reez455g/kerangka/python-atexit-network-io

Network I/O in Python atexit handlers fails with async-backed clients

Symptom

An atexit-registered handler that calls a sync client method (which internally does loop.run_until_complete(...) over aiohttp/asyncio) fails with:

RuntimeError: cannot schedule new futures after shutdown

The rest of the program worked fine; only the exit-time call fails. Often accompanied by Unclosed client session / Unclosed connector warnings.

Cause

By the time regular atexit callbacks run, the interpreter has already begun tearing down thread pools / the asyncio default executor that the client's sync wrapper needs (e.g. for DNS resolution via run_in_executor). Any client built as _run_async(coro) -> loop.run_until_complete(coro) is affected (seen with hindsight-client 0.8.4).

Fix pattern

1. Call the finalizer explicitly at the end of the main code path (__main__, CLI entrypoint, context-manager __exit__), before interpreter teardown starts.
2. Keep the atexit.register(finalize) as a safety net only, and make the finalizer idempotent with a module-level _finalized flag so the explicit call and the atexit call don't double-run:

_finalized = False
def finalize_session():
    global _finalized
    if _finalized:
        return
    _finalized = True
    ...  # network I/O here
if __name__ == "__main__":
    main()
    finalize_session()  # explicit — atexit is too late for network I/O

1. If the finalizer must never crash the program, wrap the network calls in try/except with a [WARN] log (graceful degradation).

Verification

Run the entrypoint and confirm the exit-time write actually landed (e.g. the record is recallable from the store) — absence of a traceback is NOT proof, since swallowed exceptions can hide the failed write.