tzinfo lost when concatenating multiindex arrays

[sleibman]

This may be related to #3950 though my symptoms are a bit different. In that issue, the tzinfo was lost during the creation of a MultiIndex array. In my case, I'm able to create a MultiIndex array with intact tzinfo, but tzinfo gets lost when I concatenate one or more such arrays.

Here's an example. Note the lack of tzinfo in the final line:

In [47]: import pandas
In [48]: import datetime
In [49]: import pytz
In [50]: array = pandas.DataFrame({'a':[datetime.datetime(2014,1,1,tzinfo=pytz.UTC), datetime.datetime(2014,1,2,tzinfo=pytz.UTC)], 'b':[1, 2], 'c':[3, 4]})
In [51]: a2 = array.set_index('a')
In [52]: a3 = array.set_index(['a', 'b'])
In [53]: a2_concatenated = pandas.concat([a2])
In [54]: a3_concatenated = pandas.concat([a3])
In [55]: array.iloc[0,0]
Out[55]: datetime.datetime(2014, 1, 1, 0, 0, tzinfo=<UTC>)
In [56]: a2.index[0]
Out[56]: Timestamp('2014-01-01 00:00:00+0000', tz='UTC')
In [57]: a3.index[0]
Out[57]: (Timestamp('2014-01-01 00:00:00+0000', tz='UTC'), 1)
In [58]: a2_concatenated.index[0]
Out[58]: Timestamp('2014-01-01 00:00:00+0000', tz='UTC')
In [59]: a3_concatenated.index[0]
Out[59]: (Timestamp('2014-01-01 00:00:00', tz=None), 1)

[sleibman]

Sorry, I should have also pointed out that I am using a slightly old version:

In [60]: pandas.__version__
Out[60]: '0.12.0'

I have not yet tested with any 0.13 versions.

[dsm054]

I can reproduce in 0.13.1-395-g391ed32.

[jreback]

@sleibman if you or @dsm054 would like to take a stab would be appreciated!

[sleibman]

I may be able to supply a fix but haven't dug into the underlying code yet.
An additional observation: Even just calling reset_index() on a MultiIndex array causes the tz offset to disappear:

In [70]: a3
Out[70]: 
                             c
a                         b   
2014-01-01 00:00:00+00:00 1  3
2014-01-02 00:00:00+00:00 2  4

In [71]: a3.reset_index()
Out[71]: 
                    a  b  c
0 2014-01-01 00:00:00  1  3
1 2014-01-02 00:00:00  2  4