Plotting plugin ideas

[abkfenris]

During the IOOS DMAC meeting we chatted about generating plotting with Xpublish.

I think we could build on the EDR plugin, as it can already can manage selection and supports various output plugins.

It might however be better to steal the EDR slicing & selection syntax (or use ERDDAP or others), and then extend it to support various plot formatting options. We can probably refactor things in the EDR plugin so that another plugin can use the existing slicing and selection functionality without having to re-invent it.

One question is do we want to support multiple types of plotting and how would we want to do that? Matplotlib, Holoviews, Altair/Vega, Plotly...

We could have separate plugins, and let each plotting library really do their own thing, or we could have a plugin of plugins like EDR is, and have some standardization.

More along the line of a plotting plugin of plugins with shared EDR-like slicing and selection:

• /datasets/fvcom/plot/matplotlib.svg?parameter-name=hs&coords=POINT%28-69.66+43.75%29&p_style=xkcd&p_width=700&p_height=400
• /datasets/fvcom/plot/hvplot.json?parameter-name=hs&coords=POINT%28-69.66+43.75%29p_width=700&p_height=400&p_id=my-plot

Somewhat relevant links:

• Bokeh (Holoviews) JSON output - https://docs.bokeh.org/en/latest/docs/user_guide/output/embed.html#json-items
• Holoviews static export - https://holoviews.org/user_guide/Exporting_and_Archiving.html#exporting-specific-files
• Saving Altair (Vega-lite) charts - https://altair-viz.github.io/user_guide/saving_charts.html

Any thoughts?

[wachsylon]

I am also testing plotting with xpublish and holoviews (hvplot). Here are some issues that I have:

• Idk if it is possible or a "good idea" to open another websocket for a bokeh server, I for myself did not want to go this way. I save the generated plots in a BytesIO object and return a streaming response for that. I could not render this generated plot in notebooks on client side with panel because of a version conflict in the bokeh-javascript versions. only thing that worked was iframing the plot which is... likely not desired.
• I am using the .hvplot command for the datasets. It seems that there are some blocking functions used deep in the code. I had troubles to make it work in an async framework with dask cluster and dask arrays in the background. I now load the data before the plot with an async dask cluster and then apply the .hvplot function in a sync framework still with the use_dask=True kwarg which gives me a good performance.
• Building a front end probably ends up in setting up another server, a bokeh or panel server which raises the question why to have a backend api at all. A static frontend html file would have to embed all widget states, right? So for my very large datasets, I would have to include 100k of possible time steps alone if all subsets should be possible.