DataFrame.sum() inconsistent behavior treating full row/col of NANs

[siyantan]

Seeing an odd issue. When doing a sum and the full column or row is NAN, I expect NAN as the result. Recently my system has been giving a sum of zero if the full row is NAN, unless skipna=FALSE is specified (but of course that screws up rows with only one NAN, for example).

I would have thought skipna=TRUE, which is default, should give NAN if the full row is NAN per the docs.

This seems to be dependant on some combination of package versions but I cannot figure out why. I reinstalled completely to a fresh Anaconda 2.3.0 amd64 and the behaviour is the same.

See screenshot. For In[6] I expect index [0] to be NAN, not zero.

My package list, for reference:

_license 1.1 py27_0
alabaster 0.7.3 py27_0
anaconda 2.3.0 np19py27_0
argcomplete 0.8.9 py27_0
astropy 1.0.3 np19py27_0
babel 1.3 py27_0
backports.ssl-match-hostname 3.4.0.2
bcolz 0.9.0 np19py27_0
bdateutil 0.1
beautiful-soup 4.3.2 py27_1
beautifulsoup4 4.3.2
binstar 0.11.0 py27_0
bitarray 0.8.1 py27_1
blaze 0.8.0
blaze-core 0.8.0 np19py27_0
blpapi 3.5.3
blz 0.6.2 np19py27_1
bokeh 0.9.0 np19py27_0
boto 2.38.0 py27_0
bottleneck 1.0.0 np19py27_0
cdecimal 2.3 py27_1
certifi 14.05.14 py27_0
cffi 1.1.0 py27_0
clyent 0.3.4 py27_0
colorama 0.3.3 py27_0
conda 3.15.1 py27_0
conda-build 1.14.1 py27_0
conda-env 2.3.0 py27_0
configobj 5.0.6 py27_0
cryptography 0.9.1 py27_0
cvxopt 1.1.7
cython 0.22.1 py27_0
cytoolz 0.7.3 py27_0
datashape 0.4.5 np19py27_0
dateutil 2.4.1 py27_0
decorator 3.4.2 py27_0
docutils 0.12 py27_1
dynd-python 0.6.5 np19py27_0
enum34 1.0.4 py27_0
fastcache 1.0.2 py27_0
flask 0.10.1 py27_1
funcsigs 0.4 py27_0
gevent 1.0.1 py27_0
gevent-websocket 0.9.3 py27_0
greenlet 0.4.7 py27_0
grin 1.2.1 py27_2
h5py 2.5.0 np19py27_1
hdf5 1.8.15.1 2
holidays 0.3.1
idna 2.0 py27_0
ipaddress 1.0.7 py27_0
ipython 3.2.0 py27_0
ipython-notebook 3.2.0 py27_0
ipython-qtconsole 3.2.0 py27_0
itsdangerous 0.24 py27_0
jdcal 1.0 py27_0
jedi 0.8.1 py27_0
jinja2 2.7.3 py27_2
jsonschema 2.4.0 py27_0
launcher 1.0.0 1
llvmlite 0.5.0 py27_0
lxml 3.4.4 py27_0
lz4 0.7.0
markupsafe 0.23 py27_0
matplotlib 1.4.3 np19py27_1
menuinst 1.0.4 py27_0
mistune 0.5.1 py27_1
mock 1.0.1 py27_0
multipledispatch 0.4.7 py27_0
networkx 1.9.1 py27_0
nltk 3.0.3 np19py27_0
node-webkit 0.10.1 0
nose 1.3.7 py27_0
numba 0.19.1 np19py27_0
numexpr 2.4.3 np19py27_0
numpy 1.9.2 py27_0
odo 0.3.2 np19py27_0
openpyxl 1.8.5 py27_0
pandas 0.14.1 np19py27_0
patsy 0.3.0 np19py27_0
pep8 1.6.2 py27_0
pillow 2.8.2 py27_0
pip 7.1.0 py27_0
ply 3.6 py27_0
psutil 2.2.1 py27_0
py 1.4.27 py27_0
pyasn1 0.1.7 py27_0
pycosat 0.6.1 py27_0
pycparser 2.14 py27_0
pycrypto 2.6.1 py27_3
pyflakes 0.9.2 py27_0
pygments 2.0.2 py27_0
pymongo 2.8 py27_0
pyodbc 3.0.7 py27_0
pyopenssl 0.15.1 py27_1
pyparsing 2.0.3 py27_0
pyqt 4.10.4 py27_1
pyreadline 2.0 py27_0
pytables 3.2.0 np19py27_0
pytest 2.7.1 py27_0
python 2.7.10 0
python-dateutil 2.4.2 py27_0
pytz 2015.4 py27_0
pywin32 219 py27_0
pyyaml 3.11 py27_1
pyzmq 14.7.0 py27_0
requests 2.7.0 py27_0
rope 0.9.4 py27_1
runipy 0.1.3 py27_0
scikit-image 0.11.3 np19py27_0
scikit-learn 0.16.1 np19py27_0
scipy 0.16.0 np19py27_0
setuptools 18.0.1 py27_0
six 1.9.0 py27_0
snowballstemmer 1.2.0 py27_0
sockjs-tornado 1.0.1 py27_0
sphinx 1.3.1 py27_0
sphinx-rtd-theme 0.1.7
sphinx_rtd_theme 0.1.7 py27_0
spyder 2.3.5.2 py27_0
spyder-app 2.3.5.2 py27_0
sqlalchemy 1.0.5 py27_0
ssl_match_hostname 3.4.0.2 py27_0
statsmodels 0.6.1 np19py27_0
sympy 0.7.6 py27_0
tables 3.2.0
toolz 0.7.2 py27_0
tornado 4.2 py27_0
ujson 1.33 py27_0
unicodecsv 0.9.4 py27_0
werkzeug 0.10.4 py27_0
xlrd 0.9.3 py27_0
xlsxwriter 0.7.3 py27_0
xlwings 0.3.5 py27_0
xlwt 1.0.0 py27_0
zlib 1.2.8 0

[jreback]

dupe of #9422

this was a change in bottleneck 1.0. see that issue for the soln, which is pretty straightforward.