Direct delegation of Series methods to ExtensionArrays

[xhochy]

During the implementation of non-numpy backed ExtensionArrays I quite often run into the case where it is simpler for me to write a complete re-implementation of the method defined on pd.Series instead of using the current implementation that only delegates part of the work. It would probably make sense to introduce some sort of delegation mechanism, either we continue the delegation like in

pandas/pandas/core/base.py

Line 1041 in 4274b84

def unique(self):

or we could possibly add really general interface like NumPy's __array_ufunc__: https://docs.scipy.org/doc/numpy/reference/arrays.classes.html#numpy.class.__array_ufunc__

My use case where this arises currently is coming from #21296 and pd.Series.argsort but I expect that there will be much more cases in this direction while I continue to implement the ExtensionArray interface for Arrow Arrays.

[jorisvandenbossche]

Quick comment: for the argsort case, I think this could be solved by changing np.argsort(values, ..) to values.argsort(..) in the Series.argsort implementation? (if this is blocking you, fix certainly welcome)

But indeed, we should discuss this more in general.

[TomAugspurger]

In the abstract, I'm also interested in this. The set of methods that are dispatched to currently is pretty ad-hoc (essentially enough to get df.groupby('extension_array').mean() working :)

[jbrockmendel]

2 thoughts here

1. Since the OP we've added many private EA methods that we dispatch to under the hood (EA._where, EA._putmask, EA._quantile). We could address many of these cases by leaning heavily on that pattern.
2. Implementing something like __pandas_ufunc__ or __pandas_priority__ might be helpful for eg REGR: comparison op with dask data structure fails #38946

[jbrockmendel]

I don't think there's any appetite for adding an __array_ufunc__-like mechanism, but we are definitely moving in the direction of more methods being defined on the EAs and being directly delegated to.