ExtensionArray.value_counts

[TomAugspurger]

Is this required? Right now

pandas/pandas/core/algorithms.py

Lines 708 to 711 in a03d953

	if is_extension_array_dtype(values) or is_sparse(values):
	# handle Categorical and sparse,
	result = Series(values)._values.value_counts(dropna=dropna)

assumes that it's implemented, but we don't document / implement it by default, so we end up with an AttributeError at runtime.

[jorisvandenbossche]

Personally I would remove it from the EA interface. value_counts is a very useful method, but IMO does not really fit on an array (returns a Series, while I think the arrays should be somewhat independent from pandas Series concepts), and is also rather easily to do based on _factorize/_values_for_factorize for EAs as a default implementation?

[TomAugspurger]

Yes, the main sticking point though is the index. Once we have official support for extension indexes, we should be able to update algorithms.value_counts to do the right thing.

[jreback]

nog against removing value_counts for the time being either until we have ExtensionIndex , which really comes up a lot when you groupby things :.

[jbrockmendel]

agreed on removing from EAs. It introduces a nasty dependency of the EA code on index/series code.

[jbrockmendel]

and is also rather easily to do based on _factorize/_values_for_factorize for EAs as a default implementation?

I tried this the other day and found I had to do a lot of special-casing for Categorical and SparseArray. Has anyone else tried this?