Limit on downloads/space/bandwidth

[pgleeson]

Having trouble installing packages on v2dev. pip installs are killed after ~390MB. Also slower than on v2 live:

[pgleeson]

A quick solution, use --no-cache-dir, e.g. pip install torch==1.11.0 --no-cache-dir

[pgleeson]

Still an issue. Should be possible to install large python packages if there is space on the workspace...

[filippomc]

One major issue would be that once installed they disappear in the next run as it uses ephemeral space in the current pod. Do we want to find a way to use the workspace directory instead and add it to the paths?

[filippomc]

Wonder what's the best implementation of this.

1. Leave it manual but make explicit to the user, they can use something like the pip install --target=/opt/workspace/mylib package_name, which would also give the chance to install in the user's directory
2. By default, pip install goes to /opt/workspace/mylibs
3. Thinking bigger, we might also think about "prepopulated" packages we can attach to workspaces. For instance, we could prepare a "Machine learning" add-on, which technically speaking will attach a volume including pytorch and other potentially related libraries as a read-only volume to attach to workspaces when the flag is selected