Fix ORC reader for empty DataFrame/Table

cpp/src/io/orc/orc.cpp

@@ -550,14 +550,14 @@ std::vector<int> metadata::select_columns(std::vector<std::string> use_names,
     }
   } else {
     // For now, only select all leaf nodes
-    for (int i = 0; i < get_num_columns(); ++i) {
+    for (int i = 1; i < get_num_columns(); ++i) {
       if (ff.types[i].subtypes.empty()) {
         selection.emplace_back(i);
         if (ff.types[i].kind == orc::TIMESTAMP) { has_timestamp_column = true; }
       }
     }
   }
-  CUDF_EXPECTS(selection.size() > 0, "Filtered out all columns");
+  CUDF_EXPECTS(use_names.empty() or selection.size() > 0, "Filtered out all columns");
 
   return selection;
 }

cpp/src/io/orc/reader_impl.cu

@@ -425,6 +425,9 @@ table_with_metadata reader::impl::read(size_type skip_rows,
   std::vector<std::unique_ptr<column>> out_columns;
   table_metadata out_metadata;
 
+  // There are no columns in table
+  if (_selected_columns.size() == 0) return {std::make_unique<table>(), std::move(out_metadata)};
+
   // Select only stripes required (aka row groups)
   const auto selected_stripes = _metadata->select_stripes(stripes, skip_rows, num_rows);
 

python/cudf/cudf/tests/test_orc.py

@@ -738,3 +738,19 @@ def test_nanoseconds_overflow():
 
     pyarrow_got = pa.orc.ORCFile(buffer).read()
     assert_eq(expected.to_pandas(), pyarrow_got.to_pandas())
+
+
+def test_empty_dataframe():
+    buffer = BytesIO()
+    expected = cudf.DataFrame()
+    expected.to_orc(buffer)
+
+    # Raise error if column name is mentioned, but it doesn't exist.
+    with pytest.raises(RuntimeError):
+        cudf.read_orc(buffer, columns=["a"])
+
+    got_df = cudf.read_orc(buffer)
+    expected_pdf = pd.read_orc(buffer)
+
+    assert_eq(expected, got_df)
+    assert_eq(expected_pdf, got_df)