HDF5 and Stata I/O are broken on some architectures

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

Everything that has a run=False xfail in here should also be in
the run-and-ignore set in debian/tests/numbatests

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 was previously seen to fail 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
Bug: partly pandas-dev/pandas#54396
Forwarded: no


Gbp-Pq: Name xfail_tests_nonintel_io.patch

Author: Debian Science Team

pandas/_testing/_warnings.py

@@ -12,6 +12,7 @@
     cast,
 )
 import warnings
+import platform
 
 from pandas.compat import PY311
 
@@ -186,6 +187,8 @@ def _assert_caught_no_extra_warnings(
                 # pyproject.toml errors on EncodingWarnings in pandas
                 # Ignore EncodingWarnings from other libraries
                 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]))) or (actual_warning.category==RuntimeWarning and "invalid value encountered" in str(actual_warning.message) and 'mips' in platform.uname()[4]):
+                continue
             extra_warnings.append(
                 (
                     actual_warning.category.__name__,

pandas/io/pytables.py

@@ -24,6 +24,10 @@
     overload,
 )
 import warnings
+import platform
+import sys
+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 (((platform.uname()[4].startswith('arm') or platform.uname()[4].startswith('aarch')) and sys.maxsize<2**33) or not is_platform_little_endian()) else False
 
 import numpy as np
 
@@ -559,6 +563,8 @@ def __init__(
         fletcher32: bool = False,
         **kwargs,
     ) -> None:
+        if warn_hdf_platform:
+            warnings.warn(warn_hdf_platform)
         if "format" in kwargs:
             raise ValueError("format is not a defined argument for HDFStore")
 
@@ -780,7 +786,10 @@ def flush(self, fsync: bool = False) -> None:
             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

@@ -30,6 +30,8 @@
     cast,
 )
 import warnings
+from pandas.compat import is_platform_little_endian
+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 is_platform_little_endian() else False
 
 import numpy as np
 
@@ -976,6 +978,8 @@ def __init__(self) -> None:
         # 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(
             [(i, np.dtype(f"S{i}")) for i in range(1, 245)]
             + [

pandas/tests/io/pytables/test_append.py

@@ -21,6 +21,10 @@
     _maybe_remove,
     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_cpu
 
@@ -275,6 +279,7 @@ def test_append_all_nans(setup_path):
             tm.assert_frame_equal(store["df2"], df, check_index_type=True)
 
 
+@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:
         # column oriented

pandas/tests/io/pytables/test_file_handling.py

@@ -26,6 +26,10 @@
     ensure_clean_store,
     tables,
 )
+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
 
 from pandas.io import pytables
 from pandas.io.pytables import Term
@@ -267,6 +271,7 @@ def test_complibs(tmp_path, lvl, lib):
                 assert node.filters.complib == lib
 
 
+@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"
 )
@@ -299,6 +304,7 @@ 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(tmp_path, setup_path, dtype, val):
     enc = "latin-1"
     nan_rep = ""

pandas/tests/io/pytables/test_read.py

@@ -6,7 +6,7 @@
 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 (
@@ -167,6 +167,7 @@ def test_pytables_native2_read(datapath):
         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):
     # GH 24510
     # legacy table with fixed format written in Python 2
@@ -182,6 +183,7 @@ def test_legacy_table_fixed_format_read_py2(datapath):
         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):
     # GH 31750
     # legacy table with fixed format and datetime64 column written in Python 2
@@ -364,6 +366,7 @@ def test_read_hdf_series_mode_r(tmp_path, format, setup_path):
 
 @pytest.mark.filterwarnings(r"ignore:Period with BDay freq is deprecated:FutureWarning")
 @pytest.mark.filterwarnings(r"ignore:PeriodDtype\[B\] is deprecated:FutureWarning")
+@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

@@ -29,6 +29,10 @@
     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_cpu
 
@@ -759,6 +763,7 @@ 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.default_rng(2).standard_normal((50, 100)))
     df.index = [f"{c:3d}" for c in df.index]

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 (
@@ -308,6 +309,7 @@ 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):
     # legacy from < 0.17.0
     # 8260
@@ -360,6 +362,7 @@ def test_read_with_where_tz_aware_index(tmp_path, 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):
     # The test HDF5 file was created in Python 2, but could not be read in
     # Python 3.

pandas/tests/io/test_common.py

@@ -17,7 +17,7 @@
 
 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
@@ -300,11 +300,11 @@ 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

@@ -32,6 +32,8 @@
     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():

pyproject.toml

@@ -480,6 +480,7 @@ filterwarnings = [
   "error:::pandas",
   "error::ResourceWarning",
   "error::pytest.PytestUnraisableExceptionWarning",
+  "ignore:Non-x86 system detected:UserWarning:pandas",
   # TODO(PY311-minimum): Specify EncodingWarning
   # Ignore 3rd party EncodingWarning but raise on pandas'
   "ignore:.*encoding.* argument not specified",