Match PyArrow's row count for read_parquet with columns=[]

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

@@ -80,7 +80,7 @@ dependencies:
 - polars>=1.35,<1.45
 - pre-commit
 - psutil
-- pyarrow>=19.0.0,<26.0.0
+- pyarrow>=19.0.0
 - pytables
 - pytest-benchmark
 - pytest-cases>=3.8.2

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

@@ -80,7 +80,7 @@ dependencies:
 - polars>=1.35,<1.45
 - pre-commit
 - psutil
-- pyarrow>=19.0.0,<26.0.0
+- pyarrow>=19.0.0
 - pytables
 - pytest-benchmark
 - pytest-cases>=3.8.2

conda/environments/all_cuda-133_arch-aarch64.yaml

@@ -80,7 +80,7 @@ dependencies:
 - polars>=1.35,<1.45
 - pre-commit
 - psutil
-- pyarrow>=19.0.0,<26.0.0
+- pyarrow>=19.0.0
 - pytables
 - pytest-benchmark
 - pytest-cases>=3.8.2

conda/environments/all_cuda-133_arch-x86_64.yaml

@@ -80,7 +80,7 @@ dependencies:
 - polars>=1.35,<1.45
 - pre-commit
 - psutil
-- pyarrow>=19.0.0,<26.0.0
+- pyarrow>=19.0.0
 - pytables
 - pytest-benchmark
 - pytest-cases>=3.8.2

conda/recipes/cudf/recipe.yaml

@@ -106,7 +106,7 @@ requirements:
     - numba-cuda-mlir >=0.3.0
     - numba >=0.60.0,<0.65.0
     - numpy >=2.0,<3.0
-    - pyarrow>=19.0.0,<26.0.0
+    - pyarrow>=19.0.0
     - libcudf =${{ version }}
     - pylibcudf =${{ version }}
     - ${{ pin_compatible("rmm", upper_bound="x.x") }}

conda/recipes/pylibcudf/recipe.yaml

@@ -108,7 +108,7 @@ requirements:
     - nvtx >=0.2.1
   run_constraints:
     - numpy >=2.0,<3.0
-    - pyarrow>=19.0.0,<26.0.0
+    - pyarrow>=19.0.0
   ignore_run_exports:
     from_package:
       - cuda-cudart-dev

dependencies.yaml

@@ -828,10 +828,10 @@ dependencies:
     common:
       - output_types: [conda]
         packages:
-          - pyarrow>=19.0.0,<26.0.0
+          - pyarrow>=19.0.0
       - output_types: [requirements, constraints, pyproject]
         packages:
-          - pyarrow>=19.0.0,<26.0.0
+          - pyarrow>=19.0.0
   # 'cuda_version' intentionally does not contain fallback entries... we want
   # a loud error if an unsupported 'cuda' value is passed
   cuda_version:

python/cudf/cudf/core/_compat.py

@@ -1,8 +1,10 @@
-# 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
 
 import pandas as pd
+import pyarrow as pa
 from packaging import version
 
 PANDAS_CURRENT_SUPPORTED_VERSION = version.parse("3.0.2")
 PANDAS_VERSION = version.parse(pd.__version__)
+PYARROW_GE_26 = version.parse(pa.__version__) >= version.parse("26.0.0")

python/cudf/cudf/io/parquet.py

@@ -25,6 +25,7 @@
 from pylibcudf import expressions as plc_expr
 
 from cudf.api.types import is_list_like
+from cudf.core._compat import PYARROW_GE_26
 from cudf.core.column import (
     ColumnBase,
     access_columns,
@@ -1395,11 +1396,14 @@ def _read_parquet(
             nrows = -1
         if skip_rows is None:
             skip_rows = 0
+        empty_columns = columns is not None and len(columns) == 0
         if get_option("io.parquet.low_memory"):
             # If filters are used, we can't rely on a RangeIndex
             # (row count may be reduced)
             allow_range_index = (
-                filters is None and columns is not None and len(columns) != 0
+                filters is None
+                and columns is not None
+                and (len(columns) != 0 or PYARROW_GE_26)
             )
 
             options = (
@@ -1465,10 +1469,23 @@ def _read_parquet(
                 nrows=nrows,
                 skip_rows=skip_rows,
             )
+            if (
+                empty_columns
+                and PYARROW_GE_26
+                and filters is None
+                and len(df.index) == 0
+            ):
+                df.index = RangeIndex(
+                    _num_rows(
+                        filepaths_or_buffers, row_groups, skip_rows, nrows
+                    )
+                )
             return df
         else:
             allow_range_index = True
-            if columns is not None and len(columns) == 0 or filters:
+            if (
+                columns is not None and len(columns) == 0 and not PYARROW_GE_26
+            ) or filters:
                 allow_range_index = False
 
             options = (
@@ -1505,6 +1522,17 @@ def _read_parquet(
                 nrows=nrows,
                 skip_rows=skip_rows,
             )
+            if (
+                empty_columns
+                and PYARROW_GE_26
+                and filters is None
+                and len(df.index) == 0
+            ):
+                df.index = RangeIndex(
+                    _num_rows(
+                        filepaths_or_buffers, row_groups, skip_rows, nrows
+                    )
+                )
             return df
     else:
         if (
@@ -2479,6 +2507,26 @@ def _get_stat_freq(
     return result
 
 
+def _num_rows(
+    filepaths_or_buffers, row_groups, skip_rows: int, nrows: int
+) -> int:
+    metadata = plc.io.parquet_metadata.read_parquet_metadata(
+        plc.io.SourceInfo(filepaths_or_buffers)
+    )
+    if row_groups is not None:
+        row_group_rows = metadata.rowgroup_metadata()
+        file_offsets = itertools.accumulate(
+            metadata.num_rowgroups_per_file(), initial=0
+        )
+        return sum(
+            row_group_rows[offset + row_group]["num_rows"]
+            for offset, groups in zip(file_offsets, row_groups, strict=False)
+            for row_group in groups
+        )
+    num_rows = max(metadata.num_rows() - skip_rows, 0)
+    return num_rows if nrows < 0 else min(num_rows, nrows)
+
+
 def _process_metadata(
     df: DataFrame,
     names: Sequence[Hashable],

python/cudf/cudf/tests/input_output/test_parquet.py

@@ -22,6 +22,7 @@
 from pyarrow import parquet as pq
 
 import cudf
+from cudf.core._compat import PYARROW_GE_26
 from cudf.io.parquet import (
     ParquetDatasetWriter,
     ParquetWriter,
@@ -3427,6 +3428,37 @@ def test_parquet_columns_and_range_index(columns):
     assert_eq(expected, got, check_index_type=True)
 
 
+@pytest.mark.parametrize("low_memory", [True, False])
+@pytest.mark.parametrize("index", [True, False, None])
+@pytest.mark.parametrize(
+    "kwargs, num_rows",
+    [
+        ({}, 12),
+        ({"skip_rows": 3}, 9),
+        ({"nrows": 5}, 5),
+        ({"skip_rows": 3, "nrows": 5}, 5),
+        ({"row_groups": [[0, 2], [1]]}, 6),
+    ],
+)
+def test_parquet_empty_columns_num_rows(
+    tmp_path, low_memory, index, kwargs, num_rows
+):
+    paths = []
+    for i in range(2):
+        path = str(tmp_path / f"{i}.parquet")
+        pd.DataFrame({"a": range(6)}).to_parquet(
+            path, index=index, row_group_size=2
+        )
+        paths.append(path)
+
+    with cudf.option_context("io.parquet.low_memory", low_memory):
+        got = cudf.read_parquet(paths, columns=[], **kwargs)
+
+    expected_rows = num_rows if index or PYARROW_GE_26 else 0
+    assert got.shape == (expected_rows, 0)
+    assert len(got.index) == expected_rows
+
+
 def test_parquet_nested_struct_list():
     buffer = BytesIO()
     data = {

python/cudf/pyproject.toml

@@ -31,7 +31,7 @@ dependencies = [
     "nvtx>=0.2.1",
     "packaging",
     "pandas>=3.0.0,<3.1.0",
-    "pyarrow>=19.0.0,<26.0.0",
+    "pyarrow>=19.0.0",
     "pylibcudf==26.12.*,>=0.0.0a0",
     "rich",
     "rmm==26.12.*,>=0.0.0a0",

python/pylibcudf/pyproject.toml

@@ -43,14 +43,14 @@ test = [
     "numba-cuda>=0.22.2,<0.31.0",
     "numba>=0.60.0,<0.65.0",
     "pandas>=3.0.0,<3.1.0",
-    "pyarrow>=19.0.0,<26.0.0",
+    "pyarrow>=19.0.0",
     "pytest-cov",
     "pytest-xdist",
     "pytest<9.1.0",
     "xxhash",
 ] # This list was generated by `rapids-dependency-file-generator`. To make changes, edit ../../dependencies.yaml and run `rapids-dependency-file-generator`.
 pyarrow = [
-    "pyarrow>=19.0.0,<26.0.0",
+    "pyarrow>=19.0.0",
 ] # This list was generated by `rapids-dependency-file-generator`. To make changes, edit ../../dependencies.yaml and run `rapids-dependency-file-generator`.
 numpy = [
     "numpy>=2.0,<3.0",