Holoviews: hvplot.image(rasterize=True) not working with holoviews>=1.13.4

[scottyhq]

After upgrading holoviews from 1.13.3 to 1.13.4 hvplot.xarray.image(rasterize=True, dynamic=True) no longer works. displayed image resolution is not updated upon zooming into an image plot

ALL software version info

(linux or macos)
conda create -n holoviews python=3.7 holoviews xarray rasterio hvplot datashader jupyter

holoviews                 1.13.4             pyh9f0ad1d_0    conda-forge
bokeh                     2.2.2            py37hc8dfbb8_0    conda-forge
hvplot                    0.6.0              pyh9f0ad1d_0    conda-forge
datashader                0.11.1             pyh9f0ad1d_0    conda-forge

Description of expected behavior and the observed behavior

Zooming into a subregion of an image with raster with rasterize=True, dynamic=True should update displayed resolution

Complete, minimal, self-contained example code that reproduces the issue

This uses an example image from AWS public data

import os
import xarray as xr
import hvplot.xarray
os.environ['GDAL_DISABLE_READDIR_ON_OPEN']='EMPTY_DIR'
os.environ['AWS_NO_SIGN_REQUEST']='YES'
href = 's3://sentinel-s1-rtc-indigo/tiles/RTC/1/IW/10/T/ET/2020/S1B_20200106_10TET_ASC/Gamma0_VV.tif'
da = xr.open_rasterio(href)
da.hvplot.image(rasterize=True, dynamic=True, clim=(0,0.4), aspect='equal', frame_width=600, cmap='gray')

Stack traceback and/or browser JavaScript console output

No tracebacks, not sure how to output console messages.

downgrading to holoviews 1.13.3 things work as expected. cc @rsignell-usgs

[scottyhq]

downgrading is not ideal because we'd be back to this issue that was solved in 1.13.4 https://github.com/holoviz/hvplot/issues/462

[rsignell-usgs]

@scottyhq , thanks for reporting this. I was thinking that somehow I broke it. Luckily I didn't have a live demo today.
@jbednar , do you think holoviz will stabilize at some point, or are we always likely to feel this pain due to the complexity of the package dependencies?

[jbednar]

Good question. If you have any influence, try to get @jsignell to come back to the project so that things will run more smoothly. :-) Meanwhile, I can't comment on what this particular issue indicates about stability, since I can't see what the problem could be.

I don't have rasterio in my usual environment but this problem appears to be reproducible with one of the stock examples:

import xarray as xr
import hvplot.xarray

air = xr.tutorial.open_dataset('air_temperature').load().air
air2d = air.sel(time='2013-06-01 12:00')
air2d.hvplot.image(width=400, rasterize=True)

Here the image updates with either rasterize or datashade for hv=1.13.3, but for hv=1.13.4 or hv=1.14.0a2, it only updates for datashade=True, not rasterize=True. The same happens for rasterize(air2d.hvplot.image(width=400)), so I've moved it from hvPlot to HoloViews, where I'm assuming the problem originates (and also since it depends on the hv version and not the hvPlot version).

I didn't see any relevant warnings or errors on the console. @philippjfr ?

[jbednar]

Also, the behavior even when it is working looks dubious on that example; it jumps around more than the resolution of the image would indicate to me. I think there is an off-by-one error somewhere in the raster rendering code, which wouldn't be visible for a large enough image but is revealed by this low-resolution one.

[philippjfr]

Fixed in https://github.com/holoviz/holoviews/pull/4654, will release 1.13.5 tomorrow.

[philippjfr]

Also I really don't understand why tests didn't catch this one but I'll add some more to that PR.

[scottyhq]

thanks for the speedy fix @philippjfr , much appreciated!!

[rsignell-usgs]

Unfortunately with holoviews=1.13.5, it seems we are back to the wacky zoom problems again: https://github.com/holoviz/holoviews/issues/4052#issuecomment-717194601