Any tips for debugging connection pool timeouts?

[RobertCraigie]

Hi, I'm a maintainer of the openai-python package which recently switched over to using httpx internally 🙏

We're seeing some reports from users regarding connection pool timeouts: openai/openai-python#769. Unfortunately no users have been able to share a reproduction yet and I haven't managed to reproduce the problem myself.

From looking through debug logs for httpx and httpcore I haven't seen anything explicitly mentioning the connection pool handling, just when connections are started / closed. I've also gone through our code and didn't find any obvious place where we're leaving connections open.

Any tips for debugging this would be greatly appreciated!

[karpetrosyan]

Hi @RobertCraigie!

Here are some suggestions:

• Update to the latest httpx and httpcore versions
• Do not handle "asyncio.Cancelled" exceptions (Sometimes it can break the client's state).
• Use Add support for connection pool monitoring httpcore#702 branch for connection pool monitoring.

If there is a timeout issue, it comes from httpcore, so you can see the existing httpcore issues and maybe you can find something interesting.

[RobertCraigie]

Thanks Kar, any idea when that PR would be merged / released?

[karpetrosyan]

I can't really say when it's going to be released, but I think it can't help you without the minimum code that can reproduce the issue.

[RobertCraigie]

Okay thanks, your branch has been very helpful in debugging :)

I still haven't managed to reproduce the reported connection pool timeouts but I am seeing some strange behaviour with read timeouts randomly happening - #2941