Make plotly an optional dependency

[timmens]

With optimagic 0.5.3 we have added multiple backends for our plotting functions. Because of that I would argue that we do not need a mandatory plotting library as a dependency.

Example: Matplotlib users can now use matplotlib as the optimagic plotting backend, but plotly is installed into their environment anyways.

[janosg]

I think we need a default plotting backend that is a mandatory dependency. Otherwise casual users would get annoying error messages. But it should probably be matplotlib.

[timmens]

I see the UX concern, but I’m still in favor of not having a mandatory plotting dependency. We can pick a default backend when one is available and show a clear, actionable error the first time someone tries to plot without one installed (I also don't think that would be too annoying, especially since it would only happen once). Some other arguments:

• Many widely used libraries follow this pattern already (e.g. pandas, scikit-learn).
• Making matplotlib mandatory weakens the point of having multiple backends: users who only want plotly still pay the dependency cost of matplotlib.
• The extra dependency weight can be non-trivial for CI, servers, containers, and headless environments.

Ultimately, I do not mind that much however.

[janosg]

I'm really not sure what's the right decision here. I understand all the points and they make absolute sense for more advanced users but having matplotlib in the environment rarely causes any noticeable problems but going through the getting started tutorial and getting an annoying error at one of our main selling points could deter beginners.

[janosg]

I looked at how pandas does it. There the default plotting backend is matplotlib. So if I call df.plot in an environment that has no matplotlib but e.g. plotly installed I would get an import error for matplotlib unless I explicitly change the backend to plotly. Would you replicate this or would you say we should try to detect installed backends?

[timmens]

I have a slight preference towards replicating pandas behavior.

[janosg]

After a short discussion we decided:

1. There is no mandatory plotting dependency
2. We have an adaptive default that checks which plotting libraries are installed; The order of priority is [plotly, matplotlib, ...]. Plotly comes first because if it is installed it is likely that a user wants to use it whereas matplotlib might have come in as a dependency of other packages. Moreover, it provides the best backwards compatibility.

Examples

• No plotting library installed -> NotInstalledError
• Plotly and matplotlib installed -> plotly is used
• Plotly installed -> plotly is used
• Matplotlib installed -> matplotlib is used