enable ASAN/UBSAN in pandas CI (#55102)

* enable ASAN/UBSAN in pandas CI

* try input

* try removing sanitize

* try no CFLAGS

* try GH string substituion

* change flags in build script

* quotes

* update script run

* single_cpu updates

* asan checks for datetime funcs

* try smaller config

* checkpoint

* bool fixup

* reverts

* known UB marker

* Finished marking tests with known UB

* dedicated CI job

* identifier fix

* fixes

* more test skip

* try quotes

* simplify ci

* try CFLAGS

* preload args

* skip single_cpu tests

* wording

* removed unneeded marker

* float set implementations

* Revert "float set implementations"

This reverts commit 6266422.

* change marker name

* dedicated actions file

* consolidated into matrix

* fixup

* typos

* fixups

* add qt?

* intentional UB with verbose

* disable pytest-xdist

* original issue

* remove UB

* Revert "remove UB"

This reverts commit 677da0e.

* merge fixup

* remove UB

---------

Co-authored-by: Thomas Li <[email protected]>

.github/actions/build_pandas/action.yml

@@ -4,6 +4,12 @@ inputs:
   editable:
     description: Whether to build pandas in editable mode (default true)
     default: true
+  meson_args:
+    description: Extra flags to pass to meson
+    required: false
+  cflags_adds:
+    description: Items to append to the CFLAGS variable
+    required: false
 runs:
   using: composite
   steps:
@@ -24,11 +30,12 @@ runs:
 
     - name: Build Pandas
       run: |
+        export CFLAGS="$CFLAGS ${{ inputs.cflags_adds }}"
         if [[ ${{ inputs.editable }} == "true" ]]; then
-          pip install -e . --no-build-isolation -v --no-deps \
+          pip install -e . --no-build-isolation -v --no-deps ${{ inputs.meson_args }} \
             --config-settings=setup-args="--werror"
         else
-          pip install . --no-build-isolation -v --no-deps \
+          pip install . --no-build-isolation -v --no-deps ${{ inputs.meson_args }} \
             --config-settings=setup-args="--werror"
         fi
       shell: bash -el {0}

.github/actions/run-tests/action.yml

@@ -1,9 +1,16 @@
 name: Run tests and report results
+inputs:
+  preload:
+    description: Preload arguments for sanitizer
+    required: false
+  asan_options:
+    description: Arguments for Address Sanitizer (ASAN)
+    required: false
 runs:
   using: composite
   steps:
     - name: Test
-      run: ci/run_tests.sh
+      run: ${{ inputs.asan_options }} ${{ inputs.preload }} ci/run_tests.sh
       shell: bash -el {0}
 
     - name: Publish test results

.github/workflows/unit-tests.yml

@@ -96,6 +96,14 @@ jobs:
           - name: "Pyarrow Nightly"
             env_file: actions-311-pyarrownightly.yaml
             pattern: "not slow and not network and not single_cpu"
+          - name: "ASAN / UBSAN"
+            env_file: actions-311-sanitizers.yaml
+            pattern: "not slow and not network and not single_cpu and not skip_ubsan"
+            asan_options: "ASAN_OPTIONS=detect_leaks=0"
+            preload: LD_PRELOAD=$(gcc -print-file-name=libasan.so)
+            meson_args: --config-settings=setup-args="-Db_sanitize=address,undefined"
+            cflags_adds: -fno-sanitize-recover=all
+            pytest_workers: -1  # disable pytest-xdist as it swallows stderr from ASAN
       fail-fast: false
     name: ${{ matrix.name || format('ubuntu-latest {0}', matrix.env_file) }}
     env:
@@ -105,7 +113,7 @@ jobs:
       PANDAS_COPY_ON_WRITE: ${{ matrix.pandas_copy_on_write || '0' }}
       PANDAS_CI: ${{ matrix.pandas_ci || '1' }}
       TEST_ARGS: ${{ matrix.test_args || '' }}
-      PYTEST_WORKERS: 'auto'
+      PYTEST_WORKERS: ${{ matrix.pytest_workers || 'auto' }}
       PYTEST_TARGET: ${{ matrix.pytest_target || 'pandas' }}
       # Clipboard tests
       QT_QPA_PLATFORM: offscreen
@@ -174,16 +182,25 @@ jobs:
     - name: Build Pandas
       id: build
       uses: ./.github/actions/build_pandas
+      with:
+        meson_args: ${{ matrix.meson_args }}
+        cflags_adds: ${{ matrix.cflags_adds }}
 
     - name: Test (not single_cpu)
       uses: ./.github/actions/run-tests
       if: ${{ matrix.name != 'Pypy' }}
+      with:
+        preload: ${{ matrix.preload }}
+        asan_options: ${{ matrix.asan_options }}
       env:
         # Set pattern to not single_cpu if not already set
         PATTERN: ${{ env.PATTERN == '' && 'not single_cpu' || matrix.pattern }}
 
     - name: Test (single_cpu)
       uses: ./.github/actions/run-tests
+      with:
+        preload: ${{ matrix.preload }}
+        asan_options: ${{ matrix.asan_options }}
       env:
         PATTERN: 'single_cpu'
         PYTEST_WORKERS: 0

ci/deps/actions-311-sanitizers.yaml

@@ -0,0 +1,32 @@
+name: pandas-dev
+channels:
+  - conda-forge
+dependencies:
+  - python=3.11
+
+  # build dependencies
+  - versioneer[toml]
+  - cython>=0.29.33
+  - meson[ninja]=1.2.1
+  - meson-python=0.13.1
+
+  # test dependencies
+  - pytest>=7.3.2
+  - pytest-cov
+  - pytest-xdist>=2.2.0
+  - pytest-localserver>=0.7.1
+  - pytest-qt>=4.2.0
+  - boto3
+  - hypothesis>=6.46.1
+  - pyqt>=5.15.9
+
+  # required dependencies
+  - python-dateutil
+  - numpy<2
+  - pytz
+
+  # pandas dependencies
+  - pip
+
+  - pip:
+    - "tzdata>=2022.7"

pandas/tests/frame/test_constructors.py

@@ -3206,6 +3206,7 @@ def test_from_out_of_bounds_ns_datetime(
         assert item.asm8.dtype == exp_dtype
         assert dtype == exp_dtype
 
+    @pytest.mark.skip_ubsan
     def test_out_of_s_bounds_datetime64(self, constructor):
         scalar = np.datetime64(np.iinfo(np.int64).max, "D")
         result = constructor(scalar)
@@ -3241,6 +3242,7 @@ def test_from_out_of_bounds_ns_timedelta(
         assert item.asm8.dtype == exp_dtype
         assert dtype == exp_dtype
 
+    @pytest.mark.skip_ubsan
     @pytest.mark.parametrize("cls", [np.datetime64, np.timedelta64])
     def test_out_of_s_bounds_timedelta64(self, constructor, cls):
         scalar = cls(np.iinfo(np.int64).max, "D")

pandas/tests/groupby/test_cumulative.py

@@ -60,6 +60,7 @@ def test_groupby_cumprod():
     tm.assert_series_equal(actual, expected)
 
 
+@pytest.mark.skip_ubsan
 def test_groupby_cumprod_overflow():
     # GH#37493 if we overflow we return garbage consistent with numpy
     df = DataFrame({"key": ["b"] * 4, "value": 100_000})

pandas/tests/io/parser/common/test_float.py

@@ -40,7 +40,14 @@ def test_scientific_no_exponent(all_parsers_all_precisions):
     tm.assert_frame_equal(df_roundtrip, df)
 
 
-@pytest.mark.parametrize("neg_exp", [-617, -100000, -99999999999999999])
+@pytest.mark.parametrize(
+    "neg_exp",
+    [
+        -617,
+        -100000,
+        pytest.param(-99999999999999999, marks=pytest.mark.skip_ubsan),
+    ],
+)
 def test_very_negative_exponent(all_parsers_all_precisions, neg_exp):
     # GH#38753
     parser, precision = all_parsers_all_precisions
@@ -51,6 +58,7 @@ def test_very_negative_exponent(all_parsers_all_precisions, neg_exp):
     tm.assert_frame_equal(result, expected)
 
 
+@pytest.mark.skip_ubsan
 @xfail_pyarrow  # AssertionError: Attributes of DataFrame.iloc[:, 0] are different
 @pytest.mark.parametrize("exp", [999999999999999999, -999999999999999999])
 def test_too_many_exponent_digits(all_parsers_all_precisions, exp, request):

pandas/tests/scalar/timedelta/methods/test_round.py

@@ -61,6 +61,7 @@ def test_round_invalid(self):
             with pytest.raises(ValueError, match=msg):
                 t1.round(freq)
 
+    @pytest.mark.skip_ubsan
     def test_round_implementation_bounds(self):
         # See also: analogous test for Timestamp
         # GH#38964
@@ -86,6 +87,7 @@ def test_round_implementation_bounds(self):
         with pytest.raises(OutOfBoundsTimedelta, match=msg):
             Timedelta.max.round("s")
 
+    @pytest.mark.skip_ubsan
     @given(val=st.integers(min_value=iNaT + 1, max_value=lib.i8max))
     @pytest.mark.parametrize(
         "method", [Timedelta.round, Timedelta.floor, Timedelta.ceil]

pandas/tests/scalar/timedelta/test_arithmetic.py

@@ -966,6 +966,7 @@ def test_td_op_timedelta_timedeltalike_array(self, op, arr):
 
 
 class TestTimedeltaComparison:
+    @pytest.mark.skip_ubsan
     def test_compare_pytimedelta_bounds(self):
         # GH#49021 don't overflow on comparison with very large pytimedeltas
 

pandas/tests/scalar/timedelta/test_timedelta.py

@@ -551,6 +551,7 @@ def test_timedelta_hash_equality(self):
         ns_td = Timedelta(1, "ns")
         assert hash(ns_td) != hash(ns_td.to_pytimedelta())
 
+    @pytest.mark.skip_ubsan
     @pytest.mark.xfail(
         reason="pd.Timedelta violates the Python hash invariant (GH#44504).",
     )

pandas/tests/scalar/timestamp/methods/test_tz_localize.py

@@ -25,6 +25,7 @@
 
 
 class TestTimestampTZLocalize:
+    @pytest.mark.skip_ubsan
     def test_tz_localize_pushes_out_of_bounds(self):
         # GH#12677
         # tz_localize that pushes away from the boundary is OK

pandas/tests/scalar/timestamp/test_constructors.py

@@ -822,6 +822,7 @@ def test_barely_out_of_bounds(self):
         with pytest.raises(OutOfBoundsDatetime, match=msg):
             Timestamp("2262-04-11 23:47:16.854775808")
 
+    @pytest.mark.skip_ubsan
     def test_bounds_with_different_units(self):
         out_of_bounds_dates = ("1677-09-21", "2262-04-12")
 

pandas/tests/tools/test_to_datetime.py

@@ -1140,6 +1140,7 @@ def test_to_datetime_dt64s_out_of_ns_bounds(self, cache, dt, errors):
         assert ts.unit == "s"
         assert ts.asm8 == dt
 
+    @pytest.mark.skip_ubsan
     def test_to_datetime_dt64d_out_of_bounds(self, cache):
         dt64 = np.datetime64(np.iinfo(np.int64).max, "D")
 

pyproject.toml

@@ -523,6 +523,7 @@ markers = [
   "db: tests requiring a database (mysql or postgres)",
   "clipboard: mark a pd.read_clipboard test",
   "arm_slow: mark a test as slow for arm64 architecture",
+  "skip_ubsan: Tests known to fail UBSAN check",
 ]
 
 [tool.mypy]

scripts/tests/data/deps_minimum.toml

@@ -382,6 +382,7 @@ markers = [
   "db: tests requiring a database (mysql or postgres)",
   "clipboard: mark a pd.read_clipboard test",
   "arm_slow: mark a test as slow for arm64 architecture",
+  "skip_ubsan: tests known to invoke undefined behavior",
 ]
 
 [tool.mypy]