Increase or adjust scope of default HTTP timeout

[zanieb]

Our current HTTP timeout is not sufficient for some large packages e.g. torch

• Failed to download distributions with large requirements file #1920
• Repeated timeouts in GitHub Actions fetching wheel for large packages #1912

Increasing the timeout has a downside for people that have network problems while downloading smaller packages, e.g. it will wait longer before surfacing a problem.

What does pip use for a default timeout?

[MichaReiser]

I would expect that a server only sends an HTTP Timeout response if the connection is idle for too long, which makes me wonder if increasing the timeout solves the problem or only hides the root cause.

[zanieb]

Yeah good point, I wonder if we're setting the wrong timeout? We shouldn't be enforcing this timeout when data transfer is actively occurring.

[samypr100]

FWIW the default in PIP is 15 seconds https://github.com/pypa/pip/blob/main/src/pip/_internal/cli/cmdoptions.py#L294, so I'm not sure uv's 30? second default is at fault.

[konstin]

Could it be that we need to reduce the number of parallel requests? It could also be that they timeout because some cpu intensive task is blocking the main thread, esp. with something large like pytorch.

[zanieb]

It seems like the next steps are to:

1. Determine what exactly the timeout we are configuring applies to
2. Create a simple MRE
3. Determine what is happening with the request when the error is raised