What plotting backends are worth maintaining long-term?

[aloctavodia]

With the redesign of ArviZ-plots, it is much easier to maintain alternative plotting backends. Still, some tasks may not be trivial. In particular, we have observed that bokeh is usually the trickiest of the three. I'm not sure what bokeh brings to arviz plots that Plotly doesn't; maybe others can provide some insight. Additionally, in terms of popularity, it seems that plotly is much popular than bokeh (~6 times more downloads on pypi stats).

Some questions... Should we keep supporting bokeh? Why? And in more general terms, what should be the criteria to consider having a specific plotting backend? What other backend, if any, should we support?

[jessegrabowski]

My 2 cents is that I exclusively use matplotlib, but I'm very interested in transitioning to using plotly more. I haven't ever used bokeh, and so wouldn't notice if you dropped it

[ahartikainen]

Never been a fan of plotly. I mostly use matplotlib.

I think recently holoviews has taken more of my time than pure bokeh.

[fonnesbeck]

+1 for replacing Bokeh with Plotly. Having one interactive backend is always nice. Could consider Altair as an alternative.

[tomicapretto]

I would go for Matplotlib and Plotly (one static and one interactive)

[jordandeklerk]

I've only ever used Matplolib with no issues. Never used Bokeh so I'm not sure what that brings to the table. I've used Plotly for other visualizations not involving Bayesian modeling and didn't hate it. I think it could be nice as an option if needed.

[theorashid]

I always use matplotlib, never bokeh. I would back up Chris to consider Altair as an alternative to plotly.

[ahartikainen]

One thing before making decicions, is to make sure that these libraries supports handling of large data.

I personally work with datashader+bokeh (actually through holoviews) most of the time when I work with large data. Datashader does support atleast matplotlib. Would make sense to check that Plotly has similar support.

Altair is nice, but it can not handle large data.

And when we talk about large data we mostly mean scatter / parallel coordinates plots (multiple plots + connected with selection) where you can have your full mcmc sample visualized at the same time.

[schlich]

would love to see altair support, especially for its marimo compatibility!

It seems to do OK with larger datasets with the vegafusion add-on

[juanitorduz]

+1 to Matplotlib.

Altair is the polars plotting backed so It seems its getting a lot of momentum (https://github.com/pola-rs/polars/blob/py-1.32.0/py-polars/polars/dataframe/frame.py#L631-L726) . I like it better than plotly even though I do not do interactive plots a lot.

[schlich]

Never been a fan of plotly. I mostly use matplotlib.

I think recently holoviews has taken more of my time than pure bokeh.

speaking of holoviews, holoviz might be the closest think to a 'central' library that can translate to and from various formats. In addition to being built on grammar of graphics, it might be the most generalizable/extendable. It currently supports the same 3 plotting backends that Arviz does and can take pandas/polars/xarray/dask and more as input

[OriolAbril]

And in more general terms, what should be the criteria to consider having a specific plotting backend?

I think the main point is the active maintainers agree with maintaining it, which will generally mean there are backends that would be potentially nice to have but not worth the extra work. IMO, bokeh is borderline here right now many common operations require much more effort in bokeh (e.g. we can't set the figure size or the ratios between columns/rows when creating a figure, sharing axis is also 100% manual on our side, it has no figure level legends...).

A secondary point is being compatible with our workflow and assumptions. One of the key features of ArviZ is taking xarray Datasets as input which means there are a couple places where we consciously differ from grammar of graphics/dataframe based declarative plotting. Libraries that are too closely linked to that might not be compatible with arviz-plots or require a level of wrangling data into dataframes that it is better to offer helpers like dataset_to_dataframe to add to .to_dataset and let users interact with those libraries directly.

Altair/vega might be in this category (note I have never used either of them). I did look at holoviz a few years back and it only supported plotting with DataArrays (some examples do say "xarray dataset" and use _ds in the variable name, but it is still a DataArray), so I decided the simplification on the things it helped with wasn't worth the complication added on the things it didn't support. That being said, it would be nice to be proved wrong if it helps simplify and extend ArviZ