Improve the context of asyncio slow log reporting

Ideas
1

Hi all,

Wanted to kick off a discussion for an improvement to CPython and asyncio that I view as critical to enabling debugging for server environments.

My issue is this: when debug mode is enabled for the Python event loop, is it somehow possible to get slow logs in the event loop implementation to tell us more about the context the task is running in?

The motivation is that in many asyncio server frameworks like FastAPI, it is critical to not block the event loop but using asyncio’s debug mode provides output that’s not helpful.

Consider the following FastAPI server:

# server.py
# run with PYTHONASYNCIODEBUG=1 uvicorn server:app --reload --port 8000 --host 0.0.0.0
import time

from fastapi import FastAPI

app = FastAPI()


@app.get("/")
async def root():
    time.sleep(2)  # intentionally block event loop for 2 seconds
    return {"message": "Hello World"}

If you curl http://127.0.0.1:8000/, you’ll see something along the lines of:

Executing <Task finished name='Task-4' coro=<RequestResponseCycle.run_asgi() done, defined at /Users/mike.sukmanowsky/code/z/z/.venv/lib/python3.13/site-packages/uvicorn/protocols/http/h11_impl.py:401> result=None created at /Users/mike.sukmanowsky/code/z/z/.venv/lib/python3.13/site-packages/uvicorn/protocols/http/h11_impl.py:250> took 2.006 seconds

This is consistent with what uvloop and other event loop implementations do here.

But from this output I cannot determine:

  1. What endpoint this occurred on
  2. The full stack trace of my user-defined code that might point me to the culprit that’s blocking the loop

For ASGI frameworks like FastAPI, blocking the event loop effectively means death for concurrency.

I’m unaware of any way to make outputs more useful here, but I’m very much hoping either 1) I’ve missed an obvious way to do this or 2) this inspires some discussion that could lead to changes that’ll eventually help all Python asyncio users better troubleshoot issues like this.

Thanks!
Mike

2 Likes
2

I’d love some sort of protocol where event loops could be supplied some sort of callback, to be invoked if a step of a task takes too long. For example, I imagine a tracing library could mark the current span as interesting, to examine later.

2 Likes