API: alignment behaviour of Series as assignment value in indexing

[jorisvandenbossche]

Follow up on #37427

Currently, there is an inconsistency between __getitem__ and loc related to how the RHS value in an assignment gets re-aligned or not.

In this example, the values in Series get just assigned as is, ignoring the index of the series (only requiring that the length matches the number of values in the indexer):

s = pd.Series(range(5))
idx = np.array([1, 4]) 

>>> s[idx] = Series([10, 11])   
>>> s   
0     0
1    10
2     2
3     3
4    11
dtype: int64

However, with loc we have a different behaviour: the Series gets basically aligned with s first, and then this realigned series is also indexed with idx, and those values get set:

s = pd.Series(range(5))
idx = np.array([1, 4]) 

>>> s.loc[idx] = Series([10, 11]) 
>>> s 
0     0.0
1    11.0
2     2.0
3     3.0
4     NaN
dtype: float64

you could write what basically happens more explicitly as:

s.loc[idx] = Series([10, 11]).reindex(s.index)[idx] 

To have the same behaviour as __getitem__, you need to assign with an array instead of series:

s = pd.Series(range(5))
idx = np.array([1, 4]) 

>>> s.loc[idx] = Series([10, 11]).values    
>>> s 
0     0
1    10
2     2
3     3
4    11
dtype: int64

[jbrockmendel]

cc @phofl i think this has a bearing on one of the topics you're working on

[phofl]

@jbrockmendel not really working on this right now. The underlying cause is the same as in #30439. Have to define how to move forward and if this is expected or a bug.