[REVIEW] Fix/read parquet for empty DataFrame

CHANGELOG.md

@@ -111,6 +111,7 @@
 - PR #6259 Fix compilation error with GCC 8
 - PR #6258 Pin libcudf conda recipe to boost 1.72.0
 - PR #6264 Remove include statement for missing rmm/mr/device/default_memory_resource.hpp file
+- PR #6294 Fix read parquet key error when reading empty pandas DataFrame with cudf
 - PR #6285 Removed unsafe `reinterpret_cast` and implicit pointer-to-bool casts
 - PR #6281 Fix unreachable code warning in datetime.cuh
 - PR #6286 Fix `read_csv` `int32` overflow 

python/cudf/cudf/_lib/parquet.pyx

@@ -231,9 +231,13 @@ cpdef read_parquet(filepaths_or_buffers, columns=None, row_groups=None,
                 column_names.remove(index_col)
 
         for col in column_names:
+            try:
+                data = cols_dtype_map[col]
+            except KeyError:
+                data = cols_dtype_map.get(col, None)
             df._data[col] = cudf.core.column.column_empty(
                 row_count=0,
-                dtype=np.dtype(cols_dtype_map[col])
+                dtype=np.dtype(data)
             )
 
     # Set the index column

python/cudf/cudf/tests/test_parquet.py

@@ -210,6 +210,35 @@ def test_parquet_reader_basic(parquet_file, columns, engine):
     assert_eq(expect, got, check_categorical=False)
 
 
+@pytest.mark.filterwarnings("ignore:Using CPU")
+@pytest.mark.parametrize("engine", ["cudf"])
+def test_parquet_reader_empty_pandas_dataframe(tmpdir, engine):
+    df = pd.DataFrame()
+
+    fname = tmpdir.join("test_pq_reader_empty_pandas_dataframe.parquet")
+
+    df.to_parquet(fname)
+    assert os.path.exists(fname)
+
+    expect = pd.read_parquet(fname)
+    got = cudf.read_parquet(fname, engine="cudf")
+
+    if len(expect) == 0:
+        expect = expect.reset_index(drop=True)
+        got = got.reset_index(drop=True)
+        if "col_category" in expect.columns:
+            expect["col_category"] = expect["col_category"].astype("category")
+
+    # PANDAS returns category objects whereas cuDF returns hashes
+    if engine == "cudf":
+        if "col_category" in expect.columns:
+            expect = expect.drop(columns=["col_category"])
+        if "col_category" in got.columns:
+            got = got.drop(columns=["col_category"])
+
+    assert_eq(expect, got, check_categorical=False)
+
+
 @pytest.mark.parametrize("has_null", [False, True])
 @pytest.mark.parametrize("strings_to_categorical", [False, True, None])
 def test_parquet_reader_strings(tmpdir, strings_to_categorical, has_null):