0.9.1 test failures on big endian machines

[juliantaylor]

see
https://launchpadlibrarian.net/123637321/buildlog_ubuntu-raring-powerpc.pandas_0.9.1-1ubuntu1_FAILEDTOBUILD.txt.gz

https://buildd.debian.org/status/package.php?p=pandas&suite=experimental

======================================================================
ERROR: test_fperr_robustness (pandas.stats.tests.test_moments.TestMoments)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/build/buildd/pandas-0.9.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/stats/tests/test_moments.py", line 175, in test_fperr_robustness
    result = mom.rolling_sum(arr, 2)
  File "/build/buildd/pandas-0.9.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/stats/moments.py", line 458, in f
    freq=freq, time_rule=time_rule, **kwargs)
  File "/build/buildd/pandas-0.9.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/stats/moments.py", line 261, in _rolling_moment
    result = np.apply_along_axis(calc, axis, values)
  File "/usr/lib/python2.7/dist-packages/numpy/lib/shape_base.py", line 80, in apply_along_axis
    res = func1d(arr[tuple(i.tolist())],*args)
  File "/build/buildd/pandas-0.9.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/stats/moments.py", line 258, in <lambda>
    calc = lambda x: func(x, window, minp=minp, **kwargs)
  File "/build/buildd/pandas-0.9.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/stats/moments.py", line 456, in call_cython
    return func(arg, window, minp, **kwds)
  File "moments.pyx", line 156, in pandas.lib.roll_sum (pandas/src/tseries.c:77996)
ValueError: Little-endian buffer not supported on big-endian compiler

======================================================================
FAIL: test_from_M8_structured (pandas.tseries.tests.test_timeseries.TestLegacySupport)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/build/buildd/pandas-0.9.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/tseries/tests/test_timeseries.py", line 1917, in test_from_M8_structured
    self.assertEqual(df['Date'][0], dates[0][0])
AssertionError: <Timestamp: 1976-02-12 15:25:34.986240> != datetime.datetime(2012, 9, 9, 0, 0)

[juliantaylor]

the first should probably be skipped on foreign endian machines, or the input data swapped before trying frombuffer

the second can be reproduced on little endian with:

import pandas
import numpy as np
from datetime import datetime
d = datetime(2012, 9, 9, 0, 0)
arr = np.array([d], dtype='>M8[us]')
arr[0] == d
df = pandas.DataFrame(arr)
df[0][0] == d
print df[0][0]
print arr[0]

[juliantaylor]

the problem is arr.view(np.int64) in cast_to_nanoseconds
it does not take the endianess into account:

In [15]: a = np.array([1347148800000000], dtype=">i8")

In [16]: a.view(np.int64)
Out[16]: array([192986734986240])

[ghost]

The first test has a comment saying it should be removed when 2.5 is no longer supported (which has already happend).

The second test creates data specifically in little endian, regardless of the base endianess. I think it might be the test
rather then the code which does the wrong thing.

@juliantaylor, does the following raise an error on your test system?

from pandas import *
dates = [ (datetime(2012, 9, 9, 0, 0),
           datetime(2012, 9, 8, 15, 10))]
arr = np.array(dates,
               dtype=[('Date', 'M8[us]'), ('Forecasting', 'M8[us]')])
df = DataFrame(arr)

assert(df['Date'][0]== dates[0][0])

[juliantaylor]

I agree its probably more likely a problem of the test than the code.
But it might be useful to catch wrong endianess at a high level before you get wrong results.

unfortunately I don't have direct access to a big endian machine, but your code is essentially what happens in x86 were the tests succeed.

[ghost]

#2359 needs testing.