BLD: Nosetests failures on 32-bit (test_first_last_nth_dtypes, test_iloc_getitem_array)

[stephenwlin]

I'm getting them on two different 32-bit ubuntu machines updated to master (0.11.0.dev-5f2ec15)...I'm guessing these are 32-bit specific since they're dtype related:

======================================================================
FAIL: test_first_last_nth_dtypes (pandas.tests.test_groupby.TestGroupBy)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/stephen/Projects/pandas/pandas/tests/test_groupby.py", line 176, in test_first_last_nth_dtypes
    assert_frame_equal(first, expected, check_names=False)
  File "/home/stephen/Projects/pandas/pandas/util/testing.py", line 217, in assert_frame_equal
    check_less_precise=check_less_precise)
  File "/home/stephen/Projects/pandas/pandas/util/testing.py", line 185, in assert_series_equal
    assert(left.dtype == right.dtype)
AssertionError

======================================================================
FAIL: test_iloc_getitem_array (pandas.tests.test_indexing.TestIndexing)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/stephen/Projects/pandas/pandas/tests/test_indexing.py", line 345, in test_iloc_getitem_array
    self.check_result('array like', 'iloc', s.index, 'ix', { 0 : [2,4,6], 1 : [3,6,9], 2: [4,8,12] }, typs = ['ints'])
  File "/home/stephen/Projects/pandas/pandas/tests/test_indexing.py", line 249, in check_result
    _eq(t, o, a, obj, key1, k2)
  File "/home/stephen/Projects/pandas/pandas/tests/test_indexing.py", line 206, in _eq
    raise AssertionError(_print('type error'))
AssertionError: None

I can look into this and resolve this myself if no one else does, but just wanted to throw it out there first (knee deep in other stuff atm...)

[jreback]

this is master on 32 bit - latest commit?

these have 2 do with int64 being casted

I'll take a look

weird that Travis does not fail

[jreback]

e04a239

seems to be working on 32 and 64 bit.....

which commit is the issue on ?

[jreback]

ahah it DID show up on a fresh install of 32-bit! really odd...looking

[jreback]

@stephenwlin

the groupby error I have fixed, was not specifying the correct platform comparison (using np.int_ rather than issubclass np.integer),

the 2nd is more complicated (failing in the test_indexing)

lots of places to step thru, but ultiatemly failing on an ndarray take!
I have reduced to this case, maybe an edge case on the new optimization?

In fact this is failing (but only 32-bit!)

from pandas.core.index import Int64Index
x = Int64Index([1, 2, 3], dtype='int64')
y = Int64Index([0, 3, 6, 9], dtype='int64')
y.take(x)
*** TypeError: array cannot be safely cast to required type

[jreback]

#3144 should fix the groupby issue

[jreback]

This works...so maybe the platform indexer conversion?
(which is called on the indexer before being passed in)

This is what take is actually doing.

y.take(com._ensure_platform_int(x))

In [20]: y.take(np.array([1,2,3]))
Out[20]: Int64Index([3, 6, 9], dtype=int64)

In [21]: np.array([1,2,3])
Out[21]: array([1, 2, 3])

In [22]: np.array([1,2,3]).dtype
Out[22]: dtype('int32')

[jreback]

ok....I modified Int64Index to actually use its dtype parameter if its passed (which as far as I can tell is ONLY when an astype in _ensure_platform_int), and added a couple of tests in test_common.py.

I think maybe this 'worked', on 32-bit except in a pathological case, which is weird why a fresh install doesn't work.

[jreback]

@stephenwlin

ok I update master, confirmed fresh installs on 32/64 are ok.....pls verify when you have a chance

[stephenwlin]

yep, ok now, still getting that test_operators_timedelta64 one reported at #3094 though

[jreback]

really, on fresh 32-bit build, on latest master?

[jreback]

what numpy do you use?

[jreback]

hmm...2.7 with 1.7.0 numpy not tested on travis.....I think that is the culprit!

[jreback]

I think I can install that

[jreback]

and of course that's there they messed with timedelta64 (well they were broken in 1.6.2)

[jreback]

repro the error....

[stephenwlin]

should probably follow-up on #3094 and delete all this here

[jreback]

@stephenwlin sorry meant here