| 1 | #include <Processors/Transforms/MergingAggregatedMemoryEfficientTransform.h> |
| 2 | |
| 3 | #include <Processors/ISimpleTransform.h> |
| 4 | #include <Interpreters/Aggregator.h> |
| 5 | #include <Processors/ResizeProcessor.h> |
| 6 | |
| 7 | namespace DB |
| 8 | { |
| 9 | |
| 10 | struct ChunksToMerge : public ChunkInfo |
| 11 | { |
| 12 | std::unique_ptr<Chunks> chunks; |
| 13 | Int32 bucket_num = -1; |
| 14 | bool is_overflows = false; |
| 15 | }; |
| 16 | |
| 17 | GroupingAggregatedTransform::GroupingAggregatedTransform( |
| 18 | const Block & , size_t num_inputs_, AggregatingTransformParamsPtr params_) |
| 19 | : IProcessor(InputPorts(num_inputs_, header_), { Block() }) |
| 20 | , num_inputs(num_inputs_) |
| 21 | , params(std::move(params_)) |
| 22 | , last_bucket_number(num_inputs, -1) |
| 23 | , read_from_input(num_inputs, false) |
| 24 | { |
| 25 | } |
| 26 | |
| 27 | void GroupingAggregatedTransform::readFromAllInputs() |
| 28 | { |
| 29 | auto in = inputs.begin(); |
| 30 | for (size_t i = 0; i < num_inputs; ++i, ++in) |
| 31 | { |
| 32 | if (in->isFinished()) |
| 33 | continue; |
| 34 | |
| 35 | if (read_from_input[i]) |
| 36 | continue; |
| 37 | |
| 38 | in->setNeeded(); |
| 39 | |
| 40 | if (!in->hasData()) |
| 41 | return; |
| 42 | |
| 43 | auto chunk = in->pull(); |
| 44 | read_from_input[i] = true; |
| 45 | addChunk(std::move(chunk), i); |
| 46 | } |
| 47 | |
| 48 | read_from_all_inputs = true; |
| 49 | } |
| 50 | |
| 51 | void GroupingAggregatedTransform::pushData(Chunks chunks, Int32 bucket, bool is_overflows) |
| 52 | { |
| 53 | auto & output = outputs.front(); |
| 54 | |
| 55 | auto info = std::make_shared<ChunksToMerge>(); |
| 56 | info->bucket_num = bucket; |
| 57 | info->is_overflows = is_overflows; |
| 58 | info->chunks = std::make_unique<Chunks>(std::move(chunks)); |
| 59 | |
| 60 | Chunk chunk; |
| 61 | chunk.setChunkInfo(std::move(info)); |
| 62 | output.push(std::move(chunk)); |
| 63 | } |
| 64 | |
| 65 | bool GroupingAggregatedTransform::tryPushTwoLevelData() |
| 66 | { |
| 67 | auto try_push_by_iter = [&](auto batch_it) |
| 68 | { |
| 69 | if (batch_it == chunks_map.end()) |
| 70 | return false; |
| 71 | |
| 72 | Chunks & cur_chunks = batch_it->second; |
| 73 | if (cur_chunks.empty()) |
| 74 | { |
| 75 | chunks_map.erase(batch_it); |
| 76 | return false; |
| 77 | } |
| 78 | |
| 79 | pushData(std::move(cur_chunks), batch_it->first, false); |
| 80 | chunks_map.erase(batch_it); |
| 81 | return true; |
| 82 | }; |
| 83 | |
| 84 | if (all_inputs_finished) |
| 85 | { |
| 86 | /// Chunks are sorted by bucket. |
| 87 | while (!chunks_map.empty()) |
| 88 | if (try_push_by_iter(chunks_map.begin())) |
| 89 | return true; |
| 90 | } |
| 91 | else |
| 92 | { |
| 93 | for (; next_bucket_to_push < current_bucket; ++next_bucket_to_push) |
| 94 | if (try_push_by_iter(chunks_map.find(next_bucket_to_push))) |
| 95 | return true; |
| 96 | } |
| 97 | |
| 98 | return false; |
| 99 | } |
| 100 | |
| 101 | bool GroupingAggregatedTransform::tryPushSingleLevelData() |
| 102 | { |
| 103 | if (single_level_chunks.empty()) |
| 104 | return false; |
| 105 | |
| 106 | pushData(std::move(single_level_chunks), -1, false); |
| 107 | return true; |
| 108 | } |
| 109 | |
| 110 | bool GroupingAggregatedTransform::tryPushOverflowData() |
| 111 | { |
| 112 | if (overflow_chunks.empty()) |
| 113 | return false; |
| 114 | |
| 115 | pushData(std::move(overflow_chunks), -1, true); |
| 116 | return true; |
| 117 | } |
| 118 | |
| 119 | IProcessor::Status GroupingAggregatedTransform::prepare() |
| 120 | { |
| 121 | /// Check can output. |
| 122 | auto & output = outputs.front(); |
| 123 | |
| 124 | if (output.isFinished()) |
| 125 | { |
| 126 | for (auto & input : inputs) |
| 127 | input.close(); |
| 128 | |
| 129 | chunks_map.clear(); |
| 130 | last_bucket_number.clear(); |
| 131 | return Status::Finished; |
| 132 | } |
| 133 | |
| 134 | /// Read first time from each input to understand if we have two-level aggregation. |
| 135 | if (!read_from_all_inputs) |
| 136 | { |
| 137 | readFromAllInputs(); |
| 138 | if (!read_from_all_inputs) |
| 139 | return Status::NeedData; |
| 140 | } |
| 141 | |
| 142 | /// Convert single level to two levels if have two-level input. |
| 143 | if (has_two_level && !single_level_chunks.empty()) |
| 144 | return Status::Ready; |
| 145 | |
| 146 | /// Check can push (to avoid data caching). |
| 147 | if (!output.canPush()) |
| 148 | { |
| 149 | for (auto & input : inputs) |
| 150 | input.setNotNeeded(); |
| 151 | |
| 152 | return Status::PortFull; |
| 153 | } |
| 154 | |
| 155 | bool pushed_to_output = false; |
| 156 | |
| 157 | /// Output if has data. |
| 158 | if (has_two_level) |
| 159 | pushed_to_output = tryPushTwoLevelData(); |
| 160 | |
| 161 | auto need_input = [this](size_t input_num) |
| 162 | { |
| 163 | if (last_bucket_number[input_num] < current_bucket) |
| 164 | return true; |
| 165 | |
| 166 | return expect_several_chunks_for_single_bucket_per_source && last_bucket_number[input_num] == current_bucket; |
| 167 | }; |
| 168 | |
| 169 | /// Read next bucket if can. |
| 170 | for (; ; ++current_bucket) |
| 171 | { |
| 172 | bool finished = true; |
| 173 | bool need_data = false; |
| 174 | |
| 175 | auto in = inputs.begin(); |
| 176 | for (size_t input_num = 0; input_num < num_inputs; ++input_num, ++in) |
| 177 | { |
| 178 | if (in->isFinished()) |
| 179 | continue; |
| 180 | |
| 181 | finished = false; |
| 182 | |
| 183 | if (!need_input(input_num)) |
| 184 | continue; |
| 185 | |
| 186 | in->setNeeded(); |
| 187 | |
| 188 | if (!in->hasData()) |
| 189 | { |
| 190 | need_data = true; |
| 191 | continue; |
| 192 | } |
| 193 | |
| 194 | auto chunk = in->pull(); |
| 195 | addChunk(std::move(chunk), input_num); |
| 196 | |
| 197 | if (has_two_level && !single_level_chunks.empty()) |
| 198 | return Status::Ready; |
| 199 | |
| 200 | if (!in->isFinished() && need_input(input_num)) |
| 201 | need_data = true; |
| 202 | } |
| 203 | |
| 204 | if (finished) |
| 205 | { |
| 206 | all_inputs_finished = true; |
| 207 | break; |
| 208 | } |
| 209 | |
| 210 | if (need_data) |
| 211 | return Status::NeedData; |
| 212 | } |
| 213 | |
| 214 | if (pushed_to_output) |
| 215 | return Status::PortFull; |
| 216 | |
| 217 | if (has_two_level) |
| 218 | { |
| 219 | if (tryPushTwoLevelData()) |
| 220 | return Status::PortFull; |
| 221 | |
| 222 | /// Sanity check. If new bucket was read, we should be able to push it. |
| 223 | if (!all_inputs_finished) |
| 224 | throw Exception("GroupingAggregatedTransform has read new two-level bucket, but couldn't push it." , |
| 225 | ErrorCodes::LOGICAL_ERROR); |
| 226 | } |
| 227 | else |
| 228 | { |
| 229 | if (!all_inputs_finished) |
| 230 | throw Exception("GroupingAggregatedTransform should have read all chunks for single level aggregation, " |
| 231 | "but not all of the inputs are finished." , ErrorCodes::LOGICAL_ERROR); |
| 232 | |
| 233 | if (tryPushSingleLevelData()) |
| 234 | return Status::PortFull; |
| 235 | } |
| 236 | |
| 237 | /// If we haven't pushed to output, then all data was read. Push overflows if have. |
| 238 | if (tryPushOverflowData()) |
| 239 | return Status::PortFull; |
| 240 | |
| 241 | output.finish(); |
| 242 | return Status::Finished; |
| 243 | } |
| 244 | |
| 245 | void GroupingAggregatedTransform::addChunk(Chunk chunk, size_t input) |
| 246 | { |
| 247 | auto & info = chunk.getChunkInfo(); |
| 248 | if (!info) |
| 249 | throw Exception("Chunk info was not set for chunk in GroupingAggregatedTransform." , ErrorCodes::LOGICAL_ERROR); |
| 250 | |
| 251 | auto * agg_info = typeid_cast<const AggregatedChunkInfo *>(info.get()); |
| 252 | if (!agg_info) |
| 253 | throw Exception("Chunk should have AggregatedChunkInfo in GroupingAggregatedTransform." , ErrorCodes::LOGICAL_ERROR); |
| 254 | |
| 255 | Int32 bucket = agg_info->bucket_num; |
| 256 | bool is_overflows = agg_info->is_overflows; |
| 257 | |
| 258 | if (is_overflows) |
| 259 | overflow_chunks.emplace_back(std::move(chunk)); |
| 260 | else if (bucket < 0) |
| 261 | single_level_chunks.emplace_back(std::move(chunk)); |
| 262 | else |
| 263 | { |
| 264 | chunks_map[bucket].emplace_back(std::move(chunk)); |
| 265 | has_two_level = true; |
| 266 | last_bucket_number[input] = bucket; |
| 267 | } |
| 268 | } |
| 269 | |
| 270 | void GroupingAggregatedTransform::work() |
| 271 | { |
| 272 | if (!single_level_chunks.empty()) |
| 273 | { |
| 274 | auto & = getOutputs().front().getHeader(); |
| 275 | auto block = header.cloneWithColumns(single_level_chunks.back().detachColumns()); |
| 276 | single_level_chunks.pop_back(); |
| 277 | auto blocks = params->aggregator.convertBlockToTwoLevel(block); |
| 278 | |
| 279 | for (auto & cur_block : blocks) |
| 280 | { |
| 281 | Int32 bucket = cur_block.info.bucket_num; |
| 282 | chunks_map[bucket].emplace_back(Chunk(cur_block.getColumns(), cur_block.rows())); |
| 283 | } |
| 284 | } |
| 285 | } |
| 286 | |
| 287 | |
| 288 | MergingAggregatedBucketTransform::MergingAggregatedBucketTransform(AggregatingTransformParamsPtr params_) |
| 289 | : ISimpleTransform({}, params_->getHeader(), false), params(std::move(params_)) |
| 290 | { |
| 291 | setInputNotNeededAfterRead(true); |
| 292 | } |
| 293 | |
| 294 | void MergingAggregatedBucketTransform::transform(Chunk & chunk) |
| 295 | { |
| 296 | auto & info = chunk.getChunkInfo(); |
| 297 | auto * chunks_to_merge = typeid_cast<const ChunksToMerge *>(info.get()); |
| 298 | |
| 299 | if (!chunks_to_merge) |
| 300 | throw Exception("MergingAggregatedSimpleTransform chunk must have ChunkInfo with type ChunksToMerge." , |
| 301 | ErrorCodes::LOGICAL_ERROR); |
| 302 | |
| 303 | auto = params->aggregator.getHeader(false); |
| 304 | |
| 305 | BlocksList blocks_list; |
| 306 | for (auto & cur_chunk : *chunks_to_merge->chunks) |
| 307 | { |
| 308 | auto & cur_info = cur_chunk.getChunkInfo(); |
| 309 | if (!cur_info) |
| 310 | throw Exception("Chunk info was not set for chunk in MergingAggregatedBucketTransform." , |
| 311 | ErrorCodes::LOGICAL_ERROR); |
| 312 | |
| 313 | auto * agg_info = typeid_cast<const AggregatedChunkInfo *>(cur_info.get()); |
| 314 | if (!agg_info) |
| 315 | throw Exception("Chunk should have AggregatedChunkInfo in MergingAggregatedBucketTransform." , |
| 316 | ErrorCodes::LOGICAL_ERROR); |
| 317 | |
| 318 | Block block = header.cloneWithColumns(cur_chunk.detachColumns()); |
| 319 | block.info.is_overflows = agg_info->is_overflows; |
| 320 | block.info.bucket_num = agg_info->bucket_num; |
| 321 | |
| 322 | blocks_list.emplace_back(std::move(block)); |
| 323 | } |
| 324 | |
| 325 | auto res_info = std::make_shared<AggregatedChunkInfo>(); |
| 326 | res_info->is_overflows = chunks_to_merge->is_overflows; |
| 327 | res_info->bucket_num = chunks_to_merge->bucket_num; |
| 328 | chunk.setChunkInfo(std::move(res_info)); |
| 329 | |
| 330 | auto block = params->aggregator.mergeBlocks(blocks_list, params->final); |
| 331 | size_t num_rows = block.rows(); |
| 332 | chunk.setColumns(block.getColumns(), num_rows); |
| 333 | } |
| 334 | |
| 335 | |
| 336 | SortingAggregatedTransform::SortingAggregatedTransform(size_t num_inputs_, AggregatingTransformParamsPtr params_) |
| 337 | : IProcessor(InputPorts(num_inputs_, params_->getHeader()), {params_->getHeader()}) |
| 338 | , num_inputs(num_inputs_) |
| 339 | , params(std::move(params_)) |
| 340 | , last_bucket_number(num_inputs, -1) |
| 341 | , is_input_finished(num_inputs, false) |
| 342 | { |
| 343 | } |
| 344 | |
| 345 | bool SortingAggregatedTransform::tryPushChunk() |
| 346 | { |
| 347 | auto & output = outputs.front(); |
| 348 | |
| 349 | if (chunks.empty()) |
| 350 | return false; |
| 351 | |
| 352 | /// Chunk with min current bucket. |
| 353 | auto it = chunks.begin(); |
| 354 | auto cur_bucket = it->first; |
| 355 | |
| 356 | /// Check that can push it |
| 357 | for (size_t input = 0; input < num_inputs; ++input) |
| 358 | if (!is_input_finished[input] && last_bucket_number[input] < cur_bucket) |
| 359 | return false; |
| 360 | |
| 361 | output.push(std::move(it->second)); |
| 362 | chunks.erase(it); |
| 363 | return true; |
| 364 | } |
| 365 | |
| 366 | void SortingAggregatedTransform::addChunk(Chunk chunk, size_t from_input) |
| 367 | { |
| 368 | auto & info = chunk.getChunkInfo(); |
| 369 | if (!info) |
| 370 | throw Exception("Chunk info was not set for chunk in SortingAggregatedTransform." , ErrorCodes::LOGICAL_ERROR); |
| 371 | |
| 372 | auto * agg_info = typeid_cast<const AggregatedChunkInfo *>(info.get()); |
| 373 | if (!agg_info) |
| 374 | throw Exception("Chunk should have AggregatedChunkInfo in SortingAggregatedTransform." , ErrorCodes::LOGICAL_ERROR); |
| 375 | |
| 376 | Int32 bucket = agg_info->bucket_num; |
| 377 | bool is_overflows = agg_info->is_overflows; |
| 378 | |
| 379 | if (is_overflows) |
| 380 | overflow_chunk = std::move(chunk); |
| 381 | else |
| 382 | { |
| 383 | if (chunks[bucket]) |
| 384 | throw Exception("SortingAggregatedTransform already got bucket with number " + toString(bucket), |
| 385 | ErrorCodes::LOGICAL_ERROR); |
| 386 | |
| 387 | chunks[bucket] = std::move(chunk); |
| 388 | last_bucket_number[from_input] = bucket; |
| 389 | } |
| 390 | } |
| 391 | |
| 392 | IProcessor::Status SortingAggregatedTransform::prepare() |
| 393 | { |
| 394 | /// Check can output. |
| 395 | auto & output = outputs.front(); |
| 396 | |
| 397 | if (output.isFinished()) |
| 398 | { |
| 399 | for (auto & input : inputs) |
| 400 | input.close(); |
| 401 | |
| 402 | chunks.clear(); |
| 403 | last_bucket_number.clear(); |
| 404 | return Status::Finished; |
| 405 | } |
| 406 | |
| 407 | /// Check can push (to avoid data caching). |
| 408 | if (!output.canPush()) |
| 409 | { |
| 410 | for (auto & input : inputs) |
| 411 | input.setNotNeeded(); |
| 412 | |
| 413 | return Status::PortFull; |
| 414 | } |
| 415 | |
| 416 | /// Push if have min version. |
| 417 | bool pushed_to_output = tryPushChunk(); |
| 418 | |
| 419 | bool need_data = false; |
| 420 | bool all_finished = true; |
| 421 | |
| 422 | /// Try read anything. |
| 423 | auto in = inputs.begin(); |
| 424 | for (size_t input_num = 0; input_num < num_inputs; ++input_num, ++in) |
| 425 | { |
| 426 | if (in->isFinished()) |
| 427 | { |
| 428 | is_input_finished[input_num] = true; |
| 429 | continue; |
| 430 | } |
| 431 | |
| 432 | //all_finished = false; |
| 433 | |
| 434 | in->setNeeded(); |
| 435 | |
| 436 | if (!in->hasData()) |
| 437 | { |
| 438 | need_data = true; |
| 439 | all_finished = false; |
| 440 | continue; |
| 441 | } |
| 442 | |
| 443 | auto chunk = in->pull(); |
| 444 | addChunk(std::move(chunk), input_num); |
| 445 | |
| 446 | if (in->isFinished()) |
| 447 | { |
| 448 | is_input_finished[input_num] = true; |
| 449 | } |
| 450 | else |
| 451 | { |
| 452 | /// If chunk was pulled, then we need data from this port. |
| 453 | need_data = true; |
| 454 | all_finished = false; |
| 455 | } |
| 456 | } |
| 457 | |
| 458 | if (pushed_to_output) |
| 459 | return Status::PortFull; |
| 460 | |
| 461 | if (tryPushChunk()) |
| 462 | return Status::PortFull; |
| 463 | |
| 464 | if (need_data) |
| 465 | return Status::NeedData; |
| 466 | |
| 467 | if (!all_finished) |
| 468 | throw Exception("SortingAggregatedTransform has read bucket, but couldn't push it." , |
| 469 | ErrorCodes::LOGICAL_ERROR); |
| 470 | |
| 471 | if (overflow_chunk) |
| 472 | { |
| 473 | output.push(std::move(overflow_chunk)); |
| 474 | return Status::PortFull; |
| 475 | } |
| 476 | |
| 477 | output.finish(); |
| 478 | return Status::Finished; |
| 479 | } |
| 480 | |
| 481 | |
| 482 | Processors createMergingAggregatedMemoryEfficientPipe( |
| 483 | Block , |
| 484 | AggregatingTransformParamsPtr params, |
| 485 | size_t num_inputs, |
| 486 | size_t num_merging_processors) |
| 487 | { |
| 488 | Processors processors; |
| 489 | processors.reserve(num_merging_processors + 2); |
| 490 | |
| 491 | auto grouping = std::make_shared<GroupingAggregatedTransform>(header, num_inputs, params); |
| 492 | processors.emplace_back(std::move(grouping)); |
| 493 | |
| 494 | if (num_merging_processors <= 1) |
| 495 | { |
| 496 | /// --> GroupingAggregated --> MergingAggregatedBucket --> |
| 497 | auto transform = std::make_shared<MergingAggregatedBucketTransform>(params); |
| 498 | connect(processors.back()->getOutputs().front(), transform->getInputPort()); |
| 499 | |
| 500 | processors.emplace_back(std::move(transform)); |
| 501 | return processors; |
| 502 | } |
| 503 | |
| 504 | /// --> --> MergingAggregatedBucket --> |
| 505 | /// --> GroupingAggregated --> ResizeProcessor --> MergingAggregatedBucket --> SortingAggregated --> |
| 506 | /// --> --> MergingAggregatedBucket --> |
| 507 | |
| 508 | auto resize = std::make_shared<ResizeProcessor>(Block(), 1, num_merging_processors); |
| 509 | connect(processors.back()->getOutputs().front(), resize->getInputs().front()); |
| 510 | processors.emplace_back(std::move(resize)); |
| 511 | |
| 512 | auto sorting = std::make_shared<SortingAggregatedTransform>(num_merging_processors, params); |
| 513 | auto out = processors.back()->getOutputs().begin(); |
| 514 | auto in = sorting->getInputs().begin(); |
| 515 | |
| 516 | for (size_t i = 0; i < num_merging_processors; ++i, ++in, ++out) |
| 517 | { |
| 518 | auto transform = std::make_shared<MergingAggregatedBucketTransform>(params); |
| 519 | transform->setStream(i); |
| 520 | connect(*out, transform->getInputPort()); |
| 521 | connect(transform->getOutputPort(), *in); |
| 522 | processors.emplace_back(std::move(transform)); |
| 523 | } |
| 524 | |
| 525 | processors.emplace_back(std::move(sorting)); |
| 526 | return processors; |
| 527 | } |
| 528 | |
| 529 | } |
| 530 | |