Add viewport downsample algorithm

[hoxbro]

Simple proof-of-concept of how we can use downsample to only send the data in the viewport to the frontend.

import holoviews as hv
import numpy as np
from holoviews.operation.downsample import downsample1d

hv.extension("bokeh")
downsample1d(hv.Curve(np.random.rand(1000)), algorithm="viewport").opts(xlim=(0, 10))

screenrecord-2023-12-08_19.17.53.mp4

[codecov-commenter]

Codecov Report

Attention: 21 lines in your changes are missing coverage. Please review.

Comparison is base (231cd71) 88.54% compared to head (1de1336) 88.53%.

Files	Patch %	Lines
holoviews/operation/downsample.py	22.22%	14 Missing ⚠️
holoviews/tests/core/data/test_cudfinterface.py	0.00%	5 Missing ⚠️
holoviews/core/data/cudf.py	66.66%	1 Missing ⚠️
holoviews/core/data/dask.py	90.90%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6017      +/-   ##
==========================================
- Coverage   88.54%   88.53%   -0.02%     
==========================================
  Files         314      314              
  Lines       65872    65942      +70     
==========================================
+ Hits        58327    58379      +52     
- Misses       7545     7563      +18     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[droumis]

@hoxbro . This would be a really fantastic addition. I want to make sure that it works for all the relevant use cases that we are encountering right now.

First, it looks like any type of downsampling does not work with subcoordinate_y=True

Code

import holoviews as hv
import numpy as np
from holoviews.operation.downsample import downsample1d
hv.extension("bokeh")

c1 = hv.Curve(np.random.rand(1000), label='c1').opts(subcoordinate_y=True)
c2 = hv.Curve(np.random.rand(1000), label='c2').opts(subcoordinate_y=True)

curves = hv.Overlay(c1*c2).opts(xlim=(5, 10))
downsample1d(curves, algorithm="viewport")

Screen.Recording.2023-12-18.at.3.08.26.PM.mov

[hoxbro]

subcoordinate_y=True should now work:

screenrecord-2023-12-19_10.23.57.mp4

[hoxbro]

I have added a workaround when having an x limit set on the element (downsample1d(hv.Curve(np.ones(1_000)), algorithm="viewport").opts(xlim=(5, 10))). This is not very elegant, but I can't spend more time on it. I have talked with Philipp, and did not know a better implementation at the top of his head, but maybe he has a better workaround, which at least does not need to have a separate algorithm to handle it.

The problem is the downsampling works by looking at the stream x_range. The first run will be None, meaning the viewport algorithm will send all data to the browser. When the plot shows up in the browser a stream event will trigger and change the x_range to the xlim range. This is not as big of a problem for the other algorithms as they do an actual downsample to the width pixel size (400 by default).

[hoxbro]

We can't send an empty plot either, as this will make the y-axis between 0 and 1, didn't notice this before as I was using random data.

So I'm now sending 400 points (200 from the start and 200 from the end) to get an approximation of the range.

[droumis]

#6061 is relevant, given that this PR is primarily about slicing

[philippjfr]

Set the stream value during the stream Callback initialization:

    def initialize(self, plot_id=None):
        super().initialize(plot_id)
        for stream in self.streams:
            msg = self._process_msg({})
            stream.update(**msg)

[philippjfr]

Looks good, thanks!