Rasterized NdOverlay uses ImageStack only on an anti-aliased Curve

[maximlt]

Reviewing the modernized Gerrymandering example, I noticed some issues with rasterizing a scatter plot. Some of them are hvPlot issues, but it looks like there's also one issue in HoloViews.

In this example, I create an NdOverlay of Scatter elements and rasterize it with an aggregator. I expected the returned DynamicMap to hold an ImageStack but it doesn't.

import datashader as ds
import holoviews as hv
import pandas as pd
from holoviews.operation.datashader import rasterize
hv.extension('bokeh')

df = pd.DataFrame(
    {
        "time": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
        "value": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
        "by": ["d", "d", "d", "a", "a", "b", "b", "b", "b", "b"],
    }
)
plots = {k: hv.Scatter(g, 'time', 'value') for k, g in df.groupby('by')}
ndo = hv.NdOverlay(plots, kdims='by')
r = rasterize(ndo, aggregator=ds.count_cat('by'))
r

:DynamicMap   []
   :NdOverlay   [by]
      :Image   [time,value]   (time_value Count)

I can get an ImageStack if:

• It's an NdOverlay of Curve elements
• line_width is not None

plots = {k: hv.Curve(g, 'time', 'value') for k, g in df.groupby('by')}
ndo = hv.NdOverlay(plots, kdims='by')
r = rasterize(ndo, aggregator=ds.count_cat('by'), line_width=1)
r

:DynamicMap   []
   :ImageStack   [time,value]   (a,b,d)

In aggregate._process, commenting out this if statement that looks like to be an optimization path, I can obtain an ImageStack in the first example.

holoviews/holoviews/operation/datashader.py

Lines 349 to 354 in f100137

	if overlay_aggregate.applies(element, agg_fn, line_width=self.p.line_width):
	params = dict(
	{p: v for p, v in self.param.values().items() if p != 'name'},
	dynamic=False, **{p: v for p, v in self.p.items()
	if p not in ('name', 'dynamic')})
	return overlay_aggregate(element, **params)

@philippjfr do you happen to know if the optimization provided by overlay_aggregate is still relevant? Datashader changed quite a lot over the past couple of years.

cc @jlstevens @hoxbro