Fixes Unsupported column type error due to empty list columns in Nest…

…ed JSON reader (#11897)

Fixes `Unsupported column type` error during cudf column creation in Nested JSON reader due to empty list column.

During json tree creation, Empty list column does not have `device_json_column` child because it does have any rows, or a type.
This PR fixes the issue by creating an empty column as element child column. The list column still retains the null, and empty list information.

Authors:
  - Karthikeyan (https://github.com/karthikeyann)

Approvers:
  - Mike Wilson (https://github.com/hyperbolic2346)
  - Vyas Ramasubramani (https://github.com/vyasr)

URL: #11897

cpp/src/io/json/json_column.cu

@@ -403,7 +403,7 @@ void make_device_json_column(device_span<SymbolT const> input,
     std::string name   = "";
     auto parent_col_id = column_parent_ids[this_col_id];
     if (parent_col_id == parent_node_sentinel || column_categories[parent_col_id] == NC_LIST) {
-      name = "element";
+      name = list_child_name;
     } else if (column_categories[parent_col_id] == NC_FN) {
       auto field_name_col_id = parent_col_id;
       parent_col_id          = column_parent_ids[parent_col_id];
@@ -689,19 +689,24 @@ std::pair<std::unique_ptr<column>, std::vector<column_name_info>> device_json_co
       size_type num_rows = json_col.child_offsets.size() - 1;
       std::vector<column_name_info> column_names{};
       column_names.emplace_back("offsets");
-      column_names.emplace_back(json_col.child_columns.begin()->first);
+      column_names.emplace_back(
+        json_col.child_columns.empty() ? list_child_name : json_col.child_columns.begin()->first);
 
       // Note: json_col modified here, reuse the memory
       auto offsets_column = std::make_unique<column>(
         data_type{type_id::INT32}, num_rows + 1, json_col.child_offsets.release());
       // Create children column
       auto [child_column, names] =
-        device_json_column_to_cudf_column(json_col.child_columns.begin()->second,
-                                          d_input,
-                                          options,
-                                          get_child_schema(json_col.child_columns.begin()->first),
-                                          stream,
-                                          mr);
+        json_col.child_columns.empty()
+          ? std::pair<std::unique_ptr<column>,
+                      std::vector<column_name_info>>{std::make_unique<column>(), {}}
+          : device_json_column_to_cudf_column(
+              json_col.child_columns.begin()->second,
+              d_input,
+              options,
+              get_child_schema(json_col.child_columns.begin()->first),
+              stream,
+              mr);
       column_names.back().children      = names;
       auto [result_bitmask, null_count] = make_validity(json_col);
       return {make_lists_column(num_rows,

cpp/src/io/json/nested_json.hpp

@@ -104,6 +104,9 @@ enum node_t : NodeT {
  */
 enum class json_col_t : char { ListColumn, StructColumn, StringColumn, Unknown };
 
+// Default name for a list's child column
+constexpr auto list_child_name{"element"};
+
 /**
  * @brief Intermediate representation of data from a nested JSON input
  */

cpp/src/io/json/nested_json_gpu.cu

@@ -1162,9 +1162,6 @@ void make_json_column(json_column& root_column,
   // Range of encapsulating function that parses to internal columnar data representation
   CUDF_FUNC_RANGE();
 
-  // Default name for a list's child column
-  std::string const list_child_name = "element";
-
   // Parse the JSON and get the token stream
   const auto [d_tokens_gpu, d_token_indices_gpu] = get_token_stream(d_input, options, stream, mr);
 
@@ -1286,7 +1283,7 @@ void make_json_column(json_column& root_column,
    * (b) a list, the selected child column corresponds to single child column of
    * the list column. In this case, the child column may not exist yet.
    */
-  auto get_selected_column = [&list_child_name](std::stack<tree_node>& current_data_path) {
+  auto get_selected_column = [](std::stack<tree_node>& current_data_path) {
     json_column* selected_col = current_data_path.top().current_selected_col;
 
     // If the node does not have a selected column yet
@@ -1680,20 +1677,24 @@ std::pair<std::unique_ptr<column>, std::vector<column_name_info>> json_column_to
       size_type num_rows = json_col.child_offsets.size();
       std::vector<column_name_info> column_names{};
       column_names.emplace_back("offsets");
-      column_names.emplace_back(json_col.child_columns.begin()->first);
+      column_names.emplace_back(
+        json_col.child_columns.empty() ? list_child_name : json_col.child_columns.begin()->first);
 
       rmm::device_uvector<json_column::row_offset_t> d_offsets =
         cudf::detail::make_device_uvector_async(json_col.child_offsets, stream, mr);
       auto offsets_column =
         std::make_unique<column>(data_type{type_id::INT32}, num_rows, d_offsets.release());
       // Create children column
       auto [child_column, names] =
-        json_column_to_cudf_column(json_col.child_columns.begin()->second,
-                                   d_input,
-                                   options,
-                                   get_child_schema(json_col.child_columns.begin()->first),
-                                   stream,
-                                   mr);
+        json_col.child_columns.empty()
+          ? std::pair<std::unique_ptr<column>,
+                      std::vector<column_name_info>>{std::make_unique<column>(), {}}
+          : json_column_to_cudf_column(json_col.child_columns.begin()->second,
+                                       d_input,
+                                       options,
+                                       get_child_schema(json_col.child_columns.begin()->first),
+                                       stream,
+                                       mr);
       column_names.back().children      = names;
       auto [result_bitmask, null_count] = make_validity(json_col);
       return {make_lists_column(num_rows - 1,

cpp/tests/io/json_test.cpp

@@ -813,7 +813,6 @@ TEST_P(JsonReaderDualTest, JsonLinesObjectsOutOfOrder)
                                  cudf::test::strings_column_wrapper({"aaa", "bbb"}));
 }
 
-/*
 // currently, the json reader is strict about having non-empty input.
 TEST_F(JsonReaderTest, EmptyFile)
 {
@@ -824,14 +823,17 @@ TEST_F(JsonReaderTest, EmptyFile)
   }
 
   cudf::io::json_reader_options in_options =
-    cudf::io::json_reader_options::builder(cudf::io::source_info{filepath}).lines(true);
+    cudf::io::json_reader_options::builder(cudf::io::source_info{filepath})
+      .lines(true)
+      .experimental(true);
   auto result = cudf::io::read_json(in_options);
 
   const auto view = result.tbl->view();
   EXPECT_EQ(0, view.num_columns());
 }
 
 // currently, the json reader is strict about having non-empty input.
+// experimental reader supports empty input
 TEST_F(JsonReaderTest, NoDataFile)
 {
   auto filepath = temp_env->get_temp_dir() + "NoDataFile.csv";
@@ -841,13 +843,14 @@ TEST_F(JsonReaderTest, NoDataFile)
   }
 
   cudf::io::json_reader_options in_options =
-    cudf::io::json_reader_options::builder(cudf::io::source_info{filepath}).lines(true);
+    cudf::io::json_reader_options::builder(cudf::io::source_info{filepath})
+      .lines(true)
+      .experimental(true);
   cudf::io::table_with_metadata result = cudf::io::read_json(in_options);
 
   const auto view = result.tbl->view();
   EXPECT_EQ(0, view.num_columns());
 }
-*/
 
 TEST_F(JsonReaderTest, ArrowFileSource)
 {

cpp/tests/io/json_tree.cpp

@@ -773,7 +773,11 @@ std::vector<std::string> json_lines_list = {
  { "a": { "y" : 6, "z": [] }}
  { "a": { "y" : 6, "z": [2, 3, 4, 5] }}
  { "a": { "z": [4], "y" : 6 }}
- { "a" : { "x" : 8, "y": 9 }, "b" : {"x": 10 , "z": 11 }} )"};
+ { "a" : { "x" : 8, "y": 9 }, "b" : {"x": 10 , "z": 11 }} )",
+  // empty list, row.
+  R"( {"a" : [], "b" : {}}
+ {"a" : []}
+ {"b" : {}})"};
 INSTANTIATE_TEST_SUITE_P(Mixed_And_Records,
                          JsonTreeTraversalTest,
                          ::testing::Combine(::testing::Values(false),

python/cudf/cudf/tests/test_json.py

@@ -649,6 +649,24 @@ def test_json_nested_data():
     assert df.to_arrow().equals(pa_table_pdf)
 
 
+def test_json_empty_types():
+    json_str = """ {}
+    {"a": [], "b": {}}
+    {"a": []}
+    {"b": {}}
+    {"c": {"d": []}}
+    {"e": [{}]}
+    """
+    df = cudf.read_json(
+        StringIO(json_str),
+        engine="cudf_experimental",
+        orient="records",
+        lines=True,
+    )
+    pdf = pd.read_json(StringIO(json_str), orient="records", lines=True)
+    assert_eq(df, pdf)
+
+
 def test_json_types_data():
     # 0:<0:string,1:float>
     # 1:list<int>