BLD: sparc test_decimalDecodeTest - {u'a': 4.56} + {u'a': 4.5600000000000005}

[yarikoptic]

======================================================================
FAIL: test_decimalDecodeTest (pandas.io.tests.test_json.test_ujson.UltraJSONTests)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/yoh/deb/gits/pkg-exppsy/pandas/debian/tmp/usr/lib/python2.7/dist-packages/pandas/io/tests/test_json/test_ujson.py", line 102, in test_decimalDecodeTest
    self.assertEqual(sut, decoded)
AssertionError: {u'a': 4.56} != {u'a': 4.5600000000000005}
- {u'a': 4.56}
+ {u'a': 4.5600000000000005}

on 0.12.0~rc1+git43-g7b2eaa4-1

[jreback]

cc @Komnomnomnom

[Komnomnomnom]

I'm tempted to remove this test. It's an upstream test that was originally testing for failure when round-trip serialising {u'a': 4.56}, but it turned out the test only passed on 64-bit machines. I modified it to pass on 32-bit but it's fudgy to begin with.

The test after this one verifies that {u'a': 4.56} is properly round-trip serialised when the precise_float flag is passed, IMO that's enough for testing this particular scenario.

Any thoughts on this?

[jreback]

FYI....the sparcs are little-endian which causes some issues in a couple of non-related tests. Does this make a difference?

[Komnomnomnom]

I'm no expert but I guess it might be what's causing the precision difference. However imo this test is faulty regardless, it is currently known that round-trip serialisation of {u'a': 4.56} will

• fail on 64-bit arch ==> test PASS
• succeed on 32-bit arch ==> test PASS after modifications by me
• fail on 32-bit(?) sparc ==> test FAIL

Given that there's another test which verifies that round-trip serialisation of {u'a': 4.56} will succeed with the precise_float flag turned on and this appears to pass on all platforms I'd argue for removing this test rather than fudging it again to pass on a new architecture.

I'll push the same change / idea up to ultrajson (where perhaps this discussion would be more appropriate) but I just wanted to get a consensus here first.

[jreback]

alright....do a quick PR to remove/skip this test then?

[jreback]

@yarikoptic

actually I think this is still open...hold off a bit on retest

[yarikoptic]

no problem -- I can totally do that ;)

On Sun, 14 Jul 2013, jreback wrote:

[1]@yarikoptic

actually I think this is still open...hold off a bit on retest

Yaroslav O. Halchenko, Ph.D.
http://neuro.debian.net http://www.pymvpa.org http://www.fail2ban.org
Senior Research Associate, Psychological and Brain Sciences Dept.
Dartmouth College, 419 Moore Hall, Hinman Box 6207, Hanover, NH 03755
Phone: +1 (603) 646-9834 Fax: +1 (603) 646-1419
WWW: http://www.linkedin.com/in/yarik

[jreback]

@yarikoptic why don't u retest master and let us know
thxs

[jreback]

this was closed by #4223