BUG: Unexpected behavior from DataFrame.index.levels

[8one6]

Perhaps this is me misunderstanding the way Pandas handles indices for related objects behind the scenes, but I found the following behavior very unintuitive:

import pandas as pd
import numpy as np

np.random.seed(0)

idx = pd.MultiIndex.from_product([['John', 'Josh', 'Alex'], list('abcde')], 
                                 names=['Person', 'Letter'])
large = pd.DataFrame(data=np.random.randn(15, 2), 
                     index=idx, 
                     columns=['one', 'two'])
small = large.loc[['Jo'==d[0:2] for d in large.index.get_level_values('Person')]]

print small.index.levels[0]
print large.index.levels[0]

This returns:

Index([u'Alex', u'John', u'Josh'], dtype='object')
Index([u'Alex', u'John', u'Josh'], dtype='object')

rather than the expected

Index([u'John', u'Josh'], dtype='object')
Index([u'Alex', u'John', u'Josh'], dtype='object')

I could get the results I expected by running

small.index.get_level_values('Person').unique()
large.index.get_level_values('Person').unique()

but could someone explain why the behavior I'm seeing above with DataFrame.index.levels is an appropriate result rather than a bug?

[jreback]

see discussion here:

#2770

[8one6]

Would it make sense to say somewhere in the docs that the .unique() method can be run on the kind of data you get back from calling the .get_level_values(levelname) method?

[jreback]

.get_level_values() returns an Index, and unique() is a method of an Index. So I would expect this to be natural. if you think its unclear, pls submit a pull-request for a doc update.

[jreback]

xref to #2770