| 1 | #include <queue> |
| 2 | #include <iomanip> |
| 3 | #include <sstream> |
| 4 | |
| 5 | #include <DataStreams/MergingSortedBlockInputStream.h> |
| 6 | #include <DataStreams/ColumnGathererStream.h> |
| 7 | |
| 8 | |
| 9 | namespace DB |
| 10 | { |
| 11 | |
| 12 | namespace ErrorCodes |
| 13 | { |
| 14 | extern const int LOGICAL_ERROR; |
| 15 | extern const int NUMBER_OF_COLUMNS_DOESNT_MATCH; |
| 16 | } |
| 17 | |
| 18 | |
| 19 | MergingSortedBlockInputStream::MergingSortedBlockInputStream( |
| 20 | const BlockInputStreams & inputs_, const SortDescription & description_, |
| 21 | size_t max_block_size_, UInt64 limit_, WriteBuffer * out_row_sources_buf_, bool quiet_, bool average_block_sizes_) |
| 22 | : description(description_), max_block_size(max_block_size_), limit(limit_), quiet(quiet_) |
| 23 | , average_block_sizes(average_block_sizes_), source_blocks(inputs_.size()) |
| 24 | , cursors(inputs_.size()), out_row_sources_buf(out_row_sources_buf_) |
| 25 | { |
| 26 | children.insert(children.end(), inputs_.begin(), inputs_.end()); |
| 27 | header = children.at(0)->getHeader(); |
| 28 | num_columns = header.columns(); |
| 29 | } |
| 30 | |
| 31 | void MergingSortedBlockInputStream::init(MutableColumns & merged_columns) |
| 32 | { |
| 33 | /// Read the first blocks, initialize the queue. |
| 34 | if (first) |
| 35 | { |
| 36 | first = false; |
| 37 | |
| 38 | for (size_t i = 0; i < source_blocks.size(); ++i) |
| 39 | { |
| 40 | SharedBlockPtr & shared_block_ptr = source_blocks[i]; |
| 41 | |
| 42 | if (shared_block_ptr.get()) |
| 43 | continue; |
| 44 | |
| 45 | shared_block_ptr = new detail::SharedBlock(children[i]->read()); |
| 46 | |
| 47 | const size_t rows = shared_block_ptr->rows(); |
| 48 | |
| 49 | if (rows == 0) |
| 50 | continue; |
| 51 | |
| 52 | if (expected_block_size < rows) |
| 53 | expected_block_size = std::min(rows, max_block_size); |
| 54 | |
| 55 | cursors[i] = SortCursorImpl(*shared_block_ptr, description, i); |
| 56 | shared_block_ptr->all_columns = cursors[i].all_columns; |
| 57 | shared_block_ptr->sort_columns = cursors[i].sort_columns; |
| 58 | has_collation |= cursors[i].has_collation; |
| 59 | } |
| 60 | |
| 61 | if (has_collation) |
| 62 | initQueue(queue_with_collation); |
| 63 | else |
| 64 | initQueue(queue_without_collation); |
| 65 | } |
| 66 | |
| 67 | /// Let's check that all source blocks have the same structure. |
| 68 | for (const SharedBlockPtr & shared_block_ptr : source_blocks) |
| 69 | { |
| 70 | if (!*shared_block_ptr) |
| 71 | continue; |
| 72 | |
| 73 | assertBlocksHaveEqualStructure(*shared_block_ptr, header, getName()); |
| 74 | } |
| 75 | |
| 76 | merged_columns.resize(num_columns); |
| 77 | for (size_t i = 0; i < num_columns; ++i) |
| 78 | { |
| 79 | merged_columns[i] = header.safeGetByPosition(i).column->cloneEmpty(); |
| 80 | merged_columns[i]->reserve(expected_block_size); |
| 81 | } |
| 82 | } |
| 83 | |
| 84 | |
| 85 | template <typename TSortCursor> |
| 86 | void MergingSortedBlockInputStream::initQueue(std::priority_queue<TSortCursor> & queue) |
| 87 | { |
| 88 | for (size_t i = 0; i < cursors.size(); ++i) |
| 89 | if (!cursors[i].empty()) |
| 90 | queue.push(TSortCursor(&cursors[i])); |
| 91 | } |
| 92 | |
| 93 | |
| 94 | Block MergingSortedBlockInputStream::readImpl() |
| 95 | { |
| 96 | if (finished) |
| 97 | return {}; |
| 98 | |
| 99 | if (children.size() == 1) |
| 100 | return children[0]->read(); |
| 101 | |
| 102 | MutableColumns merged_columns; |
| 103 | |
| 104 | init(merged_columns); |
| 105 | if (merged_columns.empty()) |
| 106 | return {}; |
| 107 | |
| 108 | if (has_collation) |
| 109 | merge(merged_columns, queue_with_collation); |
| 110 | else |
| 111 | merge(merged_columns, queue_without_collation); |
| 112 | |
| 113 | return header.cloneWithColumns(std::move(merged_columns)); |
| 114 | } |
| 115 | |
| 116 | |
| 117 | template <typename TSortCursor> |
| 118 | void MergingSortedBlockInputStream::fetchNextBlock(const TSortCursor & current, std::priority_queue<TSortCursor> & queue) |
| 119 | { |
| 120 | size_t order = current->order; |
| 121 | size_t size = cursors.size(); |
| 122 | |
| 123 | if (order >= size || &cursors[order] != current.impl) |
| 124 | throw Exception("Logical error in MergingSortedBlockInputStream" , ErrorCodes::LOGICAL_ERROR); |
| 125 | |
| 126 | while (true) |
| 127 | { |
| 128 | source_blocks[order] = new detail::SharedBlock(children[order]->read()); |
| 129 | |
| 130 | if (!*source_blocks[order]) |
| 131 | break; |
| 132 | |
| 133 | if (source_blocks[order]->rows()) |
| 134 | { |
| 135 | cursors[order].reset(*source_blocks[order]); |
| 136 | queue.push(TSortCursor(&cursors[order])); |
| 137 | source_blocks[order]->all_columns = cursors[order].all_columns; |
| 138 | source_blocks[order]->sort_columns = cursors[order].sort_columns; |
| 139 | break; |
| 140 | } |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | |
| 145 | bool MergingSortedBlockInputStream::MergeStopCondition::checkStop() const |
| 146 | { |
| 147 | if (!count_average) |
| 148 | return sum_rows_count == max_block_size; |
| 149 | |
| 150 | if (sum_rows_count == 0) |
| 151 | return false; |
| 152 | |
| 153 | size_t average = sum_blocks_granularity / sum_rows_count; |
| 154 | return sum_rows_count >= average; |
| 155 | } |
| 156 | |
| 157 | template |
| 158 | void MergingSortedBlockInputStream::fetchNextBlock<SortCursor>(const SortCursor & current, std::priority_queue<SortCursor> & queue); |
| 159 | |
| 160 | template |
| 161 | void MergingSortedBlockInputStream::fetchNextBlock<SortCursorWithCollation>(const SortCursorWithCollation & current, std::priority_queue<SortCursorWithCollation> & queue); |
| 162 | |
| 163 | |
| 164 | template <typename TSortCursor> |
| 165 | void MergingSortedBlockInputStream::merge(MutableColumns & merged_columns, std::priority_queue<TSortCursor> & queue) |
| 166 | { |
| 167 | size_t merged_rows = 0; |
| 168 | |
| 169 | MergeStopCondition stop_condition(average_block_sizes, max_block_size); |
| 170 | /** Increase row counters. |
| 171 | * Return true if it's time to finish generating the current data block. |
| 172 | */ |
| 173 | auto count_row_and_check_limit = [&, this](size_t current_granularity) |
| 174 | { |
| 175 | ++total_merged_rows; |
| 176 | if (limit && total_merged_rows == limit) |
| 177 | { |
| 178 | // std::cerr << "Limit reached\n"; |
| 179 | cancel(false); |
| 180 | finished = true; |
| 181 | return true; |
| 182 | } |
| 183 | |
| 184 | ++merged_rows; |
| 185 | stop_condition.addRowWithGranularity(current_granularity); |
| 186 | return stop_condition.checkStop(); |
| 187 | }; |
| 188 | |
| 189 | /// Take rows in required order and put them into `merged_columns`, while the rows are no more than `max_block_size` |
| 190 | while (!queue.empty()) |
| 191 | { |
| 192 | TSortCursor current = queue.top(); |
| 193 | size_t current_block_granularity = current->rows; |
| 194 | queue.pop(); |
| 195 | |
| 196 | while (true) |
| 197 | { |
| 198 | /** And what if the block is totally less or equal than the rest for the current cursor? |
| 199 | * Or is there only one data source left in the queue? Then you can take the entire block on current cursor. |
| 200 | */ |
| 201 | if (current->isFirst() && (queue.empty() || current.totallyLessOrEquals(queue.top()))) |
| 202 | { |
| 203 | // std::cerr << "current block is totally less or equals\n"; |
| 204 | |
| 205 | /// If there are already data in the current block, we first return it. We'll get here again the next time we call the merge function. |
| 206 | if (merged_rows != 0) |
| 207 | { |
| 208 | //std::cerr << "merged rows is non-zero\n"; |
| 209 | queue.push(current); |
| 210 | return; |
| 211 | } |
| 212 | |
| 213 | /// Actually, current->order stores source number (i.e. cursors[current->order] == current) |
| 214 | size_t source_num = current->order; |
| 215 | |
| 216 | if (source_num >= cursors.size()) |
| 217 | throw Exception("Logical error in MergingSortedBlockInputStream" , ErrorCodes::LOGICAL_ERROR); |
| 218 | |
| 219 | for (size_t i = 0; i < num_columns; ++i) |
| 220 | merged_columns[i] = (*std::move(source_blocks[source_num]->getByPosition(i).column)).mutate(); |
| 221 | |
| 222 | // std::cerr << "copied columns\n"; |
| 223 | |
| 224 | merged_rows = merged_columns.at(0)->size(); |
| 225 | |
| 226 | /// Limit output |
| 227 | if (limit && total_merged_rows + merged_rows > limit) |
| 228 | { |
| 229 | merged_rows = limit - total_merged_rows; |
| 230 | for (size_t i = 0; i < num_columns; ++i) |
| 231 | { |
| 232 | auto & column = merged_columns[i]; |
| 233 | column = (*column->cut(0, merged_rows)).mutate(); |
| 234 | } |
| 235 | |
| 236 | cancel(false); |
| 237 | finished = true; |
| 238 | } |
| 239 | |
| 240 | /// Write order of rows for other columns |
| 241 | /// this data will be used in grather stream |
| 242 | if (out_row_sources_buf) |
| 243 | { |
| 244 | RowSourcePart row_source(source_num); |
| 245 | for (size_t i = 0; i < merged_rows; ++i) |
| 246 | out_row_sources_buf->write(row_source.data); |
| 247 | } |
| 248 | |
| 249 | //std::cerr << "fetching next block\n"; |
| 250 | |
| 251 | total_merged_rows += merged_rows; |
| 252 | fetchNextBlock(current, queue); |
| 253 | return; |
| 254 | } |
| 255 | |
| 256 | // std::cerr << "total_merged_rows: " << total_merged_rows << ", merged_rows: " << merged_rows << "\n"; |
| 257 | // std::cerr << "Inserting row\n"; |
| 258 | for (size_t i = 0; i < num_columns; ++i) |
| 259 | merged_columns[i]->insertFrom(*current->all_columns[i], current->pos); |
| 260 | |
| 261 | if (out_row_sources_buf) |
| 262 | { |
| 263 | /// Actually, current.impl->order stores source number (i.e. cursors[current.impl->order] == current.impl) |
| 264 | RowSourcePart row_source(current->order); |
| 265 | out_row_sources_buf->write(row_source.data); |
| 266 | } |
| 267 | |
| 268 | if (!current->isLast()) |
| 269 | { |
| 270 | // std::cerr << "moving to next row\n"; |
| 271 | current->next(); |
| 272 | |
| 273 | if (queue.empty() || !(current.greater(queue.top()))) |
| 274 | { |
| 275 | if (count_row_and_check_limit(current_block_granularity)) |
| 276 | { |
| 277 | // std::cerr << "pushing back to queue\n"; |
| 278 | queue.push(current); |
| 279 | return; |
| 280 | } |
| 281 | |
| 282 | /// Do not put the cursor back in the queue, but continue to work with the current cursor. |
| 283 | // std::cerr << "current is still on top, using current row\n"; |
| 284 | continue; |
| 285 | } |
| 286 | else |
| 287 | { |
| 288 | // std::cerr << "next row is not least, pushing back to queue\n"; |
| 289 | queue.push(current); |
| 290 | } |
| 291 | } |
| 292 | else |
| 293 | { |
| 294 | /// We get the next block from the corresponding source, if there is one. |
| 295 | // std::cerr << "It was last row, fetching next block\n"; |
| 296 | fetchNextBlock(current, queue); |
| 297 | } |
| 298 | |
| 299 | break; |
| 300 | } |
| 301 | |
| 302 | if (count_row_and_check_limit(current_block_granularity)) |
| 303 | return; |
| 304 | } |
| 305 | |
| 306 | cancel(false); |
| 307 | finished = true; |
| 308 | } |
| 309 | |
| 310 | |
| 311 | void MergingSortedBlockInputStream::readSuffixImpl() |
| 312 | { |
| 313 | if (quiet) |
| 314 | return; |
| 315 | |
| 316 | const BlockStreamProfileInfo & profile_info = getProfileInfo(); |
| 317 | double seconds = profile_info.total_stopwatch.elapsedSeconds(); |
| 318 | |
| 319 | std::stringstream message; |
| 320 | message << std::fixed << std::setprecision(2) |
| 321 | << "Merge sorted " << profile_info.blocks << " blocks, " << profile_info.rows << " rows" |
| 322 | << " in " << seconds << " sec." ; |
| 323 | |
| 324 | if (seconds) |
| 325 | message << ", " |
| 326 | << profile_info.rows / seconds << " rows/sec., " |
| 327 | << profile_info.bytes / 1000000.0 / seconds << " MB/sec." ; |
| 328 | |
| 329 | LOG_DEBUG(log, message.str()); |
| 330 | } |
| 331 | |
| 332 | } |
| 333 | |