Use an alternative to diskcache

[glemaitre]

I just look at diskcache repository (https://github.com/grantjenks/python-diskcache).

It looks pretty unmaintained. I'm wondering how much we depend on it and if at the time of choosing this dependency, some other alternative came up that are better maintained?

ping @thomass-dev @auguste-probabl @rouk1

[auguste-probabl]

This has been an IRL debate for a while; every time the result is that we'll change when it becomes necessary. Meanwhile, shelve is quite similar to diskcache and in the standard library.

[rouk1]

I would advocate for a file based home made storage.

• stores primitive stuff as pickl dump
• store models as skops dump
• store dataframes as apache arrow
• ...

All this in a .skrore folder with file safe name and a clean hierarchy per version.

This would allow

• power user to just load there data manually as @adrinjalali suggested
• to specify a simple yet powerful exchange standard (it's just a hierarchy and some media types)
• to avoid sql migrations

[adrinjalali]

I like the suggestion, I'd be curious on which of those primitive objects can't be skore dumped though.

[auguste-probabl]

We currently default to pickle for unrecognized objects, including unstabilized skore objects like EstimatorReport.

[thomass-dev]

Leaving from diskcache is one of the next major feature 😄 , and i think it will be planned with the road to HUB PoC.
As said @auguste-probabl , it's been on the table for a long time but not prioritized as it works in state.