pd.Series assignment with pd.IndexSlice results in NaN values despite matching indices

pandas MultiIndexes are sometimes a bit buggy, and this feels like one of those circumstances. If you modify s2.index to match s.index, the assignment works:

In [155]: s2.index = pd.MultiIndex.from_product([['a'], ['X'], ['u', 'v']], names=['one', 'two', 'three'])

In [156]: s2
Out[156]:
one  two  three
a    X    u        3
          v       -3
Name: four, dtype: int64

In [157]: s
Out[157]:
one  two  three
a    X    u         1
          v         2
b    Y    u         4
a    Z    u        20
Name: four, dtype: int64

In [158]: s.loc[:, 'X', :] = s2

In [159]: s
Out[159]:
one  two  three
a    X    u         3
          v        -3
b    Y    u         4
a    Z    u        20
Name: four, dtype: int64

Probably worth searching for similar issues in https://github.com/pandas-dev/pandas/issues and adding it as a new one if it's not already there.

One other option in the meantime is to use .unstack() to reshape your data to do the assignment:

In [181]: s = s.unstack('two')

In [182]: s['X'].loc[s2.index] = s2

In [183]: s.stack().swaplevel(1,2).sort_index()
Out[183]:
one  two  three
a    X    u         3.0
          v        -3.0
     Z    u        20.0
b    Y    u         4.0
dtype: float64