Make pylibcudf.Table.columns() return a tuple

python/cudf/cudf/core/_internals/copying.py

@@ -19,7 +19,7 @@ def gather(
     columns: Sequence[ColumnBase],
     gather_map: NumericalColumn,
     nullify: bool = False,
-) -> list[plc.Column]:
+) -> tuple[plc.Column, ...]:
     with access_columns(
         *columns, gather_map, mode="read", scope="internal"
     ) as (*columns, gather_map):
@@ -38,7 +38,7 @@ def scatter(
     scatter_map: NumericalColumn,
     target_columns: list[ColumnBase],
     bounds_check: bool = True,
-):
+) -> tuple[plc.Column, ...]:
     """
     Scattering source into target as per the scatter map.
     `source` can be a list of scalars, or a list of columns. The number of
@@ -51,7 +51,7 @@ def scatter(
         raise ValueError("Mismatched number of source and target columns.")
 
     if len(sources) == 0:
-        return []
+        return ()
 
     if bounds_check:
         n_rows = len(target_columns[0])
@@ -80,7 +80,7 @@ def scatter(
 
 def columns_split(
     input_columns: Sequence[ColumnBase], splits: list[int]
-) -> list[list[plc.Column]]:
+) -> list[tuple[plc.Column, ...]]:
     with access_columns(
         *input_columns, mode="read", scope="internal"
     ) as input_columns:

python/cudf/cudf/core/_internals/sorting.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2020-2026, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2020-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 from __future__ import annotations
 
@@ -143,7 +143,7 @@ def sort_by_key(
     na_position: Iterable[Literal["first", "last"]],
     *,
     stable: bool,
-) -> list[plc.Column]:
+) -> tuple[plc.Column, ...]:
     """
     Sort a table by given keys
 
@@ -165,8 +165,8 @@ def sort_by_key(
 
     Returns
     -------
-    list[Column]
-        list of value columns sorted by keys
+    tuple[Column, ...]
+        value columns sorted by keys
     """
     column_order, null_precedence = ordering(ascending, na_position)
     func = (

python/cudf/cudf/core/frame.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2020-2026, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2020-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
 from __future__ import annotations
@@ -249,7 +249,7 @@ def _drop_duplicates_columns(
         keys: list[int],
         keep: Literal["first", "last", False],
         nulls_are_equal: bool,
-    ) -> list[plc.Column]:
+    ) -> tuple[plc.Column, ...]:
         """Core stable_distinct implementation shared by Index and IndexedFrame."""
         _keep_options = {
             "first": plc.stream_compaction.DuplicateKeepOption.KEEP_FIRST,
@@ -277,7 +277,7 @@ def _drop_nulls_columns(
         keys: list[int],
         how: Literal["any", "all"],
         thresh: int | None = None,
-    ) -> list[plc.Column]:
+    ) -> tuple[plc.Column, ...]:
         """Core drop_nulls implementation shared by Index and IndexedFrame."""
         if how not in {"any", "all"}:
             raise ValueError("how must be 'any' or 'all'")

python/cudf/cudf/core/indexed_frame.py

@@ -3455,7 +3455,7 @@ def _split(self, splits: list[int], keep_index: bool = True) -> list[Self]:
             )
 
             def split_with_dtypes(
-                split: list[plc.Column],
+                split: tuple[plc.Column, ...],
             ) -> list[ColumnBase]:
                 return [
                     ColumnBase.create(col, dtype)

python/cudf/cudf/io/parquet.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2019-2026, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2019-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 from __future__ import annotations
 
@@ -1419,13 +1419,13 @@ def _read_parquet(
             column_names = tbl_w_meta.column_names(include_children=False)
             child_names = tbl_w_meta.child_names
             per_file_user_data = tbl_w_meta.per_file_user_data
-            concatenated_columns = tbl_w_meta.tbl.columns()
+            concatenated_columns = tbl_w_meta.tbl.release()
 
             # save memory
             del tbl_w_meta
 
             while reader.has_next():
-                columns = reader.read_chunk().tbl.columns()
+                columns = reader.read_chunk().tbl.release()
                 # Iterate in reverse to avoid O(n²) cost from popping
                 for i in range(len(concatenated_columns) - 1, -1, -1):
                     concatenated_columns[i] = plc.concatenate.concatenate(

python/cudf_polars/cudf_polars/dsl/expressions/rolling.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION & AFFILIATES.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 # TODO: remove need for this
 # ruff: noqa: D101
@@ -422,6 +422,7 @@ def _(  # type: ignore[no-untyped-def]
 
         # Instead of calling self._gather_columns, let's call plc.copying.gather directly
         # since we need plc.Column objects, not cudf_polars Column objects
+        val_cols: Sequence[plc.Column]
         if order_index is not None:
             plc_cols = [
                 ne.value.children[0].evaluate(df, context=ExecutionContext.FRAME).obj
@@ -694,15 +695,16 @@ def _build_groupby_requests(
                 eval_cols.append(child.evaluate(df, context=ExecutionContext.FRAME).obj)
                 val_nodes.append((ne, val))
 
+        gathered_cols: Sequence[plc.Column] = eval_cols
         if order_index is not None and eval_cols:
-            eval_cols = plc.copying.gather(
+            gathered_cols = plc.copying.gather(
                 plc.Table(eval_cols),
                 order_index,
                 plc.copying.OutOfBoundsPolicy.NULLIFY,
                 stream=df.stream,
             ).columns()
 
-        gathered_iter = iter(eval_cols)
+        gathered_iter = iter(gathered_cols)
         for ne in named_exprs:
             val = ne.value
             if isinstance(val, expr.Len):

python/cudf_polars/cudf_polars/dsl/expressions/string.py

@@ -671,7 +671,7 @@ def do_evaluate(
                     max_splits - 1,
                     stream=df.stream,
                 )
-                children = plc_table.columns()
+                children = plc_table.release()
                 ref_column = children[0]
                 if (remainder := n + int(not is_split_n) - len(children)) > 0:
                     # Reach expected number of splits by padding with nulls

python/cudf_polars/cudf_polars/dsl/ir.py

@@ -822,9 +822,9 @@ def read_csv_header(
                 chunk = reader.read_chunk()
                 # TODO: Nested column names
                 names = chunk.column_names(include_children=False)
-                concatenated_columns = chunk.tbl.columns()
+                concatenated_columns = chunk.tbl.release()
                 while reader.has_next():
-                    columns = reader.read_chunk().tbl.columns()
+                    columns = reader.read_chunk().tbl.release()
                     # Discard columns while concatenating to reduce memory footprint.
                     # Reverse order to avoid O(n^2) list popping cost.
                     for i in reversed(range(len(concatenated_columns))):
@@ -2328,7 +2328,7 @@ def _reorder_maps(
         *,
         left_primary: bool = True,
         stream: Stream,
-    ) -> list[plc.Column]:
+    ) -> tuple[plc.Column, ...]:
         """
         Reorder gather maps to satisfy polars join order restrictions.
 
@@ -2355,7 +2355,7 @@ def _reorder_maps(
 
         Returns
         -------
-        list[plc.Column]
+        tuple[plc.Column, ...]
             Reordered left and right gather maps.
 
         Notes

python/pylibcudf/pylibcudf/column.pyi

@@ -50,7 +50,7 @@ class Column:
         mask: Span | None,
         null_count: int,
         offset: int,
-        children: list[Column],
+        children: Iterable[Column],
         validate: bool = True,
     ) -> None: ...
     def type(self) -> DataType: ...
@@ -94,7 +94,10 @@ class Column:
     ) -> Column: ...
     @staticmethod
     def from_rmm_buffer(
-        buff: DeviceBuffer, dtype: DataType, size: int, children: list[Column]
+        buff: DeviceBuffer,
+        dtype: DataType,
+        size: int,
+        children: Iterable[Column],
     ) -> Column: ...
     def to_arrow(
         self,

python/pylibcudf/pylibcudf/column.pyx

@@ -333,7 +333,7 @@ cdef class Column:
         The number of null rows in the column.
     offset : int
         The offset into the data buffer where the column's data begins.
-    children : list
+    children : Iterable[Column]
         The children of this column if it is a compound column type.
     validate : bool, default True
         Whether to validate that data and mask satisfy Span protocol.
@@ -343,8 +343,9 @@ cdef class Column:
     def __init__(
         self, DataType data_type not None, size_type size, object data,
         object mask, size_type null_count, size_type offset,
-        list children, bint validate=True
+        children, bint validate=True
     ):
+        children = list(children)
         if not all(isinstance(c, Column) for c in children):
             raise ValueError("All children must be pylibcudf Column objects")
 
@@ -600,7 +601,7 @@ cdef class Column:
         DeviceBuffer buff,
         DataType dtype,
         size_type size,
-        list children,
+        children,
     ):
         """
         Create a Column from an RMM DeviceBuffer.
@@ -613,14 +614,15 @@ cdef class Column:
             The number of rows in the column.
         dtype : DataType
             The type of the data in the buffer.
-        children : list
-            List of child columns.
+        children : Iterable[Column]
+            The child columns.
 
         Notes
         -----
         To provide a mask and null count, use `Column.with_mask` after
         this method.
         """
+        children = list(children)
         if plc_is_fixed_width(dtype) and len(children) != 0:
             raise ValueError("Fixed-width types must have zero children.")
         elif dtype.id() == type_id.STRING and len(children) != 1:

python/pylibcudf/pylibcudf/concatenate.pxd

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
 from .table cimport Table
@@ -10,4 +10,4 @@ from rmm.pylibrmm.memory_resource cimport DeviceMemoryResource
 # unify the column and table paths without using runtime dispatch instead. In this case
 # we choose to prioritize API consistency over performance, so we use the same function
 # with a bit of runtime dispatch overhead.
-cpdef concatenate(list objects, object stream = *, DeviceMemoryResource mr=*)
+cpdef concatenate(objects, object stream = *, DeviceMemoryResource mr=*)

python/pylibcudf/pylibcudf/concatenate.pyi

@@ -1,14 +1,16 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
+from collections.abc import Sequence
+
 from rmm.pylibrmm.memory_resource import DeviceMemoryResource
 
 from pylibcudf.column import Column
 from pylibcudf.table import Table
 from pylibcudf.utils import CudaStreamLike
 
 def concatenate[ColumnOrTable: (Column, Table)](
-    objects: list[ColumnOrTable],
+    objects: Sequence[ColumnOrTable],
     stream: CudaStreamLike | None = None,
     mr: DeviceMemoryResource | None = None,
 ) -> ColumnOrTable: ...

python/pylibcudf/pylibcudf/concatenate.pyx

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
 from libcpp.memory cimport unique_ptr
@@ -20,13 +20,13 @@ from cuda.bindings.cyruntime cimport cudaStream_t
 
 __all__ = ["concatenate"]
 
-cpdef concatenate(list objects, object stream=None, DeviceMemoryResource mr=None):
+cpdef concatenate(objects, object stream=None, DeviceMemoryResource mr=None):
     """Concatenate columns or tables.
 
     Parameters
     ----------
-    objects : Union[List[Column], List[Table]]
-        The list of Columns or Tables to concatenate.
+    objects : Sequence[Column] | Sequence[Table]
+        The Columns or Tables to concatenate.
     stream : Stream | None
         CUDA stream on which to perform the operation.
     mr : DeviceMemoryResource | None

python/pylibcudf/pylibcudf/io/types.pyi

@@ -87,7 +87,7 @@ class TableWithMetadata:
         self, tbl: Table, column_names: list[ColumnNameSpec]
     ) -> None: ...
     @property
-    def columns(self) -> list[Column]: ...
+    def columns(self) -> tuple[Column, ...]: ...
     @overload
     def column_names(self, include_children: Literal[False]) -> list[str]: ...
     @overload

python/pylibcudf/pylibcudf/io/types.pyx

@@ -349,7 +349,7 @@ cdef class TableWithMetadata:
     @property
     def columns(self):
         """
-        Return a list containing the columns of the table
+        Return a tuple containing the columns of the table
         """
         return self.tbl.columns()
 

python/pylibcudf/pylibcudf/null_mask.pxd

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
 from pylibcudf.libcudf.types cimport mask_state, size_type
@@ -30,9 +30,9 @@ cpdef DeviceBuffer create_null_mask(
     DeviceMemoryResource mr=*
 )
 
-cpdef tuple bitmask_and(list columns, object stream = *, DeviceMemoryResource mr=*)
+cpdef tuple bitmask_and(columns, object stream = *, DeviceMemoryResource mr=*)
 
-cpdef tuple bitmask_or(list columns, object stream = *, DeviceMemoryResource mr=*)
+cpdef tuple bitmask_or(columns, object stream = *, DeviceMemoryResource mr=*)
 
 cpdef size_type null_count(
     object bitmask,

python/pylibcudf/pylibcudf/null_mask.pyi

@@ -1,6 +1,8 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
+from collections.abc import Sequence
+
 from rmm.pylibrmm.device_buffer import DeviceBuffer
 from rmm.pylibrmm.memory_resource import DeviceMemoryResource
 
@@ -29,12 +31,12 @@ def create_null_mask(
     mr: DeviceMemoryResource | None = None,
 ) -> DeviceBuffer: ...
 def bitmask_and(
-    columns: list[Column],
+    columns: Sequence[Column],
     stream: CudaStreamLike | None = None,
     mr: DeviceMemoryResource | None = None,
 ) -> tuple[DeviceBuffer, int]: ...
 def bitmask_or(
-    columns: list[Column],
+    columns: Sequence[Column],
     stream: CudaStreamLike | None = None,
     mr: DeviceMemoryResource | None = None,
 ) -> tuple[DeviceBuffer, int]: ...