[FR] External Data Support

[blink1073]

What's the problem this feature will solve?

Project Jupyter uses data_files for discovery of runtime extensions. This feature was deprecated without a viable alternative.

Describe the solution you'd like

The flit project recently added support for external_data, which is a simplified and constrained version of data_files. We could support an equivalent feature in setuptools, with caveats about when it is appropriate to use it.

Alternative Solutions

Project Jupyter and other current consumers of data_files could instead switch to using a flit backend, which we have been actively exploring.

Additional context

Project Jupyter uses data_files to allow libraries to provide shared setting and static assets files that are consumed by Jupyter at runtime. We have explored using entry_points for this purpose, but it did not scale well.

Code of Conduct

• I agree to follow the PSF Code of Conduct

[abravalheri]

Hi @blink1073 thank you very much for putting up this issue.
I think this request is very reasonable. Is there any other requirements for the jupyter use case?
If I understood correctly the idea is to pick-up a directory and then replicate the file tree under it in the appropriate installation schema, right?

Regarding the "deprecated" status of data_files, my personal opinion is that "it's complicated"...

The behaviour of installing files to arbitrary locations as it used to do in the distutils and then in the easy_install era is definitely deprecated. But I don't see the feature itself disappearing in the future (just that users have to understand its limitations and know exactly when to use it).

Other maintainers might disagree, but I do appreciate that there are genuine use cases for it (like jupyter).

[blink1073]

Jupyter looks in <sys_prefix>/share/jupyter and <sys_prefix>/etc/jupyter for its files. As long as we can target those two directories we are fine.

I think the current semantics of data_files are what is complicated: it can target absolute paths, and it is very tricky to use for nested files. We use a helper function to make it glob-friendly.

I think with a new API we can offer a simpler, safer alternative and eventually remove data_files support altogether.

[blink1073]

cc @minrk @jasongrout @bollwyvl

[blink1073]

If there's buy-in I can take a crack at implementation.

[abravalheri]

I am sold on that.

It would be nice to hear the opinion of @jaraco.

---

For more information, there is a in-depth discussion about why jupyter needs data dirs in the discourse, this is also relevant.

If I understood correctly jupyter needs to be able to share/read/use files from extensions. These files need to be found in the disk in a reliable way, even if the package providing them has not been imported yet. The performance of searching for entry-points on disk also seems to be an issue for jypyter.