From c6c28a1221e18afd2a08550367ad79dfa92713f7 Mon Sep 17 00:00:00 2001
From: galipremsagar <sagarprem75@gmail.com>
Date: Wed, 18 Mar 2026 18:53:36 +0000
Subject: [PATCH] fix orc failures

---
 .../statistics_type_identification.cuh        |  4 +++-
 .../cudf/cudf/tests/input_output/test_orc.py  | 20 +++++++++++++++----
 2 files changed, 19 insertions(+), 5 deletions(-)

diff --git a/cpp/src/io/statistics/statistics_type_identification.cuh b/cpp/src/io/statistics/statistics_type_identification.cuh
index 825e50860438..ed7812419afe 100644
--- a/cpp/src/io/statistics/statistics_type_identification.cuh
+++ b/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>>;
 };
diff --git a/python/cudf/cudf/tests/input_output/test_orc.py b/python/cudf/cudf/tests/input_output/test_orc.py
index 0e6d224cb7d9..9a3789fa405c 100644
--- a/python/cudf/cudf/tests/input_output/test_orc.py
+++ b/python/cudf/cudf/tests/input_output/test_orc.py
@@ -607,7 +607,11 @@ def normalized_equals(value1, value2):
     if isinstance(value1, float) or isinstance(value2, float):
         return np.isclose(value1, value2)
 
-    return value1 == value2
+    try:
+        assert_eq(value1, value2)
+        return True
+    except AssertionError:
+        return False
 
 
 @pytest.mark.parametrize("stats_freq", ["STRIPE", "ROWGROUP"])
@@ -1473,8 +1477,10 @@ def test_orc_writer_lists_empty_rg():
     df = cudf.read_orc(buffer)
     assert_eq(df, cudf_in)
 
-    pdf_out = pd.read_orc(buffer)
-    assert_eq(pdf_in, pdf_out)
+    # Compare via pyarrow since pd.read_orc converts nullable integer
+    # lists to float arrays ([None] -> [nan]), losing the original types.
+    pa_out = orc.ORCFile(buffer).read()
+    assert pa_out.equals(cudf_in.to_arrow())
 
 
 def test_statistics_sum_overflow():
@@ -1647,7 +1653,13 @@ def run_orc_columns_and_index_param(index_obj, index, columns):
     expected = pd.read_orc(buffer, columns=columns)
     got = cudf.read_orc(buffer, columns=columns)
 
-    assert_eq(expected, got, check_index_type=True)
+    # When columns is an empty list, pandas uses dtype='object' for
+    # the empty column Index while cudf uses dtype='str'. Avoid
+    # checking the column index type in that case.
+    check_col_type = columns is None or len(columns) > 0
+    assert_eq(
+        expected, got, check_index_type=True, check_column_type=check_col_type
+    )
 
 
 @pytest.mark.parametrize("index_obj", [None, [10, 11, 12], ["x", "y", "z"]])