[docs] Issue on `deploying-on-slurm.rst`

Documentation Problem/Question/Comment

This SLURM example seems to assume that /tmp is a directory shared by all the nodes. I think this is not generally true.

While testing this example, I got errors of the kind “could not connect to IPC socket”, because the nodes did not share the same temp directory, which contain the communication sockets. To make it work, I had to set --temp-dir to a shared location.

(Created directly from the docs)

[richardliaw]

Yeah, we need to clean up these docs.

temp_dir doesn’t need to be shared. In fact, nothing actually needs to be shared; files can be moved around using rsync.

[lemairecarl]

Doesn’t the temp_dir contain the communication sockets? If that’s the case, the folder obviously needs to be shared.

Since I’m not using Ray myself (I was doing support for someone else), I don’t think I’m qualified to drill down on this and/or edit the docs. I’m withdrawing.