Constrain directory in FileStorage for safety?

[jas-ho]

For discussion: FileStorage currently allows arbitrary paths as directory. This could end up with surprising effects if users mistype.

It might be safer to constrain the storage location to lie within a specified part of the file system. Similarly to how all things the transformers library writes to disk end up inside .cache/huggingface/hub.

One option for handling this functionality is to use https://github.com/cthoyt/pystow (which is used e.g. by the excellent PyKeen library)

[gavento]

Hey Jason, thanks for the idea. Not sure we could use it that well here, though.

What directory limitations would you suggest? I think many are unfortunately a bit too limiting:

• Storing in a local subdir in jupyter notebooks (e.g. logs) - seems low risk, mistype creates a new local dir
• Storing in a mounted gdrive in google colab - some special path may be needed anyway
• Storing in a running experiment app - then you likely want the results to be local to your experiment run.

[jas-ho]

Storing in a local subdir in jupyter notebooks (e.g. logs) - seems low risk, mistype creates a new local dir

I don't think we currently restrict to local subdirs, right? I could just accidentally type an absolute path (not saying this is super likely .. but we don't know how this ends up being used)

[jas-ho]

thanks for the discussion points @gavento ! This discussion might in fact be related to a question how different experiments should collect their logs and whether it should be up to the user to create a new storage for every experiment or whether we want an abstraction "experiment" in between "storage" and the different contexts

[gavento]

I'm going to close this for now. Storage of context logs in different environments will come up more, but 1. I think that having a standard global directory (e.g. ~/.config/interlab/...) is not a good default (even though possibly useful for someone) and 2. integrations with experiment instrumentation (like Hydra, e.g. proposed in #30) are better based on concrete use-case or concrete framework.