Port or wrap the Python datashader package for use in R

[timchurches]

For background on datashader, see http://datashader.readthedocs.io/en/latest/

datashader addresses a similar problem to that addressed by one of the projects at the 2016 Australian RopenSci Unconf, but in a much elegant and general way that avoids all the messy overhead of breaking data into map swatches: https://github.com/ropensci/auunconf/issues/38

For discussion of datashader from the point of view of an R user, see http://www.win-vector.com/blog/2017/03/datashader-is-a-big-deal/

There would appear to be several possible approaches:

• leverage datashader as implemented in Python via the rbokeh wrapper around bokeh (although data shader can be used outside the bokeh framework)
• pass data to the Python data shader machinery via shared memory, as suggested by one of the data shader developers: https://github.com/bokeh/datashader/issues/304
• replicate core datashader functionality, rewriting the Numpy/Numba functions in C++ via Rcpp

Orthogonal to these approaches might be wrapping data shader output via htmlwidgets, easy for use in Shiny apps (which is the ultimate goal - there may be short-cuts to that?).

[goldingn]

Have you considered a reticulate import? An R wrapper around the import would probably be needed, but it should solve the technical problems of making R & python play together nicely.

[timchurches]

The best thing about datashader is that it is interactive, and will recalculate pixel-level aggregates in even ginormous on-the-fly. But that requires close integration with the javascript interface to the bokeh chart object in python. I suspect that a complete port to R would be needed to obtain the performance required. That's doable, but not in two days. In the meantime, datashader is a good reason to use a bit of python and the Jupyter interface.

[harryprince]

@timchurches the best option might implement the same functions in RStudio-Addin with Shiny like esquisse R package.

[timchurches]

@harryprince The hard bit is porting the on-the-fly viewport aggregation functions from dataShader. The C++ code is portable, but what to port it to? A ggplot2 extension, but ggplot2 isn't designed for interactive graphics, seems like it would be difficult.