uvicorn Child process died

[ch-sander]

shmarql/Dockerfile

Line 39 in 8811fdd

CMD ["uvicorn", "--workers", "4", "--host", "0.0.0.0", "--port", "8000", "app:app", "--log-level", "debug" ]

When loading from store (and only if building the Docker image locally) the worker fails (snippet):

shmarql_dev  | 2024-08-25 21:22:53 DEBUG    Opening store from /data/store
shmarql_dev  | 2024-08-25 21:22:53 DEBUG    Opening store from /data/store
shmarql_dev  | 2024-08-25 21:22:53 DEBUG    Opening store read-only
shmarql_dev  | 2024-08-25 21:22:53 DEBUG    Opening store read-only
shmarql_dev  | 2024-08-25 21:22:53 DEBUG    Opening store from /data/store
shmarql_dev  | 2024-08-25 21:22:53 DEBUG    Opening store from /data/store
shmarql_dev  | 2024-08-25 21:22:53 DEBUG    Opening store read-only
shmarql_dev  | 2024-08-25 21:22:53 DEBUG    Opening store read-only
shmarql_dev  | INFO:     Waiting for child process [8]
shmarql_dev  | INFO:     Child process [8] died
shmarql_dev  | INFO:     Waiting for child process [9]
shmarql_dev  | INFO:     Child process [9] died

Loading from ttl works.

It also works if I run just one worker per CMD ["uvicorn", "--workers", "1", "--host", "0.0.0.0", "--port", "8000", "app:app", "--log-level", "debug" ]

Very odd. Any idea why that is?

Maybe the uvicorn library has changed meanwhile (and is on an older version in origin docker image at ghcr.io/epoz/shmarql:latest)?

shmarql/requirements.txt

Line 3 in 8811fdd

uvicorn>=0.11.3

[ch-sander]

Running uvicorn 0.29.0 with CPython 3.9.7 on Linux for ghcr.io/epoz/shmarql:latest. If changed in

shmarql/requirements.txt

Line 3 in 8811fdd

uvicorn>=0.11.3

, problem solved.

Still, something to look out for.

[epoz]

In uvicorn 0.30 a new multiprocess manager was released, and this caused breakage in shmarql with the uvicorn parent process just dying.
I had to pin the uvicorn to 0.29 at the time, as all newer versions were showing this behaviour.

It looks like newer versions have now fixed this, so we can probably start using the newer versions, but we will have to pin to something a bit higher. To be tested...

[ahirner]

I see it pinned to 0.29.

It looks like newer versions have now fixed this, so we can probably start using the newer versions, but we will have to pin to something a bit higher. To be tested...

Did you test higher versions @epoz ?

[epoz]

@ahirner - not extensively, there was no need for me to use a higher version yet.

Please note that I have started working on a re-boot of shmarql in fasthtml which brings several improvements.

May I ask, are you using shmarql for anything that I should be aware of with regards to breaking changes, or backwards compatibility? To date I thought there was not that many other users.

[ahirner]

Thanks! We are see the same error in a django app of ours. 0.29 can't restart workers (ignores SIGHUP), workers with later versions die. I was mainly looking for hints how to reproduce that, unfortunately no user of shmarql :)

PS: as we speak, we found that granian doesn't suffer from dying workes and can restart them gracefully.

[epoz]

Ah, clear...
FastAPI did work with later versions of uvicorn when I tested it informally, so yes, it seemed this was sorted. But to be on the safe side I left the version pinned in my repo.