Remove all preprocessing from PlotAccessor.__call__

[TomAugspurger]

In #28647, we discovered that PlotAccessor.__call___ does quite a bit of preprocessing of the data before handing it off to the backend.

pandas/pandas/plotting/_core.py

Lines 728 to 794 in 3992126

	# The original data structured can be transformed before passed to the
	# backend. For example, for DataFrame is common to set the index as the
	# `x` parameter, and return a Series with the parameter `y` as values.
	data = self._parent.copy()
	if isinstance(data, ABCSeries):
	kwargs["reuse_plot"] = True
	if kind in self._dataframe_kinds:
	if isinstance(data, ABCDataFrame):
	return plot_backend.plot(data, x=x, y=y, kind=kind, **kwargs)
	else:
	raise ValueError(
	("plot kind {} can only be used for data frames").format(kind)
	)
	elif kind in self._series_kinds:
	if isinstance(data, ABCDataFrame):
	if y is None and kwargs.get("subplots") is False:
	msg = "{} requires either y column or 'subplots=True'"
	raise ValueError(msg.format(kind))
	elif y is not None:
	if is_integer(y) and not data.columns.holds_integer():
	y = data.columns[y]
	# converted to series actually. copy to not modify
	data = data[y].copy()
	data.index.name = y
	elif isinstance(data, ABCDataFrame):
	data_cols = data.columns
	if x is not None:
	if is_integer(x) and not data.columns.holds_integer():
	x = data_cols[x]
	elif not isinstance(data[x], ABCSeries):
	raise ValueError("x must be a label or position")
	data = data.set_index(x)
	if y is not None:
	# check if we have y as int or list of ints
	int_ylist = is_list_like(y) and all(is_integer(c) for c in y)
	int_y_arg = is_integer(y) or int_ylist
	if int_y_arg and not data.columns.holds_integer():
	y = data_cols[y]
	label_kw = kwargs["label"] if "label" in kwargs else False
	for kw in ["xerr", "yerr"]:
	if kw in kwargs and (
	isinstance(kwargs[kw], str) or is_integer(kwargs[kw])
	):
	try:
	kwargs[kw] = data[kwargs[kw]]
	except (IndexError, KeyError, TypeError):
	pass
	# don't overwrite
	data = data[y].copy()
	if isinstance(data, ABCSeries):
	label_name = label_kw or y
	data.name = label_name
	else:
	match = is_list_like(label_kw) and len(label_kw) == len(y)
	if label_kw and not match:
	raise ValueError(
	"label should be list-like and same length as y"
	)
	label_name = label_kw or data.columns
	data.columns = label_name
	return plot_backend.plot(data, kind=kind, **kwargs)

Depending on the plot kind and backend, some of this preprocesing may not be appropriate. The backend might want to see the "raw" data. That's not to say we shouldn't do any preprocessing though. We'll need some feedback from backend authors on what should be done before handing it off.

[TomAugspurger]

cc @jsignell

[jsignell]

Thanks for writing this up Tom. I think in particular the issue for hvplot was that the preprocessing step was assuming that the the columns of interest can be known on the pandas side. This doesn't hold true for hvplot because there are additional kwargs (groupby, by, row, ...) that accept column names.