API: Add entrypoint for plotting

[TomAugspurger]

Libraries, including pandas, register backends via entrypoints.

xref #26747

cc @datapythonista @jakevdp @philippjfr

[datapythonista]

Couple of comments.

[jreback]

this seems like a very complex solution

why is not the simpler

plotting.backend = ‘altair.pandas_plot’
plotting.backend = ‘pandas._matplotlib.plot’

not much more straightforward here?

eg this is a standard import as passed to importlib

[TomAugspurger]

We want to decouple the backend name (altair) from the implementation. Altair doesn't want and shouldn't need a top-level altair.pandas_plot just to implement this backend (which may even be implemented in a separate package).

[jreback]

and how does my suggestion not do this is? it is fully decoupled

[TomAugspurger]

I guess I don't understand your suggestion then. You say

plotting.backend = ‘altair.pandas_plot’

which seems to require a top-level altair.pandas_plot?

[jreback]

not at all
these are examples of what you could pass; it could be called anything
that’s the point - we don’t care what it’s called
and no pandas code would have to change for this to be added

[jakevdp]

The reason I proposed this is because I believe it's cleaner for the implementation's code path to be separate from how the user specifies it. My goal would be for a user to write

plotting.backend = 'altair'

and for this to cause the plotting backend to use Altair.

Before this solution, that would require me to pollute the Altair package's top-level namespace with APIs that are irrelevant to users of the package. I think entrypoints is a clean solution to that problem (decoupling of specification from code paths), and one that is in common use among packages in the jupyter ecosystem.

[jakevdp]

I just want to avoid a situation where users have to remember, "for Matplotlib use plotting.backend = 'matplotlib'", "for Bokeh, use plotting.backend = 'bokeh._pandas_api", for Altair use plotting.backend = 'altair_pandas_backend'"...

Seems like it would lead to more confusion than necessary.

[TomAugspurger]

Thanks @jakevdp. Agreed entirely.

[jreback]

ok I see your point @jakevdp and I agree that is nice to de-couple the user supplied key to plotting.backend, but then again it does couple the api of the package to pandas directly, meaning we then need to make an update in pandas to actually use a new package / api.

So this is nice, I would then also allow a fully qualified module.function call to succeed

[TomAugspurger]

it does couple the api of the package to pandas directly, meaning we then need to make an update in pandas to actually use a new package / api.

I may be misunderstanding, but I don't see how the change to use entrypoints affects that. If you're saying that we'll need to update pandas so that set_option('plotting.backend', 'altair') works, then that's not correct. The only change required after this is that altair (or pdvega, whatever package) adds an EntryPoint for pandas_plotting_backends in its setup.py.

Basically, pandas adds an EntryPoint "group" in our setup.py. And third party libraries add items to that group in their setup.pys. Pandas and the library don't need to talk to each other directly to register a backend. We talk through pkg_resources.

So this is nice, I would then also allow a fully qualified module.function call to succeed

We do allow that (see test_register_import).

[jreback]

Basically, pandas adds an EntryPoint "group" in our setup.py. And third party libraries add items to that group in their setup.pys. Pandas and the library don't need to talk to each other directly to register a backend. We talk through pkg_resources.

ok so if this works then that's great. +1.

[TomAugspurger]

Yep, this should work out just fine.

[jreback]

@datapythonista ok on this?

[datapythonista]

lgtm, just couple of small things in the docstring.

[TomAugspurger]

Thanks, fixed.

[jreback]

thanks @TomAugspurger

[jbrockmendel]

this is failing for me locally

[TomAugspurger]

What are you seeing?

[TomAugspurger]

Oh, I'm guessing it's a KeyError: 'pandas_plotting_backends'?

You'll need to re-run python -m pip install -e . in your pandas directory. This adds the entrypoint.

[jbrockmendel]

Yep that does fix it, thanks