DataFrame.ix losing row ordering when index has duplicates

[dalejung]

import pandas as pd

ind = ['A', 'A', 'B', 'C']i
df = pd.DataFrame({'test':range(len(ind))}, index=ind)

rows = ['C', 'B']
res = df.ix[rows]
assert rows == list(res.index) # fails

The problem is that the resulting DataFrame keeps the ordering of the df.index and not the rows key. You'll notice that the rows key doesn't reference a duplicate value.

[jreback]

thanks for the catch, this is a case that though worked, was using a set like indexer so the order was not guaranteeed - provided an opportunity to refactor a bit...PR coming soon

[jreback]

Unique

In [1]: df=DataFrame(randn(5,3),index=list('ABCDE'))

In [2]: df.ix[['A']]
Out[2]: 
          0         1        2
A -1.048431 -0.435366  0.33573

In [3]: df.ix[['A','G']]
Out[3]: 
          0         1        2
A -1.048431 -0.435366  0.33573
G       NaN       NaN      NaN

Duplicate

In [4]: dfnu=DataFrame(randn(5,3),index=list('AABCD'))

In [5]: dfnu.ix[['A']]
Out[5]: 
          0         1         2
A  0.039932  1.049630 -2.647776
A -0.213537  0.747972 -0.830574

In [7]: dfnu.ix[['B','A','E']]
Out[7]: 
          0         1         2
B  0.292704 -1.396854 -0.414920
A  0.039932  1.049630 -2.647776
A -0.213537  0.747972 -0.830574

@dalejung @y-p @wesm
ok...behavior fixed, but what do you think about the last case
e.g. selecting something that doesn't exist (but at least 1 value exists)
in the unique case you get equivalent of reindexing, should I fix the duplicate case to do
the same?

[ghost]

re:

In [41]: dfnu=DataFrame(randn(4,3),index=list('ABCD'))

In [42]: dfnu.ix[['E']]
Out[42]: 
    0   1   2
E NaN NaN NaN

In [43]: dfnu=DataFrame(randn(5,3),index=list('AABCD'))

In [44]: dfnu.ix[['E']]
Out[44]: 
Empty DataFrame
Columns: [0, 1, 2]
Index: []

yeah, that is inconsistent.

[dalejung]

I think for consistency sake it should be the same. To be honest, I don't have a use case for indexing a non-existent label or an iterable key that contains a duplicate. I came across the bug when a source file upstream had a duplicate row.

Thanks for the quick patch.

[jreback]

np...we have been fixing duplicate indicies lately (again not there is that much use for them), but they should work....will be merged soon