[BUG] Incorrect legend color with different glyphs with same CDS and a CDSView

Software version info

• python: 3.6.10
• bokeh: 2.0.1
• OS: macOS 10.14.6

Description of expected behavior and the observed behavior

I’m plotting circles and segments with different colors from the same CDS. All circles are plotted, but only some segments are plotted, using a CDSView.

I expect the legend to show the correct symbol and color, but one of the colors doesn’t match the source.

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

from bokeh.plotting import figure
from bokeh.models import CDSView, ColumnDataSource, BooleanFilter
from bokeh.io import show

p = figure()

ds = ColumnDataSource(data={
    'color': ["blue", "red"],
    'x0': [0, 0],
    'y0': [1, 0],
    'x1': [1, 1],
    'y1': [1, 0]
})

# I also tried with an IndexFilter and CustomJSFilter and got the same result
view = CDSView(source=ds, filters=[BooleanFilter([False, True])])

p.circle(
	x='x0', y='y0',
	color='color',
	legend_field='color',
	source=ds)

p.segment(
	x0='x0', y0='y0', x1='x1', y1='y1',
	color='color',
	legend_field='color',
	source=ds,
	view=view)

show(p)

Stack traceback and/or browser JavaScript console output

No logs in python or the JavaScript console output.

Screenshots or screencasts of the bug in action

Note the red line across the blue dot in the legend

[bryevdv]

I’ve left a note in the other issue to make sure to validate the above analysis when the other issue is addressed.

[bryevdv]

And actually looking in to things a bit more closely I think we will be limited in what we can do. Basically all renderers that specify the same legend_field are grouped together on the Python side. We can probably make the BokehJS side respect one view for one renderer pretty easily, and that is how I will proceed for the other issue. But it’s not even clear what the correct behavior should be when different renderers in the same legend map out to different groups. In the general case there would be no guarantee there is even any overlap like above. This just wildly breaks very deep assumptions about how legends are expected to be, and has never be defined, and I think making it do anything well-defined would be a rather heavy lift for a rather obscure use case, i.e. not justifiable in terms of time/effort, given other priorities. I think in this situation it would be better to construct a legend manually however you want it using legend items with explicit indices.