feature using a dummy FastAPI app.

I checked the code where this gets implemented here. So my understanding is that the parent process will periodically ping each child process, with a 5s timeout. If child process does not pong, it will get killed.

I was wondering what could cause child process not to pong. I created an API method in FastAPI:

@app.get("/async_wait")
async def async_wait():
    while True:
        time.sleep(100)

This method is intentionally implemented in a way to block the event loop. I expect the main thread of the child process to sleep forever and cannot respond pong. But it seems that it's not working as expected. Is my understanding incorrect? How can I simulate a stuck child process? Thanks

cc @abersheeran

[papa99do]

I realised that the blocking async_wait function is not going to stop pong since always_pong is running in a daemon thread in the worker process, and the GIL is released when running time.sleep(100) so the daemon thread always gets chance to run pong.
I'll give this a try:

def long_running_task():
    while True:
        pass  # Busy loop