Merge pandas3 feature branch into main

ci/cudf_pandas_scripts/pandas-tests/run.sh

@@ -26,7 +26,7 @@ LIBCUDF_WHEELHOUSE=$(RAPIDS_PY_WHEEL_NAME="libcudf_${RAPIDS_PY_CUDA_SUFFIX}" rap
 PYLIBCUDF_WHEELHOUSE=$(rapids-download-from-github "$(rapids-package-name "wheel_python" pylibcudf --stable --cuda "$RAPIDS_CUDA_VERSION")")
 
 # echo to expand wildcard before adding `[extra]` requires for pip
-python -m pip install \
+rapids-pip-retry install \
   "$(echo "${CUDF_WHEELHOUSE}"/cudf_"${RAPIDS_PY_CUDA_SUFFIX}"*.whl)[test,pandas-tests]" \
   "$(echo "${LIBCUDF_WHEELHOUSE}"/libcudf_"${RAPIDS_PY_CUDA_SUFFIX}"*.whl)" \
   "$(echo "${PYLIBCUDF_WHEELHOUSE}"/pylibcudf_"${RAPIDS_PY_CUDA_SUFFIX}"*.whl)"

ci/test_narwhals.sh

@@ -44,6 +44,12 @@ python -c "import narwhals; print(narwhals.show_versions())"
 # test_first_last_expr_over_order_by[cudf] - now passing but tests expect failure
 # test_cast_datetime_tz_aware[cudf]: Passes as of https://github.com/rapidsai/cudf/pull/21451
 # test_nested_structures: Passes as of https://github.com/rapidsai/cudf/pull/21514
+# test_group_by_depth_1_agg_bool_ops[cudf-*]: xpass(strict) - cudf groupby any/all on nullable bool now passes
+# test_replace_time_zone[cudf], test_replace_time_zone_series[cudf]: xpass(strict) - cudf tz replace + strftime now correct
+# test_convert_time_zone[cudf], test_convert_time_zone_series[cudf], test_convert_time_zone_from_none[cudf]: xpass(strict) - cudf tz convert + strftime now correct
+# test_contains_case_sensitive[cudf], test_contains_series_case_sensitive[cudf], test_contains_literal[cudf], test_contains_series_literal[cudf]:
+#   cudf's default ``str`` dtype mirrors pandas 3 ``str`` (fills nulls with False on ``str.contains``); narwhals
+#   expects ``None`` because it treats cudf strings as nullable. Compatibility mismatch, not a cudf bug.
 TESTS_THAT_NEED_NARWHALS_FIX_FOR_CUDF=" \
 test_to_numpy[cudf] or \
 test_fill_null_strategies_with_limit_as_none[cudf] or \
@@ -69,7 +75,14 @@ test_first_last_expr_over_order_by[cudf] or \
 test_series_rfloordiv_by_zero[cudf-0] or \
 test_expr_rfloordiv_by_zero[cudf-0] or \
 test_cast_datetime_tz_aware[cudf] or \
-test_nested_structures \
+test_nested_structures or \
+(test_group_by_depth_1_agg_bool_ops and cudf) or \
+(test_replace_time_zone and cudf) or \
+(test_convert_time_zone and cudf) or \
+(test_contains_case_sensitive and cudf) or \
+(test_contains_series_case_sensitive and cudf) or \
+(test_contains_literal and cudf) or \
+(test_contains_series_literal and cudf) \
 "
 
 rapids-logger "Run narwhals tests for cuDF"
@@ -174,6 +187,8 @@ test_pandas_object_series \
 # test_check_row_order_nested_only[pandas]: Nested dtype row ordering check not raising NotImplementedError as expected
 # test_cast_string: String casting issues with cudf.pandas
 # test_contains_case_insensitive[pandas], test_contains_series_case_insensitive[pandas]: String contains with case_insensitive returns False instead of None for nulls
+# test_fill_null_pandas_downcast: asserts dtype is 'object' after fill_null on [True, None]; cudf.pandas reports 'bool'
+#   because cudf represents nullable bool natively (not as object) — fundamental design difference, not fixable in cudf
 TESTS_THAT_NEED_NARWHALS_FIX_FOR_CUDF_PANDAS=" \
 test_dtypes or \
 test_explode_multiple_cols or \
@@ -193,7 +208,8 @@ test_sqrt_dtype_pandas or \
 test_check_row_order_nested_only[pandas] or \
 test_cast_string or \
 (test_contains_case_insensitive and pandas) or \
-(test_contains_series_case_insensitive and pandas) \
+(test_contains_series_case_insensitive and pandas) or \
+test_fill_null_pandas_downcast \
 "
 
 PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 \

ci/test_python_cudf.sh

@@ -31,13 +31,6 @@ version_gte() {
   [ "$2" = "$(echo -e "$2\n$1" | sort -V | head -n1)" ]
 }
 
-# If the RAPIDS_PY_VERSION is 3.13, set CUDF_TEST_COPY_ON_WRITE to '1' to enable copy-on-write tests.
-if version_gte "${RAPIDS_PY_VERSION}" "3.13"; then
-  echo "Running tests with CUDF_TEST_COPY_ON_WRITE enabled"
-  export CUDF_TEST_COPY_ON_WRITE='1'
-else
-  echo "Running tests with CUDF_TEST_COPY_ON_WRITE disabled"
-fi
 rapids-logger "pytest cudf"
 timeout 40m ./ci/run_cudf_pytests.sh \
   --junitxml="${RAPIDS_TESTS_DIR}/junit-cudf.xml" \

conda/environments/all_cuda-129_arch-aarch64.yaml

@@ -65,7 +65,7 @@ dependencies:
 - nvtx>=0.2.1
 - openpyxl
 - packaging
-- pandas>=2.0,<2.4.0
+- pandas>=3.0.0,<3.1.0
 - pandoc
 - polars>=1.35,<1.39
 - pre-commit

conda/environments/all_cuda-129_arch-x86_64.yaml

@@ -65,7 +65,7 @@ dependencies:
 - nvtx>=0.2.1
 - openpyxl
 - packaging
-- pandas>=2.0,<2.4.0
+- pandas>=3.0.0,<3.1.0
 - pandoc
 - polars>=1.35,<1.39
 - pre-commit

conda/environments/all_cuda-132_arch-aarch64.yaml

@@ -65,7 +65,7 @@ dependencies:
 - nvtx>=0.2.1
 - openpyxl
 - packaging
-- pandas>=2.0,<2.4.0
+- pandas>=3.0.0,<3.1.0
 - pandoc
 - polars>=1.35,<1.39
 - pre-commit

conda/environments/all_cuda-132_arch-x86_64.yaml

@@ -65,7 +65,7 @@ dependencies:
 - nvtx>=0.2.1
 - openpyxl
 - packaging
-- pandas>=2.0,<2.4.0
+- pandas>=3.0.0,<3.1.0
 - pandoc
 - polars>=1.35,<1.39
 - pre-commit

conda/recipes/cudf/recipe.yaml

@@ -88,7 +88,7 @@ requirements:
     - cuda-version =${{ cuda_version }}
   run:
     - python
-    - pandas >=2.0,<2.4.0
+    - pandas >=3.0.0,<3.1.0
     - cupy >=13.6.0,!=14.0.0
     - numba-cuda >=0.22.2,<0.29.0
     - numba >=0.60.0,<0.65.0

cpp/src/io/statistics/statistics_type_identification.cuh

@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2021-2024, NVIDIA CORPORATION.
+ * SPDX-FileCopyrightText: Copyright (c) 2021-2026, NVIDIA CORPORATION.
  * SPDX-License-Identifier: Apache-2.0
  */
 
@@ -37,9 +37,11 @@ struct conversion_map;
 template <is_int96_timestamp INT96>
 struct conversion_map<io_file_format::ORC, INT96> {
   using types = std::tuple<std::pair<cudf::timestamp_s, cudf::timestamp_ns>,
+                           std::pair<cudf::timestamp_ms, cudf::timestamp_ns>,
                            std::pair<cudf::timestamp_us, cudf::timestamp_ns>,
                            std::pair<cudf::timestamp_ns, cudf::timestamp_ns>,
                            std::pair<cudf::duration_s, cudf::duration_ns>,
+                           std::pair<cudf::duration_ms, cudf::duration_ns>,
                            std::pair<cudf::duration_us, cudf::duration_ns>,
                            std::pair<cudf::duration_ns, cudf::duration_ns>>;
 };

cpp/tests/rolling/rolling_test.cpp

@@ -811,7 +811,7 @@ TYPED_TEST_SUITE(RollingTest, cudf::test::FixedWidthTypesWithoutFixedPoint);
 // simple example from Pandas docs
 TYPED_TEST(RollingTest, SimpleStatic)
 {
-  // https://pandas.pydata.org/pandas-docs/version/2.3.3/reference/api/pandas.DataFrame.rolling.html
+  // https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.rolling.html
   auto const col_data              = cudf::test::make_type_param_vector<TypeParam>({0, 1, 2, 0, 4});
   const std::vector<bool> col_mask = {1, 1, 1, 0, 1};
 
@@ -958,7 +958,7 @@ TYPED_TEST(RollingTest, NegativeWindowSizes)
 // simple example from Pandas docs:
 TYPED_TEST(RollingTest, SimpleDynamic)
 {
-  // https://pandas.pydata.org/pandas-docs/version/2.3.3/reference/api/pandas.DataFrame.rolling.html
+  // https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.rolling.html
   auto const col_data              = cudf::test::make_type_param_vector<TypeParam>({0, 1, 2, 0, 4});
   const std::vector<bool> col_mask = {1, 1, 1, 0, 1};
 

dependencies.yaml

@@ -783,7 +783,7 @@ dependencies:
       - output_types: [conda, requirements, pyproject]
         packages:
           - fsspec>=0.6.0
-          - &pandas pandas>=2.0,<2.4.0
+          - &pandas pandas>=3.0.0,<3.1.0
   run_pylibcudf:
     common:
       - output_types: [conda, requirements, pyproject]
@@ -901,11 +901,11 @@ dependencies:
           - matrix: {dependencies: "oldest"}
             packages:
               - numba==0.60.0
-              - pandas==2.0.*
+              - pandas==3.0.*
               - numba-cuda==0.22.2,<0.29.0
           - matrix: {dependencies: "latest"}
             packages:
-              - pandas==2.3.3
+              - pandas==3.0.3
           - matrix:
             packages:
       - output_types: conda

docs/cudf/source/conf.py

@@ -327,7 +327,7 @@ def clean_all_xml_files(path):
     "numpy": ("https://numpy.org/doc/stable", None),
     # Temporarily disable nitpick warnings for pandas: https://github.com/pandas-dev/pandas/issues/64584
     # "pandas": (
-    #     "https://pandas.pydata.org/pandas-docs/version/2.3.3/",
+    #     "https://pandas.pydata.org/pandas-docs/stable/",
     #     None,
     # ),
     "polars": ("https://docs.pola.rs/api/python/stable/", None),

docs/cudf/source/cudf_pandas/faq.md

@@ -195,7 +195,7 @@ This means that automatic conversion between GPU and CPU types and automatic fal
 There are a few known limitations that you should be aware of:
 
 - Because fallback involves copying data from GPU to CPU and back,
-  [value mutability](https://pandas.pydata.org/pandas-docs/version/2.3.3/getting_started/overview.html#mutability-and-copying-of-data)
+  [value mutability](https://pandas.pydata.org/pandas-docs/stable/getting_started/overview.html#mutability-and-copying-of-data)
   of Pandas objects is not always guaranteed. You should follow the
   pandas recommendation to favor immutable operations.
 - For performance reasons, joins and join-based operations are not

docs/cudf/source/user_guide/api_docs/dataframe.rst

@@ -107,7 +107,6 @@ Function application, GroupBy & window
 
    DataFrame.agg
    DataFrame.apply
-   DataFrame.applymap
    DataFrame.groupby
    DataFrame.map
    DataFrame.pipe
@@ -167,9 +166,7 @@ Reindexing / selection / label manipulation
    DataFrame.drop_duplicates
    DataFrame.duplicated
    DataFrame.equals
-   DataFrame.first
    DataFrame.head
-   DataFrame.last
    DataFrame.reindex
    DataFrame.rename
    DataFrame.reset_index
@@ -189,7 +186,6 @@ Missing data handling
 .. autosummary::
    :toctree: api/
 
-   DataFrame.backfill
    DataFrame.bfill
    DataFrame.dropna
    DataFrame.ffill
@@ -200,7 +196,6 @@ Missing data handling
    DataFrame.nans_to_nulls
    DataFrame.notna
    DataFrame.notnull
-   DataFrame.pad
    DataFrame.replace
 
 Reshaping, sorting, transposing
@@ -252,6 +247,8 @@ Metadata
    :toctree: api/
 
    DataFrame.attrs
+   DataFrame.flags
+   DataFrame.set_flags
 
 Serialization / IO / conversion
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

docs/cudf/source/user_guide/api_docs/index_objects.rst

@@ -51,13 +51,6 @@ Modifying and computations
    Index.drop_duplicates
    Index.equals
    Index.factorize
-   Index.is_boolean
-   Index.is_categorical
-   Index.is_floating
-   Index.is_integer
-   Index.is_interval
-   Index.is_numeric
-   Index.is_object
    Index.min
    Index.max
    Index.rename

docs/cudf/source/user_guide/api_docs/series.rst

@@ -175,10 +175,8 @@ Reindexing / selection / label manipulation
    Series.drop_duplicates
    Series.duplicated
    Series.equals
-   Series.first
    Series.head
    Series.isin
-   Series.last
    Series.reindex
    Series.rename
    Series.reset_index
@@ -195,7 +193,6 @@ Missing data handling
 .. autosummary::
    :toctree: api/
 
-   Series.backfill
    Series.bfill
    Series.dropna
    Series.ffill
@@ -206,7 +203,6 @@ Missing data handling
    Series.nans_to_nulls
    Series.notna
    Series.notnull
-   Series.pad
    Series.replace
 
 Reshaping, sorting
@@ -243,6 +239,8 @@ Metadata
    :toctree: api/
 
    Series.attrs
+   Series.flags
+   Series.set_flags
 
 Accessors
 ---------

docs/cudf/source/user_guide/groupby.md

@@ -9,7 +9,7 @@ myst:
 # GroupBy
 
 cuDF supports a small (but important) subset of Pandas' [groupby
-API](https://pandas.pydata.org/pandas-docs/version/2.3.3/user_guide/groupby.html).
+API](https://pandas.pydata.org/pandas-docs/stable/user_guide/groupby.html).
 
 ## Summary of supported operations
 
@@ -28,7 +28,7 @@ API](https://pandas.pydata.org/pandas-docs/version/2.3.3/user_guide/groupby.html
    equivalent Pandas function. See the section on
    [apply](#groupby-apply) for more details.
 9. `GroupBy.pipe` similar to
-   [Pandas](https://pandas.pydata.org/pandas-docs/version/2.3.3/user_guide/groupby.html#piping-function-calls).
+   [Pandas](https://pandas.pydata.org/pandas-docs/stable/user_guide/groupby.html#piping-function-calls).
 
 ## Grouping
 
@@ -271,4 +271,4 @@ a
   4     4     5     9
 ```
 
-[describe]: https://pandas.pydata.org/pandas-docs/version/2.3.3/user_guide/groupby.html#flexible-apply
+[describe]: https://pandas.pydata.org/pandas-docs/stable/user_guide/groupby.html#flexible-apply

python/cudf/cudf/_typing.py

@@ -6,7 +6,7 @@
 
 import numpy as np
 import pandas as pd
-from packaging.version import parse
+from pandas.api.typing import NoDefault  # noqa: F401
 
 import pylibcudf as plc
 
@@ -53,8 +53,3 @@ def __call__(
 ]
 
 Axis = Literal[0, 1, "index", "columns"]
-
-if parse(pd.__version__) >= parse("3.0.0"):
-    from pandas.api.typing import NoDefault
-else:
-    NoDefault = Any

python/cudf/cudf/api/extensions/accessor.py

@@ -1,10 +1,8 @@
-# SPDX-FileCopyrightText: Copyright (c) 2020-2025, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2020-2026, NVIDIA CORPORATION.
 # SPDX-License-Identifier: Apache-2.0
 
 import warnings
 
-from pandas.core.accessor import CachedAccessor
-
 from cudf.core.dataframe import DataFrame
 from cudf.core.index import Index
 from cudf.core.series import Series
@@ -25,8 +23,8 @@
 
     Notes
     -----
-    The `{klass}` object will be passed to your custom accessor upon first
-    invocation. And will be cached for future calls.
+    The `{klass}` object will be passed to your custom accessor upon each
+    invocation.
 
     If the data passed to your accessor is of wrong datatype, you should
     raise an `AttributeError` in consistent with other cudf methods.
@@ -132,12 +130,25 @@
 )
 
 
+class _Accessor:
+    """Custom property-like object. A descriptor for accessors."""
+
+    def __init__(self, name, accessor):
+        self._name = name
+        self._accessor = accessor
+
+    def __get__(self, obj, cls):
+        if obj is None:
+            return self._accessor
+        return self._accessor(obj)
+
+
 def _register_accessor(name, cls):
     def decorator(accessor):
         if hasattr(cls, name):
             msg = f"Attribute {name} will be overridden in {cls.__name__}"
             warnings.warn(msg)
-        cached_accessor = CachedAccessor(name, accessor)
+        cached_accessor = _Accessor(name, accessor)
         cls._accessors.add(name)
         setattr(cls, name, cached_accessor)