sym_diff failure on 3.4?

[dsm054]

test_symmetric_diff is failing after trying out pandas on a fresh 3.4 pull. It passes on 3.3 with the same numpy trunk version. It seems to have to do with the nan section:

>>> from pandas import Index
>>> import numpy as np
>>> idx1 = Index([1, 2, np.nan])
>>> idx2 = Index([0, 1, np.nan])
>>> result = idx1.sym_diff(idx2)
>>> expected = Index([0.0, np.nan, 2.0, np.nan])  # oddness with nans
>>> nans = pd.isnull(expected)
>>> result
Float64Index([0.0, nan, nan, 2.0], dtype='object')
>>> expected
Float64Index([0.0, nan, 2.0, nan], dtype='object')
>>> nans
array([False,  True, False,  True], dtype=bool)
>>> result[nans]
Float64Index([nan, 2.0], dtype='object')

Version info (basically a fresh 3.4 build with only numpy dev installed):

>>> pd.show_versions()

INSTALLED VERSIONS
------------------
commit: None
python: 3.4.0.candidate.1
python-bits: 32
OS: Linux
OS-release: 3.8.0-35-generic
machine: i686
processor: i686
byteorder: little
LC_ALL: None
LANG: en_CA.UTF-8

pandas: 0.13.1-282-g9564ead
Cython: None
numpy: 1.9.0.dev-2d6ea6e
scipy: None
statsmodels: None
IPython: None
sphinx: None
patsy: None
scikits.timeseries: None
dateutil: 2.2
pytz: 2013.9
bottleneck: None
tables: None
numexpr: None
matplotlib: None
openpyxl: None
xlrd: None
xlwt: None
xlsxwriter: None
lxml: None
bs4: None
html5lib: None
bq: None
apiclient: None
rpy2: None
sqlalchemy: None
pymysql: None
psycopg2: None

[jreback]

cc @TomAugspurger

yeh...Tom will have a look when he tries out 3.4

its has to do with nan ordering. weird

[jreback]

@dsm054 you can take a look if you'd like as well!

right now I only tests 3.4 on windows...and has been coming up since this was merged in

#6016

[dsm054]

Not sure how I missed that in the search. Should we close this as a dup?

[jreback]

wasn't a 'direct' issue, just a comment...so this is fine

[TomAugspurger]

I haven't been able to get a 3.4 virtualenv running yet. Does machine with 3.4 have a newer version of numpy as well?

On Feb 22, 2014, at 11:46 AM, "jreback" <notifications@github.commailto:notifications@github.com> wrote:

cc @TomAugspurgerhttps://github.com/TomAugspurger

yeh...Tom will have a look when he tries out 3.4

its has to do with nan ordering. weird

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/6444#issuecomment-35808943.

[dsm054]

What's the "expected" (= desired, really) behaviour for sorting an array with nans?

[jreback]

@TomAugspurger you can use numpy 1.8.

[jreback]

hmm...if you sort with a stable (mergesort), it should leave the nans alone (and you can't tell if 2 nans switch place). though IIRC their was an issue on the sort ordering....

[dsm054]

I had another look at this and I'm starting to think there's nothing wrong here. It's simply that in Python 3.3 we have

>>> set(Float64Index([0.0, np.nan, np.nan, 2.0], dtype='object'))
{0.0, nan, 2.0, nan}

and in 3.4 we have

>>> set(Float64Index([0.0, np.nan, np.nan, 2.0], dtype='object'))
{0.0, nan, nan, 2.0}

and we were never promised otherwise. If sorting preserves nan location (although annoyingly, the presence of nans breaks the sorting even of the non-nan elements in both lists and ndarrays), then since sets are unordered, we have no reason to expect our expected result. If we want to impose Series-style "push nans to the back" sorting, we can, but right now I think it's just that the test is too sensitive.

[TomAugspurger]

Agreed. I think I'll rewrite the test to make sure the count of the nans is correct, and ignore the order.

[jreback]

ok

I suspect maybe 3.4 changed some sort of hashing scheme though
eg they are now more pseudo random whereas 3.3 it is not turned on

this is some of security thing I think

you can prob test this by seeing if the ordering of a dict keys is the same in 3.3 vs 3.4

[jreback]

http://bugs.python.org/issue19183

is prob the culprit