be/src/format/arrow/arrow_row_batch.cpp
Line | Count | Source |
1 | | // Licensed to the Apache Software Foundation (ASF) under one |
2 | | // or more contributor license agreements. See the NOTICE file |
3 | | // distributed with this work for additional information |
4 | | // regarding copyright ownership. The ASF licenses this file |
5 | | // to you under the Apache License, Version 2.0 (the |
6 | | // "License"); you may not use this file except in compliance |
7 | | // with the License. You may obtain a copy of the License at |
8 | | // |
9 | | // http://www.apache.org/licenses/LICENSE-2.0 |
10 | | // |
11 | | // Unless required by applicable law or agreed to in writing, |
12 | | // software distributed under the License is distributed on an |
13 | | // "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY |
14 | | // KIND, either express or implied. See the License for the |
15 | | // specific language governing permissions and limitations |
16 | | // under the License. |
17 | | |
18 | | #include "format/arrow/arrow_row_batch.h" |
19 | | |
20 | | #include <arrow/array/util.h> |
21 | | #include <arrow/buffer.h> |
22 | | #include <arrow/extension/parquet_variant.h> |
23 | | #include <arrow/extension/uuid.h> |
24 | | #include <arrow/io/memory.h> |
25 | | #include <arrow/ipc/writer.h> |
26 | | #include <arrow/record_batch.h> |
27 | | #include <arrow/result.h> |
28 | | #include <arrow/status.h> |
29 | | #include <arrow/type.h> |
30 | | #include <arrow/type_fwd.h> |
31 | | #include <arrow/util/key_value_metadata.h> |
32 | | #include <glog/logging.h> |
33 | | #include <stdint.h> |
34 | | |
35 | | #include <algorithm> |
36 | | #include <cstdlib> |
37 | | #include <memory> |
38 | | #include <utility> |
39 | | #include <vector> |
40 | | |
41 | | #include "core/block/block.h" |
42 | | #include "core/data_type/data_type_agg_state.h" |
43 | | #include "core/data_type/data_type_array.h" |
44 | | #include "core/data_type/data_type_map.h" |
45 | | #include "core/data_type/data_type_struct.h" |
46 | | #include "core/data_type/data_type_variant_v2.h" |
47 | | #include "core/data_type/define_primitive_type.h" |
48 | | #include "exprs/vexpr.h" |
49 | | #include "exprs/vexpr_context.h" |
50 | | #include "format/arrow/arrow_block_convertor.h" |
51 | | #include "format/arrow/arrow_utils.h" |
52 | | #include "runtime/descriptors.h" |
53 | | |
54 | | namespace doris { |
55 | | |
56 | 43 | Status register_arrow_variant_extension() { |
57 | | // Remote Flight readers must restore the extension before decoding the result schema. |
58 | 43 | static const auto status = [] { |
59 | 1 | if (arrow::GetExtensionType("arrow.parquet.variant") != nullptr) { |
60 | 0 | return arrow::Status::OK(); |
61 | 0 | } |
62 | 1 | return arrow::RegisterExtensionType( |
63 | 1 | std::static_pointer_cast<arrow::ExtensionType>(arrow::extension::variant( |
64 | 1 | arrow::struct_({arrow::field("metadata", arrow::binary(), false), |
65 | 1 | arrow::field("value", arrow::binary(), false)})))); |
66 | 1 | }(); |
67 | 43 | return status.ok() ? Status::OK() : Status::InternalError(status.ToString()); |
68 | 43 | } |
69 | | |
70 | | // Keep the exhaustive type mapping together so protocol-specific overrides remain visible. |
71 | | // NOLINTNEXTLINE(readability-function-size) |
72 | | Status DorisArrowSchemaConvertor::convert_to_arrow_type( |
73 | 1.66k | const DataTypePtr& origin_type, std::shared_ptr<arrow::DataType>* result) const { |
74 | 1.66k | auto type = get_serialized_type(origin_type); |
75 | 1.66k | switch (type->get_primitive_type()) { |
76 | 0 | case TYPE_NULL: |
77 | 0 | *result = arrow::null(); |
78 | 0 | break; |
79 | 24 | case TYPE_TINYINT: |
80 | 24 | *result = arrow::int8(); |
81 | 24 | break; |
82 | 24 | case TYPE_SMALLINT: |
83 | 24 | *result = arrow::int16(); |
84 | 24 | break; |
85 | 70 | case TYPE_INT: |
86 | 70 | *result = arrow::int32(); |
87 | 70 | break; |
88 | 26 | case TYPE_BIGINT: |
89 | 26 | *result = arrow::int64(); |
90 | 26 | break; |
91 | 26 | case TYPE_FLOAT: |
92 | 26 | *result = arrow::float32(); |
93 | 26 | break; |
94 | 50 | case TYPE_DOUBLE: |
95 | 50 | *result = arrow::float64(); |
96 | 50 | break; |
97 | 2 | case TYPE_TIMEV2: |
98 | 2 | *result = arrow::float64(); |
99 | 2 | break; |
100 | 63 | case TYPE_IPV4: |
101 | | // ipv4 is uint32, but parquet not uint32, it's will be convert to int64 |
102 | | // so use int32 directly |
103 | 63 | *result = arrow::int32(); |
104 | 63 | break; |
105 | 53 | case TYPE_IPV6: |
106 | 53 | *result = arrow::utf8(); |
107 | 53 | break; |
108 | 52 | case TYPE_UUID: |
109 | 52 | *result = arrow::extension::uuid(); |
110 | 52 | break; |
111 | 82 | case TYPE_LARGEINT: |
112 | 88 | case TYPE_VARCHAR: |
113 | 94 | case TYPE_CHAR: |
114 | 128 | case TYPE_DATE: |
115 | 154 | case TYPE_DATETIME: |
116 | 311 | case TYPE_STRING: |
117 | 324 | case TYPE_JSONB: |
118 | 324 | *result = arrow::utf8(); |
119 | 324 | break; |
120 | 30 | case TYPE_DATEV2: |
121 | 30 | *result = std::make_shared<arrow::Date32Type>(); |
122 | 30 | break; |
123 | 11 | case TYPE_TIMESTAMP_NS: |
124 | | // TIMESTAMP_NS is stored as signed epoch nanoseconds, but its SQL type has no timezone. |
125 | 11 | *result = std::make_shared<arrow::TimestampType>(arrow::TimeUnit::NANO); |
126 | 11 | break; |
127 | 18 | case TYPE_TIMESTAMPTZ: |
128 | 118 | case TYPE_DATETIMEV2: { |
129 | 118 | arrow::TimeUnit::type time_unit; |
130 | 118 | if (type->get_scale() > 3) { |
131 | 81 | time_unit = arrow::TimeUnit::MICRO; |
132 | 81 | } else if (type->get_scale() > 0) { |
133 | 9 | time_unit = arrow::TimeUnit::MILLI; |
134 | 28 | } else { |
135 | 28 | time_unit = arrow::TimeUnit::SECOND; |
136 | 28 | } |
137 | 118 | *result = std::make_shared<arrow::TimestampType>( |
138 | 118 | time_unit, timestamp_timezone(type->get_primitive_type())); |
139 | 118 | break; |
140 | 18 | } |
141 | 14 | case TYPE_DECIMALV2: |
142 | 62 | case TYPE_DECIMAL32: |
143 | 98 | case TYPE_DECIMAL64: |
144 | 122 | case TYPE_DECIMAL128I: |
145 | 122 | *result = std::make_shared<arrow::Decimal128Type>(type->get_precision(), type->get_scale()); |
146 | 122 | break; |
147 | 32 | case TYPE_DECIMAL256: |
148 | 32 | *result = std::make_shared<arrow::Decimal256Type>(type->get_precision(), type->get_scale()); |
149 | 32 | break; |
150 | 26 | case TYPE_BOOLEAN: |
151 | 26 | *result = arrow::boolean(); |
152 | 26 | break; |
153 | 264 | case TYPE_ARRAY: { |
154 | 264 | const auto* type_arr = assert_cast<const DataTypeArray*>(remove_nullable(type).get()); |
155 | 264 | std::shared_ptr<arrow::DataType> item_type; |
156 | 264 | RETURN_IF_ERROR(convert_to_arrow_type(type_arr->get_nested_type(), &item_type)); |
157 | | // Arrow stores metadata on fields, so implicit child fields lose the Doris logical type. |
158 | 264 | *result = std::make_shared<arrow::ListType>(make_child_field( |
159 | 264 | "item", item_type, true, type_arr->get_nested_type()->get_primitive_type())); |
160 | 264 | break; |
161 | 264 | } |
162 | 165 | case TYPE_MAP: { |
163 | 165 | const auto* type_map = assert_cast<const DataTypeMap*>(remove_nullable(type).get()); |
164 | 165 | std::shared_ptr<arrow::DataType> key_type; |
165 | 165 | std::shared_ptr<arrow::DataType> val_type; |
166 | 165 | RETURN_IF_ERROR(convert_to_arrow_type(type_map->get_key_type(), &key_type)); |
167 | 165 | RETURN_IF_ERROR(convert_to_arrow_type(type_map->get_value_type(), &val_type)); |
168 | 165 | auto key_field = make_child_field("key", key_type, false, |
169 | 165 | type_map->get_key_type()->get_primitive_type()); |
170 | 165 | auto value_field = make_child_field("value", val_type, true, |
171 | 165 | type_map->get_value_type()->get_primitive_type()); |
172 | 165 | *result = std::make_shared<arrow::MapType>(key_field, value_field); |
173 | 165 | break; |
174 | 165 | } |
175 | 155 | case TYPE_STRUCT: { |
176 | 155 | const auto* type_struct = assert_cast<const DataTypeStruct*>(remove_nullable(type).get()); |
177 | 155 | std::vector<std::shared_ptr<arrow::Field>> fields; |
178 | 482 | for (size_t i = 0; i < type_struct->get_elements().size(); i++) { |
179 | 327 | std::shared_ptr<arrow::DataType> field_type; |
180 | 327 | RETURN_IF_ERROR(convert_to_arrow_type(type_struct->get_element(i), &field_type)); |
181 | 327 | fields.push_back(make_child_field(type_struct->get_element_name(i), field_type, |
182 | 327 | type_struct->get_element(i)->is_nullable(), |
183 | 327 | type_struct->get_element(i)->get_primitive_type())); |
184 | 327 | } |
185 | 155 | *result = std::make_shared<arrow::StructType>(fields); |
186 | 155 | break; |
187 | 155 | } |
188 | 8 | case TYPE_VARIANT: |
189 | 8 | *result = arrow::utf8(); |
190 | 8 | break; |
191 | 2 | case TYPE_QUANTILE_STATE: |
192 | 6 | case TYPE_BITMAP: |
193 | 10 | case TYPE_HLL: { |
194 | 10 | *result = arrow::binary(); |
195 | 10 | break; |
196 | 6 | } |
197 | 13 | case TYPE_VARBINARY: { |
198 | 13 | *result = arrow::binary(); |
199 | 13 | break; |
200 | 6 | } |
201 | 0 | default: |
202 | 0 | return Status::InvalidArgument("Unknown primitive type({}) convert to Arrow type", |
203 | 0 | type->get_name()); |
204 | 1.66k | } |
205 | 1.66k | return Status::OK(); |
206 | 1.66k | } |
207 | | |
208 | | // Logical types sharing Arrow storage need the same marker at the root and every child field. |
209 | | std::shared_ptr<arrow::Field> create_arrow_field_with_metadata( |
210 | | const std::string& field_name, const std::shared_ptr<arrow::DataType>& arrow_type, |
211 | 1.26k | bool is_nullable, PrimitiveType primitive_type) { |
212 | 1.26k | const char* type_name; |
213 | 1.26k | switch (primitive_type) { |
214 | 41 | case TYPE_UUID: |
215 | 41 | type_name = "UUID"; |
216 | 41 | break; |
217 | 55 | case TYPE_IPV4: |
218 | 55 | type_name = "IPV4"; |
219 | 55 | break; |
220 | 45 | case TYPE_IPV6: |
221 | 45 | type_name = "IPV6"; |
222 | 45 | break; |
223 | 69 | case TYPE_LARGEINT: |
224 | 69 | type_name = "LARGEINT"; |
225 | 69 | break; |
226 | 8 | case TYPE_JSONB: |
227 | 8 | type_name = "JSON"; |
228 | 8 | break; |
229 | 30 | case TYPE_VARIANT: |
230 | 30 | type_name = "VARIANT"; |
231 | 30 | break; |
232 | 1.02k | default: |
233 | 1.02k | return std::make_shared<arrow::Field>(field_name, arrow_type, is_nullable); |
234 | 1.26k | } |
235 | 248 | auto metadata = arrow::KeyValueMetadata::Make({"doris_type"}, {type_name}); |
236 | 248 | return std::make_shared<arrow::Field>(field_name, arrow_type, is_nullable, metadata); |
237 | 1.26k | } |
238 | | |
239 | | Status DorisArrowSchemaConvertor::get_arrow_schema_from_block( |
240 | 67 | const Block& block, std::shared_ptr<arrow::Schema>* result) const { |
241 | 67 | std::vector<std::shared_ptr<arrow::Field>> fields; |
242 | 331 | for (const auto& type_and_name : block) { |
243 | 331 | std::shared_ptr<arrow::DataType> arrow_type; |
244 | 331 | RETURN_IF_ERROR(convert_to_arrow_type(type_and_name.type, &arrow_type)); |
245 | 331 | auto field = make_field(type_and_name.name, arrow_type, type_and_name.type->is_nullable(), |
246 | 331 | type_and_name.type->get_primitive_type()); |
247 | 331 | fields.push_back(field); |
248 | 331 | } |
249 | 67 | *result = arrow::schema(std::move(fields)); |
250 | 67 | return Status::OK(); |
251 | 67 | } |
252 | | |
253 | 17 | Status DorisArrowSchemaConvertor::get_arrow_schema(std::shared_ptr<arrow::Schema>* result) const { |
254 | 17 | return get_arrow_schema_from_block(_header, result); |
255 | 17 | } |
256 | | |
257 | 109 | std::string DorisArrowSchemaConvertor::timestamp_timezone(PrimitiveType) const { |
258 | | // Arrow clients expect a timezone name rather than the ISO-8601 UTC alias. |
259 | 109 | return _timezone == "Z" ? "UTC" : _timezone; |
260 | 109 | } |
261 | | |
262 | | Status ArrowFlightSchemaConvertor::convert_to_arrow_type( |
263 | 307 | const DataTypePtr& type, std::shared_ptr<arrow::DataType>* result) const { |
264 | | // Flight always uses native Variant, including recursively converted children. |
265 | 307 | if (type->get_primitive_type() == TYPE_VARIANT) { |
266 | | // Reject by type before reading rows, including empty results and nested legacy leaves. |
267 | 41 | if (dynamic_cast<const DataTypeVariantV2*>(remove_nullable(type).get()) == nullptr) { |
268 | 0 | return Status::NotSupported( |
269 | 0 | "Native Arrow Flight output only supports Variant V2, not legacy Variant; " |
270 | 0 | "cast the result to STRING for text output"); |
271 | 0 | } |
272 | 41 | RETURN_IF_ERROR(register_arrow_variant_extension()); |
273 | 41 | *result = arrow::extension::variant( |
274 | 41 | arrow::struct_({arrow::field("metadata", arrow::binary(), false), |
275 | 41 | arrow::field("value", arrow::binary(), false)})); |
276 | 41 | return Status::OK(); |
277 | 41 | } |
278 | 266 | return DorisArrowSchemaConvertor::convert_to_arrow_type(type, result); |
279 | 307 | } |
280 | | |
281 | 17 | std::string ArrowFlightSchemaConvertor::timestamp_timezone(PrimitiveType type) const { |
282 | | // DATETIMEV2 is wall-clock time; TIMESTAMPTZ must still describe an instant. |
283 | 17 | return type == TYPE_DATETIMEV2 ? "" : DorisArrowSchemaConvertor::timestamp_timezone(type); |
284 | 17 | } |
285 | | |
286 | | std::shared_ptr<arrow::Field> DorisArrowSchemaConvertor::make_field( |
287 | | const std::string& name, const std::shared_ptr<arrow::DataType>& type, bool nullable, |
288 | 379 | PrimitiveType primitive) const { |
289 | 379 | return create_arrow_field_with_metadata(name, type, nullable, primitive); |
290 | 379 | } |
291 | | |
292 | | std::shared_ptr<arrow::Field> DorisArrowSchemaConvertor::make_child_field( |
293 | | const std::string& name, const std::shared_ptr<arrow::DataType>& type, bool nullable, |
294 | 890 | PrimitiveType primitive) const { |
295 | 890 | return create_arrow_field_with_metadata(name, type, nullable, primitive); |
296 | 890 | } |
297 | | |
298 | | std::shared_ptr<arrow::Field> LegacyArrowFlightSchemaConvertor::make_field( |
299 | | const std::string& name, const std::shared_ptr<arrow::DataType>& type, bool nullable, |
300 | 27 | PrimitiveType primitive) const { |
301 | 27 | if (primitive == TYPE_JSONB || primitive == TYPE_VARIANT) { |
302 | 2 | return arrow::field(name, type, nullable); |
303 | 2 | } |
304 | 25 | return DorisArrowSchemaConvertor::make_field(name, type, nullable, primitive); |
305 | 27 | } |
306 | | |
307 | | std::shared_ptr<arrow::Field> LegacyArrowFlightSchemaConvertor::make_child_field( |
308 | | const std::string& name, const std::shared_ptr<arrow::DataType>& type, bool nullable, |
309 | 31 | PrimitiveType) const { |
310 | 31 | return arrow::field(name, type, nullable); |
311 | 31 | } |
312 | | |
313 | | Status DorisArrowSchemaConvertor::get_arrow_schema_from_expr_ctxs( |
314 | 14 | const VExprContextSPtrs& output_vexpr_ctxs, std::shared_ptr<arrow::Schema>* result) const { |
315 | 14 | std::vector<std::shared_ptr<arrow::Field>> fields; |
316 | 64 | for (int i = 0; i < output_vexpr_ctxs.size(); i++) { |
317 | 50 | std::shared_ptr<arrow::DataType> arrow_type; |
318 | 50 | auto root_expr = output_vexpr_ctxs.at(i)->root(); |
319 | 50 | RETURN_IF_ERROR(convert_to_arrow_type(root_expr->data_type(), &arrow_type)); |
320 | 50 | auto field_name = root_expr->is_slot_ref() && !root_expr->expr_label().empty() |
321 | 50 | ? root_expr->expr_label() |
322 | 50 | : fmt::format("{}_{}", root_expr->data_type()->get_name(), i); |
323 | 50 | auto field = make_field(field_name, arrow_type, root_expr->is_nullable(), |
324 | 50 | root_expr->data_type()->get_primitive_type()); |
325 | 50 | fields.push_back(std::move(field)); |
326 | 50 | } |
327 | 14 | *result = arrow::schema(std::move(fields)); |
328 | 14 | return Status::OK(); |
329 | 14 | } |
330 | | |
331 | 4 | Status serialize_record_batch(const arrow::RecordBatch& record_batch, std::string* result) { |
332 | | // create sink memory buffer outputstream with the computed capacity |
333 | 4 | int64_t capacity; |
334 | 4 | arrow::Status a_st = arrow::ipc::GetRecordBatchSize(record_batch, &capacity); |
335 | 4 | if (!a_st.ok()) { |
336 | 0 | return Status::InternalError("GetRecordBatchSize failure, reason: {}", a_st.ToString()); |
337 | 0 | } |
338 | 4 | auto sink_res = arrow::io::BufferOutputStream::Create(capacity, arrow::default_memory_pool()); |
339 | 4 | if (!sink_res.ok()) { |
340 | 0 | return Status::InternalError("create BufferOutputStream failure, reason: {}", |
341 | 0 | sink_res.status().ToString()); |
342 | 0 | } |
343 | 4 | std::shared_ptr<arrow::io::BufferOutputStream> sink = sink_res.ValueOrDie(); |
344 | | // create RecordBatch Writer |
345 | 4 | auto res = arrow::ipc::MakeStreamWriter(sink.get(), record_batch.schema()); |
346 | 4 | if (!res.ok()) { |
347 | 0 | return Status::InternalError("open RecordBatchStreamWriter failure, reason: {}", |
348 | 0 | res.status().ToString()); |
349 | 0 | } |
350 | | // write RecordBatch to memory buffer outputstream |
351 | 4 | std::shared_ptr<arrow::ipc::RecordBatchWriter> record_batch_writer = res.ValueOrDie(); |
352 | 4 | a_st = record_batch_writer->WriteRecordBatch(record_batch); |
353 | 4 | if (!a_st.ok()) { |
354 | 0 | return Status::InternalError("write record batch failure, reason: {}", a_st.ToString()); |
355 | 0 | } |
356 | 4 | a_st = record_batch_writer->Close(); |
357 | 4 | if (!a_st.ok()) { |
358 | 0 | return Status::InternalError("Close failed, reason: {}", a_st.ToString()); |
359 | 0 | } |
360 | 4 | auto finish_res = sink->Finish(); |
361 | 4 | if (!finish_res.ok()) { |
362 | 0 | return Status::InternalError("allocate result buffer failure, reason: {}", |
363 | 0 | finish_res.status().ToString()); |
364 | 0 | } |
365 | 4 | *result = finish_res.ValueOrDie()->ToString(); |
366 | | // close the sink |
367 | 4 | a_st = sink->Close(); |
368 | 4 | if (!a_st.ok()) { |
369 | 0 | return Status::InternalError("Close failed, reason: {}", a_st.ToString()); |
370 | 0 | } |
371 | 4 | return Status::OK(); |
372 | 4 | } |
373 | | |
374 | 14 | Status serialize_arrow_schema(std::shared_ptr<arrow::Schema>* schema, std::string* result) { |
375 | | // Schema RPC readers only consume the IPC schema. Building an empty batch would require |
376 | | // nested extension builders, which Arrow does not provide for ARRAY/MAP/STRUCT<VARIANT>. |
377 | 14 | std::shared_ptr<arrow::io::BufferOutputStream> sink; |
378 | 14 | RETURN_DORIS_STATUS_IF_RESULT_ERROR(sink, arrow::io::BufferOutputStream::Create()); |
379 | 14 | std::shared_ptr<arrow::ipc::RecordBatchWriter> writer; |
380 | 14 | RETURN_DORIS_STATUS_IF_RESULT_ERROR(writer, arrow::ipc::MakeStreamWriter(sink.get(), *schema)); |
381 | 14 | RETURN_DORIS_STATUS_IF_ERROR(writer->Close()); |
382 | 14 | std::shared_ptr<arrow::Buffer> buffer; |
383 | 14 | RETURN_DORIS_STATUS_IF_RESULT_ERROR(buffer, sink->Finish()); |
384 | 14 | *result = buffer->ToString(); |
385 | 14 | return Status::OK(); |
386 | 14 | } |
387 | | |
388 | | } // namespace doris |