Controlling dask behaviour when plotting using `xarray.hvplot.image`

hvPlot
1

Hi!

I am quite new to Holoviz but “seasoned” on dask/xarray/numpy.

I’m trying to visualize a large GeoTiff/Zarr geographical dataset using hvplot, but I’m having trouble to understand how to control the dask side of things.

Given that:


data = rioxarray.open_rasterio(path, chunks='auto')

test_data = data.isel(x=range(100_000), y=range(100_000)).persist()
# note that chunking is untouched, here being (272, 100_000)

I noticed that if I call the following

img = test_data.hvplot.image(
    x='x', 
    y='y', 
    crs=data.rio.crs.to_proj4(), 
    rasterize=True, 
    width=700, 
    height=500, 
    cmap='magma', 
    tiles=True, 
    aggregator='mode')
img

I get the cluster doing some initial work that includes two finalize_from_value tasks (half the data each), then performs some extra work as I interact with the plot

However, if I try to rechunk to something that I believe should make more sense for this visualization (lets say dict(x=10_000, y=10_000)) I get the two finalize_from_value tasks become a single one and failing due to memory limit on the single worker.

Now, that is quite the opposite of what I was expecting and I was wondering how you would control the partitioning of that part of the workload (I guess is the rasterization?)