FIX: Scalar timedelta NaT results raise

[ischwabacher]

Currently, coercion of scalar results from float to timedelta64[ns] passes through int, which raises when attempting to coerce NaN to NaT.

To reproduce:

import pandas as pd
import numpy as np
pd.Series([np.timedelta64('NaT')]).sum()
# TypeError: reduction operation 'sum' not allowed for this dtype

[jreback]

needs tests (put in tests/test_series.py, after the timdelta operations sections. Need to systematically tests these types of ops (both with and w/o NaT). Might also be some tests in tseries/tests/test_timeseries.py/test_timedeltas.py

e.g. sum,mean,max,min, and that raises on things like var,std

[ischwabacher]

Oh, the problem is that test_nanops.py only covers NaN and ignores NaT. Hrm. Any bets on how much breakage this change is going to reveal?

[ischwabacher]

It looks like np.timedelta64('NaT') creates an object of dtype m8 instead of m8[ns], and the isnull can't handle that. Instead of accepting m8, it would probably be better to coerce np.timedelta64('NaT') to np.timedelta64('NaT', 'ns') whenever it's being added to a Pandas object, if that's feasible.

[jreback]

you don't handle m8 like that
it's an integer array; use tslib.iNaT

[ischwabacher]

I don't understand what you mean by that. I will see if I can get back to it sometime over the weekend.

[ischwabacher]

OK, clearly I need to wait until I have time to really read this and figure out what's going on.

[jreback]

have a look at tseries/test_timedeltas.py. timedeltas are an odd duck. You generally work on an int64 based series, then take a timedelta64[ns] view. note that almost all tests are skipped for numpy < 1.7 (as its broken) for timedeltas.

[jreback]

@ischwabacher status?

[ischwabacher]

Sorry, hectic time at home. I will try to make time for this soon.

[jreback]

closing in favor of #8268

this is already fixed in master with Timedelta refactor