Add plotly-based visualization tool

[ihincks]

What is the expected enhancement?

If plotly were added as an option for visualizing a graph, we would get nice interactive features in Jupyter notebooks like showing edge and node information on hover, zooming to regions of busy graphs etc.

There would be performance limits since plotly uses a relatively heavy JSON format+plotting engine. For example, I expect it would choke on a graph with tens of thousands of nodes. But when there are fewer than 1000 nodes, I'd expect it to be a nicer experience than matplotlib.

[IvanIsCoding]

I think someone should make a Plotly integration for us, that would be great. And I see that you have done so in #1554. Congratulations, that is cool! I was not aware that Plotly had dedicated networks support.

With that being said, there is the second aspect of merging #1554 that is maintaining. I personally would not like to maintain it. And this comes from my experience supporting graphviz and matplotlib.

graphviz support was a bare minimum for us. We initially implemented a conversion to the DOT format and delegated most of the work to the dot CLI. There are still some bugs (mostly from escaping strings), but overall it is a low-maintenance system.

Matplotlib has been a different story. I think supporting Matplotlib is very Pythonic and it comes handy for documentation. But there are more issues like #1497. Testing Matplotlib is also not our forte (#1316).

So Plotly falls in the same category. It would be a nice to have, but it is unclear to me if I want to test it and support all its features. Do not get me wrong, it should be done. But I want to keep rustworkx to something I can maintain.

In general, we merge most PRs implementing algorithms because honestly they are not too hard to maintain. As long as the implementation is correct and the tests are good. For example, seldom do I need to worry for Katz Centrality, it's code that is "finished". I found the same not to be true about the plotting libraries and most Python dependencies we have except for NumPy.

This all is to say: the answer to #1554 is not concluded but I am leaning towards "no".

[IvanIsCoding]

Now for a productive next step: I think you should publish a new Python library called either rustworkx-plotly, rustworkx-viz or just rxviz. Make it depend on rustworkx and plotly. #1554 does not access any internal APIs to rustworkx so it could all live separately.

Even with a separate library, I would be happy to add rustworkx-plotly to the README and perhaps even to have a guide like https://www.rustworkx.org/tutorial/introduction.html discussing interactive plots.

[ihincks]

Thanks for your response. Your concern about maintenance is fair, especially when robust testing is challenging.

Please note that plotly, like matplotlib, does not have any layout/planar embedding capabilities that I know about. For example, in the docs you linked to, it's just using scatter plots and line plots to draw the graph (not unlike my PR), and is letting networkx define coordinates.

Along these lines, in case it sways your position at all, I should point out more explicitly that #1554 calls into the existing DOT conversion format rustworkx implements, and also calls into rustworkx's implementation of a spring layout. It then uses these coordinates to draw circles and lines with plotly.

[IvanIsCoding]

I agree that the implementation for the layouts is very clean. 512 lines total for the contribution is quite optimal. It is good code, it is quite useful on its own.

But it has some items that I personally dread: https://github.com/ihincks/rustworkx/blob/b3745a77bef3c60972455b7d4c5d9434933c719c/rustworkx/visualization/plotly.py#L349

My least favorite part of rustworkx are the Matplotlib bugs. And if I add Plotly to a release, people kind of expect it to work and to have some sort of support. Our release cycles have slowed down over the years, so fixing the reported bugs also becomes more painful.

My offer for the future of the code still stands. I think:

• The repository could live in https://github.com/qiskit-community
• We make the rustworkx documentation depend on a pinned version of rustworkx-plotly. We can demo the results in the documentation
• With separate releases, we can iterate rustworkx-plotly at a faster pace and it becomes clearer to people what to expect

Let me know what you think.