Computing time difference with timezone in Series with one row gives wrong result

When there are only one row containing timestamps with timezone in DataFrame, computing time difference between columns gives incorrect result - always '0 days':

In [133]: import pandas as pd

In [134]: df = pd.DataFrame(columns=['s', 'f'])

In [135]: df.s = [pd.Timestamp('2016-02-08 13:43:14.605000', tz='America/Sao_Paulo')]

In [136]: df.f = [pd.Timestamp('2016-02-10 13:43:14.605000', tz='America/Sao_Paulo')]

In [137]: df
Out[137]:
                                 s                                f
0 2016-02-08 13:43:14.605000-02:00 2016-02-10 13:43:14.605000-02:00

In [138]: df.f - df.s    # should give '2 days'
Out[138]:
0   0 days
Name: f, dtype: timedelta64[ns]

In [139]: df.s - df.f    # should give '-2 days'
Out[139]:
0   0 days
Name: s, dtype: timedelta64[ns]

If there are more than one row, result is correct:

In [140]: import pandas as pd

In [141]: df = pd.DataFrame(columns=['s', 'f'])

In [142]: df.s = [pd.Timestamp('2016-02-08 13:43:14.605000', tz='America/Sao_Paulo'), pd.Timestamp('2016-02-08 13:43:14.605000', tz='America/Sao_P aulo')]

In [143]: df.f = [pd.Timestamp('2016-02-10 13:43:14.605000', tz='America/Sao_Paulo'), pd.Timestamp('2016-02-11 15:43:14.605000', tz='America/Sao_P aulo')]

In [144]: df
Out[144]:
                                 s                                f
0 2016-02-08 13:43:14.605000-02:00 2016-02-10 13:43:14.605000-02:00
1 2016-02-08 13:43:14.605000-02:00 2016-02-11 15:43:14.605000-02:00

In [145]: df.f - df.s
Out[145]:
0   2 days 00:00:00
1   3 days 02:00:00
Name: f, dtype: timedelta64[ns]

One row containing timestamps in naive datetime also gives correct result:

In [147]: import pandas as pd

In [148]: df = pd.DataFrame(columns=['s', 'f'])

In [149]: df.s = [pd.Timestamp('2016-02-08 13:43:14.605000')]

In [150]: df.f = [pd.Timestamp('2016-02-10 13:43:14.605000')]

In [151]: df
Out[151]:
                        s                       f
0 2016-02-08 13:43:14.605 2016-02-10 13:43:14.605

In [152]: df.f - df.s
Out[152]:
0   2 days
dtype: timedelta64[ns]

Installed versions:

In [162]: pd.show_versions()

INSTALLED VERSIONS
------------------
commit: None
python: 2.7.6.final.0
python-bits: 64
OS: Linux
OS-release: 3.13.0-49-generic
machine: x86_64
processor: x86_64
byteorder: little
LC_ALL: None
LANG: en_US.UTF-8

pandas: 0.17.1
nose: None
pip: 8.0.2
setuptools: 18.2
Cython: None
numpy: 1.10.4
scipy: None
statsmodels: None
IPython: 4.0.3
sphinx: None
patsy: None
dateutil: 2.4.2
pytz: 2015.7
blosc: None
bottleneck: None
tables: None
numexpr: None
matplotlib: None
openpyxl: None
xlrd: None
xlwt: None
xlsxwriter: None
lxml: None
bs4: None
html5lib: None
httplib2: 0.9.2
apiclient: 1.4.2
sqlalchemy: None
pymysql: None
psycopg2: 2.6.1 (dt dec pq3 ext lo64)
Jinja2: None

[jreback]

hmm that does look like a bug.

[gfyoung]

Where is "-" being implemented in for Series? The sub() and subtract() methods work perfectly fine. I suspect it's somewhere hidden here, but if someone with a better understanding could enlighten, that would be great!

[jreback]

@gfyoung that's the right place. you should just step thru.

[gfyoung]

pandas is trolling me.

I was able to track the bug down to this line here. For some strange reason, the lvalue (and also the rvalue) become NaT which is why the difference becomes zero because the np.int64 view value is just np.iinfo(np.int64).min. However, my attempts to replicate that error directly were strangely not very successful.

What I did was this, using a local installation of the codebase:

1. Right before this line, I placed a raise Exception(lvalues, rvalues). Afterwards, if you walk through the code presented in the issue, you should be able to confirm that they are both DateTimeIndex objects with the correct times.
2. I then removed the Exception, and right before this line, I placed a raise Exception(lvalues, rvalues). Afterwards, if you walk through the code presented in the issue, you should be able to confirm that both are converted to NaT.
3. I then removed the Exception, and right before this line. I assigned attributes old_lvalues and old_rvalues to the _TimeOp attribute.
4. I then ran this code:

import operator
import pandas as pd

df = pd.DataFrame(columns=['s', 'f'])

df.s = [pd.Timestamp('2016-02-08 13:43:14.605000', tz='America/Sao_Paulo')]
df.f = [pd.Timestamp('2016-02-10 13:43:14.605000', tz='America/Sao_Paulo')]

t = _TimeOp.maybe_convert_for_time_op(df.s, df.f, '__sub__', operator.sub)

print t.lvalue, t.rvalue
print t.old_lvalues.tz_localize(None), t.old_rvalues.tz_localize(None)

You should be able to see my confusion.

[jreback]

that IS odd.

[gfyoung]

@jreback @polart : Also, the title should be changed, as it's really an issue with Series arithmetic operations and has nothing to do actually with DataFrame.