Weird assignment behaviour with MultiIndex

[mvds314]

I came across some weird behavior when using assignment with a multi-index in Pandas 0.17.1

import pandas as pd
lvl0=range(2)
lvl1 = range(3)
index=pd.MultiIndex.from_product([lvl0,lvl1])
columns=['A','B']
df1=pd.DataFrame(data=1, columns=columns, index=index)
df2 = pd.DataFrame(data=2, index=lvl1, columns=['A'])

#this returns a dataframe without a multi index
print(df1.loc[0,df2.columns]) 

#but when you try an assignment it doesn't work
df1.loc[0,df2.columns]=df2
print(df1)

I debugged the pandas package a little bit and I think I know what is going. When you call df1.loc[0,df2.columns] it uses some getitem method. But, df1.loc[0,df2.columns]=df2 uses some setitem method. When you set a break point in indexing.py in _NDFrameIndexer._get_setitem_indexer, then you see that the key that is being looked up in the index of the DataFrame is (0,['A']). Obviously, this key cannot be found, since A is a column. The result is that nothing is assigned.

I think this is a bug because there is inconsistent behaviour between setitem and getitem: when df1.loc[0,df2.columns] returns a DataFrame without a multindex (which isn't a copy) then assignment should be possible.

[jreback]

xref #6699

Multiindex assignment is a bit buggy.

You can do this

In [47]: df1 = pd.DataFrame(data=1, columns=columns, index=index)

In [48]: df1
Out[48]: 
     A  B
0 0  1  1
  1  1  1
  2  1  1
1 0  1  1
  1  1  1
  2  1  1

In [49]: df1.loc[0,'A'] = df2.values

In [50]: df1
Out[50]: 
     A  B
0 0  2  1
  1  2  1
  2  2  1
1 0  1  1
  1  1  1
  2  1  1

I would expect this should work (and doesn't)

In [51]: df1 = pd.DataFrame(data=1, columns=columns, index=index)
In [53]: df2['A']                   
Out[53]: 
0    2
1    2
2    2
Name: A, dtype: int64

In [54]: df1.loc[0,'A']
Out[54]: 
0    1
1    1
2    1
Name: A, dtype: int64

In [55]: df1.loc[0,'A'] = df2['A']  

In [56]: df1
Out[56]: 
       A  B
0 0  NaN  1
  1  NaN  1
  2  NaN  1
1 0  1.0  1
  1  1.0  1
  2  1.0  1

however, not sure that assignment to a different shape should actually work, e.g. df2 is a dataframe. I would expect this to raise.

pull-requests are welcome to fix.

[jreback]

furthermore, pls looks thru other multiindex assignment issues to see if you can find some related ones.

[mroeschke]

Looks like this works now and I believe we have a test for this so closing