Add try/except for expected null-schema error in read_parquet

python/dask_cudf/dask_cudf/io/parquet.py

@@ -93,14 +93,36 @@ def _read_paths(
                 )
 
             # Use cudf to read in data
-            df = cudf.read_parquet(
-                paths_or_fobs,
-                engine="cudf",
-                columns=columns,
-                row_groups=row_groups if row_groups else None,
-                strings_to_categorical=strings_to_categorical,
-                **kwargs,
-            )
+            try:
+                df = cudf.read_parquet(
+                    paths_or_fobs,
+                    engine="cudf",
+                    columns=columns,
+                    row_groups=row_groups if row_groups else None,
+                    strings_to_categorical=strings_to_categorical,
+                    **kwargs,
+                )
+            except RuntimeError as err:
+                # TODO: Remove try/except after null-schema issue is resolved
+                # (See: https://github.com/rapidsai/cudf/issues/12702)
+                if len(paths_or_fobs) > 1:
+                    df = cudf.concat(
+                        [
+                            cudf.read_parquet(
+                                pof,
+                                engine="cudf",
+                                columns=columns,
+                                row_groups=row_groups[i]
+                                if row_groups
+                                else None,
+                                strings_to_categorical=strings_to_categorical,
+                                **kwargs,
+                            )
+                            for i, pof in enumerate(paths_or_fobs)
+                        ]
+                    )
+                else:
+                    raise err
 
         if partitions and partition_keys is None:
 

python/dask_cudf/dask_cudf/io/tests/test_parquet.py

@@ -502,3 +502,17 @@ def test_check_file_size(tmpdir):
     cudf.DataFrame({"a": np.arange(1000)}).to_parquet(fn)
     with pytest.warns(match="large parquet file"):
         dask_cudf.read_parquet(fn, check_file_size=1).compute()
+
+
+def test_nullable_schema_mismatch(tmpdir):
+    # See: https://github.com/rapidsai/cudf/issues/12702
+    path0 = str(tmpdir.join("test.0.parquet"))
+    path1 = str(tmpdir.join("test.1.parquet"))
+    cudf.DataFrame.from_dict({"a": [1, 2, 3]}).to_parquet(path0)
+    cudf.DataFrame.from_dict({"a": [4, 5, None]}).to_parquet(path1)
+    with dask.config.set({"dataframe.backend": "cudf"}):
+        ddf = dd.read_parquet(
+            [path0, path1], split_row_groups=2, aggregate_files=True
+        )
+        expect = pd.read_parquet([path0, path1])
+    dd.assert_eq(ddf, expect, check_index=False)