ENH: maybe_dispatch_to_dunder_op handle min/max

[jbrockmendel]

I'm running up against a case where I need np.maximum.reduce(period_array) to work; itd be nice if that could be handled by maybe_dispatch to call PeriodArray.max

cc @jorisvandenbossche @TomAugspurger

[TomAugspurger]

To clarify, np.maximum(PeriodArray) works fine, right? It's the .reduce part that isn't handled?

[jbrockmendel]

np.maximum(PeriodArray) breaks because it expects multiple arguments, np.maximum(period_array, period_array) works, while np.maximum(period_index, period_index) fails with ValueError: ufunc 'maximum' not supported for the PeriodIndex

But yes, it is specifically the .reduce that is coming up at the moment

[TomAugspurger]

while np.maximum(period_index, period_index) fails with ValueError: ufunc 'maximum' not supported for the PeriodIndex

That would be solved with Index having an __array_ufunc__. I wounder if Series.__array_ufunc__ can be moved to to IndexOpsMixin. I suppose alignment of Series but not Index makes that tricky.

Can you provide a failing example for reduce on period array?

In [8]: np.maximum.reduce(pd.array(['2000', '2001'], dtype='Period[D]'))
Out[8]: Period('2001-01-01', 'D')

[jbrockmendel]

Can you provide a failing example for reduce on period array?

Not on master. In the relevant branch I have implemented DatetimeLikeArrayMixin.__array_ufunc__ (to get np.isfinite(datetimeindex_tzaware) to work... yak shaving)