HDF5 and Stata I/O are broken on some architectures

Fix some issues, warn on use and xfail tests for the remainder

armhf TestHDF5Store::test*encoding only sometimes crashes
(1.1.3+dfsg-1 passed on build but failed autopkgtest)

HDF5 and Stata are known to fail on big-endian architectures
Stata also fails on qemu-ppc64el, but not real ppc64el

Author: Andreas Tille <[email protected]>, Graham Inggs <[email protected]>, Yaroslav Halchenko <[email protected]>, Rebecca N. Palmer <[email protected]>
Bug-Debian: https://bugs.debian.org/877419
Forwarded: no


Gbp-Pq: Name xfail_tests_nonintel_io.patch

Author: Debian Science Team

pandas/_testing/_warnings.py

@@ -8,6 +8,7 @@
     cast,
 )
 import warnings
+import platform
 
 
 @contextmanager
@@ -149,6 +150,8 @@ def _assert_caught_no_extra_warnings(
                 # FIXME: kludge because pytest.filterwarnings does not
                 #  suppress these, xref GH#38630
                 continue
+            if actual_warning.category==UserWarning and "Non-x86 system detected" in str(actual_warning.message) and not bool(re.match('i.?86|x86',platform.uname()[4])):
+                continue
 
             extra_warnings.append(
                 (

pandas/io/pytables.py

@@ -23,6 +23,10 @@
     cast,
 )
 import warnings
+import platform
+import re
+from pandas.compat import is_platform_little_endian
+warn_hdf_platform = "Non-x86 system detected, HDF(5) format I/O may give wrong results (particularly on files created with older versions) or crash - https://bugs.debian.org/877419" if not bool(re.match('i.?86|x86',platform.uname()[4])) else False
 
 import numpy as np
 
@@ -565,6 +569,8 @@ def __init__(
         fletcher32: bool = False,
         **kwargs,
     ):
+        if warn_hdf_platform:
+            warnings.warn(warn_hdf_platform)
 
         if "format" in kwargs:
             raise ValueError("format is not a defined argument for HDFStore")
@@ -776,7 +782,10 @@ def flush(self, fsync: bool = False):
             self._handle.flush()
             if fsync:
                 with suppress(OSError):
-                    os.fsync(self._handle.fileno())
+                    if is_platform_little_endian():
+                        os.fsync(self._handle.fileno())
+                    else:
+                        os.sync() # due to a pytables bad-cast bug, fileno is invalid on 64-bit big-endian
 
     def get(self, key: str):
         """

pandas/io/stata.py

@@ -25,6 +25,9 @@
     cast,
 )
 import warnings
+import platform
+import re
+warn_stata_platform = "Non-x86 system detected, Stata format I/O may give wrong results (particularly on strings) - https://bugs.debian.org/877419" if not bool(re.match('i.?86|x86',platform.uname()[4])) else False
 
 from dateutil.relativedelta import relativedelta
 import numpy as np
@@ -896,6 +899,8 @@ def __init__(self):
         # NOTE: the byte type seems to be reserved for categorical variables
         # with a label, but the underlying variable is -127 to 100
         # we're going to drop the label and cast to int
+        if warn_stata_platform:
+            warnings.warn(warn_stata_platform)
         self.DTYPE_MAP = dict(
             list(zip(range(1, 245), [np.dtype("a" + str(i)) for i in range(1, 245)]))
             + [

pandas/tests/io/pytables/test_append.py

@@ -24,6 +24,10 @@
     ensure_clean_path,
     ensure_clean_store,
 )
+import platform
+import re
+import sys
+is_crashing_arch=bool((platform.uname()[4].startswith('arm') or platform.uname()[4].startswith('aarch')) and sys.maxsize<2**33) # meant for armhf, though this form will also skip on armel - uname = kernel arch
 
 pytestmark = pytest.mark.single
 
@@ -277,6 +281,7 @@ def test_append_all_nans(setup_path):
         tm.assert_frame_equal(store["df2"], df)
 
 
+@pytest.mark.xfail(condition=is_crashing_arch,reason="https://bugs.debian.org/790925",strict=False,run=False)
 def test_append_frame_column_oriented(setup_path):
     with ensure_clean_store(setup_path) as store:
 

pandas/tests/io/pytables/test_file_handling.py

@@ -25,6 +25,10 @@
     PossibleDataLossError,
     Term,
 )
+import platform
+import re
+import sys
+is_crashing_arch=bool((platform.uname()[4].startswith('arm') or platform.uname()[4].startswith('aarch')) and sys.maxsize<2**33) # meant for armhf, though this form will also skip on armel - uname = kernel arch
 
 pytestmark = pytest.mark.single
 
@@ -270,6 +274,7 @@ def test_complibs(setup_path):
             h5table.close()
 
 
+@pytest.mark.xfail(condition=is_crashing_arch,reason="https://bugs.debian.org/790925",strict=False,run=False)
 @pytest.mark.skipif(
     not is_platform_little_endian(), reason="reason platform is not little endian"
 )
@@ -303,6 +308,7 @@ def test_encoding(setup_path):
     ],
 )
 @pytest.mark.parametrize("dtype", ["category", object])
+@pytest.mark.xfail(condition=is_crashing_arch,reason="https://bugs.debian.org/790925",strict=False,run=False)
 def test_latin_encoding(setup_path, dtype, val):
     enc = "latin-1"
     nan_rep = ""

pandas/tests/io/pytables/test_read.py

@@ -5,7 +5,7 @@
 import pytest
 
 from pandas._libs.tslibs import Timestamp
-from pandas.compat import is_platform_windows
+from pandas.compat import is_platform_windows, is_platform_little_endian
 
 import pandas as pd
 from pandas import (
@@ -155,6 +155,7 @@ def test_pytables_native2_read(datapath, setup_path):
         assert isinstance(d1, DataFrame)
 
 
+@pytest.mark.xfail(condition=not is_platform_little_endian(),reason="known failure of hdf on non-little endian",strict=False,raises=AttributeError)
 def test_legacy_table_fixed_format_read_py2(datapath, setup_path):
     # GH 24510
     # legacy table with fixed format written in Python 2
@@ -170,6 +171,7 @@ def test_legacy_table_fixed_format_read_py2(datapath, setup_path):
         tm.assert_frame_equal(expected, result)
 
 
+@pytest.mark.xfail(condition=not is_platform_little_endian(),reason="known failure of hdf on non-little endian",strict=False,raises=AttributeError)
 def test_legacy_table_fixed_format_read_datetime_py2(datapath, setup_path):
     # GH 31750
     # legacy table with fixed format and datetime64 column written in Python 2
@@ -319,6 +321,7 @@ def test_read_hdf_series_mode_r(format, setup_path):
     tm.assert_series_equal(result, series)
 
 
+@pytest.mark.xfail(condition=not is_platform_little_endian(),reason="known failure of hdf on non-little endian",strict=False,raises=AttributeError)
 def test_read_py2_hdf_file_in_py3(datapath):
     # GH 16781
 

pandas/tests/io/pytables/test_store.py

@@ -39,6 +39,10 @@
     HDFStore,
     read_hdf,
 )
+import platform
+import re
+import sys
+is_crashing_arch=bool((platform.uname()[4].startswith('arm') or platform.uname()[4].startswith('aarch')) and sys.maxsize<2**33) # meant for armhf, though this form will also skip on armel - uname = kernel arch
 
 pytestmark = pytest.mark.single
 
@@ -789,6 +793,7 @@ def test_start_stop_fixed(setup_path):
         df.iloc[8:10, -2] = np.nan
 
 
+@pytest.mark.xfail(condition=is_crashing_arch,reason="https://bugs.debian.org/790925",strict=False,run=False)
 def test_select_filter_corner(setup_path):
 
     df = DataFrame(np.random.randn(50, 100))

pandas/tests/io/pytables/test_timezones.py

@@ -8,6 +8,7 @@
 
 from pandas._libs.tslibs.timezones import maybe_get_tz
 import pandas.util._test_decorators as td
+from pandas.compat import is_platform_little_endian
 
 import pandas as pd
 from pandas import (
@@ -304,6 +305,7 @@ def test_store_timezone(setup_path):
         tm.assert_frame_equal(result, df)
 
 
+@pytest.mark.xfail(condition=not is_platform_little_endian(),reason="known failure of hdf on non-little endian",strict=False,raises=AttributeError)
 def test_legacy_datetimetz_object(datapath, setup_path):
     # legacy from < 0.17.0
     # 8260
@@ -356,6 +358,7 @@ def test_read_with_where_tz_aware_index(setup_path):
         tm.assert_frame_equal(result, expected)
 
 
+@pytest.mark.xfail(condition=not is_platform_little_endian(),reason="known failure of hdf on non-little endian",strict=False,raises=AttributeError)
 def test_py2_created_with_datetimez(datapath, setup_path):
     # The test HDF5 file was created in Python 2, but could not be read in
     # Python 3.

pandas/tests/io/test_common.py

@@ -15,7 +15,7 @@
 
 import pytest
 
-from pandas.compat import is_platform_windows
+from pandas.compat import is_platform_windows, is_platform_little_endian
 import pandas.util._test_decorators as td
 
 import pandas as pd
@@ -244,11 +244,11 @@ def test_read_expands_user_home_dir(
                 "pyarrow",
                 ("io", "data", "feather", "feather-0_3_1.feather"),
             ),
-            (
+            pytest.param(
                 pd.read_hdf,
                 "tables",
                 ("io", "data", "legacy_hdf", "datetimetz_object.h5"),
-            ),
+            marks=pytest.mark.xfail(condition=not is_platform_little_endian(),reason="known failure of hdf on non-little endian",strict=False,raises=AttributeError)),
             (pd.read_stata, "os", ("io", "data", "stata", "stata10_115.dta")),
             (pd.read_sas, "os", ("io", "sas", "data", "test1.sas7bdat")),
             (pd.read_json, "os", ("io", "json", "data", "tsframe_v012.json")),

pandas/tests/io/test_stata.py

@@ -33,6 +33,8 @@
     read_stata,
 )
 
+from pandas.compat import is_platform_little_endian
+pytestmark = pytest.mark.xfail(condition=not is_platform_little_endian(),reason="known failure of test_stata on non-little endian",strict=False)
 
 @pytest.fixture()
 def mixed_frame():
@@ -207,7 +209,7 @@ def test_read_dta2(self):
             # parsed_113 = self.read_dta(self.dta2_113)
 
             # Remove resource warnings
-            w = [x for x in w if x.category is UserWarning]
+            w = [x for x in w if x.category is UserWarning and not "Non-x86 system detected" in str(x.message)]
 
             # should get warning for each call to read_dta
             assert len(w) == 3
@@ -469,7 +471,7 @@ def test_read_write_dta12(self, version):
                 warnings.simplefilter("always", InvalidColumnName)
                 original.to_stata(path, None, version=version)
                 # should get a warning for that format.
-                assert len(w) == 1
+                assert len([x for x in w if not "Non-x86 system detected" in str(x.message)]) == 1
 
             written_and_read_again = self.read_dta(path)
             tm.assert_frame_equal(written_and_read_again.set_index("index"), formatted)
@@ -1793,8 +1795,9 @@ def test_encoding_latin1_118(self):
 the string values returned are correct."""
         with tm.assert_produces_warning(UnicodeWarning) as w:
             encoded = read_stata(self.dta_encoding_118)
-            assert len(w) == 151
-            assert w[0].message.args[0] == msg
+            w2 = [x for x in w if not "Non-x86 system detected" in str(x.message)]
+            assert len(w2) == 151
+            assert w2[0].message.args[0] == msg
 
         expected = DataFrame([["Düsseldorf"]] * 151, columns=["kreis1849"])
         tm.assert_frame_equal(encoded, expected)