be/src/storage/segment/vertical_segment_writer.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 "storage/segment/vertical_segment_writer.h" |
19 | | |
20 | | #include <crc32c/crc32c.h> |
21 | | #include <gen_cpp/olap_file.pb.h> |
22 | | #include <gen_cpp/segment_v2.pb.h> |
23 | | #include <parallel_hashmap/phmap.h> |
24 | | |
25 | | #include <algorithm> |
26 | | #include <cassert> |
27 | | #include <memory> |
28 | | #include <ostream> |
29 | | #include <string> |
30 | | #include <unordered_map> |
31 | | #include <unordered_set> |
32 | | #include <utility> |
33 | | |
34 | | #include "cloud/config.h" |
35 | | #include "common/cast_set.h" |
36 | | #include "common/compiler_util.h" // IWYU pragma: keep |
37 | | #include "common/config.h" |
38 | | #include "common/logging.h" // LOG |
39 | | #include "common/status.h" |
40 | | #include "core/assert_cast.h" |
41 | | #include "core/block/block.h" |
42 | | #include "core/block/column_with_type_and_name.h" |
43 | | #include "core/column/column_nullable.h" |
44 | | #include "core/column/column_string.h" |
45 | | #include "core/column/column_vector.h" |
46 | | #include "core/data_type/data_type.h" |
47 | | #include "core/data_type/data_type_factory.hpp" |
48 | | #include "core/data_type/data_type_number.h" // IWYU pragma: keep |
49 | | #include "core/types.h" |
50 | | #include "exec/common/variant_util.h" |
51 | | #include "io/fs/file_writer.h" |
52 | | #include "io/fs/local_file_system.h" |
53 | | #include "runtime/exec_env.h" |
54 | | #include "runtime/memory/mem_tracker.h" |
55 | | #include "storage/data_dir.h" |
56 | | #include "storage/index/index_file_writer.h" |
57 | | #include "storage/index/inverted/inverted_index_desc.h" |
58 | | #include "storage/index/inverted/inverted_index_fs_directory.h" |
59 | | #include "storage/index/primary_key_index.h" |
60 | | #include "storage/index/short_key_index.h" |
61 | | #include "storage/iterator/olap_data_convertor.h" |
62 | | #include "storage/key_coder.h" |
63 | | #include "storage/mow/historical_row_fetcher.h" |
64 | | #include "storage/mow/key_probe.h" |
65 | | #include "storage/olap_common.h" |
66 | | #include "storage/partial_update_info.h" |
67 | | #include "storage/row_cursor.h" // RowCursor // IWYU pragma: keep |
68 | | #include "storage/rowset/rowset_fwd.h" |
69 | | #include "storage/rowset/rowset_writer_context.h" // RowsetWriterContext |
70 | | #include "storage/rowset/segment_creator.h" |
71 | | #include "storage/segment/column_writer.h" // ColumnWriter |
72 | | #include "storage/segment/encoding_info.h" |
73 | | #include "storage/segment/external_col_meta_util.h" |
74 | | #include "storage/segment/historical_row_retriever.h" |
75 | | #include "storage/segment/page_io.h" |
76 | | #include "storage/segment/page_pointer.h" |
77 | | #include "storage/segment/segment_loader.h" |
78 | | #include "storage/segment/variant/variant_ext_meta_writer.h" |
79 | | #include "storage/tablet/base_tablet.h" |
80 | | #include "storage/tablet/tablet_schema.h" |
81 | | #include "storage/transform/block_transform.h" |
82 | | #include "storage/utils.h" |
83 | | #include "util/coding.h" |
84 | | #include "util/debug_points.h" |
85 | | #include "util/faststring.h" |
86 | | #include "util/json/path_in_data.h" |
87 | | #include "util/jsonb/serialize.h" |
88 | | namespace doris::segment_v2 { |
89 | | |
90 | | using namespace ErrorCode; |
91 | | |
92 | | static constexpr const char* k_segment_magic = "D0R1"; |
93 | | static constexpr uint32_t k_segment_magic_length = 4; |
94 | | |
95 | 140 | inline std::string vertical_segment_writer_mem_tracker_name(uint32_t segment_id) { |
96 | 140 | return "VerticalSegmentWriter:Segment-" + std::to_string(segment_id); |
97 | 140 | } |
98 | | |
99 | 8 | static ColumnBitmap* get_mutable_skip_bitmap_column(Block* block, size_t skip_bitmap_col_idx) { |
100 | 8 | auto skip_bitmap_column = |
101 | 8 | IColumn::mutate(std::move(block->get_by_position(skip_bitmap_col_idx).column)); |
102 | 8 | auto* skip_bitmap_column_ptr = assert_cast<ColumnBitmap*>(skip_bitmap_column.get()); |
103 | 8 | block->replace_by_position(skip_bitmap_col_idx, std::move(skip_bitmap_column)); |
104 | 8 | return skip_bitmap_column_ptr; |
105 | 8 | } |
106 | | |
107 | | VerticalSegmentWriter::VerticalSegmentWriter(io::FileWriter* file_writer, uint32_t segment_id, |
108 | | TabletSchemaSPtr tablet_schema, BaseTabletSPtr tablet, |
109 | | DataDir* data_dir, |
110 | | const VerticalSegmentWriterOptions& opts, |
111 | | IndexFileWriter* index_file_writer) |
112 | 140 | : _segment_id(segment_id), |
113 | 140 | _tablet_schema(std::move(tablet_schema)), |
114 | 140 | _tablet(std::move(tablet)), |
115 | 140 | _data_dir(data_dir), |
116 | 140 | _opts(opts), |
117 | 140 | _file_writer(file_writer), |
118 | 140 | _index_file_writer(index_file_writer), |
119 | 140 | _mem_tracker(std::make_unique<MemTracker>( |
120 | 140 | vertical_segment_writer_mem_tracker_name(segment_id))), |
121 | 140 | _key_encoder(*_tablet_schema, _is_mow()), |
122 | 140 | _mow_context(std::move(opts.mow_ctx)), |
123 | 140 | _block_aggregator(*this) { |
124 | 140 | CHECK_NOTNULL(file_writer); |
125 | 140 | _num_short_key_columns = _tablet_schema->num_short_key_columns(); |
126 | 140 | } |
127 | | |
128 | 140 | VerticalSegmentWriter::~VerticalSegmentWriter() { |
129 | 140 | _mem_tracker->release(_mem_tracker->consumption()); |
130 | 140 | } |
131 | | |
132 | | void VerticalSegmentWriter::_init_column_meta(ColumnMetaPB* meta, uint32_t column_id, |
133 | | const TabletColumn& column, |
134 | 1.99k | const ColumnWriterOptions& opts) { |
135 | 1.99k | meta->set_column_id(column_id); |
136 | 1.99k | meta->set_type(int(column.type())); |
137 | 1.99k | meta->set_length(cast_set<int32_t>(column.length())); |
138 | 1.99k | meta->set_encoding(EncodingInfo::resolve_default_encoding(opts.storage_format, column)); |
139 | 1.99k | meta->set_compression(_opts.compression_type); |
140 | 1.99k | meta->set_is_nullable(column.is_nullable()); |
141 | 1.99k | meta->set_default_value(column.default_value()); |
142 | 1.99k | meta->set_precision(column.precision()); |
143 | 1.99k | meta->set_frac(column.frac()); |
144 | 1.99k | if (column.has_path_info()) { |
145 | 32 | column.path_info_ptr()->to_protobuf(meta->mutable_column_path_info(), |
146 | 32 | column.parent_unique_id()); |
147 | 32 | } |
148 | 1.99k | meta->set_unique_id(column.unique_id()); |
149 | 2.21k | for (uint32_t i = 0; i < column.get_subtype_count(); ++i) { |
150 | 216 | _init_column_meta(meta->add_children_columns(), column_id, column.get_sub_column(i), opts); |
151 | 216 | } |
152 | 1.99k | if (column.is_variant_type()) { |
153 | 32 | meta->set_variant_max_subcolumns_count(column.variant_max_subcolumns_count()); |
154 | 32 | meta->set_variant_enable_doc_mode(column.variant_enable_doc_mode()); |
155 | 32 | } |
156 | 1.99k | meta->set_result_is_nullable(column.get_result_is_nullable()); |
157 | 1.99k | meta->set_function_name(column.get_aggregation_name()); |
158 | 1.99k | meta->set_be_exec_version(column.get_be_exec_version()); |
159 | 1.99k | } |
160 | | |
161 | | Status VerticalSegmentWriter::_create_column_writer(uint32_t cid, const TabletColumn& column, |
162 | 1.77k | const TabletSchemaSPtr& tablet_schema) { |
163 | 1.77k | ColumnWriterOptions opts; |
164 | 1.77k | opts.meta = _footer.add_columns(); |
165 | 1.77k | opts.storage_format = tablet_schema->storage_format(); |
166 | | |
167 | 1.77k | _init_column_meta(opts.meta, cid, column, opts); |
168 | | |
169 | | // now we create zone map for key columns in AGG_KEYS or all column in UNIQUE_KEYS or DUP_KEYS |
170 | | // except for columns whose type don't support zone map. |
171 | 1.77k | opts.need_zone_map = column.is_key() || tablet_schema->keys_type() != KeysType::AGG_KEYS; |
172 | 1.77k | opts.need_bloom_filter = column.is_bf_column(); |
173 | 1.77k | if (opts.need_bloom_filter) { |
174 | 24 | opts.bf_options.fpp = |
175 | 24 | tablet_schema->has_bf_fpp() ? tablet_schema->bloom_filter_fpp() : 0.05; |
176 | 24 | } |
177 | 1.77k | auto* tablet_index = tablet_schema->get_ngram_bf_index(column.unique_id()); |
178 | 1.77k | if (tablet_index) { |
179 | 8 | opts.need_bloom_filter = true; |
180 | 8 | opts.is_ngram_bf_index = true; |
181 | | //narrow convert from int32_t to uint8_t and uint16_t which is dangerous |
182 | 8 | auto gram_size = tablet_index->get_gram_size(); |
183 | 8 | auto gram_bf_size = tablet_index->get_gram_bf_size(); |
184 | 8 | if (gram_size > 256 || gram_size < 1) { |
185 | 0 | return Status::NotSupported("Do not support ngram bloom filter for ngram_size: ", |
186 | 0 | gram_size); |
187 | 0 | } |
188 | 8 | if (gram_bf_size > 65535 || gram_bf_size < 64) { |
189 | 0 | return Status::NotSupported("Do not support ngram bloom filter for bf_size: ", |
190 | 0 | gram_bf_size); |
191 | 0 | } |
192 | 8 | opts.gram_size = cast_set<uint8_t>(gram_size); |
193 | 8 | opts.gram_bf_size = cast_set<uint16_t>(gram_bf_size); |
194 | 8 | } |
195 | | |
196 | 1.77k | bool skip_inverted_index = false; |
197 | 1.77k | if (_opts.rowset_ctx != nullptr) { |
198 | | // skip write inverted index for index compaction column |
199 | 1.77k | skip_inverted_index = |
200 | 1.77k | _opts.rowset_ctx->columns_to_do_index_compaction.contains(column.unique_id()); |
201 | 1.77k | } |
202 | | // skip write inverted index on load if skip_write_index_on_load is true |
203 | 1.77k | if (_opts.write_type == DataWriteType::TYPE_DIRECT && |
204 | 1.77k | tablet_schema->skip_write_index_on_load()) { |
205 | 0 | skip_inverted_index = true; |
206 | 0 | } |
207 | 1.77k | if (!skip_inverted_index) { |
208 | 1.77k | auto inverted_indexs = tablet_schema->inverted_indexs(column); |
209 | 1.77k | if (!inverted_indexs.empty()) { |
210 | 14 | opts.inverted_indexes = inverted_indexs; |
211 | 14 | opts.need_inverted_index = true; |
212 | 14 | DCHECK(_index_file_writer != nullptr); |
213 | 14 | } |
214 | 1.77k | } |
215 | 1.77k | opts.index_file_writer = _index_file_writer; |
216 | | |
217 | 1.77k | if (const auto& index = tablet_schema->ann_index(column); index != nullptr) { |
218 | 8 | opts.ann_index = index; |
219 | 8 | opts.need_ann_index = true; |
220 | 8 | DCHECK(_index_file_writer != nullptr); |
221 | 8 | opts.index_file_writer = _index_file_writer; |
222 | 8 | } |
223 | | |
224 | 1.77k | #define DISABLE_INDEX_IF_FIELD_TYPE(TYPE) \ |
225 | 16.0k | if (column.type() == FieldType::OLAP_FIELD_TYPE_##TYPE) { \ |
226 | 232 | opts.need_zone_map = false; \ |
227 | 232 | opts.need_bloom_filter = false; \ |
228 | 232 | } |
229 | | |
230 | 1.77k | DISABLE_INDEX_IF_FIELD_TYPE(STRUCT) |
231 | 1.77k | DISABLE_INDEX_IF_FIELD_TYPE(ARRAY) |
232 | 1.77k | DISABLE_INDEX_IF_FIELD_TYPE(JSONB) |
233 | 1.77k | DISABLE_INDEX_IF_FIELD_TYPE(AGG_STATE) |
234 | 1.77k | DISABLE_INDEX_IF_FIELD_TYPE(MAP) |
235 | 1.77k | DISABLE_INDEX_IF_FIELD_TYPE(BITMAP) |
236 | 1.77k | DISABLE_INDEX_IF_FIELD_TYPE(HLL) |
237 | 1.77k | DISABLE_INDEX_IF_FIELD_TYPE(QUANTILE_STATE) |
238 | 1.77k | DISABLE_INDEX_IF_FIELD_TYPE(VARIANT) |
239 | | |
240 | 1.77k | #undef DISABLE_INDEX_IF_FIELD_TYPE |
241 | | |
242 | 1.77k | #undef CHECK_FIELD_TYPE |
243 | | |
244 | 1.77k | int64_t storage_page_size = _tablet_schema->storage_page_size(); |
245 | | // storage_page_size must be between 4KB and 10MB. |
246 | 1.77k | if (storage_page_size >= 4096 && storage_page_size <= 10485760) { |
247 | 1.77k | opts.data_page_size = storage_page_size; |
248 | 1.77k | } |
249 | 1.77k | opts.dict_page_size = _tablet_schema->storage_dict_page_size(); |
250 | 1.77k | DBUG_EXECUTE_IF("VerticalSegmentWriter._create_column_writer.storage_page_size", { |
251 | 1.77k | auto table_id = DebugPoints::instance()->get_debug_param_or_default<int64_t>( |
252 | 1.77k | "VerticalSegmentWriter._create_column_writer.storage_page_size", "table_id", |
253 | 1.77k | INT_MIN); |
254 | 1.77k | auto target_data_page_size = DebugPoints::instance()->get_debug_param_or_default<int64_t>( |
255 | 1.77k | "VerticalSegmentWriter._create_column_writer.storage_page_size", |
256 | 1.77k | "storage_page_size", INT_MIN); |
257 | 1.77k | if (table_id == INT_MIN || target_data_page_size == INT_MIN) { |
258 | 1.77k | return Status::Error<ErrorCode::INTERNAL_ERROR>( |
259 | 1.77k | "Debug point parameters missing: either 'table_id' or 'storage_page_size' not " |
260 | 1.77k | "set."); |
261 | 1.77k | } |
262 | 1.77k | if (table_id == _tablet_schema->table_id() && |
263 | 1.77k | opts.data_page_size != target_data_page_size) { |
264 | 1.77k | return Status::Error<ErrorCode::INTERNAL_ERROR>( |
265 | 1.77k | "Mismatch in 'storage_page_size': expected size does not match the current " |
266 | 1.77k | "data page size. " |
267 | 1.77k | "Expected: " + |
268 | 1.77k | std::to_string(target_data_page_size) + |
269 | 1.77k | ", Actual: " + std::to_string(opts.data_page_size) + "."); |
270 | 1.77k | } |
271 | 1.77k | }) |
272 | 1.77k | if (column.is_row_store_column()) { |
273 | | // smaller page size for row store column; encoding is already set to PLAIN / |
274 | | // PLAIN_V2 by _init_column_meta via resolve_default_encoding(). |
275 | 28 | auto page_size = _tablet_schema->row_store_page_size(); |
276 | 28 | opts.data_page_size = |
277 | 28 | (page_size > 0) ? page_size : segment_v2::ROW_STORE_PAGE_SIZE_DEFAULT_VALUE; |
278 | 28 | } |
279 | | |
280 | 1.77k | opts.rowset_ctx = _opts.rowset_ctx; |
281 | 1.77k | opts.file_writer = _file_writer; |
282 | 1.77k | opts.compression_type = _opts.compression_type; |
283 | 1.77k | opts.footer = &_footer; |
284 | 1.77k | opts.input_rs_readers = _opts.rowset_ctx->input_rs_readers; |
285 | | |
286 | 1.77k | std::unique_ptr<ColumnWriter> writer; |
287 | 1.77k | RETURN_IF_ERROR(ColumnWriter::create(opts, &column, _file_writer, &writer)); |
288 | 1.77k | RETURN_IF_ERROR(writer->init()); |
289 | 1.77k | _column_writers[cid] = std::move(writer); |
290 | 1.77k | _olap_data_convertor->add_column_data_convertor_at(column, cid); |
291 | 1.77k | return Status::OK(); |
292 | 1.77k | }; |
293 | | |
294 | 137 | Status VerticalSegmentWriter::init() { |
295 | 137 | DCHECK(_column_writers.empty()); |
296 | 137 | if (_opts.compression_type == UNKNOWN_COMPRESSION) { |
297 | 112 | _opts.compression_type = _tablet_schema->compression_type(); |
298 | 112 | } |
299 | 137 | _olap_data_convertor = std::make_unique<OlapBlockDataConvertor>(); |
300 | 137 | _olap_data_convertor->resize(_tablet_schema->num_columns()); |
301 | 137 | _column_writers.resize(_tablet_schema->num_columns()); |
302 | | // we don't need the short key index for unique key merge on write table. |
303 | 137 | if (_is_mow()) { |
304 | 46 | size_t seq_col_length = 0; |
305 | 46 | if (_tablet_schema->has_sequence_col()) { |
306 | 23 | seq_col_length = |
307 | 23 | _tablet_schema->column(_tablet_schema->sequence_col_idx()).length() + 1; |
308 | 23 | } |
309 | 46 | size_t rowid_length = 0; |
310 | 46 | if (_is_mow_with_cluster_key()) { |
311 | 9 | rowid_length = PrimaryKeyIndexReader::ROW_ID_LENGTH; |
312 | 9 | _short_key_index_builder.reset( |
313 | 9 | new ShortKeyIndexBuilder(_segment_id, _opts.num_rows_per_block)); |
314 | 9 | } |
315 | 46 | _primary_key_index_builder.reset( |
316 | 46 | new PrimaryKeyIndexBuilder(_file_writer, seq_col_length, rowid_length)); |
317 | 46 | RETURN_IF_ERROR(_primary_key_index_builder->init()); |
318 | 91 | } else { |
319 | 91 | _short_key_index_builder.reset( |
320 | 91 | new ShortKeyIndexBuilder(_segment_id, _opts.num_rows_per_block)); |
321 | 91 | } |
322 | 137 | return Status::OK(); |
323 | 137 | } |
324 | | |
325 | | Status VerticalSegmentWriter::_append_row_store_column(const Block& block, size_t row_pos, |
326 | 8 | size_t num_rows, uint32_t cid) { |
327 | 8 | DCHECK(_tablet_schema->column(cid).is_row_store_column()); |
328 | 8 | if (num_rows == 0) { |
329 | 0 | return Status::OK(); |
330 | 0 | } |
331 | 8 | DCHECK_LE(row_pos + num_rows, block.rows()); |
332 | | |
333 | 8 | auto serdes = create_data_type_serdes(block.get_data_types()); |
334 | 8 | std::unordered_set<int32_t> row_store_cids_set(_tablet_schema->row_columns_uids().begin(), |
335 | 8 | _tablet_schema->row_columns_uids().end()); |
336 | 8 | size_t end_pos = row_pos + num_rows; |
337 | 8 | size_t batch_rows = _opts.num_rows_per_block; |
338 | 8 | static constexpr size_t kRowStoreBatchBytes = 4 * 1024 * 1024; |
339 | 8 | DCHECK_GT(batch_rows, 0); |
340 | 16 | for (size_t pos = row_pos; pos < end_pos;) { |
341 | 8 | size_t max_rows = std::min(batch_rows, end_pos - pos); |
342 | 8 | auto row_column = ColumnString::create(); |
343 | 8 | auto* row_store_column = row_column.get(); |
344 | 8 | size_t rows = JsonbSerializeUtil::block_to_jsonb( |
345 | 8 | *_tablet_schema, block, *row_store_column, |
346 | 8 | cast_set<int>(_tablet_schema->num_columns()), serdes, row_store_cids_set, pos, |
347 | 8 | max_rows, kRowStoreBatchBytes); |
348 | 8 | DCHECK_GT(rows, 0); |
349 | | |
350 | 8 | auto typed_column = block.get_by_position(cid); |
351 | 8 | typed_column.column = std::move(row_column); |
352 | 8 | RETURN_IF_ERROR(_olap_data_convertor->set_source_content_with_specifid_column( |
353 | 8 | typed_column, 0, rows, cid)); |
354 | 8 | auto [status, column] = _olap_data_convertor->convert_column_data(cid); |
355 | 8 | RETURN_IF_ERROR(status); |
356 | 8 | RETURN_IF_ERROR( |
357 | 8 | _column_writers[cid]->append(column->get_nullmap(), column->get_data(), rows)); |
358 | 8 | _olap_data_convertor->clear_source_content(cid); |
359 | 8 | pos += rows; |
360 | 8 | } |
361 | 8 | return Status::OK(); |
362 | 8 | } |
363 | | |
364 | | Status VerticalSegmentWriter::_append_generated_column(const DerivedColumnGenerator& generator, |
365 | | const Block& block, size_t row_pos, |
366 | 12 | size_t num_rows, uint32_t cid) { |
367 | 12 | if (num_rows == 0) { |
368 | 0 | return Status::OK(); |
369 | 0 | } |
370 | 12 | DCHECK_LE(row_pos + num_rows, block.rows()); |
371 | | |
372 | 12 | size_t end_pos = row_pos + num_rows; |
373 | 12 | size_t batch_rows = _opts.num_rows_per_block; |
374 | 12 | static constexpr size_t kDerivedColumnBatchBytes = 4 * 1024 * 1024; |
375 | 12 | DCHECK_GT(batch_rows, 0); |
376 | 24 | for (size_t pos = row_pos; pos < end_pos;) { |
377 | 12 | size_t max_rows = std::min(batch_rows, end_pos - pos); |
378 | 12 | auto generated_column = block.get_by_position(cid).column->clone_empty(); |
379 | 12 | size_t rows = generator.generate(block, pos, max_rows, kDerivedColumnBatchBytes, |
380 | 12 | generated_column.get()); |
381 | 12 | DCHECK_GT(rows, 0); |
382 | | |
383 | 12 | auto typed_column = block.get_by_position(cid); |
384 | 12 | typed_column.column = std::move(generated_column); |
385 | 12 | RETURN_IF_ERROR(_olap_data_convertor->set_source_content_with_specifid_column( |
386 | 12 | typed_column, 0, rows, cid)); |
387 | 12 | auto [status, column] = _olap_data_convertor->convert_column_data(cid); |
388 | 12 | RETURN_IF_ERROR(status); |
389 | 12 | RETURN_IF_ERROR( |
390 | 12 | _column_writers[cid]->append(column->get_nullmap(), column->get_data(), rows)); |
391 | 12 | _olap_data_convertor->clear_source_content(cid); |
392 | 12 | pos += rows; |
393 | 12 | } |
394 | 12 | return Status::OK(); |
395 | 12 | } |
396 | | |
397 | | Status VerticalSegmentWriter::_probe_key_for_mow( |
398 | | const MowKeyProbe& probe, std::string key, std::size_t segment_pos, |
399 | | bool have_input_seq_column, bool have_delete_sign, |
400 | | const std::vector<RowsetSharedPtr>& specified_rowsets, |
401 | | std::vector<std::unique_ptr<SegmentCacheHandle>>& segment_caches, |
402 | | bool& has_default_or_nullable, std::vector<bool>& use_default_or_null_flag, |
403 | | const std::function<void(const RowLocation& loc, const RowsetSharedPtr& rowset)>& found_cb, |
404 | 60 | const std::function<Status()>& not_found_cb, PartialUpdateStats& stats) { |
405 | 60 | ProbeOutcome outcome = |
406 | 60 | DORIS_TRY(probe.probe(key, segment_pos, have_input_seq_column, have_delete_sign, |
407 | 60 | specified_rowsets, segment_caches, stats)); |
408 | 60 | if (outcome.result == KeyProbeResult::NOT_FOUND) { |
409 | 20 | if (!have_delete_sign) { |
410 | 20 | RETURN_IF_ERROR(not_found_cb()); |
411 | 20 | } |
412 | 20 | has_default_or_nullable = true; |
413 | 20 | use_default_or_null_flag.emplace_back(true); |
414 | 20 | return Status::OK(); |
415 | 20 | } |
416 | 40 | if (outcome.use_default_or_null) { |
417 | 8 | has_default_or_nullable = true; |
418 | 8 | use_default_or_null_flag.emplace_back(true); |
419 | 32 | } else { |
420 | | // partial update should not contain invisible columns |
421 | 32 | use_default_or_null_flag.emplace_back(false); |
422 | 32 | found_cb(outcome.loc, outcome.rowset); |
423 | 32 | } |
424 | 40 | return Status::OK(); |
425 | 60 | } |
426 | | |
427 | 1.26k | Status VerticalSegmentWriter::_check_column_writer_disk_capacity(size_t cid) { |
428 | 1.26k | if (_data_dir != nullptr && |
429 | 1.26k | _data_dir->reach_capacity_limit(_column_writers[cid]->estimate_buffer_size())) { |
430 | 0 | return Status::Error<DISK_REACH_CAPACITY_LIMIT>("disk {} exceed capacity limit.", |
431 | 0 | _data_dir->path_hash()); |
432 | 0 | } |
433 | 1.26k | return Status::OK(); |
434 | 1.26k | } |
435 | | |
436 | 1.77k | Status VerticalSegmentWriter::_finalize_column_writer_and_update_meta(size_t cid) { |
437 | 1.77k | RETURN_IF_ERROR(_column_writers[cid]->finish()); |
438 | 1.77k | RETURN_IF_ERROR(_column_writers[cid]->write_data()); |
439 | | |
440 | 1.77k | auto* column_meta = _column_writers[cid]->get_column_meta(); |
441 | 1.77k | column_meta->set_compressed_data_bytes( |
442 | 1.77k | _column_writers[cid]->get_total_compressed_data_pages_bytes()); |
443 | 1.77k | column_meta->set_uncompressed_data_bytes( |
444 | 1.77k | _column_writers[cid]->get_total_uncompressed_data_pages_bytes()); |
445 | 1.77k | column_meta->set_raw_data_bytes(_column_writers[cid]->get_raw_data_bytes()); |
446 | 1.77k | return Status::OK(); |
447 | 1.77k | } |
448 | | |
449 | | Status VerticalSegmentWriter::_partial_update_preconditions_check(size_t row_pos, |
450 | 20 | bool is_flexible_update) { |
451 | 20 | if (!_is_mow()) { |
452 | 0 | auto msg = fmt::format( |
453 | 0 | "Can only do partial update on merge-on-write unique table, but found: " |
454 | 0 | "keys_type={}, _opts.enable_unique_key_merge_on_write={}, tablet_id={}", |
455 | 0 | _tablet_schema->keys_type(), _opts.enable_unique_key_merge_on_write, |
456 | 0 | _tablet->tablet_id()); |
457 | 0 | DCHECK(false) << msg; |
458 | 0 | return Status::InternalError<false>(msg); |
459 | 0 | } |
460 | 20 | if (_opts.rowset_ctx->partial_update_info == nullptr) { |
461 | 0 | auto msg = |
462 | 0 | fmt::format("partial_update_info should not be nullptr, please check, tablet_id={}", |
463 | 0 | _tablet->tablet_id()); |
464 | 0 | DCHECK(false) << msg; |
465 | 0 | return Status::InternalError<false>(msg); |
466 | 0 | } |
467 | 20 | if (!is_flexible_update) { |
468 | 12 | if (!_opts.rowset_ctx->partial_update_info->is_fixed_partial_update()) { |
469 | 0 | auto msg = fmt::format( |
470 | 0 | "in fixed partial update code, but update_mode={}, please check, tablet_id={}", |
471 | 0 | _opts.rowset_ctx->partial_update_info->update_mode(), _tablet->tablet_id()); |
472 | 0 | DCHECK(false) << msg; |
473 | 0 | return Status::InternalError<false>(msg); |
474 | 0 | } |
475 | 12 | } else { |
476 | 8 | if (!_opts.rowset_ctx->partial_update_info->is_flexible_partial_update()) { |
477 | 0 | auto msg = fmt::format( |
478 | 0 | "in flexible partial update code, but update_mode={}, please check, " |
479 | 0 | "tablet_id={}", |
480 | 0 | _opts.rowset_ctx->partial_update_info->update_mode(), _tablet->tablet_id()); |
481 | 0 | DCHECK(false) << msg; |
482 | 0 | return Status::InternalError<false>(msg); |
483 | 0 | } |
484 | 8 | } |
485 | 20 | if (row_pos != 0) { |
486 | 0 | auto msg = fmt::format("row_pos should be 0, but found {}, tablet_id={}", row_pos, |
487 | 0 | _tablet->tablet_id()); |
488 | 0 | DCHECK(false) << msg; |
489 | 0 | return Status::InternalError<false>(msg); |
490 | 0 | } |
491 | 20 | return Status::OK(); |
492 | 20 | } |
493 | | |
494 | | // for partial update, we should do following steps to fill content of block: |
495 | | // 1. set block data to data convertor, and get all key_column's converted slice |
496 | | // 2. get pk of input block, and read missing columns |
497 | | // 2.1 first find key location{rowset_id, segment_id, row_id} |
498 | | // 2.2 build read plan to read by batch |
499 | | // 2.3 fill block |
500 | | // 3. set columns to data convertor and then write all columns |
501 | | Status VerticalSegmentWriter::_append_block_with_partial_content(RowsInBlock& data, |
502 | 12 | Block& full_block) { |
503 | 12 | DBUG_EXECUTE_IF("_append_block_with_partial_content.block", DBUG_BLOCK); |
504 | | |
505 | 12 | RETURN_IF_ERROR(_partial_update_preconditions_check(data.row_pos, false)); |
506 | | // create full block and fill with input columns |
507 | 12 | full_block = _tablet_schema->create_block(); |
508 | 12 | const auto& including_cids = _opts.rowset_ctx->partial_update_info->update_cids; |
509 | 12 | size_t input_id = 0; |
510 | 260 | for (auto i : including_cids) { |
511 | 260 | full_block.replace_by_position(i, data.block->get_by_position(input_id++).column); |
512 | 260 | } |
513 | | |
514 | 12 | if (_opts.rowset_ctx->write_type != DataWriteType::TYPE_COMPACTION && |
515 | 12 | _tablet_schema->num_variant_columns() > 0) { |
516 | 4 | RETURN_IF_ERROR(variant_util::parse_and_materialize_variant_columns( |
517 | 4 | full_block, *_tablet_schema, including_cids)); |
518 | 4 | } |
519 | 12 | bool have_input_seq_column = false; |
520 | | // write including columns |
521 | 12 | std::vector<IOlapColumnDataAccessor*> key_columns; |
522 | 12 | IOlapColumnDataAccessor* seq_column = nullptr; |
523 | 12 | uint32_t segment_start_pos = 0; |
524 | 260 | for (auto cid : including_cids) { |
525 | 260 | RETURN_IF_ERROR(_create_column_writer(cid, _tablet_schema->column(cid), _tablet_schema)); |
526 | 260 | RETURN_IF_ERROR(_olap_data_convertor->set_source_content_with_specifid_columns( |
527 | 260 | &full_block, data.row_pos, data.num_rows, std::vector<uint32_t> {cid})); |
528 | | // here we get segment column row num before append data. |
529 | 260 | segment_start_pos = cast_set<uint32_t>(_column_writers[cid]->get_next_rowid()); |
530 | | // olap data convertor alway start from id = 0 |
531 | 260 | auto [status, column] = _olap_data_convertor->convert_column_data(cid); |
532 | 260 | if (!status.ok()) { |
533 | 0 | return status; |
534 | 0 | } |
535 | 260 | if (cid < _key_encoder.num_sort_key_columns()) { |
536 | 240 | key_columns.push_back(column); |
537 | 240 | } else if (_tablet_schema->has_sequence_col() && |
538 | 20 | cid == _tablet_schema->sequence_col_idx()) { |
539 | 4 | seq_column = column; |
540 | 4 | have_input_seq_column = true; |
541 | 4 | } |
542 | 260 | RETURN_IF_ERROR(_column_writers[cid]->append(column->get_nullmap(), column->get_data(), |
543 | 260 | data.num_rows)); |
544 | 260 | RETURN_IF_ERROR(_finalize_column_writer_and_update_meta(cid)); |
545 | | // Don't clear source content for key columns and sequence column here, |
546 | | // as they will be used later for key encoding and _generate_primary_key_index(). |
547 | | // They will be cleared at the end of this method. |
548 | 260 | bool is_key_column = (cid < _key_encoder.num_sort_key_columns()); |
549 | 260 | bool is_seq_column = (_tablet_schema->has_sequence_col() && |
550 | 260 | cid == _tablet_schema->sequence_col_idx() && have_input_seq_column); |
551 | 260 | if (!is_key_column && !is_seq_column) { |
552 | 16 | _olap_data_convertor->clear_source_content(cid); |
553 | 16 | } |
554 | 260 | } |
555 | | |
556 | 12 | bool has_default_or_nullable = false; |
557 | 12 | std::vector<bool> use_default_or_null_flag; |
558 | 12 | use_default_or_null_flag.reserve(data.num_rows); |
559 | 12 | const auto* delete_signs = |
560 | 12 | BaseTablet::get_delete_sign_column_data(full_block, data.row_pos + data.num_rows); |
561 | | |
562 | 12 | DBUG_EXECUTE_IF("VerticalSegmentWriter._append_block_with_partial_content.sleep", |
563 | 12 | { sleep(60); }) |
564 | 12 | const std::vector<RowsetSharedPtr>& specified_rowsets = _mow_context->rowset_ptrs; |
565 | 12 | std::vector<std::unique_ptr<SegmentCacheHandle>> segment_caches(specified_rowsets.size()); |
566 | | |
567 | 12 | MowKeyProbe probe = MowKeyProbe::for_partial_update( |
568 | 12 | _tablet.get(), _tablet_schema.get(), _tablet_schema->has_sequence_col(), _mow_context, |
569 | 12 | _opts.rowset_ctx->rowset_id, _segment_id, /*flexible=*/false); |
570 | | // owns the rowset pins and the read plan for the historical read below |
571 | 12 | HistoricalRowFetcher fetcher {_opts.rowset_ctx->make_historical_row_retriever_context()}; |
572 | | |
573 | | // locate rows in base data |
574 | 12 | PartialUpdateStats stats; |
575 | | |
576 | 48 | for (size_t block_pos = data.row_pos; block_pos < data.row_pos + data.num_rows; block_pos++) { |
577 | | // block segment |
578 | | // 2 -> 0 |
579 | | // 3 -> 1 |
580 | | // 4 -> 2 |
581 | | // 5 -> 3 |
582 | | // here row_pos = 2, num_rows = 4. |
583 | 36 | size_t delta_pos = block_pos - data.row_pos; |
584 | 36 | size_t segment_pos = segment_start_pos + delta_pos; |
585 | 36 | std::string key = encode_mow_key_invalidate_cache( |
586 | 36 | _key_encoder, key_columns, seq_column, delta_pos, have_input_seq_column, |
587 | 36 | _opts.rowset_ctx->tablet_id, *_tablet_schema, _opts.write_type); |
588 | | // If the table have sequence column, and the include-cids don't contain the sequence |
589 | | // column, we need to update the primary key index builder at the end of this method. |
590 | | // At that time, we have a valid sequence column to encode the key with seq col. |
591 | 36 | if (!_tablet_schema->has_sequence_col() || have_input_seq_column) { |
592 | 24 | RETURN_IF_ERROR(_primary_key_index_builder->add_item(key)); |
593 | 24 | } |
594 | | |
595 | | // mark key with delete sign as deleted. |
596 | 36 | bool have_delete_sign = (delete_signs != nullptr && delete_signs[block_pos] != 0); |
597 | | |
598 | 36 | auto not_found_cb = [&]() { |
599 | 12 | return _opts.rowset_ctx->partial_update_info->handle_new_key( |
600 | 12 | *_tablet_schema, [&]() -> std::string { |
601 | 0 | return data.block->dump_one_line( |
602 | 0 | block_pos, cast_set<int>(_key_encoder.num_sort_key_columns())); |
603 | 0 | }); |
604 | 12 | }; |
605 | 36 | auto update_read_plan = [&](const RowLocation& loc, const RowsetSharedPtr& rowset) { |
606 | | // keep the rowset alive until the historical read below is done |
607 | 24 | fetcher.pin_rowset(rowset); |
608 | 24 | fetcher.plan_fixed_read(loc, segment_pos); |
609 | 24 | }; |
610 | 36 | RETURN_IF_ERROR(_probe_key_for_mow( |
611 | 36 | probe, std::move(key), segment_pos, have_input_seq_column, have_delete_sign, |
612 | 36 | specified_rowsets, segment_caches, has_default_or_nullable, |
613 | 36 | use_default_or_null_flag, update_read_plan, not_found_cb, stats)); |
614 | 36 | } |
615 | 12 | CHECK_EQ(use_default_or_null_flag.size(), data.num_rows); |
616 | | |
617 | 12 | if (config::enable_merge_on_write_correctness_check) { |
618 | 12 | _tablet->add_sentinel_mark_to_delete_bitmap(_mow_context->delete_bitmap.get(), |
619 | 12 | *_mow_context->rowset_ids); |
620 | 12 | } |
621 | | |
622 | | // read to fill full_block |
623 | 12 | RETURN_IF_ERROR(fetcher.fill_missing_columns(*_tablet_schema, full_block, |
624 | 12 | use_default_or_null_flag, has_default_or_nullable, |
625 | 12 | segment_start_pos, data.block)); |
626 | | |
627 | 12 | if (_tablet_schema->num_variant_columns() > 0) { |
628 | 4 | RETURN_IF_ERROR(variant_util::parse_and_materialize_variant_columns( |
629 | 4 | full_block, *_tablet_schema, _opts.rowset_ctx->partial_update_info->missing_cids)); |
630 | 4 | } |
631 | | |
632 | | // convert missing columns and send to column writer |
633 | 12 | const auto& missing_cids = _opts.rowset_ctx->partial_update_info->missing_cids; |
634 | 48 | for (auto cid : missing_cids) { |
635 | 48 | RETURN_IF_ERROR(_create_column_writer(cid, _tablet_schema->column(cid), _tablet_schema)); |
636 | 48 | if (_tablet_schema->column(cid).is_row_store_column()) { |
637 | 4 | RETURN_IF_ERROR(_append_row_store_column(full_block, data.row_pos, data.num_rows, cid)); |
638 | 4 | RETURN_IF_ERROR(_finalize_column_writer_and_update_meta(cid)); |
639 | 4 | continue; |
640 | 4 | } |
641 | 44 | RETURN_IF_ERROR(_olap_data_convertor->set_source_content_with_specifid_columns( |
642 | 44 | &full_block, data.row_pos, data.num_rows, std::vector<uint32_t> {cid})); |
643 | 44 | auto [status, column] = _olap_data_convertor->convert_column_data(cid); |
644 | 44 | if (!status.ok()) { |
645 | 0 | return status; |
646 | 0 | } |
647 | 44 | if (_tablet_schema->has_sequence_col() && !have_input_seq_column && |
648 | 44 | cid == _tablet_schema->sequence_col_idx()) { |
649 | 4 | DCHECK_EQ(seq_column, nullptr); |
650 | 4 | seq_column = column; |
651 | 4 | } |
652 | 44 | RETURN_IF_ERROR(_column_writers[cid]->append(column->get_nullmap(), column->get_data(), |
653 | 44 | data.num_rows)); |
654 | 44 | RETURN_IF_ERROR(_finalize_column_writer_and_update_meta(cid)); |
655 | | // Don't clear source content for sequence column here if it will be used later |
656 | | // in _generate_primary_key_index(). It will be cleared at the end of this method. |
657 | 44 | bool is_seq_column = (_tablet_schema->has_sequence_col() && !have_input_seq_column && |
658 | 44 | cid == _tablet_schema->sequence_col_idx()); |
659 | 44 | if (!is_seq_column) { |
660 | 40 | _olap_data_convertor->clear_source_content(cid); |
661 | 40 | } |
662 | 44 | } |
663 | | |
664 | 12 | _num_rows_updated += stats.num_rows_updated; |
665 | 12 | _num_rows_deleted += stats.num_rows_deleted; |
666 | 12 | _num_rows_new_added += stats.num_rows_new_added; |
667 | 12 | _num_rows_filtered += stats.num_rows_filtered; |
668 | 12 | if (_tablet_schema->has_sequence_col() && !have_input_seq_column) { |
669 | 4 | DCHECK_NE(seq_column, nullptr); |
670 | 4 | if (_num_rows_written != data.row_pos || |
671 | 4 | _primary_key_index_builder->num_rows() != _num_rows_written) { |
672 | 0 | return Status::InternalError( |
673 | 0 | "Correctness check failed, _num_rows_written: {}, row_pos: {}, primary key " |
674 | 0 | "index builder num rows: {}", |
675 | 0 | _num_rows_written, data.row_pos, _primary_key_index_builder->num_rows()); |
676 | 0 | } |
677 | 4 | RETURN_IF_ERROR(_generate_primary_key_index(key_columns, seq_column, data.num_rows, false)); |
678 | 4 | } |
679 | | |
680 | 12 | _num_rows_written += data.num_rows; |
681 | 12 | DCHECK_EQ(_primary_key_index_builder->num_rows(), _num_rows_written) |
682 | 0 | << "primary key index builder num rows(" << _primary_key_index_builder->num_rows() |
683 | 0 | << ") not equal to segment writer's num rows written(" << _num_rows_written << ")"; |
684 | 12 | _olap_data_convertor->clear_source_content(); |
685 | 12 | return Status::OK(); |
686 | 12 | } |
687 | | |
688 | | Status VerticalSegmentWriter::_append_block_with_flexible_partial_content(RowsInBlock& data, |
689 | 8 | Block& full_block) { |
690 | 8 | RETURN_IF_ERROR(_partial_update_preconditions_check(data.row_pos, true)); |
691 | | |
692 | | // data.block has the same schema with full_block |
693 | 8 | DCHECK(data.block->columns() == _tablet_schema->num_columns()); |
694 | | |
695 | | // create full block and fill with sort key columns |
696 | 8 | full_block = _tablet_schema->create_block(); |
697 | | |
698 | | // Use _num_rows_written instead of creating column writer 0, since all column writers |
699 | | // should have the same row count, which equals _num_rows_written. |
700 | 8 | uint32_t segment_start_pos = cast_set<uint32_t>(_num_rows_written); |
701 | | |
702 | 8 | DCHECK(_tablet_schema->has_skip_bitmap_col()); |
703 | 8 | auto skip_bitmap_col_idx = _tablet_schema->skip_bitmap_col_idx(); |
704 | | |
705 | 8 | bool has_default_or_nullable = false; |
706 | 8 | std::vector<bool> use_default_or_null_flag; |
707 | 8 | use_default_or_null_flag.reserve(data.num_rows); |
708 | | |
709 | 8 | int32_t seq_map_col_unique_id = _opts.rowset_ctx->partial_update_info->sequence_map_col_uid(); |
710 | 8 | bool schema_has_sequence_col = _tablet_schema->has_sequence_col(); |
711 | | |
712 | 8 | DBUG_EXECUTE_IF("VerticalSegmentWriter._append_block_with_flexible_partial_content.sleep", |
713 | 8 | { sleep(60); }) |
714 | 8 | const std::vector<RowsetSharedPtr>& specified_rowsets = _mow_context->rowset_ptrs; |
715 | 8 | std::vector<std::unique_ptr<SegmentCacheHandle>> segment_caches(specified_rowsets.size()); |
716 | | |
717 | | // Ensure all primary key column writers and sequence column writer are created before |
718 | | // aggregate_for_flexible_partial_update, because it internally calls convert_pk_columns |
719 | | // and convert_seq_column which need the convertors in _olap_data_convertor |
720 | 168 | for (uint32_t cid = 0; cid < _tablet_schema->num_key_columns(); ++cid) { |
721 | 160 | RETURN_IF_ERROR(_create_column_writer(cid, _tablet_schema->column(cid), _tablet_schema)); |
722 | 160 | } |
723 | 8 | if (schema_has_sequence_col) { |
724 | 4 | uint32_t cid = _tablet_schema->sequence_col_idx(); |
725 | 4 | RETURN_IF_ERROR(_create_column_writer(cid, _tablet_schema->column(cid), _tablet_schema)); |
726 | 4 | } |
727 | | |
728 | | // 1. aggregate duplicate rows in block |
729 | 8 | RETURN_IF_ERROR(_block_aggregator.aggregate_for_flexible_partial_update( |
730 | 8 | const_cast<Block*>(data.block), data.num_rows, specified_rowsets, segment_caches)); |
731 | 8 | if (data.block->rows() != data.num_rows) { |
732 | 8 | data.num_rows = data.block->rows(); |
733 | 8 | _olap_data_convertor->clear_source_content(); |
734 | 8 | } |
735 | | |
736 | | // 2. encode primary key columns |
737 | | // we can only encode primary key columns currently becasue all non-primary columns in flexible partial update |
738 | | // can have missing cells |
739 | 8 | std::vector<IOlapColumnDataAccessor*> key_columns {}; |
740 | 8 | RETURN_IF_ERROR(_block_aggregator.convert_pk_columns(const_cast<Block*>(data.block), |
741 | 8 | data.row_pos, data.num_rows, key_columns)); |
742 | | // 3. encode sequence column |
743 | | // We encode the seguence column even thought it may have invalid values in some rows because we need to |
744 | | // encode the value of sequence column in key for rows that have a valid value in sequence column during |
745 | | // lookup_raw_key. We will encode the sequence column again at the end of this method. At that time, we have |
746 | | // a valid sequence column to encode the key with seq col. |
747 | 8 | IOlapColumnDataAccessor* seq_column {nullptr}; |
748 | 8 | RETURN_IF_ERROR(_block_aggregator.convert_seq_column(const_cast<Block*>(data.block), |
749 | 8 | data.row_pos, data.num_rows, seq_column)); |
750 | | |
751 | 8 | auto* mutable_block = const_cast<Block*>(data.block); |
752 | 8 | std::vector<BitmapValue>* skip_bitmaps = |
753 | 8 | &get_mutable_skip_bitmap_column(mutable_block, skip_bitmap_col_idx)->get_data(); |
754 | 8 | const auto* delete_signs = |
755 | 8 | BaseTablet::get_delete_sign_column_data(*data.block, data.row_pos + data.num_rows); |
756 | 8 | DCHECK(delete_signs != nullptr); |
757 | | |
758 | 168 | for (std::size_t cid {0}; cid < _tablet_schema->num_key_columns(); cid++) { |
759 | 160 | full_block.replace_by_position(cid, data.block->get_by_position(cid).column); |
760 | 160 | } |
761 | | |
762 | | // 4. write primary key columns data |
763 | 168 | for (std::size_t cid {0}; cid < _tablet_schema->num_key_columns(); cid++) { |
764 | 160 | const auto& column = key_columns[cid]; |
765 | 160 | DCHECK(_column_writers[cid]->get_next_rowid() == _num_rows_written); |
766 | 160 | RETURN_IF_ERROR(_column_writers[cid]->append(column->get_nullmap(), column->get_data(), |
767 | 160 | data.num_rows)); |
768 | 160 | DCHECK(_column_writers[cid]->get_next_rowid() == _num_rows_written + data.num_rows); |
769 | 160 | RETURN_IF_ERROR(_finalize_column_writer_and_update_meta(cid)); |
770 | 160 | } |
771 | | |
772 | | // 5. genreate read plan |
773 | 8 | FlexibleReadPlan read_plan {_tablet_schema->has_row_store_for_all_columns()}; |
774 | 8 | PartialUpdateStats stats; |
775 | 8 | RETURN_IF_ERROR(_generate_flexible_read_plan( |
776 | 8 | read_plan, data, segment_start_pos, schema_has_sequence_col, seq_map_col_unique_id, |
777 | 8 | skip_bitmaps, key_columns, seq_column, delete_signs, specified_rowsets, segment_caches, |
778 | 8 | has_default_or_nullable, use_default_or_null_flag, stats)); |
779 | 8 | CHECK_EQ(use_default_or_null_flag.size(), data.num_rows); |
780 | | |
781 | 8 | if (config::enable_merge_on_write_correctness_check) { |
782 | 8 | _tablet->add_sentinel_mark_to_delete_bitmap(_mow_context->delete_bitmap.get(), |
783 | 8 | *_mow_context->rowset_ids); |
784 | 8 | } |
785 | | |
786 | | // 6. read according plan to fill full_block |
787 | 8 | RETURN_IF_ERROR(read_plan.fill_non_primary_key_columns( |
788 | 8 | _opts.rowset_ctx->make_historical_row_retriever_context(), _rsid_to_rowset, |
789 | 8 | *_tablet_schema, full_block, use_default_or_null_flag, has_default_or_nullable, |
790 | 8 | segment_start_pos, cast_set<uint32_t>(data.row_pos), data.block, skip_bitmaps)); |
791 | | |
792 | | // TODO(bobhan1): should we replace the skip bitmap column with empty bitmaps to reduce storage occupation? |
793 | | // this column is not needed in read path for merge-on-write table |
794 | | |
795 | | // 7. fill row store column |
796 | 56 | for (auto cid = _tablet_schema->num_key_columns(); cid < _tablet_schema->num_columns(); cid++) { |
797 | 48 | if (!_tablet_schema->column(cid).is_row_store_column()) { |
798 | 44 | continue; |
799 | 44 | } |
800 | 4 | RETURN_IF_ERROR(_create_column_writer(cast_set<uint32_t>(cid), _tablet_schema->column(cid), |
801 | 4 | _tablet_schema)); |
802 | 4 | RETURN_IF_ERROR(_append_row_store_column(full_block, data.row_pos, data.num_rows, |
803 | 4 | cast_set<uint32_t>(cid))); |
804 | 4 | RETURN_IF_ERROR(_finalize_column_writer_and_update_meta(cid)); |
805 | 4 | } |
806 | | |
807 | 8 | std::vector<uint32_t> column_ids; |
808 | 216 | for (uint32_t i = 0; i < _tablet_schema->num_columns(); ++i) { |
809 | 208 | column_ids.emplace_back(i); |
810 | 208 | } |
811 | 8 | if (_opts.rowset_ctx->write_type != DataWriteType::TYPE_COMPACTION && |
812 | 8 | _tablet_schema->num_variant_columns() > 0) { |
813 | 0 | RETURN_IF_ERROR(variant_util::parse_and_materialize_variant_columns( |
814 | 0 | full_block, *_tablet_schema, column_ids)); |
815 | 0 | } |
816 | | |
817 | | // 8. encode and write all non-primary key columns(including sequence column if exists) |
818 | 56 | for (auto cid = _tablet_schema->num_key_columns(); cid < _tablet_schema->num_columns(); cid++) { |
819 | 48 | if (_tablet_schema->column(cid).is_row_store_column()) { |
820 | 4 | continue; |
821 | 4 | } |
822 | 44 | if (cid != _tablet_schema->sequence_col_idx()) { |
823 | 40 | RETURN_IF_ERROR(_create_column_writer(cast_set<uint32_t>(cid), |
824 | 40 | _tablet_schema->column(cid), _tablet_schema)); |
825 | 40 | } |
826 | 44 | RETURN_IF_ERROR(_olap_data_convertor->set_source_content_with_specifid_column( |
827 | 44 | full_block.get_by_position(cid), data.row_pos, data.num_rows, |
828 | 44 | cast_set<uint32_t>(cid))); |
829 | 44 | auto [status, column] = _olap_data_convertor->convert_column_data(cid); |
830 | 44 | if (!status.ok()) { |
831 | 0 | return status; |
832 | 0 | } |
833 | 44 | if (cid == _tablet_schema->sequence_col_idx()) { |
834 | | // should use the latest encoded sequence column to build the primary index |
835 | 4 | seq_column = column; |
836 | 4 | } |
837 | 44 | DCHECK(_column_writers[cid]->get_next_rowid() == _num_rows_written); |
838 | 44 | RETURN_IF_ERROR(_column_writers[cid]->append(column->get_nullmap(), column->get_data(), |
839 | 44 | data.num_rows)); |
840 | 44 | DCHECK(_column_writers[cid]->get_next_rowid() == _num_rows_written + data.num_rows); |
841 | 44 | RETURN_IF_ERROR(_finalize_column_writer_and_update_meta(cid)); |
842 | 44 | } |
843 | | |
844 | 8 | _num_rows_updated += stats.num_rows_updated; |
845 | 8 | _num_rows_deleted += stats.num_rows_deleted; |
846 | 8 | _num_rows_new_added += stats.num_rows_new_added; |
847 | 8 | _num_rows_filtered += stats.num_rows_filtered; |
848 | | |
849 | 8 | if (_num_rows_written != data.row_pos || |
850 | 8 | _primary_key_index_builder->num_rows() != _num_rows_written) { |
851 | 0 | return Status::InternalError( |
852 | 0 | "Correctness check failed, _num_rows_written: {}, row_pos: {}, primary key " |
853 | 0 | "index builder num rows: {}", |
854 | 0 | _num_rows_written, data.row_pos, _primary_key_index_builder->num_rows()); |
855 | 0 | } |
856 | | |
857 | | // 9. build primary key index |
858 | 8 | RETURN_IF_ERROR(_generate_primary_key_index(key_columns, seq_column, data.num_rows, false)); |
859 | | |
860 | 8 | _num_rows_written += data.num_rows; |
861 | 8 | DCHECK_EQ(_primary_key_index_builder->num_rows(), _num_rows_written) |
862 | 0 | << "primary key index builder num rows(" << _primary_key_index_builder->num_rows() |
863 | 0 | << ") not equal to segment writer's num rows written(" << _num_rows_written << ")"; |
864 | 8 | _olap_data_convertor->clear_source_content(); |
865 | 8 | return Status::OK(); |
866 | 8 | } |
867 | | |
868 | | Status VerticalSegmentWriter::_generate_encoded_default_seq_value(const TabletSchema& tablet_schema, |
869 | | const PartialUpdateInfo& info, |
870 | 0 | std::string* encoded_value) { |
871 | 0 | const auto& seq_column = tablet_schema.column(tablet_schema.sequence_col_idx()); |
872 | 0 | auto block = tablet_schema.create_block_by_cids( |
873 | 0 | {cast_set<uint32_t>(tablet_schema.sequence_col_idx())}); |
874 | 0 | if (seq_column.has_default_value()) { |
875 | 0 | auto idx = tablet_schema.sequence_col_idx() - tablet_schema.num_key_columns(); |
876 | 0 | const auto& default_value = info.default_values[idx]; |
877 | 0 | StringRef str {default_value}; |
878 | 0 | RETURN_IF_ERROR(block.get_by_position(0).type->get_serde()->default_from_string( |
879 | 0 | str, *block.get_by_position(0).column->assert_mutable().get())); |
880 | |
|
881 | 0 | } else { |
882 | 0 | block.get_by_position(0).column->assert_mutable()->insert_default(); |
883 | 0 | } |
884 | 0 | DCHECK_EQ(block.rows(), 1); |
885 | 0 | auto olap_data_convertor = std::make_unique<OlapBlockDataConvertor>(); |
886 | 0 | olap_data_convertor->add_column_data_convertor(seq_column); |
887 | 0 | olap_data_convertor->set_source_content(&block, 0, 1); |
888 | 0 | auto [status, column] = olap_data_convertor->convert_column_data(0); |
889 | 0 | if (!status.ok()) { |
890 | 0 | return status; |
891 | 0 | } |
892 | | // include marker |
893 | 0 | _key_encoder.append_seq_suffix(encoded_value, column, 0); |
894 | 0 | return Status::OK(); |
895 | 0 | } |
896 | | |
897 | | Status VerticalSegmentWriter::_generate_flexible_read_plan( |
898 | | FlexibleReadPlan& read_plan, RowsInBlock& data, size_t segment_start_pos, |
899 | | bool schema_has_sequence_col, int32_t seq_map_col_unique_id, |
900 | | std::vector<BitmapValue>* skip_bitmaps, |
901 | | const std::vector<IOlapColumnDataAccessor*>& key_columns, |
902 | | IOlapColumnDataAccessor* seq_column, const signed char* delete_signs, |
903 | | const std::vector<RowsetSharedPtr>& specified_rowsets, |
904 | | std::vector<std::unique_ptr<SegmentCacheHandle>>& segment_caches, |
905 | | bool& has_default_or_nullable, std::vector<bool>& use_default_or_null_flag, |
906 | 8 | PartialUpdateStats& stats) { |
907 | 8 | int32_t delete_sign_col_unique_id = |
908 | 8 | _tablet_schema->column(_tablet_schema->delete_sign_idx()).unique_id(); |
909 | 8 | int32_t seq_col_unique_id = |
910 | 8 | (_tablet_schema->has_sequence_col() |
911 | 8 | ? _tablet_schema->column(_tablet_schema->sequence_col_idx()).unique_id() |
912 | 8 | : -1); |
913 | 8 | MowKeyProbe probe = MowKeyProbe::for_partial_update( |
914 | 8 | _tablet.get(), _tablet_schema.get(), _tablet_schema->has_sequence_col(), _mow_context, |
915 | 8 | _opts.rowset_ctx->rowset_id, _segment_id, /*flexible=*/true); |
916 | 32 | for (size_t block_pos = data.row_pos; block_pos < data.row_pos + data.num_rows; block_pos++) { |
917 | 24 | size_t delta_pos = block_pos - data.row_pos; |
918 | 24 | size_t segment_pos = segment_start_pos + delta_pos; |
919 | 24 | auto& skip_bitmap = skip_bitmaps->at(block_pos); |
920 | | |
921 | 24 | bool row_has_sequence_col = |
922 | 24 | (schema_has_sequence_col && !skip_bitmap.contains(seq_col_unique_id)); |
923 | 24 | std::string key = encode_mow_key_invalidate_cache( |
924 | 24 | _key_encoder, key_columns, seq_column, delta_pos, row_has_sequence_col, |
925 | 24 | _opts.rowset_ctx->tablet_id, *_tablet_schema, _opts.write_type); |
926 | | |
927 | | // mark key with delete sign as deleted. |
928 | 24 | bool have_delete_sign = |
929 | 24 | (!skip_bitmap.contains(delete_sign_col_unique_id) && delete_signs[block_pos] != 0); |
930 | | |
931 | 24 | auto not_found_cb = [&]() { |
932 | 8 | return _opts.rowset_ctx->partial_update_info->handle_new_key( |
933 | 8 | *_tablet_schema, |
934 | 8 | [&]() -> std::string { |
935 | 0 | return data.block->dump_one_line( |
936 | 0 | block_pos, cast_set<int>(_key_encoder.num_sort_key_columns())); |
937 | 0 | }, |
938 | 8 | &skip_bitmap); |
939 | 8 | }; |
940 | 24 | auto update_read_plan = [&](const RowLocation& loc, const RowsetSharedPtr& rowset) { |
941 | | // the flexible fill still reads through the writer's pin map, which the block |
942 | | // aggregator also feeds |
943 | 8 | _rsid_to_rowset.emplace(rowset->rowset_id(), rowset); |
944 | 8 | read_plan.prepare_to_read(loc, segment_pos, skip_bitmap); |
945 | 8 | }; |
946 | | |
947 | 24 | RETURN_IF_ERROR(_probe_key_for_mow(probe, std::move(key), segment_pos, row_has_sequence_col, |
948 | 24 | have_delete_sign, specified_rowsets, segment_caches, |
949 | 24 | has_default_or_nullable, use_default_or_null_flag, |
950 | 24 | update_read_plan, not_found_cb, stats)); |
951 | 24 | } |
952 | 8 | return Status::OK(); |
953 | 8 | } |
954 | | |
955 | 137 | Status VerticalSegmentWriter::batch_block(const Block* block, size_t row_pos, size_t num_rows) { |
956 | 137 | if (_opts.rowset_ctx->partial_update_info && |
957 | 137 | _opts.rowset_ctx->partial_update_info->is_partial_update() && |
958 | 137 | _opts.write_type == DataWriteType::TYPE_DIRECT && |
959 | 137 | !_opts.rowset_ctx->is_transient_rowset_writer) { |
960 | 20 | if (_opts.rowset_ctx->partial_update_info->is_flexible_partial_update()) { |
961 | 8 | if (block->columns() != _tablet_schema->num_columns()) { |
962 | 0 | return Status::InvalidArgument( |
963 | 0 | "illegal flexible partial update block columns, block columns = {}, " |
964 | 0 | "tablet_schema columns = {}", |
965 | 0 | block->dump_structure(), _tablet_schema->dump_structure()); |
966 | 0 | } |
967 | 12 | } else { |
968 | 12 | if (block->columns() < _tablet_schema->num_key_columns() || |
969 | 12 | block->columns() >= _tablet_schema->num_columns()) { |
970 | 0 | return Status::InvalidArgument(fmt::format( |
971 | 0 | "illegal partial update block columns: {}, num key columns: {}, total " |
972 | 0 | "schema columns: {}", |
973 | 0 | block->columns(), _tablet_schema->num_key_columns(), |
974 | 0 | _tablet_schema->num_columns())); |
975 | 0 | } |
976 | 12 | } |
977 | 117 | } else if (block->columns() != _tablet_schema->num_columns()) { |
978 | 0 | return Status::InvalidArgument( |
979 | 0 | "illegal block columns, block columns = {}, tablet_schema columns = {}", |
980 | 0 | block->dump_structure(), _tablet_schema->dump_structure()); |
981 | 0 | } |
982 | 137 | _batched_blocks.emplace_back(block, row_pos, num_rows); |
983 | 137 | return Status::OK(); |
984 | 137 | } |
985 | | |
986 | 137 | Status VerticalSegmentWriter::write_batch() { |
987 | 137 | if (_opts.rowset_ctx->partial_update_info && |
988 | 137 | _opts.rowset_ctx->partial_update_info->is_partial_update() && |
989 | 137 | _opts.write_type == DataWriteType::TYPE_DIRECT && |
990 | 137 | !_opts.rowset_ctx->is_transient_rowset_writer) { |
991 | 20 | bool is_flexible_partial_update = |
992 | 20 | _opts.rowset_ctx->partial_update_info->is_flexible_partial_update(); |
993 | 20 | Block full_block; |
994 | 20 | for (auto& data : _batched_blocks) { |
995 | 20 | if (is_flexible_partial_update) { |
996 | 8 | RETURN_IF_ERROR(_append_block_with_flexible_partial_content(data, full_block)); |
997 | 12 | } else { |
998 | 12 | RETURN_IF_ERROR(_append_block_with_partial_content(data, full_block)); |
999 | 12 | } |
1000 | 20 | } |
1001 | 20 | return Status::OK(); |
1002 | 20 | } |
1003 | | // The transform chain already validated, parsed variants and decided the derived |
1004 | | // (row-store) column; this writer only pumps the generator in bounded batches. |
1005 | 117 | if (_derived_column.second) { |
1006 | 12 | const auto& [cid, generator] = _derived_column; |
1007 | 12 | RETURN_IF_ERROR(_create_column_writer(cid, _tablet_schema->column(cid), _tablet_schema)); |
1008 | 12 | for (auto& data : _batched_blocks) { |
1009 | 12 | RETURN_IF_ERROR(_append_generated_column(*generator, *data.block, data.row_pos, |
1010 | 12 | data.num_rows, cid)); |
1011 | 12 | } |
1012 | 12 | RETURN_IF_ERROR(_check_column_writer_disk_capacity(cid)); |
1013 | 12 | RETURN_IF_ERROR(_finalize_column_writer_and_update_meta(cid)); |
1014 | 12 | } |
1015 | | |
1016 | 117 | std::vector<IOlapColumnDataAccessor*> key_columns; |
1017 | 117 | IOlapColumnDataAccessor* seq_column = nullptr; |
1018 | | // the key is cluster key column unique id |
1019 | 117 | std::map<uint32_t, IOlapColumnDataAccessor*> cid_to_column; |
1020 | 1.38k | for (uint32_t cid = 0; cid < _tablet_schema->num_columns(); ++cid) { |
1021 | 1.26k | if (_derived_column.second && _derived_column.first == cid) { |
1022 | 12 | continue; |
1023 | 12 | } |
1024 | 1.25k | RETURN_IF_ERROR(_create_column_writer(cid, _tablet_schema->column(cid), _tablet_schema)); |
1025 | 1.25k | for (auto& data : _batched_blocks) { |
1026 | 1.25k | RETURN_IF_ERROR(_olap_data_convertor->set_source_content_with_specifid_columns( |
1027 | 1.25k | data.block, data.row_pos, data.num_rows, std::vector<uint32_t> {cid})); |
1028 | | |
1029 | | // convert column data from engine format to storage layer format |
1030 | 1.25k | auto [status, column] = _olap_data_convertor->convert_column_data(cid); |
1031 | 1.25k | if (!status.ok()) { |
1032 | 0 | return status; |
1033 | 0 | } |
1034 | 1.25k | if (cid < _tablet_schema->num_key_columns()) { |
1035 | 670 | key_columns.push_back(column); |
1036 | 670 | } |
1037 | 1.25k | if (_tablet_schema->has_sequence_col() && cid == _tablet_schema->sequence_col_idx()) { |
1038 | 15 | seq_column = column; |
1039 | 15 | } |
1040 | 1.25k | auto column_unique_id = _tablet_schema->column(cid).unique_id(); |
1041 | 1.25k | if (_is_mow_with_cluster_key() && |
1042 | 1.25k | std::find(_tablet_schema->cluster_key_uids().begin(), |
1043 | 193 | _tablet_schema->cluster_key_uids().end(), |
1044 | 193 | column_unique_id) != _tablet_schema->cluster_key_uids().end()) { |
1045 | 162 | cid_to_column[column_unique_id] = column; |
1046 | 162 | } |
1047 | 1.25k | RETURN_IF_ERROR(_column_writers[cid]->append(column->get_nullmap(), column->get_data(), |
1048 | 1.25k | data.num_rows)); |
1049 | 1.25k | _olap_data_convertor->clear_source_content(); |
1050 | 1.25k | } |
1051 | 1.25k | RETURN_IF_ERROR(_check_column_writer_disk_capacity(cid)); |
1052 | 1.25k | RETURN_IF_ERROR(_finalize_column_writer_and_update_meta(cid)); |
1053 | 1.25k | } |
1054 | | |
1055 | 117 | for (auto& data : _batched_blocks) { |
1056 | 117 | _olap_data_convertor->set_source_content(data.block, data.row_pos, data.num_rows); |
1057 | 117 | RETURN_IF_ERROR(_generate_key_index(data, key_columns, seq_column, cid_to_column)); |
1058 | 117 | _olap_data_convertor->clear_source_content(); |
1059 | 117 | _num_rows_written += data.num_rows; |
1060 | 117 | } |
1061 | | |
1062 | 117 | _batched_blocks.clear(); |
1063 | | // The generator snapshots the batched blocks' rows; it must not survive them. |
1064 | 117 | _derived_column = {}; |
1065 | 117 | return Status::OK(); |
1066 | 117 | } |
1067 | | |
1068 | | Status VerticalSegmentWriter::_generate_key_index( |
1069 | | RowsInBlock& data, std::vector<IOlapColumnDataAccessor*>& key_columns, |
1070 | | IOlapColumnDataAccessor* seq_column, |
1071 | 117 | std::map<uint32_t, IOlapColumnDataAccessor*>& cid_to_column) { |
1072 | | // find all row pos for short key indexes |
1073 | 117 | std::vector<size_t> short_key_pos; |
1074 | | // We build a short key index every `_opts.num_rows_per_block` rows. Specifically, we |
1075 | | // build a short key index using 1st rows for first block and `_short_key_row_pos - _row_count` |
1076 | | // for next blocks. |
1077 | 117 | if (_short_key_row_pos == 0 && _num_rows_written == 0) { |
1078 | 117 | short_key_pos.push_back(0); |
1079 | 117 | } |
1080 | 117 | while (_short_key_row_pos + _opts.num_rows_per_block < _num_rows_written + data.num_rows) { |
1081 | 0 | _short_key_row_pos += _opts.num_rows_per_block; |
1082 | 0 | short_key_pos.push_back(_short_key_row_pos - _num_rows_written); |
1083 | 0 | } |
1084 | 117 | if (_is_mow_with_cluster_key()) { |
1085 | | // 1. generate primary key index |
1086 | 9 | RETURN_IF_ERROR(_generate_primary_key_index(key_columns, seq_column, data.num_rows, true)); |
1087 | | // 2. generate short key index (use cluster key) |
1088 | 9 | std::vector<IOlapColumnDataAccessor*> short_key_columns; |
1089 | 162 | for (const auto& cid : _tablet_schema->cluster_key_uids()) { |
1090 | 162 | short_key_columns.push_back(cid_to_column[cid]); |
1091 | 162 | } |
1092 | 9 | RETURN_IF_ERROR(_generate_short_key_index(short_key_columns, data.num_rows, short_key_pos)); |
1093 | 108 | } else if (_is_mow()) { |
1094 | 17 | RETURN_IF_ERROR(_generate_primary_key_index(key_columns, seq_column, data.num_rows, false)); |
1095 | 91 | } else { // other tables |
1096 | 91 | RETURN_IF_ERROR(_generate_short_key_index(key_columns, data.num_rows, short_key_pos)); |
1097 | 91 | } |
1098 | 117 | return Status::OK(); |
1099 | 117 | } |
1100 | | |
1101 | | Status VerticalSegmentWriter::_generate_primary_key_index( |
1102 | | const std::vector<IOlapColumnDataAccessor*>& primary_key_columns, |
1103 | 38 | IOlapColumnDataAccessor* seq_column, size_t num_rows, bool need_sort) { |
1104 | 38 | if (!need_sort) { // mow table without cluster key |
1105 | 29 | std::string last_key; |
1106 | 111 | for (size_t pos = 0; pos < num_rows; pos++) { |
1107 | 82 | std::string key = encode_mow_key_invalidate_cache( |
1108 | 82 | _key_encoder, primary_key_columns, seq_column, pos, |
1109 | 82 | _tablet_schema->has_sequence_col(), _opts.rowset_ctx->tablet_id, |
1110 | 82 | *_tablet_schema, _opts.write_type); |
1111 | 82 | DCHECK(key.compare(last_key) > 0) |
1112 | 0 | << "found duplicate key or key is not sorted! current key: " << key |
1113 | 0 | << ", last key: " << last_key; |
1114 | 82 | RETURN_IF_ERROR(_primary_key_index_builder->add_item(key)); |
1115 | 82 | last_key = std::move(key); |
1116 | 82 | } |
1117 | 29 | } else { // mow table with cluster key |
1118 | | // 1. generate primary keys in memory |
1119 | 9 | std::vector<std::string> primary_keys; |
1120 | 37 | for (uint32_t pos = 0; pos < num_rows; pos++) { |
1121 | 28 | std::string key = _key_encoder.full_encode_primary_keys(primary_key_columns, pos); |
1122 | 28 | MowKeyProbe::maybe_invalidate_row_cache(_opts.rowset_ctx->tablet_id, *_tablet_schema, |
1123 | 28 | _opts.write_type, key); |
1124 | 28 | if (_tablet_schema->has_sequence_col()) { |
1125 | 16 | _key_encoder.append_seq_suffix(&key, seq_column, pos); |
1126 | 16 | } |
1127 | 28 | _key_encoder.append_rowid_suffix(&key, pos); |
1128 | 28 | primary_keys.emplace_back(std::move(key)); |
1129 | 28 | } |
1130 | | // 2. sort primary keys |
1131 | 9 | std::sort(primary_keys.begin(), primary_keys.end()); |
1132 | | // 3. write primary keys index |
1133 | 9 | std::string last_key; |
1134 | 28 | for (const auto& key : primary_keys) { |
1135 | 28 | DCHECK(key.compare(last_key) > 0) |
1136 | 0 | << "found duplicate key or key is not sorted! current key: " << key |
1137 | 0 | << ", last key: " << last_key; |
1138 | 28 | RETURN_IF_ERROR(_primary_key_index_builder->add_item(key)); |
1139 | 28 | last_key = key; |
1140 | 28 | } |
1141 | 9 | } |
1142 | 38 | return Status::OK(); |
1143 | 38 | } |
1144 | | |
1145 | | Status VerticalSegmentWriter::_generate_short_key_index( |
1146 | | std::vector<IOlapColumnDataAccessor*>& key_columns, size_t num_rows, |
1147 | 100 | const std::vector<size_t>& short_key_pos) { |
1148 | 100 | _set_min_key(_key_encoder.full_encode(key_columns, 0)); |
1149 | 100 | _set_max_key(_key_encoder.full_encode(key_columns, num_rows - 1)); |
1150 | 100 | DCHECK(Slice(_max_key.data(), _max_key.size()) |
1151 | 0 | .compare(Slice(_min_key.data(), _min_key.size())) >= 0) |
1152 | 0 | << "key is not sorted! min key: " << _min_key << ", max key: " << _max_key; |
1153 | | |
1154 | 100 | key_columns.resize(_num_short_key_columns); |
1155 | 100 | std::string last_key; |
1156 | 100 | for (const auto pos : short_key_pos) { |
1157 | 100 | std::string key = _key_encoder.encode_short_keys(key_columns, pos); |
1158 | 100 | DCHECK(key.compare(last_key) >= 0) |
1159 | 0 | << "key is not sorted! current key: " << key << ", last key: " << last_key; |
1160 | 100 | RETURN_IF_ERROR(_short_key_index_builder->add_item(key)); |
1161 | 100 | last_key = std::move(key); |
1162 | 100 | } |
1163 | 100 | return Status::OK(); |
1164 | 100 | } |
1165 | | |
1166 | | // TODO(lingbin): Currently this function does not include the size of various indexes, |
1167 | | // We should make this more precise. |
1168 | 34 | uint64_t VerticalSegmentWriter::_estimated_remaining_size() { |
1169 | | // footer_size(4) + checksum(4) + segment_magic(4) |
1170 | 34 | uint64_t size = 12; |
1171 | 34 | if (_is_mow_with_cluster_key()) { |
1172 | 1 | size += _primary_key_index_builder->size() + _short_key_index_builder->size(); |
1173 | 33 | } else if (_is_mow()) { |
1174 | 24 | size += _primary_key_index_builder->size(); |
1175 | 24 | } else { |
1176 | 9 | size += _short_key_index_builder->size(); |
1177 | 9 | } |
1178 | | |
1179 | | // update the mem_tracker of segment size |
1180 | 34 | _mem_tracker->consume(size - _mem_tracker->consumption()); |
1181 | 34 | return size; |
1182 | 34 | } |
1183 | | |
1184 | 137 | Status VerticalSegmentWriter::finalize_columns_index(uint64_t* index_size) { |
1185 | 137 | uint64_t index_start = _file_writer->bytes_appended(); |
1186 | | // Record the common index range for cloud index-only file-cache preload. |
1187 | | // This VerticalSegmentWriter path is used when cloud load, compaction, or schema change flushes |
1188 | | // a whole block through SegmentCreator with enable_vertical_segment_writer enabled. |
1189 | 137 | RETURN_IF_ERROR(_write_ordinal_index()); |
1190 | 137 | RETURN_IF_ERROR(_write_zone_map()); |
1191 | 137 | RETURN_IF_ERROR(_write_inverted_index()); |
1192 | 137 | RETURN_IF_ERROR(_write_ann_index()); |
1193 | 137 | RETURN_IF_ERROR(_write_bloom_filter_index()); |
1194 | | |
1195 | 137 | *index_size = _file_writer->bytes_appended() - index_start; |
1196 | 137 | if (_is_mow_with_cluster_key()) { |
1197 | 9 | RETURN_IF_ERROR(_write_short_key_index()); |
1198 | 9 | *index_size = _file_writer->bytes_appended() - index_start; |
1199 | 9 | RETURN_IF_ERROR(_write_primary_key_index()); |
1200 | 9 | *index_size += _primary_key_index_builder->disk_size(); |
1201 | 128 | } else if (_is_mow()) { |
1202 | 37 | RETURN_IF_ERROR(_write_primary_key_index()); |
1203 | | // IndexedColumnWriter write data pages mixed with segment data, we should use |
1204 | | // the stat from primary key index builder. |
1205 | 37 | *index_size += _primary_key_index_builder->disk_size(); |
1206 | 91 | } else { |
1207 | 91 | RETURN_IF_ERROR(_write_short_key_index()); |
1208 | 91 | *index_size = _file_writer->bytes_appended() - index_start; |
1209 | 91 | } |
1210 | 137 | uint64_t file_index_end = _file_writer->bytes_appended(); |
1211 | 137 | _index_file_cache_info.add_index_range(index_start, file_index_end - index_start); |
1212 | | |
1213 | | // reset all column writers and data_conveter |
1214 | 137 | clear(); |
1215 | | |
1216 | 137 | return Status::OK(); |
1217 | 137 | } |
1218 | | |
1219 | | Status VerticalSegmentWriter::finalize_footer(uint64_t* segment_file_size, |
1220 | 137 | SegmentIndexFileCacheInfo* index_file_cache_info) { |
1221 | 137 | uint64_t footer_start = _file_writer->bytes_appended(); |
1222 | 137 | RETURN_IF_ERROR(_write_footer()); |
1223 | | // finish |
1224 | 137 | RETURN_IF_ERROR(_file_writer->close(true)); |
1225 | 137 | *segment_file_size = _file_writer->bytes_appended(); |
1226 | | // The closed size completes the preload range recorded above. SegmentIndexFileCacheLoader |
1227 | | // later decides whether this is a remote cloud rowset that should actually be preloaded. |
1228 | 137 | _index_file_cache_info.segment_file_size = *segment_file_size; |
1229 | 137 | _index_file_cache_info.add_index_range(footer_start, *segment_file_size - footer_start); |
1230 | 137 | if (index_file_cache_info != nullptr) { |
1231 | 137 | *index_file_cache_info = _index_file_cache_info; |
1232 | 137 | } |
1233 | 137 | if (*segment_file_size == 0) { |
1234 | 0 | return Status::Corruption("Bad segment, file size = 0"); |
1235 | 0 | } |
1236 | 137 | return Status::OK(); |
1237 | 137 | } |
1238 | | |
1239 | | Status VerticalSegmentWriter::finalize(uint64_t* segment_file_size, uint64_t* index_size, |
1240 | 137 | SegmentIndexFileCacheInfo* index_file_cache_info) { |
1241 | 137 | MonotonicStopWatch timer; |
1242 | 137 | timer.start(); |
1243 | | // check disk capacity |
1244 | 137 | if (_data_dir != nullptr && |
1245 | 137 | _data_dir->reach_capacity_limit((int64_t)_estimated_remaining_size())) { |
1246 | 0 | return Status::Error<DISK_REACH_CAPACITY_LIMIT>("disk {} exceed capacity limit.", |
1247 | 0 | _data_dir->path_hash()); |
1248 | 0 | } |
1249 | 137 | _row_count = _num_rows_written; |
1250 | 137 | _num_rows_written = 0; |
1251 | | // write index |
1252 | 137 | RETURN_IF_ERROR(finalize_columns_index(index_size)); |
1253 | | // write footer |
1254 | 137 | RETURN_IF_ERROR(finalize_footer(segment_file_size, index_file_cache_info)); |
1255 | | |
1256 | 137 | if (timer.elapsed_time() > 5000000000L) { |
1257 | 0 | LOG(INFO) << "segment flush consumes a lot time_ns " << timer.elapsed_time() |
1258 | 0 | << ", segmemt_size " << *segment_file_size; |
1259 | 0 | } |
1260 | 137 | return Status::OK(); |
1261 | 137 | } |
1262 | | |
1263 | 137 | void VerticalSegmentWriter::clear() { |
1264 | 1.77k | for (auto& column_writer : _column_writers) { |
1265 | 1.77k | column_writer.reset(); |
1266 | 1.77k | } |
1267 | 137 | _column_writers.clear(); |
1268 | 137 | _olap_data_convertor.reset(); |
1269 | 137 | } |
1270 | | |
1271 | | // write ordinal index after data has been written |
1272 | 137 | Status VerticalSegmentWriter::_write_ordinal_index() { |
1273 | 1.77k | for (auto& column_writer : _column_writers) { |
1274 | 1.77k | RETURN_IF_ERROR(column_writer->write_ordinal_index()); |
1275 | 1.77k | } |
1276 | 137 | return Status::OK(); |
1277 | 137 | } |
1278 | | |
1279 | 137 | Status VerticalSegmentWriter::_write_zone_map() { |
1280 | 1.77k | for (auto& column_writer : _column_writers) { |
1281 | 1.77k | RETURN_IF_ERROR(column_writer->write_zone_map()); |
1282 | 1.77k | } |
1283 | 137 | return Status::OK(); |
1284 | 137 | } |
1285 | | |
1286 | 137 | Status VerticalSegmentWriter::_write_inverted_index() { |
1287 | 1.77k | for (auto& column_writer : _column_writers) { |
1288 | 1.77k | RETURN_IF_ERROR(column_writer->write_inverted_index()); |
1289 | 1.77k | } |
1290 | 137 | return Status::OK(); |
1291 | 137 | } |
1292 | | |
1293 | 137 | Status VerticalSegmentWriter::_write_ann_index() { |
1294 | 1.77k | for (auto& column_writer : _column_writers) { |
1295 | 1.77k | RETURN_IF_ERROR(column_writer->write_ann_index()); |
1296 | 1.77k | } |
1297 | 137 | return Status::OK(); |
1298 | 137 | } |
1299 | | |
1300 | 137 | Status VerticalSegmentWriter::_write_bloom_filter_index() { |
1301 | 1.77k | for (auto& column_writer : _column_writers) { |
1302 | 1.77k | RETURN_IF_ERROR(column_writer->write_bloom_filter_index()); |
1303 | 1.77k | } |
1304 | 137 | return Status::OK(); |
1305 | 137 | } |
1306 | | |
1307 | 100 | Status VerticalSegmentWriter::_write_short_key_index() { |
1308 | 100 | std::vector<Slice> body; |
1309 | 100 | PageFooterPB footer; |
1310 | 100 | RETURN_IF_ERROR(_short_key_index_builder->finalize(_row_count, &body, &footer)); |
1311 | 100 | PagePointer pp; |
1312 | | // short key index page is not compressed right now |
1313 | 100 | RETURN_IF_ERROR(PageIO::write_page(_file_writer, body, footer, &pp)); |
1314 | 100 | pp.to_proto(_footer.mutable_short_key_index_page()); |
1315 | 100 | return Status::OK(); |
1316 | 100 | } |
1317 | | |
1318 | 46 | Status VerticalSegmentWriter::_write_primary_key_index() { |
1319 | 46 | CHECK_EQ(_primary_key_index_builder->num_rows(), _row_count); |
1320 | 46 | return _primary_key_index_builder->finalize(_footer.mutable_primary_key_index_meta()); |
1321 | 46 | } |
1322 | | |
1323 | 137 | Status VerticalSegmentWriter::_write_footer() { |
1324 | 137 | _footer.set_num_rows(_row_count); |
1325 | | |
1326 | | // Decide whether to externalize ColumnMetaPB by tablet default, and stamp footer version |
1327 | | |
1328 | 137 | if (_tablet_schema->storage_format() == TabletStorageFormatPB::TABLET_STORAGE_FORMAT_V3) { |
1329 | 32 | _footer.set_version(SEGMENT_FOOTER_VERSION_V3_EXT_COL_META); |
1330 | 32 | VLOG_DEBUG << "use external column meta"; |
1331 | | // External ColumnMetaPB writing (optional) |
1332 | 32 | RETURN_IF_ERROR(ExternalColMetaUtil::write_external_column_meta( |
1333 | 32 | _file_writer, &_footer, _opts.compression_type, |
1334 | 32 | [this](const std::vector<Slice>& slices) { return _write_raw_data(slices); })); |
1335 | 32 | } |
1336 | | |
1337 | | // Footer := SegmentFooterPB, FooterPBSize(4), FooterPBChecksum(4), MagicNumber(4) |
1338 | 137 | VLOG_DEBUG << "footer " << _footer.DebugString(); |
1339 | 137 | std::string footer_buf; |
1340 | 137 | if (!_footer.SerializeToString(&footer_buf)) { |
1341 | 0 | return Status::InternalError("failed to serialize segment footer"); |
1342 | 0 | } |
1343 | | |
1344 | 137 | faststring fixed_buf; |
1345 | | // footer's size |
1346 | 137 | put_fixed32_le(&fixed_buf, cast_set<uint32_t>(footer_buf.size())); |
1347 | | // footer's checksum |
1348 | 137 | uint32_t checksum = crc32c::Crc32c(footer_buf.data(), footer_buf.size()); |
1349 | 137 | put_fixed32_le(&fixed_buf, checksum); |
1350 | | // Append magic number. we don't write magic number in the header because |
1351 | | // that will need an extra seek when reading |
1352 | 137 | fixed_buf.append(k_segment_magic, k_segment_magic_length); |
1353 | | |
1354 | 137 | std::vector<Slice> slices {footer_buf, fixed_buf}; |
1355 | 137 | return _write_raw_data(slices); |
1356 | 137 | } |
1357 | | |
1358 | 761 | Status VerticalSegmentWriter::_write_raw_data(const std::vector<Slice>& slices) { |
1359 | 761 | RETURN_IF_ERROR(_file_writer->appendv(&slices[0], slices.size())); |
1360 | 761 | return Status::OK(); |
1361 | 761 | } |
1362 | | |
1363 | 137 | Slice VerticalSegmentWriter::min_encoded_key() { |
1364 | 137 | return (_primary_key_index_builder == nullptr) ? Slice(_min_key.data(), _min_key.size()) |
1365 | 137 | : _primary_key_index_builder->min_key(); |
1366 | 137 | } |
1367 | 137 | Slice VerticalSegmentWriter::max_encoded_key() { |
1368 | 137 | return (_primary_key_index_builder == nullptr) ? Slice(_max_key.data(), _max_key.size()) |
1369 | 137 | : _primary_key_index_builder->max_key(); |
1370 | 137 | } |
1371 | | |
1372 | 0 | void VerticalSegmentWriter::_set_min_max_key(const Slice& key) { |
1373 | 0 | if (UNLIKELY(_is_first_row)) { |
1374 | 0 | _min_key.append(key.get_data(), key.get_size()); |
1375 | 0 | _is_first_row = false; |
1376 | 0 | } |
1377 | 0 | if (key.compare(_max_key) > 0) { |
1378 | 0 | _max_key.clear(); |
1379 | 0 | _max_key.append(key.get_data(), key.get_size()); |
1380 | 0 | } |
1381 | 0 | } |
1382 | | |
1383 | 100 | void VerticalSegmentWriter::_set_min_key(const Slice& key) { |
1384 | 100 | if (UNLIKELY(_is_first_row)) { |
1385 | 100 | _min_key.append(key.get_data(), key.get_size()); |
1386 | 100 | _is_first_row = false; |
1387 | 100 | } |
1388 | 100 | } |
1389 | | |
1390 | 100 | void VerticalSegmentWriter::_set_max_key(const Slice& key) { |
1391 | 100 | _max_key.clear(); |
1392 | 100 | _max_key.append(key.get_data(), key.get_size()); |
1393 | 100 | } |
1394 | | |
1395 | | } // namespace doris::segment_v2 |