Area/Bar stacking implementation

[stas-sl]

This is related to holoviz/hvplot#551, but I think it is a more broad issue. For bar charts specifying columns order works as expected (except reversed legend), but for area chart this order is not respected.

df = pd.DataFrame({'a': [2, 3, 1], 'b': [1, 2, 3], 'c': [3, 2, 1]})
df.hvplot.bar(y=['b', 'c', 'a'], stacked=True)

df = pd.DataFrame({'a': [2, 3, 1], 'b': [1, 2, 3], 'c': [3, 2, 1]})
df.hvplot.area(y=['b', 'c', 'a'], stacked=True)

I believe this is due rather different implementations. For area it goes inside Area.stack and has df.groupby(level=levels) which causes sorting implicitly:

holoviews/holoviews/element/chart.py

Lines 237 to 264 in b6ff2d1

	@classmethod
	def stack(cls, areas, baseline_name='Baseline'):
	"""
	Stacks an (Nd)Overlay of Area or Curve Elements by offsetting
	their baselines. To stack a HoloMap or DynamicMap use the map
	method.
	"""
	if not len(areas):
	return areas
	is_overlay = isinstance(areas, Overlay)
	if is_overlay:
	areas = NdOverlay({i: el for i, el in enumerate(areas)})
	df = areas.dframe(multi_index=True)
	levels = list(range(areas.ndims))
	vdim = areas.last.vdims[0]
	vdims = [vdim, baseline_name]
	baseline = None
	stacked = areas.clone(shared_data=False)
	for key, sdf in df.groupby(level=levels):
	sdf = sdf.droplevel(levels).reindex(index=df.index.unique(-1), fill_value=0)
	if baseline is None:
	sdf[baseline_name] = 0
	else:
	sdf[vdim.name] = sdf[vdim.name] + baseline
	sdf[baseline_name] = baseline
	baseline = sdf[vdim.name]
	stacked[key] = areas[key].clone(sdf, vdims=vdims)
	return Overlay(stacked.values()) if is_overlay else stacked

Whereas for bar charts it is handled on plotting level in BarPlot.get_stack:

holoviews/holoviews/plotting/bokeh/chart.py

Lines 828 to 848 in b6ff2d1

	def get_stack(self, xvals, yvals, baselines, sign='positive'):
	"""
	Iterates over a x- and y-values in a stack layer
	and appropriately offsets the layer on top of the
	previous layer.
	"""
	bottoms, tops = [], []
	for x, y in zip(xvals, yvals):
	baseline = baselines[x][sign]
	if sign == 'positive':
	bottom = baseline
	top = bottom+y
	baseline = top
	else:
	top = baseline
	bottom = top+y
	baseline = bottom
	baselines[x][sign] = baseline
	bottoms.append(bottom)
	tops.append(top)
	return bottoms, tops

Also I'm not sure why not to just use ad hoc bokeh methods vbar_stack and varea_stack for this.

[jbednar]

Would adding sort=False to the groupby call be enough to fix it?

[stas-sl]

Yes, it indeed solves the issue! I didn't know there is such option.

df = pd.DataFrame({'a': [2, 3, 1], 'b': [1, 2, 3], 'c': [3, 2, 1]})
df.hvplot.area(y=['b', 'c', 'a'], stacked=True)

Though on a higher level maybe it worth considering unifying stacking for both these charts? Ideally it seems reasonable to extract this functionality in a common helper method and make use of bokeh methods, or at least to have it on the same level, whereas currently bar chart is stacked on plotting/bokeh level, while area chart is stacked on a more abstract level before specific backend, which might cause inconsistencies when using other backends.

[jbednar]

Happy to consider a PR either just making this tiny sort=False fix or making things more consistent, which seems like more work.

[stas-sl]

I'll proceed with the quick fix option for now, as indeed, refactoring requires much deeper understanding of the library.