Implement min, argmin, max, argmax on ExtensionArrays?

[TomAugspurger]

We already have ExtensionArray.argsort as part of the EA interface, so we should be able to to do min, argmin, max, and argmax as a composition of argsort and __getitem__. Do we want to do this?

[jreback]

i think this is a nice idea; shows a nice benefit of shared EA infrastructure

[TomAugspurger]

One thing I notice though, our docs on _values_for_argsort don't mention missing data at all. I think we may need to adjust the API there to return a tuple of (values, mask), where the mask indicates missing values?

…

On Fri, Dec 21, 2018 at 8:33 AM Jeff Reback ***@***.***> wrote: i think this is a nice idea; shows a nice benefit of shared EA infrastructure — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#24382 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABQHIglNkqMNqDzj8ZLtFDBLjHGe4yT5ks5u7PE-gaJpZM4Zd_jY> .

[makbigc]

When nan is invloved, argsort behaves in two ways.

The element of index 1 is always nan in the following.

1. Putting the nan in the beginning

In [27]: arr = integer_array([1, np.nan, 2])                                    

In [28]: arr                                                                    
Out[28]: 
<IntegerArray>
[1, NaN, 2]
Length: 3, dtype: Int64

In [29]: arr.argsort()                                                          
Out[29]: array([1, 0, 2])

2. Putting the nan in the end

In [20]: arr1 = integer_array([1, np.nan, 0], dtype='uint8')                    

In [21]: arr1                                                                   
Out[21]: 
<IntegerArray>
[1, NaN, 0]
Length: 3, dtype: UInt8

In [22]: arr1.argsort()                                                         
Out[22]: array([2, 0, 1])

In [23]: idx = pd.Index([1, np.nan, 2])                                         

In [24]: arr = idx.array                                                        

In [25]: arr                                                                    
Out[25]: 
<PandasArray>
[1.0, nan, 2.0]
Length: 3, dtype: float64

In [26]: arr.argsort()                                                          
Out[26]: array([0, 2, 1])

Should we standardize where nan to be placed? _values_for_argsort returns different value for nan in different EA.

[jorisvandenbossche]

@makbigc there indeed still some inconsistencies there ... I opened #21801 for that some time ago. Can you move your comment there?