/root/doris/be/src/olap/base_tablet.cpp
Line | Count | Source (jump to first uncovered line) |
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 "olap/base_tablet.h" |
19 | | |
20 | | #include <fmt/format.h> |
21 | | #include <rapidjson/prettywriter.h> |
22 | | |
23 | | #include "common/logging.h" |
24 | | #include "common/status.h" |
25 | | #include "olap/calc_delete_bitmap_executor.h" |
26 | | #include "olap/delete_bitmap_calculator.h" |
27 | | #include "olap/memtable.h" |
28 | | #include "olap/partial_update_info.h" |
29 | | #include "olap/primary_key_index.h" |
30 | | #include "olap/rowid_conversion.h" |
31 | | #include "olap/rowset/beta_rowset.h" |
32 | | #include "olap/rowset/rowset.h" |
33 | | #include "olap/rowset/rowset_fwd.h" |
34 | | #include "olap/rowset/rowset_reader.h" |
35 | | #include "olap/tablet_fwd.h" |
36 | | #include "olap/txn_manager.h" |
37 | | #include "service/point_query_executor.h" |
38 | | #include "util/bvar_helper.h" |
39 | | #include "util/crc32c.h" |
40 | | #include "util/debug_points.h" |
41 | | #include "util/doris_metrics.h" |
42 | | #include "vec/common/assert_cast.h" |
43 | | #include "vec/common/schema_util.h" |
44 | | #include "vec/data_types/data_type_factory.hpp" |
45 | | #include "vec/jsonb/serialize.h" |
46 | | |
47 | | namespace doris { |
48 | | using namespace ErrorCode; |
49 | | |
50 | | namespace { |
51 | | |
52 | | bvar::LatencyRecorder g_tablet_commit_phase_update_delete_bitmap_latency( |
53 | | "doris_pk", "commit_phase_update_delete_bitmap"); |
54 | | bvar::LatencyRecorder g_tablet_lookup_rowkey_latency("doris_pk", "tablet_lookup_rowkey"); |
55 | | bvar::Adder<uint64_t> g_tablet_pk_not_found("doris_pk", "lookup_not_found"); |
56 | | bvar::PerSecond<bvar::Adder<uint64_t>> g_tablet_pk_not_found_per_second( |
57 | | "doris_pk", "lookup_not_found_per_second", &g_tablet_pk_not_found, 60); |
58 | | bvar::LatencyRecorder g_tablet_update_delete_bitmap_latency("doris_pk", "update_delete_bitmap"); |
59 | | |
60 | | static bvar::Adder<size_t> g_total_tablet_num("doris_total_tablet_num"); |
61 | | |
62 | | Status _get_segment_column_iterator(const BetaRowsetSharedPtr& rowset, uint32_t segid, |
63 | | const TabletColumn& target_column, |
64 | | SegmentCacheHandle* segment_cache_handle, |
65 | | std::unique_ptr<segment_v2::ColumnIterator>* column_iterator, |
66 | 0 | OlapReaderStatistics* stats) { |
67 | 0 | RETURN_IF_ERROR(SegmentLoader::instance()->load_segments(rowset, segment_cache_handle, true)); |
68 | | // find segment |
69 | 0 | auto it = std::find_if( |
70 | 0 | segment_cache_handle->get_segments().begin(), |
71 | 0 | segment_cache_handle->get_segments().end(), |
72 | 0 | [&segid](const segment_v2::SegmentSharedPtr& seg) { return seg->id() == segid; }); |
73 | 0 | if (it == segment_cache_handle->get_segments().end()) { |
74 | 0 | return Status::NotFound(fmt::format("rowset {} 's segemnt not found, seg_id {}", |
75 | 0 | rowset->rowset_id().to_string(), segid)); |
76 | 0 | } |
77 | 0 | segment_v2::SegmentSharedPtr segment = *it; |
78 | 0 | RETURN_IF_ERROR(segment->new_column_iterator(target_column, column_iterator, nullptr)); |
79 | 0 | segment_v2::ColumnIteratorOptions opt { |
80 | 0 | .use_page_cache = !config::disable_storage_page_cache, |
81 | 0 | .file_reader = segment->file_reader().get(), |
82 | 0 | .stats = stats, |
83 | 0 | .io_ctx = io::IOContext {.reader_type = ReaderType::READER_QUERY, |
84 | 0 | .file_cache_stats = &stats->file_cache_stats}, |
85 | 0 | }; |
86 | 0 | RETURN_IF_ERROR((*column_iterator)->init(opt)); |
87 | 0 | return Status::OK(); |
88 | 0 | } |
89 | | |
90 | | } // namespace |
91 | | |
92 | | extern MetricPrototype METRIC_query_scan_bytes; |
93 | | extern MetricPrototype METRIC_query_scan_rows; |
94 | | extern MetricPrototype METRIC_query_scan_count; |
95 | | DEFINE_COUNTER_METRIC_PROTOTYPE_2ARG(flush_bytes, MetricUnit::BYTES); |
96 | | DEFINE_COUNTER_METRIC_PROTOTYPE_2ARG(flush_finish_count, MetricUnit::OPERATIONS); |
97 | | |
98 | 275 | BaseTablet::BaseTablet(TabletMetaSharedPtr tablet_meta) : _tablet_meta(std::move(tablet_meta)) { |
99 | 275 | _metric_entity = DorisMetrics::instance()->metric_registry()->register_entity( |
100 | 275 | fmt::format("Tablet.{}", tablet_id()), {{"tablet_id", std::to_string(tablet_id())}}, |
101 | 275 | MetricEntityType::kTablet); |
102 | 275 | INT_COUNTER_METRIC_REGISTER(_metric_entity, query_scan_bytes); |
103 | 275 | INT_COUNTER_METRIC_REGISTER(_metric_entity, query_scan_rows); |
104 | 275 | INT_COUNTER_METRIC_REGISTER(_metric_entity, query_scan_count); |
105 | 275 | INT_COUNTER_METRIC_REGISTER(_metric_entity, flush_bytes); |
106 | 275 | INT_COUNTER_METRIC_REGISTER(_metric_entity, flush_finish_count); |
107 | | |
108 | | // construct _timestamped_versioned_tracker from rs and stale rs meta |
109 | 275 | _timestamped_version_tracker.construct_versioned_tracker(_tablet_meta->all_rs_metas(), |
110 | 275 | _tablet_meta->all_stale_rs_metas()); |
111 | | |
112 | | // if !_tablet_meta->all_rs_metas()[0]->tablet_schema(), |
113 | | // that mean the tablet_meta is still no upgrade to doris 1.2 versions. |
114 | | // Before doris 1.2 version, rowset metas don't have tablet schema. |
115 | | // And when upgrade to doris 1.2 version, |
116 | | // all rowset metas will be set the tablet schmea from tablet meta. |
117 | 275 | if (_tablet_meta->all_rs_metas().empty() || !_tablet_meta->all_rs_metas()[0]->tablet_schema()) { |
118 | 240 | _max_version_schema = _tablet_meta->tablet_schema(); |
119 | 240 | } else { |
120 | 35 | _max_version_schema = |
121 | 35 | tablet_schema_with_merged_max_schema_version(_tablet_meta->all_rs_metas()); |
122 | 35 | } |
123 | 275 | DCHECK(_max_version_schema); |
124 | 275 | g_total_tablet_num << 1; |
125 | 275 | } |
126 | | |
127 | 275 | BaseTablet::~BaseTablet() { |
128 | 275 | DorisMetrics::instance()->metric_registry()->deregister_entity(_metric_entity); |
129 | 275 | g_total_tablet_num << -1; |
130 | 275 | } |
131 | | |
132 | | TabletSchemaSPtr BaseTablet::tablet_schema_with_merged_max_schema_version( |
133 | 39 | const std::vector<RowsetMetaSharedPtr>& rowset_metas) { |
134 | 39 | RowsetMetaSharedPtr max_schema_version_rs = *std::max_element( |
135 | 39 | rowset_metas.begin(), rowset_metas.end(), |
136 | 168 | [](const RowsetMetaSharedPtr& a, const RowsetMetaSharedPtr& b) { |
137 | 168 | return !a->tablet_schema() |
138 | 168 | ? true |
139 | 168 | : (!b->tablet_schema() |
140 | 168 | ? false |
141 | 168 | : a->tablet_schema()->schema_version() < |
142 | 168 | b->tablet_schema()->schema_version()); |
143 | 168 | }); |
144 | 39 | TabletSchemaSPtr target_schema = max_schema_version_rs->tablet_schema(); |
145 | 39 | if (target_schema->num_variant_columns() > 0) { |
146 | | // For variant columns tablet schema need to be the merged wide tablet schema |
147 | 0 | std::vector<TabletSchemaSPtr> schemas; |
148 | 0 | std::transform(rowset_metas.begin(), rowset_metas.end(), std::back_inserter(schemas), |
149 | 0 | [](const RowsetMetaSharedPtr& rs_meta) { return rs_meta->tablet_schema(); }); |
150 | 0 | static_cast<void>( |
151 | 0 | vectorized::schema_util::get_least_common_schema(schemas, nullptr, target_schema)); |
152 | 0 | VLOG_DEBUG << "dump schema: " << target_schema->dump_full_schema(); |
153 | 0 | } |
154 | 39 | return target_schema; |
155 | 39 | } |
156 | | |
157 | 46 | Status BaseTablet::set_tablet_state(TabletState state) { |
158 | 46 | if (_tablet_meta->tablet_state() == TABLET_SHUTDOWN && state != TABLET_SHUTDOWN) { |
159 | 0 | return Status::Error<META_INVALID_ARGUMENT>( |
160 | 0 | "could not change tablet state from shutdown to {}", state); |
161 | 0 | } |
162 | 46 | _tablet_meta->set_tablet_state(state); |
163 | 46 | return Status::OK(); |
164 | 46 | } |
165 | | |
166 | 0 | void BaseTablet::update_max_version_schema(const TabletSchemaSPtr& tablet_schema) { |
167 | 0 | std::lock_guard wrlock(_meta_lock); |
168 | | // Double Check for concurrent update |
169 | 0 | if (!_max_version_schema || |
170 | 0 | tablet_schema->schema_version() > _max_version_schema->schema_version()) { |
171 | 0 | _max_version_schema = tablet_schema; |
172 | 0 | } |
173 | 0 | } |
174 | | |
175 | 0 | Status BaseTablet::update_by_least_common_schema(const TabletSchemaSPtr& update_schema) { |
176 | 0 | std::lock_guard wrlock(_meta_lock); |
177 | 0 | CHECK(_max_version_schema->schema_version() >= update_schema->schema_version()); |
178 | 0 | TabletSchemaSPtr final_schema; |
179 | 0 | bool check_column_size = true; |
180 | 0 | RETURN_IF_ERROR(vectorized::schema_util::get_least_common_schema( |
181 | 0 | {_max_version_schema, update_schema}, _max_version_schema, final_schema, |
182 | 0 | check_column_size)); |
183 | 0 | _max_version_schema = final_schema; |
184 | 0 | VLOG_DEBUG << "dump updated tablet schema: " << final_schema->dump_structure(); |
185 | 0 | return Status::OK(); |
186 | 0 | } |
187 | | |
188 | 27 | uint32_t BaseTablet::get_real_compaction_score() const { |
189 | 27 | const auto& rs_metas = _tablet_meta->all_rs_metas(); |
190 | 27 | return std::accumulate(rs_metas.begin(), rs_metas.end(), 0, |
191 | 378 | [](uint32_t score, const RowsetMetaSharedPtr& rs_meta) { |
192 | 378 | return score + rs_meta->get_compaction_score(); |
193 | 378 | }); |
194 | 27 | } |
195 | | |
196 | | Status BaseTablet::capture_rs_readers_unlocked(const Versions& version_path, |
197 | 1 | std::vector<RowSetSplits>* rs_splits) const { |
198 | 1 | DCHECK(rs_splits != nullptr && rs_splits->empty()); |
199 | 1 | for (auto version : version_path) { |
200 | 1 | auto it = _rs_version_map.find(version); |
201 | 1 | if (it == _rs_version_map.end()) { |
202 | 0 | VLOG_NOTICE << "fail to find Rowset in rs_version for version. tablet=" << tablet_id() |
203 | 0 | << ", version='" << version.first << "-" << version.second; |
204 | |
|
205 | 0 | it = _stale_rs_version_map.find(version); |
206 | 0 | if (it == _stale_rs_version_map.end()) { |
207 | 0 | return Status::Error<CAPTURE_ROWSET_READER_ERROR>( |
208 | 0 | "fail to find Rowset in stale_rs_version for version. tablet={}, " |
209 | 0 | "version={}-{}", |
210 | 0 | tablet_id(), version.first, version.second); |
211 | 0 | } |
212 | 0 | } |
213 | 1 | RowsetReaderSharedPtr rs_reader; |
214 | 1 | auto res = it->second->create_reader(&rs_reader); |
215 | 1 | if (!res.ok()) { |
216 | 0 | return Status::Error<CAPTURE_ROWSET_READER_ERROR>( |
217 | 0 | "failed to create reader for rowset:{}", it->second->rowset_id().to_string()); |
218 | 0 | } |
219 | 1 | rs_splits->emplace_back(std::move(rs_reader)); |
220 | 1 | } |
221 | 1 | return Status::OK(); |
222 | 1 | } |
223 | | |
224 | | // snapshot manager may call this api to check if version exists, so that |
225 | | // the version maybe not exist |
226 | | RowsetSharedPtr BaseTablet::get_rowset_by_version(const Version& version, |
227 | 1 | bool find_in_stale) const { |
228 | 1 | auto iter = _rs_version_map.find(version); |
229 | 1 | if (iter == _rs_version_map.end()) { |
230 | 0 | if (find_in_stale) { |
231 | 0 | return get_stale_rowset_by_version(version); |
232 | 0 | } |
233 | 0 | return nullptr; |
234 | 0 | } |
235 | 1 | return iter->second; |
236 | 1 | } |
237 | | |
238 | 0 | RowsetSharedPtr BaseTablet::get_stale_rowset_by_version(const Version& version) const { |
239 | 0 | auto iter = _stale_rs_version_map.find(version); |
240 | 0 | if (iter == _stale_rs_version_map.end()) { |
241 | 0 | VLOG_NOTICE << "no rowset for version:" << version << ", tablet: " << tablet_id(); |
242 | 0 | return nullptr; |
243 | 0 | } |
244 | 0 | return iter->second; |
245 | 0 | } |
246 | | |
247 | | // Already under _meta_lock |
248 | 30 | RowsetSharedPtr BaseTablet::get_rowset_with_max_version() const { |
249 | 30 | Version max_version = _tablet_meta->max_version(); |
250 | 30 | if (max_version.first == -1) { |
251 | 0 | return nullptr; |
252 | 0 | } |
253 | | |
254 | 30 | auto iter = _rs_version_map.find(max_version); |
255 | 30 | if (iter == _rs_version_map.end()) { |
256 | 0 | DCHECK(false) << "invalid version:" << max_version; |
257 | 0 | return nullptr; |
258 | 0 | } |
259 | 30 | return iter->second; |
260 | 30 | } |
261 | | |
262 | 0 | Status BaseTablet::get_all_rs_id(int64_t max_version, RowsetIdUnorderedSet* rowset_ids) const { |
263 | 0 | std::shared_lock rlock(_meta_lock); |
264 | 0 | return get_all_rs_id_unlocked(max_version, rowset_ids); |
265 | 0 | } |
266 | | |
267 | | Status BaseTablet::get_all_rs_id_unlocked(int64_t max_version, |
268 | 9 | RowsetIdUnorderedSet* rowset_ids) const { |
269 | | // Ensure that the obtained versions of rowsets are continuous |
270 | 9 | Version spec_version(0, max_version); |
271 | 9 | Versions version_path; |
272 | 9 | auto st = _timestamped_version_tracker.capture_consistent_versions(spec_version, &version_path); |
273 | 9 | if (!st.ok()) [[unlikely]] { |
274 | 0 | return st; |
275 | 0 | } |
276 | | |
277 | 11 | for (auto& ver : version_path) { |
278 | 11 | if (ver.second == 1) { |
279 | | // [0-1] rowset is empty for each tablet, skip it |
280 | 9 | continue; |
281 | 9 | } |
282 | 2 | auto it = _rs_version_map.find(ver); |
283 | 2 | if (it == _rs_version_map.end()) { |
284 | 0 | return Status::Error<CAPTURE_ROWSET_ERROR, false>( |
285 | 0 | "fail to find Rowset for version. tablet={}, version={}", tablet_id(), |
286 | 0 | ver.to_string()); |
287 | 0 | } |
288 | 2 | rowset_ids->emplace(it->second->rowset_id()); |
289 | 2 | } |
290 | 9 | return Status::OK(); |
291 | 9 | } |
292 | | |
293 | 0 | Versions BaseTablet::get_missed_versions(int64_t spec_version) const { |
294 | 0 | DCHECK(spec_version > 0) << "invalid spec_version: " << spec_version; |
295 | |
|
296 | 0 | Versions existing_versions; |
297 | 0 | { |
298 | 0 | std::shared_lock rdlock(_meta_lock); |
299 | 0 | for (const auto& rs : _tablet_meta->all_rs_metas()) { |
300 | 0 | existing_versions.emplace_back(rs->version()); |
301 | 0 | } |
302 | 0 | } |
303 | 0 | return calc_missed_versions(spec_version, std::move(existing_versions)); |
304 | 0 | } |
305 | | |
306 | 2 | Versions BaseTablet::get_missed_versions_unlocked(int64_t spec_version) const { |
307 | 2 | DCHECK(spec_version > 0) << "invalid spec_version: " << spec_version; |
308 | | |
309 | 2 | Versions existing_versions; |
310 | 8 | for (const auto& rs : _tablet_meta->all_rs_metas()) { |
311 | 8 | existing_versions.emplace_back(rs->version()); |
312 | 8 | } |
313 | 2 | return calc_missed_versions(spec_version, std::move(existing_versions)); |
314 | 2 | } |
315 | | |
316 | 1 | void BaseTablet::_print_missed_versions(const Versions& missed_versions) const { |
317 | 1 | std::stringstream ss; |
318 | 1 | ss << tablet_id() << " has " << missed_versions.size() << " missed version:"; |
319 | | // print at most 10 version |
320 | 3 | for (int i = 0; i < 10 && i < missed_versions.size(); ++i) { |
321 | 2 | ss << missed_versions[i] << ","; |
322 | 2 | } |
323 | 1 | LOG(WARNING) << ss.str(); |
324 | 1 | } |
325 | | |
326 | 1 | bool BaseTablet::_reconstruct_version_tracker_if_necessary() { |
327 | 1 | double orphan_vertex_ratio = _timestamped_version_tracker.get_orphan_vertex_ratio(); |
328 | 1 | if (orphan_vertex_ratio >= config::tablet_version_graph_orphan_vertex_ratio) { |
329 | 1 | _timestamped_version_tracker.construct_versioned_tracker( |
330 | 1 | _tablet_meta->all_rs_metas(), _tablet_meta->all_stale_rs_metas()); |
331 | 1 | return true; |
332 | 1 | } |
333 | 0 | return false; |
334 | 1 | } |
335 | | |
336 | | // should use this method to get a copy of current tablet meta |
337 | | // there are some rowset meta in local meta store and in in-memory tablet meta |
338 | | // but not in tablet meta in local meta store |
339 | 0 | void BaseTablet::generate_tablet_meta_copy(TabletMeta& new_tablet_meta) const { |
340 | 0 | TabletMetaPB tablet_meta_pb; |
341 | 0 | { |
342 | 0 | std::shared_lock rdlock(_meta_lock); |
343 | 0 | _tablet_meta->to_meta_pb(&tablet_meta_pb); |
344 | 0 | } |
345 | 0 | generate_tablet_meta_copy_unlocked(new_tablet_meta); |
346 | 0 | } |
347 | | |
348 | | // this is a unlocked version of generate_tablet_meta_copy() |
349 | | // some method already hold the _meta_lock before calling this, |
350 | | // such as EngineCloneTask::_finish_clone -> tablet->revise_tablet_meta |
351 | 2 | void BaseTablet::generate_tablet_meta_copy_unlocked(TabletMeta& new_tablet_meta) const { |
352 | 2 | TabletMetaPB tablet_meta_pb; |
353 | 2 | _tablet_meta->to_meta_pb(&tablet_meta_pb); |
354 | 2 | new_tablet_meta.init_from_pb(tablet_meta_pb); |
355 | 2 | } |
356 | | |
357 | | Status BaseTablet::calc_delete_bitmap_between_segments( |
358 | | RowsetSharedPtr rowset, const std::vector<segment_v2::SegmentSharedPtr>& segments, |
359 | 0 | DeleteBitmapPtr delete_bitmap) { |
360 | 0 | size_t const num_segments = segments.size(); |
361 | 0 | if (num_segments < 2) { |
362 | 0 | return Status::OK(); |
363 | 0 | } |
364 | | |
365 | 0 | OlapStopWatch watch; |
366 | 0 | auto const rowset_id = rowset->rowset_id(); |
367 | 0 | size_t seq_col_length = 0; |
368 | 0 | if (_tablet_meta->tablet_schema()->has_sequence_col()) { |
369 | 0 | auto seq_col_idx = _tablet_meta->tablet_schema()->sequence_col_idx(); |
370 | 0 | seq_col_length = _tablet_meta->tablet_schema()->column(seq_col_idx).length() + 1; |
371 | 0 | } |
372 | 0 | size_t rowid_length = 0; |
373 | 0 | if (!_tablet_meta->tablet_schema()->cluster_key_idxes().empty()) { |
374 | 0 | rowid_length = PrimaryKeyIndexReader::ROW_ID_LENGTH; |
375 | 0 | } |
376 | |
|
377 | 0 | MergeIndexDeleteBitmapCalculator calculator; |
378 | 0 | RETURN_IF_ERROR(calculator.init(rowset_id, segments, seq_col_length, rowid_length)); |
379 | | |
380 | 0 | RETURN_IF_ERROR(calculator.calculate_all(delete_bitmap)); |
381 | | |
382 | 0 | LOG(INFO) << fmt::format( |
383 | 0 | "construct delete bitmap between segments, " |
384 | 0 | "tablet: {}, rowset: {}, number of segments: {}, bitmap size: {}, cost {} (us)", |
385 | 0 | tablet_id(), rowset_id.to_string(), num_segments, delete_bitmap->delete_bitmap.size(), |
386 | 0 | watch.get_elapse_time_us()); |
387 | 0 | return Status::OK(); |
388 | 0 | } |
389 | | |
390 | | std::vector<RowsetSharedPtr> BaseTablet::get_rowset_by_ids( |
391 | 72 | const RowsetIdUnorderedSet* specified_rowset_ids) { |
392 | 72 | std::vector<RowsetSharedPtr> rowsets; |
393 | 72 | for (auto& rs : _rs_version_map) { |
394 | 15 | if (!specified_rowset_ids || |
395 | 15 | specified_rowset_ids->find(rs.second->rowset_id()) != specified_rowset_ids->end()) { |
396 | 1 | rowsets.push_back(rs.second); |
397 | 1 | } |
398 | 15 | } |
399 | | |
400 | 72 | std::sort(rowsets.begin(), rowsets.end(), [](RowsetSharedPtr& lhs, RowsetSharedPtr& rhs) { |
401 | 0 | return lhs->end_version() > rhs->end_version(); |
402 | 0 | }); |
403 | 72 | return rowsets; |
404 | 72 | } |
405 | | |
406 | | Status BaseTablet::lookup_row_data(const Slice& encoded_key, const RowLocation& row_location, |
407 | | RowsetSharedPtr input_rowset, const TupleDescriptor* desc, |
408 | | OlapReaderStatistics& stats, std::string& values, |
409 | 0 | bool write_to_cache) { |
410 | 0 | MonotonicStopWatch watch; |
411 | 0 | size_t row_size = 1; |
412 | 0 | watch.start(); |
413 | 0 | Defer _defer([&]() { |
414 | 0 | LOG_EVERY_N(INFO, 500) << "get a single_row, cost(us):" << watch.elapsed_time() / 1000 |
415 | 0 | << ", row_size:" << row_size; |
416 | 0 | }); |
417 | |
|
418 | 0 | BetaRowsetSharedPtr rowset = std::static_pointer_cast<BetaRowset>(input_rowset); |
419 | 0 | CHECK(rowset); |
420 | 0 | const TabletSchemaSPtr tablet_schema = rowset->tablet_schema(); |
421 | 0 | SegmentCacheHandle segment_cache_handle; |
422 | 0 | std::unique_ptr<segment_v2::ColumnIterator> column_iterator; |
423 | 0 | const auto& column = *DORIS_TRY(tablet_schema->column(BeConsts::ROW_STORE_COL)); |
424 | 0 | RETURN_IF_ERROR(_get_segment_column_iterator(rowset, row_location.segment_id, column, |
425 | 0 | &segment_cache_handle, &column_iterator, &stats)); |
426 | | // get and parse tuple row |
427 | 0 | vectorized::MutableColumnPtr column_ptr = vectorized::ColumnString::create(); |
428 | 0 | std::vector<segment_v2::rowid_t> rowids {static_cast<segment_v2::rowid_t>(row_location.row_id)}; |
429 | 0 | RETURN_IF_ERROR(column_iterator->read_by_rowids(rowids.data(), 1, column_ptr)); |
430 | 0 | assert(column_ptr->size() == 1); |
431 | 0 | auto* string_column = static_cast<vectorized::ColumnString*>(column_ptr.get()); |
432 | 0 | StringRef value = string_column->get_data_at(0); |
433 | 0 | values = value.to_string(); |
434 | 0 | if (write_to_cache) { |
435 | 0 | StringRef value = string_column->get_data_at(0); |
436 | 0 | RowCache::instance()->insert({tablet_id(), encoded_key}, Slice {value.data, value.size}); |
437 | 0 | } |
438 | 0 | return Status::OK(); |
439 | 0 | } |
440 | | |
441 | | Status BaseTablet::lookup_row_key(const Slice& encoded_key, TabletSchema* latest_schema, |
442 | | bool with_seq_col, |
443 | | const std::vector<RowsetSharedPtr>& specified_rowsets, |
444 | | RowLocation* row_location, uint32_t version, |
445 | | std::vector<std::unique_ptr<SegmentCacheHandle>>& segment_caches, |
446 | | RowsetSharedPtr* rowset, bool with_rowid, |
447 | 2 | std::string* encoded_seq_value, OlapReaderStatistics* stats) { |
448 | 2 | SCOPED_BVAR_LATENCY(g_tablet_lookup_rowkey_latency); |
449 | 2 | size_t seq_col_length = 0; |
450 | | // use the latest tablet schema to decide if the tablet has sequence column currently |
451 | 2 | const TabletSchema* schema = |
452 | 2 | (latest_schema == nullptr ? _tablet_meta->tablet_schema().get() : latest_schema); |
453 | 2 | if (schema->has_sequence_col() && with_seq_col) { |
454 | 2 | seq_col_length = schema->column(schema->sequence_col_idx()).length() + 1; |
455 | 2 | } |
456 | 2 | size_t rowid_length = 0; |
457 | 2 | if (with_rowid && !schema->cluster_key_idxes().empty()) { |
458 | 0 | rowid_length = PrimaryKeyIndexReader::ROW_ID_LENGTH; |
459 | 0 | } |
460 | 2 | Slice key_without_seq = |
461 | 2 | Slice(encoded_key.get_data(), encoded_key.get_size() - seq_col_length - rowid_length); |
462 | 2 | RowLocation loc; |
463 | | |
464 | 2 | for (size_t i = 0; i < specified_rowsets.size(); i++) { |
465 | 2 | auto& rs = specified_rowsets[i]; |
466 | 2 | auto& segments_key_bounds = rs->rowset_meta()->get_segments_key_bounds(); |
467 | 2 | int num_segments = rs->num_segments(); |
468 | 2 | DCHECK_EQ(segments_key_bounds.size(), num_segments); |
469 | 2 | std::vector<uint32_t> picked_segments; |
470 | 4 | for (int i = num_segments - 1; i >= 0; i--) { |
471 | 2 | if (key_without_seq.compare(segments_key_bounds[i].max_key()) > 0 || |
472 | 2 | key_without_seq.compare(segments_key_bounds[i].min_key()) < 0) { |
473 | 0 | continue; |
474 | 0 | } |
475 | 2 | picked_segments.emplace_back(i); |
476 | 2 | } |
477 | 2 | if (picked_segments.empty()) { |
478 | 0 | continue; |
479 | 0 | } |
480 | | |
481 | 2 | if (UNLIKELY(segment_caches[i] == nullptr)) { |
482 | 1 | segment_caches[i] = std::make_unique<SegmentCacheHandle>(); |
483 | 1 | RETURN_IF_ERROR(SegmentLoader::instance()->load_segments( |
484 | 1 | std::static_pointer_cast<BetaRowset>(rs), segment_caches[i].get(), true, true)); |
485 | 1 | } |
486 | 2 | auto& segments = segment_caches[i]->get_segments(); |
487 | 2 | DCHECK_EQ(segments.size(), num_segments); |
488 | | |
489 | 2 | for (auto id : picked_segments) { |
490 | 2 | Status s = segments[id]->lookup_row_key(encoded_key, schema, with_seq_col, with_rowid, |
491 | 2 | &loc, encoded_seq_value, stats); |
492 | 2 | if (s.is<KEY_NOT_FOUND>()) { |
493 | 0 | continue; |
494 | 0 | } |
495 | 2 | if (!s.ok() && !s.is<KEY_ALREADY_EXISTS>()) { |
496 | 0 | return s; |
497 | 0 | } |
498 | 2 | if (s.ok() && _tablet_meta->delete_bitmap().contains_agg_without_cache( |
499 | 1 | {loc.rowset_id, loc.segment_id, version}, loc.row_id)) { |
500 | | // if has sequence col, we continue to compare the sequence_id of |
501 | | // all rowsets, util we find an existing key. |
502 | 0 | if (schema->has_sequence_col()) { |
503 | 0 | continue; |
504 | 0 | } |
505 | | // The key is deleted, we don't need to search for it any more. |
506 | 0 | break; |
507 | 0 | } |
508 | | // `st` is either OK or KEY_ALREADY_EXISTS now. |
509 | | // for partial update, even if the key is already exists, we still need to |
510 | | // read it's original values to keep all columns align. |
511 | 2 | *row_location = loc; |
512 | 2 | if (rowset) { |
513 | | // return it's rowset |
514 | 2 | *rowset = rs; |
515 | 2 | } |
516 | | // find it and return |
517 | 2 | return s; |
518 | 2 | } |
519 | 2 | } |
520 | 0 | g_tablet_pk_not_found << 1; |
521 | 0 | return Status::Error<ErrorCode::KEY_NOT_FOUND>("can't find key in all rowsets"); |
522 | 2 | } |
523 | | |
524 | | // if user pass a token, then all calculation works will submit to a threadpool, |
525 | | // user can get all delete bitmaps from that token. |
526 | | // if `token` is nullptr, the calculation will run in local, and user can get the result |
527 | | // delete bitmap from `delete_bitmap` directly. |
528 | | Status BaseTablet::calc_delete_bitmap(const BaseTabletSPtr& tablet, RowsetSharedPtr rowset, |
529 | | const std::vector<segment_v2::SegmentSharedPtr>& segments, |
530 | | const std::vector<RowsetSharedPtr>& specified_rowsets, |
531 | | DeleteBitmapPtr delete_bitmap, int64_t end_version, |
532 | 69 | CalcDeleteBitmapToken* token, RowsetWriter* rowset_writer) { |
533 | 69 | auto rowset_id = rowset->rowset_id(); |
534 | 69 | if (specified_rowsets.empty() || segments.empty()) { |
535 | 68 | LOG(INFO) << "skip to construct delete bitmap tablet: " << tablet->tablet_id() |
536 | 68 | << " rowset: " << rowset_id; |
537 | 68 | return Status::OK(); |
538 | 68 | } |
539 | | |
540 | 1 | OlapStopWatch watch; |
541 | 1 | for (const auto& segment : segments) { |
542 | 1 | const auto& seg = segment; |
543 | 1 | if (token != nullptr) { |
544 | 1 | RETURN_IF_ERROR(token->submit(tablet, rowset, seg, specified_rowsets, end_version, |
545 | 1 | delete_bitmap, rowset_writer)); |
546 | 1 | } else { |
547 | 0 | RETURN_IF_ERROR(tablet->calc_segment_delete_bitmap( |
548 | 0 | rowset, segment, specified_rowsets, delete_bitmap, end_version, rowset_writer)); |
549 | 0 | } |
550 | 1 | } |
551 | | |
552 | 1 | return Status::OK(); |
553 | 1 | } |
554 | | |
555 | | Status BaseTablet::calc_segment_delete_bitmap(RowsetSharedPtr rowset, |
556 | | const segment_v2::SegmentSharedPtr& seg, |
557 | | const std::vector<RowsetSharedPtr>& specified_rowsets, |
558 | | DeleteBitmapPtr delete_bitmap, int64_t end_version, |
559 | 1 | RowsetWriter* rowset_writer) { |
560 | 1 | OlapStopWatch watch; |
561 | 1 | auto rowset_id = rowset->rowset_id(); |
562 | 1 | Version dummy_version(end_version + 1, end_version + 1); |
563 | 1 | auto rowset_schema = rowset->tablet_schema(); |
564 | | |
565 | 1 | PartialUpdateInfo* partial_update_info {nullptr}; |
566 | 1 | bool is_partial_update = rowset_writer && rowset_writer->is_partial_update(); |
567 | | // `have_input_seq_column` is for fixed partial update only. For flexible partial update, we should use |
568 | | // the skip bitmap to determine wheather a row has specified the sequence column |
569 | 1 | bool have_input_seq_column = false; |
570 | | // `rids_be_overwritten` is for flexible partial update only, it records row ids that is overwritten by |
571 | | // another row with higher seqeucne value |
572 | 1 | std::set<uint32_t> rids_be_overwritten; |
573 | 1 | if (is_partial_update) { |
574 | 0 | partial_update_info = rowset_writer->get_partial_update_info().get(); |
575 | 0 | if (partial_update_info->is_fixed_partial_update() && rowset_schema->has_sequence_col()) { |
576 | 0 | std::vector<uint32_t> including_cids = |
577 | 0 | rowset_writer->get_partial_update_info()->update_cids; |
578 | 0 | have_input_seq_column = |
579 | 0 | rowset_schema->has_sequence_col() && |
580 | 0 | (std::find(including_cids.cbegin(), including_cids.cend(), |
581 | 0 | rowset_schema->sequence_col_idx()) != including_cids.cend()); |
582 | 0 | } |
583 | 0 | } |
584 | | |
585 | 1 | if (rowset_schema->num_variant_columns() > 0) { |
586 | | // During partial updates, the extracted columns of a variant should not be included in the rowset schema. |
587 | | // This is because the partial update for a variant needs to ignore the extracted columns. |
588 | | // Otherwise, the schema types in different rowsets might be inconsistent. When performing a partial update, |
589 | | // the complete variant is constructed by reading all the sub-columns of the variant. |
590 | 0 | rowset_schema = rowset_schema->copy_without_variant_extracted_columns(); |
591 | 0 | } |
592 | | // use for partial update |
593 | 1 | FixedReadPlan read_plan_ori; |
594 | 1 | FixedReadPlan read_plan_update; |
595 | 1 | int64_t conflict_rows = 0; |
596 | 1 | int64_t new_generated_rows = 0; |
597 | | |
598 | 1 | std::map<RowsetId, RowsetSharedPtr> rsid_to_rowset; |
599 | 1 | rsid_to_rowset[rowset_id] = rowset; |
600 | 1 | vectorized::Block block = rowset_schema->create_block(); |
601 | 1 | vectorized::Block ordered_block = block.clone_empty(); |
602 | 1 | uint32_t pos = 0; |
603 | | |
604 | 1 | RETURN_IF_ERROR(seg->load_pk_index_and_bf()); // We need index blocks to iterate |
605 | 1 | const auto* pk_idx = seg->get_primary_key_index(); |
606 | 1 | int total = pk_idx->num_rows(); |
607 | 1 | uint32_t row_id = 0; |
608 | 1 | int32_t remaining = total; |
609 | 1 | bool exact_match = false; |
610 | 1 | std::string last_key; |
611 | 1 | int batch_size = 1024; |
612 | | // The data for each segment may be lookup multiple times. Creating a SegmentCacheHandle |
613 | | // will update the lru cache, and there will be obvious lock competition in multithreading |
614 | | // scenarios, so using a segment_caches to cache SegmentCacheHandle. |
615 | 1 | std::vector<std::unique_ptr<SegmentCacheHandle>> segment_caches(specified_rowsets.size()); |
616 | 2 | while (remaining > 0) { |
617 | 1 | std::unique_ptr<segment_v2::IndexedColumnIterator> iter; |
618 | 1 | RETURN_IF_ERROR(pk_idx->new_iterator(&iter)); |
619 | | |
620 | 1 | size_t num_to_read = std::min(batch_size, remaining); |
621 | 1 | auto index_type = vectorized::DataTypeFactory::instance().create_data_type( |
622 | 1 | pk_idx->type_info()->type(), 1, 0); |
623 | 1 | auto index_column = index_type->create_column(); |
624 | 1 | Slice last_key_slice(last_key); |
625 | 1 | RETURN_IF_ERROR(iter->seek_at_or_after(&last_key_slice, &exact_match)); |
626 | 1 | auto current_ordinal = iter->get_current_ordinal(); |
627 | 1 | DCHECK(total == remaining + current_ordinal) |
628 | 0 | << "total: " << total << ", remaining: " << remaining |
629 | 0 | << ", current_ordinal: " << current_ordinal; |
630 | | |
631 | 1 | size_t num_read = num_to_read; |
632 | 1 | RETURN_IF_ERROR(iter->next_batch(&num_read, index_column)); |
633 | 1 | DCHECK(num_to_read == num_read) |
634 | 0 | << "num_to_read: " << num_to_read << ", num_read: " << num_read; |
635 | 1 | last_key = index_column->get_data_at(num_read - 1).to_string(); |
636 | | |
637 | | // exclude last_key, last_key will be read in next batch. |
638 | 1 | if (num_read == batch_size && num_read != remaining) { |
639 | 0 | num_read -= 1; |
640 | 0 | } |
641 | 3 | for (size_t i = 0; i < num_read; i++, row_id++) { |
642 | 2 | Slice key = Slice(index_column->get_data_at(i).data, index_column->get_data_at(i).size); |
643 | 2 | RowLocation loc; |
644 | | // calculate row id |
645 | 2 | if (!_tablet_meta->tablet_schema()->cluster_key_idxes().empty()) { |
646 | 0 | size_t seq_col_length = 0; |
647 | 0 | if (_tablet_meta->tablet_schema()->has_sequence_col()) { |
648 | 0 | seq_col_length = |
649 | 0 | _tablet_meta->tablet_schema() |
650 | 0 | ->column(_tablet_meta->tablet_schema()->sequence_col_idx()) |
651 | 0 | .length() + |
652 | 0 | 1; |
653 | 0 | } |
654 | 0 | size_t rowid_length = PrimaryKeyIndexReader::ROW_ID_LENGTH; |
655 | 0 | Slice key_without_seq = |
656 | 0 | Slice(key.get_data(), key.get_size() - seq_col_length - rowid_length); |
657 | 0 | Slice rowid_slice = |
658 | 0 | Slice(key.get_data() + key_without_seq.get_size() + seq_col_length + 1, |
659 | 0 | rowid_length - 1); |
660 | 0 | const auto* type_info = |
661 | 0 | get_scalar_type_info<FieldType::OLAP_FIELD_TYPE_UNSIGNED_INT>(); |
662 | 0 | const auto* rowid_coder = get_key_coder(type_info->type()); |
663 | 0 | RETURN_IF_ERROR(rowid_coder->decode_ascending(&rowid_slice, rowid_length, |
664 | 0 | (uint8_t*)&row_id)); |
665 | 0 | } |
666 | | // same row in segments should be filtered |
667 | 2 | if (delete_bitmap->contains({rowset_id, seg->id(), DeleteBitmap::TEMP_VERSION_COMMON}, |
668 | 2 | row_id)) { |
669 | 0 | continue; |
670 | 0 | } |
671 | | |
672 | 2 | RowsetSharedPtr rowset_find; |
673 | 2 | auto st = lookup_row_key(key, rowset_schema.get(), true, specified_rowsets, &loc, |
674 | 2 | dummy_version.first - 1, segment_caches, &rowset_find); |
675 | 2 | bool expected_st = st.ok() || st.is<KEY_NOT_FOUND>() || st.is<KEY_ALREADY_EXISTS>(); |
676 | | // It's a defensive DCHECK, we need to exclude some common errors to avoid core-dump |
677 | | // while stress test |
678 | 2 | DCHECK(expected_st || st.is<MEM_LIMIT_EXCEEDED>()) |
679 | 0 | << "unexpected error status while lookup_row_key:" << st; |
680 | 2 | if (!expected_st) { |
681 | 0 | return st; |
682 | 0 | } |
683 | 2 | if (st.is<KEY_NOT_FOUND>()) { |
684 | 0 | continue; |
685 | 0 | } |
686 | | |
687 | 2 | ++conflict_rows; |
688 | 2 | if (st.is<KEY_ALREADY_EXISTS>() && |
689 | 2 | (!is_partial_update || |
690 | 1 | (partial_update_info->is_fixed_partial_update() && have_input_seq_column))) { |
691 | | // `st.is<KEY_ALREADY_EXISTS>()` means that there exists a row with the same key and larger value |
692 | | // in seqeunce column. |
693 | | // - If the current load is not a partial update, we just delete current row. |
694 | | // - Otherwise, it means that we are doing the alignment process in publish phase due to conflicts |
695 | | // during concurrent partial updates. And there exists another load which introduces a row with |
696 | | // the same keys and larger sequence column value published successfully after the commit phase |
697 | | // of the current load. |
698 | | // - If the columns we update include sequence column, we should delete the current row becase the |
699 | | // partial update on the current row has been `overwritten` by the previous one with larger sequence |
700 | | // column value. |
701 | | // - Otherwise, we should combine the values of the missing columns in the previous row and the values |
702 | | // of the including columns in the current row into a new row. |
703 | 1 | delete_bitmap->add({rowset_id, seg->id(), DeleteBitmap::TEMP_VERSION_COMMON}, |
704 | 1 | row_id); |
705 | 1 | continue; |
706 | 1 | } |
707 | 1 | if (is_partial_update && rowset_writer != nullptr) { |
708 | | // In publish version, record rows to be deleted for concurrent update |
709 | | // For example, if version 5 and 6 update a row, but version 6 only see |
710 | | // version 4 when write, and when publish version, version 5's value will |
711 | | // be marked as deleted and it's update is losed. |
712 | | // So here we should read version 5's columns and build a new row, which is |
713 | | // consists of version 6's update columns and version 5's origin columns |
714 | | // here we build 2 read plan for ori values and update values |
715 | | |
716 | | // - for fixed partial update, we should read update columns from current load's rowset |
717 | | // and read missing columns from previous rowsets to create the final block |
718 | | // - for flexible partial update, we should read all columns from current load's rowset |
719 | | // and read non sort key columns from previous rowsets to create the final block |
720 | | // So we only need to record rows to read for both mode partial update |
721 | 0 | read_plan_ori.prepare_to_read(loc, pos); |
722 | 0 | read_plan_update.prepare_to_read(RowLocation {rowset_id, seg->id(), row_id}, pos); |
723 | | |
724 | | // For flexible partial update, we should use skip bitmap to determine wheather |
725 | | // a row has specified the sequence column. But skip bitmap should be read from the segment. |
726 | | // So we record these row ids and process and filter them in `generate_new_block_for_flexible_partial_update()` |
727 | 0 | if (st.is<KEY_ALREADY_EXISTS>() && |
728 | 0 | partial_update_info->is_flexible_partial_update()) { |
729 | 0 | rids_be_overwritten.insert(pos); |
730 | 0 | } |
731 | |
|
732 | 0 | rsid_to_rowset[rowset_find->rowset_id()] = rowset_find; |
733 | 0 | ++pos; |
734 | | |
735 | | // delete bitmap will be calculate when memtable flush and |
736 | | // publish. The two stages may see different versions. |
737 | | // When there is sequence column, the currently imported data |
738 | | // of rowset may be marked for deletion at memtablet flush or |
739 | | // publish because the seq column is smaller than the previous |
740 | | // rowset. |
741 | | // just set 0 as a unified temporary version number, and update to |
742 | | // the real version number later. |
743 | 0 | delete_bitmap->add( |
744 | 0 | {loc.rowset_id, loc.segment_id, DeleteBitmap::TEMP_VERSION_COMMON}, |
745 | 0 | loc.row_id); |
746 | 0 | delete_bitmap->add({rowset_id, seg->id(), DeleteBitmap::TEMP_VERSION_COMMON}, |
747 | 0 | row_id); |
748 | 0 | ++new_generated_rows; |
749 | 0 | continue; |
750 | 0 | } |
751 | | // when st = ok |
752 | 1 | delete_bitmap->add({loc.rowset_id, loc.segment_id, DeleteBitmap::TEMP_VERSION_COMMON}, |
753 | 1 | loc.row_id); |
754 | 1 | } |
755 | 1 | remaining -= num_read; |
756 | 1 | } |
757 | | // DCHECK_EQ(total, row_id) << "segment total rows: " << total << " row_id:" << row_id; |
758 | | |
759 | 1 | if (config::enable_merge_on_write_correctness_check) { |
760 | 1 | RowsetIdUnorderedSet rowsetids; |
761 | 1 | for (const auto& rowset : specified_rowsets) { |
762 | 1 | rowsetids.emplace(rowset->rowset_id()); |
763 | 1 | VLOG_NOTICE << "[tabletID:" << tablet_id() << "]" |
764 | 0 | << "[add_sentinel_mark_to_delete_bitmap][end_version:" << end_version << "]" |
765 | 0 | << "add:" << rowset->rowset_id(); |
766 | 1 | } |
767 | 1 | add_sentinel_mark_to_delete_bitmap(delete_bitmap.get(), rowsetids); |
768 | 1 | } |
769 | | |
770 | 1 | if (pos > 0) { |
771 | 0 | DCHECK(partial_update_info); |
772 | 0 | if (partial_update_info->is_fixed_partial_update()) { |
773 | 0 | RETURN_IF_ERROR(generate_new_block_for_partial_update( |
774 | 0 | rowset_schema, partial_update_info, read_plan_ori, read_plan_update, |
775 | 0 | rsid_to_rowset, &block)); |
776 | 0 | } else { |
777 | 0 | RETURN_IF_ERROR(generate_new_block_for_flexible_partial_update( |
778 | 0 | rowset_schema, partial_update_info, rids_be_overwritten, read_plan_ori, |
779 | 0 | read_plan_update, rsid_to_rowset, &block)); |
780 | 0 | } |
781 | 0 | RETURN_IF_ERROR(sort_block(block, ordered_block)); |
782 | 0 | RETURN_IF_ERROR(rowset_writer->flush_single_block(&ordered_block)); |
783 | 0 | if (new_generated_rows != rowset_writer->num_rows()) { |
784 | 0 | LOG_WARNING( |
785 | 0 | "{} correctness warning: new_generated_rows != flushed_rows, " |
786 | 0 | "new_generated_rows={}, flushed_rows={}, filtered_rows={}, tablet={}", |
787 | 0 | partial_update_info->partial_update_mode_str(), new_generated_rows, |
788 | 0 | rowset_writer->num_rows(), rids_be_overwritten.size(), tablet_id()); |
789 | 0 | } |
790 | 0 | LOG(INFO) << "calc segment delete bitmap for " |
791 | 0 | << partial_update_info->partial_update_mode_str() << ", tablet: " << tablet_id() |
792 | 0 | << " rowset: " << rowset_id << " seg_id: " << seg->id() |
793 | 0 | << " dummy_version: " << end_version + 1 << " rows: " << seg->num_rows() |
794 | 0 | << " conflict rows: " << conflict_rows |
795 | 0 | << " filtered rows: " << rids_be_overwritten.size() |
796 | 0 | << " new generated rows: " << new_generated_rows |
797 | 0 | << " bimap num: " << delete_bitmap->delete_bitmap.size() |
798 | 0 | << " cost: " << watch.get_elapse_time_us() << "(us)"; |
799 | 0 | return Status::OK(); |
800 | 0 | } |
801 | 1 | LOG(INFO) << "calc segment delete bitmap, tablet: " << tablet_id() << " rowset: " << rowset_id |
802 | 1 | << " seg_id: " << seg->id() << " dummy_version: " << end_version + 1 |
803 | 1 | << " rows: " << seg->num_rows() << " conflict rows: " << conflict_rows |
804 | 1 | << " bitmap num: " << delete_bitmap->delete_bitmap.size() |
805 | 1 | << " cost: " << watch.get_elapse_time_us() << "(us)"; |
806 | 1 | return Status::OK(); |
807 | 1 | } |
808 | | |
809 | 0 | Status BaseTablet::sort_block(vectorized::Block& in_block, vectorized::Block& output_block) { |
810 | 0 | vectorized::MutableBlock mutable_input_block = |
811 | 0 | vectorized::MutableBlock::build_mutable_block(&in_block); |
812 | 0 | vectorized::MutableBlock mutable_output_block = |
813 | 0 | vectorized::MutableBlock::build_mutable_block(&output_block); |
814 | |
|
815 | 0 | std::shared_ptr<RowInBlockComparator> vec_row_comparator = |
816 | 0 | std::make_shared<RowInBlockComparator>(_tablet_meta->tablet_schema()); |
817 | 0 | vec_row_comparator->set_block(&mutable_input_block); |
818 | |
|
819 | 0 | std::vector<std::unique_ptr<RowInBlock>> row_in_blocks; |
820 | 0 | DCHECK(in_block.rows() <= std::numeric_limits<int>::max()); |
821 | 0 | row_in_blocks.reserve(in_block.rows()); |
822 | 0 | for (size_t i = 0; i < in_block.rows(); ++i) { |
823 | 0 | row_in_blocks.emplace_back(std::make_unique<RowInBlock>(i)); |
824 | 0 | } |
825 | 0 | std::sort(row_in_blocks.begin(), row_in_blocks.end(), |
826 | 0 | [&](const std::unique_ptr<RowInBlock>& l, |
827 | 0 | const std::unique_ptr<RowInBlock>& r) -> bool { |
828 | 0 | auto value = (*vec_row_comparator)(l.get(), r.get()); |
829 | 0 | DCHECK(value != 0) << "value equel when sort block, l_pos: " << l->_row_pos |
830 | 0 | << " r_pos: " << r->_row_pos; |
831 | 0 | return value < 0; |
832 | 0 | }); |
833 | 0 | std::vector<uint32_t> row_pos_vec; |
834 | 0 | row_pos_vec.reserve(in_block.rows()); |
835 | 0 | for (auto& block : row_in_blocks) { |
836 | 0 | row_pos_vec.emplace_back(block->_row_pos); |
837 | 0 | } |
838 | 0 | return mutable_output_block.add_rows(&in_block, row_pos_vec.data(), |
839 | 0 | row_pos_vec.data() + in_block.rows()); |
840 | 0 | } |
841 | | |
842 | | // fetch value by row column |
843 | | Status BaseTablet::fetch_value_through_row_column(RowsetSharedPtr input_rowset, |
844 | | const TabletSchema& tablet_schema, uint32_t segid, |
845 | | const std::vector<uint32_t>& rowids, |
846 | | const std::vector<uint32_t>& cids, |
847 | 0 | vectorized::Block& block) { |
848 | 0 | MonotonicStopWatch watch; |
849 | 0 | watch.start(); |
850 | 0 | Defer _defer([&]() { |
851 | 0 | LOG_EVERY_N(INFO, 500) << "fetch_value_by_rowids, cost(us):" << watch.elapsed_time() / 1000 |
852 | 0 | << ", row_batch_size:" << rowids.size(); |
853 | 0 | }); |
854 | |
|
855 | 0 | BetaRowsetSharedPtr rowset = std::static_pointer_cast<BetaRowset>(input_rowset); |
856 | 0 | CHECK(rowset); |
857 | 0 | CHECK(tablet_schema.has_row_store_for_all_columns()); |
858 | 0 | SegmentCacheHandle segment_cache_handle; |
859 | 0 | std::unique_ptr<segment_v2::ColumnIterator> column_iterator; |
860 | 0 | OlapReaderStatistics stats; |
861 | 0 | const auto& column = *DORIS_TRY(tablet_schema.column(BeConsts::ROW_STORE_COL)); |
862 | 0 | RETURN_IF_ERROR(_get_segment_column_iterator(rowset, segid, column, &segment_cache_handle, |
863 | 0 | &column_iterator, &stats)); |
864 | | // get and parse tuple row |
865 | 0 | vectorized::MutableColumnPtr column_ptr = vectorized::ColumnString::create(); |
866 | 0 | RETURN_IF_ERROR(column_iterator->read_by_rowids(rowids.data(), rowids.size(), column_ptr)); |
867 | 0 | assert(column_ptr->size() == rowids.size()); |
868 | 0 | auto* string_column = static_cast<vectorized::ColumnString*>(column_ptr.get()); |
869 | 0 | vectorized::DataTypeSerDeSPtrs serdes; |
870 | 0 | serdes.resize(cids.size()); |
871 | 0 | std::unordered_map<uint32_t, uint32_t> col_uid_to_idx; |
872 | 0 | std::vector<std::string> default_values; |
873 | 0 | default_values.resize(cids.size()); |
874 | 0 | for (int i = 0; i < cids.size(); ++i) { |
875 | 0 | const TabletColumn& column = tablet_schema.column(cids[i]); |
876 | 0 | vectorized::DataTypePtr type = |
877 | 0 | vectorized::DataTypeFactory::instance().create_data_type(column); |
878 | 0 | col_uid_to_idx[column.unique_id()] = i; |
879 | 0 | default_values[i] = column.default_value(); |
880 | 0 | serdes[i] = type->get_serde(); |
881 | 0 | } |
882 | 0 | vectorized::JsonbSerializeUtil::jsonb_to_block(serdes, *string_column, col_uid_to_idx, block, |
883 | 0 | default_values, {}); |
884 | 0 | return Status::OK(); |
885 | 0 | } |
886 | | |
887 | | Status BaseTablet::fetch_value_by_rowids(RowsetSharedPtr input_rowset, uint32_t segid, |
888 | | const std::vector<uint32_t>& rowids, |
889 | | const TabletColumn& tablet_column, |
890 | 0 | vectorized::MutableColumnPtr& dst) { |
891 | 0 | MonotonicStopWatch watch; |
892 | 0 | watch.start(); |
893 | 0 | Defer _defer([&]() { |
894 | 0 | LOG_EVERY_N(INFO, 500) << "fetch_value_by_rowids, cost(us):" << watch.elapsed_time() / 1000 |
895 | 0 | << ", row_batch_size:" << rowids.size(); |
896 | 0 | }); |
897 | | |
898 | | // read row data |
899 | 0 | BetaRowsetSharedPtr rowset = std::static_pointer_cast<BetaRowset>(input_rowset); |
900 | 0 | CHECK(rowset); |
901 | 0 | SegmentCacheHandle segment_cache_handle; |
902 | 0 | std::unique_ptr<segment_v2::ColumnIterator> column_iterator; |
903 | 0 | OlapReaderStatistics stats; |
904 | 0 | RETURN_IF_ERROR(_get_segment_column_iterator(rowset, segid, tablet_column, |
905 | 0 | &segment_cache_handle, &column_iterator, &stats)); |
906 | 0 | RETURN_IF_ERROR(column_iterator->read_by_rowids(rowids.data(), rowids.size(), dst)); |
907 | 0 | return Status::OK(); |
908 | 0 | } |
909 | | |
910 | | const signed char* BaseTablet::get_delete_sign_column_data(const vectorized::Block& block, |
911 | 0 | size_t rows_at_least) { |
912 | 0 | if (const vectorized::ColumnWithTypeAndName* delete_sign_column = |
913 | 0 | block.try_get_by_name(DELETE_SIGN); |
914 | 0 | delete_sign_column != nullptr) { |
915 | 0 | const auto& delete_sign_col = |
916 | 0 | reinterpret_cast<const vectorized::ColumnInt8&>(*(delete_sign_column->column)); |
917 | 0 | if (delete_sign_col.size() >= rows_at_least) { |
918 | 0 | return delete_sign_col.get_data().data(); |
919 | 0 | } |
920 | 0 | } |
921 | 0 | return nullptr; |
922 | 0 | }; |
923 | | |
924 | | Status BaseTablet::generate_default_value_block(const TabletSchema& schema, |
925 | | const std::vector<uint32_t>& cids, |
926 | | const std::vector<std::string>& default_values, |
927 | | const vectorized::Block& ref_block, |
928 | 0 | vectorized::Block& default_value_block) { |
929 | 0 | auto mutable_default_value_columns = default_value_block.mutate_columns(); |
930 | 0 | for (auto i = 0; i < cids.size(); ++i) { |
931 | 0 | const auto& column = schema.column(cids[i]); |
932 | 0 | if (column.has_default_value()) { |
933 | 0 | const auto& default_value = default_values[i]; |
934 | 0 | vectorized::ReadBuffer rb(const_cast<char*>(default_value.c_str()), |
935 | 0 | default_value.size()); |
936 | 0 | RETURN_IF_ERROR(ref_block.get_by_position(i).type->from_string( |
937 | 0 | rb, mutable_default_value_columns[i].get())); |
938 | 0 | } |
939 | 0 | } |
940 | 0 | default_value_block.set_columns(std::move(mutable_default_value_columns)); |
941 | 0 | return Status::OK(); |
942 | 0 | } |
943 | | |
944 | | Status BaseTablet::generate_new_block_for_partial_update( |
945 | | TabletSchemaSPtr rowset_schema, const PartialUpdateInfo* partial_update_info, |
946 | | const FixedReadPlan& read_plan_ori, const FixedReadPlan& read_plan_update, |
947 | | const std::map<RowsetId, RowsetSharedPtr>& rsid_to_rowset, |
948 | 0 | vectorized::Block* output_block) { |
949 | | // do partial update related works |
950 | | // 1. read columns by read plan |
951 | | // 2. generate new block |
952 | | // 3. write a new segment and modify rowset meta |
953 | | // 4. mark current keys deleted |
954 | 0 | CHECK(output_block); |
955 | 0 | auto full_mutable_columns = output_block->mutate_columns(); |
956 | 0 | const auto& missing_cids = partial_update_info->missing_cids; |
957 | 0 | const auto& update_cids = partial_update_info->update_cids; |
958 | 0 | auto old_block = rowset_schema->create_block_by_cids(missing_cids); |
959 | 0 | auto update_block = rowset_schema->create_block_by_cids(update_cids); |
960 | | |
961 | | // rowid in the final block(start from 0, increase continuously) -> rowid to read in update_block |
962 | 0 | std::map<uint32_t, uint32_t> read_index_update; |
963 | | |
964 | | // read current rowset first, if a row in the current rowset has delete sign mark |
965 | | // we don't need to read values from old block |
966 | 0 | RETURN_IF_ERROR(read_plan_update.read_columns_by_plan( |
967 | 0 | *rowset_schema, update_cids, rsid_to_rowset, update_block, &read_index_update)); |
968 | 0 | size_t update_rows = read_index_update.size(); |
969 | 0 | for (auto i = 0; i < update_cids.size(); ++i) { |
970 | 0 | for (auto idx = 0; idx < update_rows; ++idx) { |
971 | 0 | full_mutable_columns[update_cids[i]]->insert_from( |
972 | 0 | *update_block.get_by_position(i).column, read_index_update[idx]); |
973 | 0 | } |
974 | 0 | } |
975 | | |
976 | | // if there is sequence column in the table, we need to read the sequence column, |
977 | | // otherwise it may cause the merge-on-read based compaction policy to produce incorrect results |
978 | 0 | const auto* __restrict new_block_delete_signs = |
979 | 0 | rowset_schema->has_sequence_col() |
980 | 0 | ? nullptr |
981 | 0 | : get_delete_sign_column_data(update_block, update_rows); |
982 | | |
983 | | // rowid in the final block(start from 0, increase, may not continuous becasue we skip to read some rows) -> rowid to read in old_block |
984 | 0 | std::map<uint32_t, uint32_t> read_index_old; |
985 | 0 | RETURN_IF_ERROR(read_plan_ori.read_columns_by_plan(*rowset_schema, missing_cids, rsid_to_rowset, |
986 | 0 | old_block, &read_index_old, |
987 | 0 | new_block_delete_signs)); |
988 | 0 | size_t old_rows = read_index_old.size(); |
989 | 0 | const auto* __restrict old_block_delete_signs = |
990 | 0 | get_delete_sign_column_data(old_block, old_rows); |
991 | | |
992 | | // build default value block |
993 | 0 | auto default_value_block = old_block.clone_empty(); |
994 | 0 | if (old_block_delete_signs != nullptr || new_block_delete_signs != nullptr) { |
995 | 0 | RETURN_IF_ERROR(BaseTablet::generate_default_value_block( |
996 | 0 | *rowset_schema, missing_cids, partial_update_info->default_values, old_block, |
997 | 0 | default_value_block)); |
998 | 0 | } |
999 | | |
1000 | 0 | CHECK(update_rows >= old_rows); |
1001 | | |
1002 | | // build full block |
1003 | 0 | for (auto i = 0; i < missing_cids.size(); ++i) { |
1004 | 0 | const auto& rs_column = rowset_schema->column(missing_cids[i]); |
1005 | 0 | auto& mutable_column = full_mutable_columns[missing_cids[i]]; |
1006 | 0 | for (auto idx = 0; idx < update_rows; ++idx) { |
1007 | | // There are two cases we don't need to read values from old data: |
1008 | | // 1. if the conflicting new row's delete sign is marked, which means the value columns |
1009 | | // of the row will not be read. So we don't need to read the missing values from the previous rows. |
1010 | | // 2. if the conflicting old row's delete sign is marked, which means that the key is not exist now, |
1011 | | // we should not read old values from the deleted data, and should use default value instead. |
1012 | | // NOTE: since now we are in the publishing phase, all data is commited |
1013 | | // before, even the `strict_mode` is true (which requires partial update |
1014 | | // load job can't insert new keys), this "new" key MUST be written into |
1015 | | // the new generated segment file. |
1016 | 0 | if (new_block_delete_signs != nullptr && new_block_delete_signs[idx]) { |
1017 | 0 | mutable_column->insert_default(); |
1018 | 0 | } else if (old_block_delete_signs != nullptr && |
1019 | 0 | old_block_delete_signs[read_index_old[idx]] != 0) { |
1020 | 0 | if (rs_column.has_default_value()) { |
1021 | 0 | mutable_column->insert_from(*default_value_block.get_by_position(i).column, 0); |
1022 | 0 | } else if (rs_column.is_nullable()) { |
1023 | 0 | assert_cast<vectorized::ColumnNullable*, TypeCheckOnRelease::DISABLE>( |
1024 | 0 | mutable_column.get()) |
1025 | 0 | ->insert_null_elements(1); |
1026 | 0 | } else { |
1027 | 0 | mutable_column->insert_default(); |
1028 | 0 | } |
1029 | 0 | } else { |
1030 | 0 | mutable_column->insert_from(*old_block.get_by_position(i).column, |
1031 | 0 | read_index_old[idx]); |
1032 | 0 | } |
1033 | 0 | } |
1034 | 0 | } |
1035 | 0 | output_block->set_columns(std::move(full_mutable_columns)); |
1036 | 0 | VLOG_DEBUG << "full block when publish: " << output_block->dump_data(); |
1037 | 0 | return Status::OK(); |
1038 | 0 | } |
1039 | | |
1040 | | Status BaseTablet::generate_new_block_for_flexible_partial_update( |
1041 | | TabletSchemaSPtr rowset_schema, const PartialUpdateInfo* partial_update_info, |
1042 | | std::set<uint32_t>& rids_be_overwritten, const FixedReadPlan& read_plan_ori, |
1043 | | const FixedReadPlan& read_plan_update, |
1044 | | const std::map<RowsetId, RowsetSharedPtr>& rsid_to_rowset, |
1045 | 0 | vectorized::Block* output_block) { |
1046 | 0 | CHECK(output_block); |
1047 | |
|
1048 | 0 | const auto& non_sort_key_cids = partial_update_info->missing_cids; |
1049 | 0 | std::vector<uint32_t> all_cids(rowset_schema->num_columns()); |
1050 | 0 | std::iota(all_cids.begin(), all_cids.end(), 0); |
1051 | 0 | auto old_block = rowset_schema->create_block_by_cids(non_sort_key_cids); |
1052 | 0 | auto update_block = rowset_schema->create_block_by_cids(all_cids); |
1053 | | |
1054 | | // rowid in the final block(start from 0, increase continuously) -> rowid to read in update_block |
1055 | 0 | std::map<uint32_t, uint32_t> read_index_update; |
1056 | | |
1057 | | // 1. read the current rowset first, if a row in the current rowset has delete sign mark |
1058 | | // we don't need to read values from old block for that row |
1059 | 0 | RETURN_IF_ERROR(read_plan_update.read_columns_by_plan(*rowset_schema, all_cids, rsid_to_rowset, |
1060 | 0 | update_block, &read_index_update)); |
1061 | 0 | size_t update_rows = read_index_update.size(); |
1062 | | |
1063 | | // TODO(bobhan1): add the delete sign optimazation here |
1064 | | // // if there is sequence column in the table, we need to read the sequence column, |
1065 | | // // otherwise it may cause the merge-on-read based compaction policy to produce incorrect results |
1066 | | // const auto* __restrict new_block_delete_signs = |
1067 | | // rowset_schema->has_sequence_col() |
1068 | | // ? nullptr |
1069 | | // : get_delete_sign_column_data(update_block, update_rows); |
1070 | | |
1071 | | // 2. read previous rowsets |
1072 | | // rowid in the final block(start from 0, increase, may not continuous becasue we skip to read some rows) -> rowid to read in old_block |
1073 | 0 | std::map<uint32_t, uint32_t> read_index_old; |
1074 | 0 | RETURN_IF_ERROR(read_plan_ori.read_columns_by_plan(*rowset_schema, non_sort_key_cids, |
1075 | 0 | rsid_to_rowset, old_block, &read_index_old)); |
1076 | 0 | size_t old_rows = read_index_old.size(); |
1077 | 0 | DCHECK(update_rows == old_rows); |
1078 | 0 | const auto* __restrict old_block_delete_signs = |
1079 | 0 | get_delete_sign_column_data(old_block, old_rows); |
1080 | 0 | DCHECK(old_block_delete_signs != nullptr); |
1081 | | |
1082 | | // 3. build default value block |
1083 | 0 | auto default_value_block = old_block.clone_empty(); |
1084 | 0 | RETURN_IF_ERROR(BaseTablet::generate_default_value_block(*rowset_schema, non_sort_key_cids, |
1085 | 0 | partial_update_info->default_values, |
1086 | 0 | old_block, default_value_block)); |
1087 | | |
1088 | | // 4. build the final block |
1089 | 0 | auto full_mutable_columns = output_block->mutate_columns(); |
1090 | 0 | DCHECK(rowset_schema->has_skip_bitmap_col()); |
1091 | 0 | auto skip_bitmap_col_idx = rowset_schema->skip_bitmap_col_idx(); |
1092 | 0 | const std::vector<BitmapValue>* skip_bitmaps = |
1093 | 0 | &(assert_cast<const vectorized::ColumnBitmap*, TypeCheckOnRelease::DISABLE>( |
1094 | 0 | update_block.get_by_position(skip_bitmap_col_idx).column->get_ptr().get()) |
1095 | 0 | ->get_data()); |
1096 | |
|
1097 | 0 | VLOG_DEBUG << fmt::format( |
1098 | 0 | "BaseTablet::generate_new_block_for_flexible_partial_update: " |
1099 | 0 | "rids_be_overwritten.size()={}", |
1100 | 0 | rids_be_overwritten.size()); |
1101 | 0 | if (rowset_schema->has_sequence_col() && !rids_be_overwritten.empty()) { |
1102 | 0 | int32_t seq_col_unique_id = |
1103 | 0 | rowset_schema->column(rowset_schema->sequence_col_idx()).unique_id(); |
1104 | | // If the row specifies the sequence column, we should delete the current row becase the |
1105 | | // flexible partial update on the current row has been `overwritten` by the previous one with larger sequence |
1106 | | // column value. |
1107 | 0 | for (auto it = rids_be_overwritten.begin(); it != rids_be_overwritten.end();) { |
1108 | 0 | auto rid = *it; |
1109 | 0 | if (!skip_bitmaps->at(rid).contains(seq_col_unique_id)) { |
1110 | 0 | VLOG_DEBUG << fmt::format( |
1111 | 0 | "BaseTablet::generate_new_block_for_flexible_partial_update: rid={} " |
1112 | 0 | "filtered", |
1113 | 0 | rid); |
1114 | 0 | ++it; |
1115 | 0 | } else { |
1116 | 0 | it = rids_be_overwritten.erase(it); |
1117 | 0 | VLOG_DEBUG << fmt::format( |
1118 | 0 | "BaseTablet::generate_new_block_for_flexible_partial_update: rid={} " |
1119 | 0 | "keeped", |
1120 | 0 | rid); |
1121 | 0 | } |
1122 | 0 | } |
1123 | 0 | } |
1124 | |
|
1125 | 0 | auto fill_one_cell = [&read_index_old](const TabletColumn& tablet_column, std::size_t idx, |
1126 | 0 | vectorized::MutableColumnPtr& new_col, |
1127 | 0 | const vectorized::IColumn& default_value_col, |
1128 | 0 | const vectorized::IColumn& old_value_col, |
1129 | 0 | const vectorized::IColumn& cur_col, bool skipped, |
1130 | 0 | const signed char* delete_sign_column_data) { |
1131 | 0 | if (skipped) { |
1132 | 0 | if (delete_sign_column_data != nullptr && |
1133 | 0 | delete_sign_column_data[read_index_old[idx]] != 0) { |
1134 | 0 | if (tablet_column.has_default_value()) { |
1135 | 0 | new_col->insert_from(default_value_col, 0); |
1136 | 0 | } else if (tablet_column.is_nullable()) { |
1137 | 0 | assert_cast<vectorized::ColumnNullable*, TypeCheckOnRelease::DISABLE>( |
1138 | 0 | new_col.get()) |
1139 | 0 | ->insert_null_elements(1); |
1140 | 0 | } else { |
1141 | 0 | new_col->insert_default(); |
1142 | 0 | } |
1143 | 0 | } else { |
1144 | 0 | new_col->insert_from(old_value_col, idx); |
1145 | 0 | } |
1146 | 0 | } else { |
1147 | 0 | new_col->insert_from(cur_col, idx); |
1148 | 0 | } |
1149 | 0 | }; |
1150 | |
|
1151 | 0 | for (std::size_t cid {0}; cid < rowset_schema->num_columns(); cid++) { |
1152 | 0 | vectorized::MutableColumnPtr& new_col = full_mutable_columns[cid]; |
1153 | 0 | const vectorized::IColumn& cur_col = *update_block.get_by_position(cid).column; |
1154 | 0 | const auto& rs_column = rowset_schema->column(cid); |
1155 | 0 | auto col_uid = rs_column.unique_id(); |
1156 | 0 | for (auto idx = 0; idx < update_rows; ++idx) { |
1157 | 0 | if (cid < rowset_schema->num_key_columns()) { |
1158 | 0 | new_col->insert_from(cur_col, idx); |
1159 | 0 | } else { |
1160 | 0 | const vectorized::IColumn& default_value_col = |
1161 | 0 | *default_value_block.get_by_position(cid - rowset_schema->num_key_columns()) |
1162 | 0 | .column; |
1163 | 0 | const vectorized::IColumn& old_value_col = |
1164 | 0 | *old_block.get_by_position(cid - rowset_schema->num_key_columns()).column; |
1165 | 0 | if (rids_be_overwritten.contains(idx)) { |
1166 | 0 | new_col->insert_from(old_value_col, idx); |
1167 | 0 | } else { |
1168 | 0 | fill_one_cell(rs_column, idx, new_col, default_value_col, old_value_col, |
1169 | 0 | cur_col, skip_bitmaps->at(idx).contains(col_uid), |
1170 | 0 | old_block_delete_signs); |
1171 | 0 | } |
1172 | 0 | } |
1173 | 0 | } |
1174 | 0 | DCHECK_EQ(full_mutable_columns[cid]->size(), update_rows); |
1175 | 0 | } |
1176 | |
|
1177 | 0 | output_block->set_columns(std::move(full_mutable_columns)); |
1178 | 0 | VLOG_DEBUG << "full block when publish: " << output_block->dump_data(); |
1179 | 0 | return Status::OK(); |
1180 | 0 | } |
1181 | | |
1182 | | Status BaseTablet::commit_phase_update_delete_bitmap( |
1183 | | const BaseTabletSPtr& tablet, const RowsetSharedPtr& rowset, |
1184 | | RowsetIdUnorderedSet& pre_rowset_ids, DeleteBitmapPtr delete_bitmap, |
1185 | | const std::vector<segment_v2::SegmentSharedPtr>& segments, int64_t txn_id, |
1186 | 3 | CalcDeleteBitmapToken* token, RowsetWriter* rowset_writer) { |
1187 | 3 | SCOPED_BVAR_LATENCY(g_tablet_commit_phase_update_delete_bitmap_latency); |
1188 | 3 | RowsetIdUnorderedSet cur_rowset_ids; |
1189 | 3 | RowsetIdUnorderedSet rowset_ids_to_add; |
1190 | 3 | RowsetIdUnorderedSet rowset_ids_to_del; |
1191 | 3 | int64_t cur_version; |
1192 | | |
1193 | 3 | std::vector<RowsetSharedPtr> specified_rowsets; |
1194 | 3 | { |
1195 | 3 | std::shared_lock meta_rlock(tablet->_meta_lock); |
1196 | 3 | cur_version = tablet->max_version_unlocked(); |
1197 | 3 | RETURN_IF_ERROR(tablet->get_all_rs_id_unlocked(cur_version, &cur_rowset_ids)); |
1198 | 3 | _rowset_ids_difference(cur_rowset_ids, pre_rowset_ids, &rowset_ids_to_add, |
1199 | 3 | &rowset_ids_to_del); |
1200 | 3 | specified_rowsets = tablet->get_rowset_by_ids(&rowset_ids_to_add); |
1201 | 3 | } |
1202 | 0 | for (const auto& to_del : rowset_ids_to_del) { |
1203 | 0 | delete_bitmap->remove({to_del, 0, 0}, {to_del, UINT32_MAX, INT64_MAX}); |
1204 | 0 | } |
1205 | | |
1206 | 3 | RETURN_IF_ERROR(calc_delete_bitmap(tablet, rowset, segments, specified_rowsets, delete_bitmap, |
1207 | 3 | cur_version, token, rowset_writer)); |
1208 | 3 | size_t total_rows = std::accumulate( |
1209 | 3 | segments.begin(), segments.end(), 0, |
1210 | 3 | [](size_t sum, const segment_v2::SegmentSharedPtr& s) { return sum += s->num_rows(); }); |
1211 | 3 | LOG(INFO) << "[Before Commit] construct delete bitmap tablet: " << tablet->tablet_id() |
1212 | 3 | << ", rowset_ids to add: " << rowset_ids_to_add.size() |
1213 | 3 | << ", rowset_ids to del: " << rowset_ids_to_del.size() |
1214 | 3 | << ", cur max_version: " << cur_version << ", transaction_id: " << txn_id |
1215 | 3 | << ", total rows: " << total_rows; |
1216 | 3 | pre_rowset_ids = cur_rowset_ids; |
1217 | 3 | return Status::OK(); |
1218 | 3 | } |
1219 | | |
1220 | | void BaseTablet::add_sentinel_mark_to_delete_bitmap(DeleteBitmap* delete_bitmap, |
1221 | 5 | const RowsetIdUnorderedSet& rowsetids) { |
1222 | 5 | for (const auto& rowsetid : rowsetids) { |
1223 | 5 | delete_bitmap->add( |
1224 | 5 | {rowsetid, DeleteBitmap::INVALID_SEGMENT_ID, DeleteBitmap::TEMP_VERSION_COMMON}, |
1225 | 5 | DeleteBitmap::ROWSET_SENTINEL_MARK); |
1226 | 5 | } |
1227 | 5 | } |
1228 | | |
1229 | | void BaseTablet::_rowset_ids_difference(const RowsetIdUnorderedSet& cur, |
1230 | | const RowsetIdUnorderedSet& pre, |
1231 | | RowsetIdUnorderedSet* to_add, |
1232 | 6 | RowsetIdUnorderedSet* to_del) { |
1233 | 6 | for (const auto& id : cur) { |
1234 | 2 | if (pre.find(id) == pre.end()) { |
1235 | 1 | to_add->insert(id); |
1236 | 1 | } |
1237 | 2 | } |
1238 | 6 | for (const auto& id : pre) { |
1239 | 1 | if (cur.find(id) == cur.end()) { |
1240 | 0 | to_del->insert(id); |
1241 | 0 | } |
1242 | 1 | } |
1243 | 6 | } |
1244 | | |
1245 | | Status BaseTablet::_capture_consistent_rowsets_unlocked( |
1246 | 2 | const std::vector<Version>& version_path, std::vector<RowsetSharedPtr>* rowsets) const { |
1247 | 2 | DCHECK(rowsets != nullptr); |
1248 | 2 | rowsets->reserve(version_path.size()); |
1249 | 4 | for (const auto& version : version_path) { |
1250 | 4 | bool is_find = false; |
1251 | 4 | do { |
1252 | 4 | auto it = _rs_version_map.find(version); |
1253 | 4 | if (it != _rs_version_map.end()) { |
1254 | 4 | is_find = true; |
1255 | 4 | rowsets->push_back(it->second); |
1256 | 4 | break; |
1257 | 4 | } |
1258 | | |
1259 | 0 | auto it_expired = _stale_rs_version_map.find(version); |
1260 | 0 | if (it_expired != _stale_rs_version_map.end()) { |
1261 | 0 | is_find = true; |
1262 | 0 | rowsets->push_back(it_expired->second); |
1263 | 0 | break; |
1264 | 0 | } |
1265 | 0 | } while (false); |
1266 | | |
1267 | 4 | if (!is_find) { |
1268 | 0 | return Status::Error<CAPTURE_ROWSET_ERROR>( |
1269 | 0 | "fail to find Rowset for version. tablet={}, version={}", tablet_id(), |
1270 | 0 | version.to_string()); |
1271 | 0 | } |
1272 | 4 | } |
1273 | 2 | return Status::OK(); |
1274 | 2 | } |
1275 | | |
1276 | | Status BaseTablet::check_delete_bitmap_correctness(DeleteBitmapPtr delete_bitmap, |
1277 | | int64_t max_version, int64_t txn_id, |
1278 | | const RowsetIdUnorderedSet& rowset_ids, |
1279 | 6 | std::vector<RowsetSharedPtr>* rowsets) { |
1280 | 6 | RowsetIdUnorderedSet missing_ids; |
1281 | 6 | for (const auto& rowsetid : rowset_ids) { |
1282 | 2 | if (!delete_bitmap->delete_bitmap.contains({rowsetid, DeleteBitmap::INVALID_SEGMENT_ID, |
1283 | 2 | DeleteBitmap::TEMP_VERSION_COMMON})) { |
1284 | 0 | missing_ids.insert(rowsetid); |
1285 | 0 | } |
1286 | 2 | } |
1287 | | |
1288 | 6 | if (!missing_ids.empty()) { |
1289 | 0 | LOG(WARNING) << "[txn_id:" << txn_id << "][tablet_id:" << tablet_id() |
1290 | 0 | << "][max_version: " << max_version |
1291 | 0 | << "] check delete bitmap correctness failed!"; |
1292 | 0 | rapidjson::Document root; |
1293 | 0 | root.SetObject(); |
1294 | 0 | rapidjson::Document required_rowsets_arr; |
1295 | 0 | required_rowsets_arr.SetArray(); |
1296 | 0 | rapidjson::Document missing_rowsets_arr; |
1297 | 0 | missing_rowsets_arr.SetArray(); |
1298 | |
|
1299 | 0 | if (rowsets != nullptr) { |
1300 | 0 | for (const auto& rowset : *rowsets) { |
1301 | 0 | rapidjson::Value value; |
1302 | 0 | std::string version_str = rowset->get_rowset_info_str(); |
1303 | 0 | value.SetString(version_str.c_str(), version_str.length(), |
1304 | 0 | required_rowsets_arr.GetAllocator()); |
1305 | 0 | required_rowsets_arr.PushBack(value, required_rowsets_arr.GetAllocator()); |
1306 | 0 | } |
1307 | 0 | } else { |
1308 | 0 | std::vector<RowsetSharedPtr> rowsets; |
1309 | 0 | { |
1310 | 0 | std::shared_lock meta_rlock(_meta_lock); |
1311 | 0 | rowsets = get_rowset_by_ids(&rowset_ids); |
1312 | 0 | } |
1313 | 0 | for (const auto& rowset : rowsets) { |
1314 | 0 | rapidjson::Value value; |
1315 | 0 | std::string version_str = rowset->get_rowset_info_str(); |
1316 | 0 | value.SetString(version_str.c_str(), version_str.length(), |
1317 | 0 | required_rowsets_arr.GetAllocator()); |
1318 | 0 | required_rowsets_arr.PushBack(value, required_rowsets_arr.GetAllocator()); |
1319 | 0 | } |
1320 | 0 | } |
1321 | 0 | for (const auto& missing_rowset_id : missing_ids) { |
1322 | 0 | rapidjson::Value miss_value; |
1323 | 0 | std::string rowset_id_str = missing_rowset_id.to_string(); |
1324 | 0 | miss_value.SetString(rowset_id_str.c_str(), rowset_id_str.length(), |
1325 | 0 | missing_rowsets_arr.GetAllocator()); |
1326 | 0 | missing_rowsets_arr.PushBack(miss_value, missing_rowsets_arr.GetAllocator()); |
1327 | 0 | } |
1328 | |
|
1329 | 0 | root.AddMember("required_rowsets", required_rowsets_arr, root.GetAllocator()); |
1330 | 0 | root.AddMember("missing_rowsets", missing_rowsets_arr, root.GetAllocator()); |
1331 | 0 | rapidjson::StringBuffer strbuf; |
1332 | 0 | rapidjson::PrettyWriter<rapidjson::StringBuffer> writer(strbuf); |
1333 | 0 | root.Accept(writer); |
1334 | 0 | std::string rowset_status_string = std::string(strbuf.GetString()); |
1335 | 0 | LOG_EVERY_SECOND(WARNING) << rowset_status_string; |
1336 | | // let it crash if correctness check failed in Debug mode |
1337 | 0 | DCHECK(false) << "delete bitmap correctness check failed in publish phase!"; |
1338 | 0 | return Status::InternalError("check delete bitmap failed!"); |
1339 | 0 | } |
1340 | 6 | return Status::OK(); |
1341 | 6 | } |
1342 | | |
1343 | | Status BaseTablet::update_delete_bitmap(const BaseTabletSPtr& self, TabletTxnInfo* txn_info, |
1344 | 3 | int64_t txn_id, int64_t txn_expiration) { |
1345 | 3 | SCOPED_BVAR_LATENCY(g_tablet_update_delete_bitmap_latency); |
1346 | 3 | RowsetIdUnorderedSet cur_rowset_ids; |
1347 | 3 | RowsetIdUnorderedSet rowset_ids_to_add; |
1348 | 3 | RowsetIdUnorderedSet rowset_ids_to_del; |
1349 | 3 | RowsetSharedPtr rowset = txn_info->rowset; |
1350 | 3 | int64_t cur_version = rowset->start_version(); |
1351 | | |
1352 | 3 | std::unique_ptr<RowsetWriter> transient_rs_writer; |
1353 | 3 | DeleteBitmapPtr delete_bitmap = txn_info->delete_bitmap; |
1354 | 3 | bool is_partial_update = |
1355 | 3 | txn_info->partial_update_info && txn_info->partial_update_info->is_partial_update(); |
1356 | 3 | if (is_partial_update) { |
1357 | 0 | transient_rs_writer = DORIS_TRY(self->create_transient_rowset_writer( |
1358 | 0 | *rowset, txn_info->partial_update_info, txn_expiration)); |
1359 | | // Partial update might generate new segments when there is conflicts while publish, and mark |
1360 | | // the same key in original segments as delete. |
1361 | | // When the new segment flush fails or the rowset build fails, the deletion marker for the |
1362 | | // duplicate key of the original segment should not remain in `txn_info->delete_bitmap`, |
1363 | | // so we need to make a copy of `txn_info->delete_bitmap` and make changes on it. |
1364 | 0 | delete_bitmap = std::make_shared<DeleteBitmap>(*(txn_info->delete_bitmap)); |
1365 | 0 | } |
1366 | | |
1367 | 3 | OlapStopWatch watch; |
1368 | 3 | std::vector<segment_v2::SegmentSharedPtr> segments; |
1369 | 3 | RETURN_IF_ERROR(std::dynamic_pointer_cast<BetaRowset>(rowset)->load_segments(&segments)); |
1370 | 3 | auto t1 = watch.get_elapse_time_us(); |
1371 | | |
1372 | 3 | { |
1373 | 3 | std::shared_lock meta_rlock(self->_meta_lock); |
1374 | | // tablet is under alter process. The delete bitmap will be calculated after conversion. |
1375 | 3 | if (self->tablet_state() == TABLET_NOTREADY) { |
1376 | 0 | LOG(INFO) << "tablet is under alter process, update delete bitmap later, tablet_id=" |
1377 | 0 | << self->tablet_id(); |
1378 | 0 | return Status::OK(); |
1379 | 0 | } |
1380 | 3 | RETURN_IF_ERROR(self->get_all_rs_id_unlocked(cur_version - 1, &cur_rowset_ids)); |
1381 | 3 | } |
1382 | 3 | auto t2 = watch.get_elapse_time_us(); |
1383 | | |
1384 | 3 | _rowset_ids_difference(cur_rowset_ids, txn_info->rowset_ids, &rowset_ids_to_add, |
1385 | 3 | &rowset_ids_to_del); |
1386 | 3 | for (const auto& to_del : rowset_ids_to_del) { |
1387 | 0 | delete_bitmap->remove({to_del, 0, 0}, {to_del, UINT32_MAX, INT64_MAX}); |
1388 | 0 | } |
1389 | | |
1390 | 3 | std::vector<RowsetSharedPtr> specified_rowsets; |
1391 | 3 | { |
1392 | 3 | std::shared_lock meta_rlock(self->_meta_lock); |
1393 | 3 | specified_rowsets = self->get_rowset_by_ids(&rowset_ids_to_add); |
1394 | 3 | } |
1395 | 3 | auto t3 = watch.get_elapse_time_us(); |
1396 | | |
1397 | | // If a rowset is produced by compaction before the commit phase of the partial update load |
1398 | | // and is not included in txn_info->rowset_ids, we can skip the alignment process of that rowset |
1399 | | // because data remains the same before and after compaction. But we still need to calculate the |
1400 | | // the delete bitmap for that rowset. |
1401 | 3 | std::vector<RowsetSharedPtr> rowsets_skip_alignment; |
1402 | 3 | if (is_partial_update) { |
1403 | 0 | int64_t max_version_in_flush_phase = |
1404 | 0 | txn_info->partial_update_info->max_version_in_flush_phase; |
1405 | 0 | DCHECK(max_version_in_flush_phase != -1); |
1406 | 0 | std::vector<RowsetSharedPtr> remained_rowsets; |
1407 | 0 | for (const auto& rowset : specified_rowsets) { |
1408 | 0 | if (rowset->end_version() <= max_version_in_flush_phase && |
1409 | 0 | rowset->produced_by_compaction()) { |
1410 | 0 | rowsets_skip_alignment.emplace_back(rowset); |
1411 | 0 | } else { |
1412 | 0 | remained_rowsets.emplace_back(rowset); |
1413 | 0 | } |
1414 | 0 | } |
1415 | 0 | if (!rowsets_skip_alignment.empty()) { |
1416 | 0 | specified_rowsets = std::move(remained_rowsets); |
1417 | 0 | } |
1418 | 0 | } |
1419 | | |
1420 | 3 | DBUG_EXECUTE_IF("BaseTablet::update_delete_bitmap.enable_spin_wait", { |
1421 | 3 | auto token = dp->param<std::string>("token", "invalid_token"); |
1422 | 3 | while (DebugPoints::instance()->is_enable("BaseTablet::update_delete_bitmap.block")) { |
1423 | 3 | auto block_dp = DebugPoints::instance()->get_debug_point( |
1424 | 3 | "BaseTablet::update_delete_bitmap.block"); |
1425 | 3 | if (block_dp) { |
1426 | 3 | auto wait_token = block_dp->param<std::string>("wait_token", ""); |
1427 | 3 | if (wait_token != token) { |
1428 | 3 | break; |
1429 | 3 | } |
1430 | 3 | } |
1431 | 3 | std::this_thread::sleep_for(std::chrono::milliseconds(50)); |
1432 | 3 | } |
1433 | 3 | }); |
1434 | | |
1435 | 3 | if (!rowsets_skip_alignment.empty()) { |
1436 | 0 | auto token = self->calc_delete_bitmap_executor()->create_token(); |
1437 | | // set rowset_writer to nullptr to skip the alignment process |
1438 | 0 | RETURN_IF_ERROR(calc_delete_bitmap(self, rowset, segments, rowsets_skip_alignment, |
1439 | 0 | delete_bitmap, cur_version - 1, token.get(), nullptr)); |
1440 | 0 | RETURN_IF_ERROR(token->wait()); |
1441 | 0 | } |
1442 | | |
1443 | | // When there is only one segment, it will be calculated in the current thread. |
1444 | | // Otherwise, it will be submitted to the thread pool for calculation. |
1445 | 3 | if (segments.size() <= 1) { |
1446 | 3 | RETURN_IF_ERROR(calc_delete_bitmap(self, rowset, segments, specified_rowsets, delete_bitmap, |
1447 | 3 | cur_version - 1, nullptr, transient_rs_writer.get())); |
1448 | | |
1449 | 3 | } else { |
1450 | 0 | auto token = self->calc_delete_bitmap_executor()->create_token(); |
1451 | 0 | RETURN_IF_ERROR(calc_delete_bitmap(self, rowset, segments, specified_rowsets, delete_bitmap, |
1452 | 0 | cur_version - 1, token.get(), |
1453 | 0 | transient_rs_writer.get())); |
1454 | 0 | RETURN_IF_ERROR(token->wait()); |
1455 | 0 | } |
1456 | | |
1457 | 3 | std::stringstream ss; |
1458 | 3 | ss << "cost(us): (load segments: " << t1 << ", get all rsid: " << t2 - t1 |
1459 | 3 | << ", get rowsets: " << t3 - t2 |
1460 | 3 | << ", calc delete bitmap: " << watch.get_elapse_time_us() - t3 << ")"; |
1461 | | |
1462 | 3 | if (config::enable_merge_on_write_correctness_check && rowset->num_rows() != 0) { |
1463 | | // only do correctness check if the rowset has at least one row written |
1464 | | // check if all the rowset has ROWSET_SENTINEL_MARK |
1465 | 3 | auto st = self->check_delete_bitmap_correctness(delete_bitmap, cur_version - 1, -1, |
1466 | 3 | cur_rowset_ids, &specified_rowsets); |
1467 | 3 | if (!st.ok()) { |
1468 | 0 | LOG(WARNING) << fmt::format("delete bitmap correctness check failed in publish phase!"); |
1469 | 0 | } |
1470 | 3 | } |
1471 | | |
1472 | 3 | if (transient_rs_writer) { |
1473 | 0 | auto t4 = watch.get_elapse_time_us(); |
1474 | 0 | DBUG_EXECUTE_IF("Tablet.update_delete_bitmap.partial_update_write_rowset_fail", { |
1475 | 0 | if (rand() % 100 < (100 * dp->param("percent", 0.5))) { |
1476 | 0 | LOG_WARNING("Tablet.update_delete_bitmap.partial_update_write_rowset random failed") |
1477 | 0 | .tag("txn_id", txn_id); |
1478 | 0 | return Status::InternalError( |
1479 | 0 | "debug update_delete_bitmap partial update write rowset random failed"); |
1480 | 0 | } |
1481 | 0 | }); |
1482 | | // build rowset writer and merge transient rowset |
1483 | 0 | RETURN_IF_ERROR(transient_rs_writer->flush()); |
1484 | 0 | RowsetSharedPtr transient_rowset; |
1485 | 0 | RETURN_IF_ERROR(transient_rs_writer->build(transient_rowset)); |
1486 | 0 | auto old_segments = rowset->num_segments(); |
1487 | 0 | rowset->merge_rowset_meta(*transient_rowset->rowset_meta()); |
1488 | 0 | auto new_segments = rowset->num_segments(); |
1489 | 0 | ss << ", " << txn_info->partial_update_info->partial_update_mode_str() |
1490 | 0 | << " flush rowset (old segment num: " << old_segments |
1491 | 0 | << ", new segment num: " << new_segments << ")" |
1492 | 0 | << ", cost:" << watch.get_elapse_time_us() - t4 << "(us)"; |
1493 | | |
1494 | | // update the shared_ptr to new bitmap, which is consistent with current rowset. |
1495 | 0 | txn_info->delete_bitmap = delete_bitmap; |
1496 | | // erase segment cache cause we will add a segment to rowset |
1497 | 0 | SegmentLoader::instance()->erase_segments(rowset->rowset_id(), rowset->num_segments()); |
1498 | 0 | } |
1499 | | |
1500 | 3 | size_t total_rows = std::accumulate( |
1501 | 3 | segments.begin(), segments.end(), 0, |
1502 | 3 | [](size_t sum, const segment_v2::SegmentSharedPtr& s) { return sum += s->num_rows(); }); |
1503 | 3 | auto t5 = watch.get_elapse_time_us(); |
1504 | 3 | RETURN_IF_ERROR(self->save_delete_bitmap(txn_info, txn_id, delete_bitmap, |
1505 | 3 | transient_rs_writer.get(), cur_rowset_ids)); |
1506 | 3 | LOG(INFO) << "[Publish] construct delete bitmap tablet: " << self->tablet_id() |
1507 | 3 | << ", rowset_ids to add: " << rowset_ids_to_add.size() |
1508 | 3 | << ", rowset_ids to del: " << rowset_ids_to_del.size() |
1509 | 3 | << ", cur version: " << cur_version << ", transaction_id: " << txn_id << "," |
1510 | 3 | << ss.str() << " , total rows: " << total_rows |
1511 | 3 | << ", update delete_bitmap cost: " << watch.get_elapse_time_us() - t5 << "(us)"; |
1512 | 3 | return Status::OK(); |
1513 | 3 | } |
1514 | | |
1515 | | void BaseTablet::calc_compaction_output_rowset_delete_bitmap( |
1516 | | const std::vector<RowsetSharedPtr>& input_rowsets, const RowIdConversion& rowid_conversion, |
1517 | | uint64_t start_version, uint64_t end_version, std::set<RowLocation>* missed_rows, |
1518 | | std::map<RowsetSharedPtr, std::list<std::pair<RowLocation, RowLocation>>>* location_map, |
1519 | 0 | const DeleteBitmap& input_delete_bitmap, DeleteBitmap* output_rowset_delete_bitmap) { |
1520 | 0 | RowLocation src; |
1521 | 0 | RowLocation dst; |
1522 | 0 | for (auto& rowset : input_rowsets) { |
1523 | 0 | src.rowset_id = rowset->rowset_id(); |
1524 | 0 | for (uint32_t seg_id = 0; seg_id < rowset->num_segments(); ++seg_id) { |
1525 | 0 | src.segment_id = seg_id; |
1526 | 0 | DeleteBitmap subset_map(tablet_id()); |
1527 | 0 | input_delete_bitmap.subset({rowset->rowset_id(), seg_id, start_version}, |
1528 | 0 | {rowset->rowset_id(), seg_id, end_version}, &subset_map); |
1529 | | // traverse all versions and convert rowid |
1530 | 0 | for (auto iter = subset_map.delete_bitmap.begin(); |
1531 | 0 | iter != subset_map.delete_bitmap.end(); ++iter) { |
1532 | 0 | auto cur_version = std::get<2>(iter->first); |
1533 | 0 | for (auto index = iter->second.begin(); index != iter->second.end(); ++index) { |
1534 | 0 | src.row_id = *index; |
1535 | 0 | if (rowid_conversion.get(src, &dst) != 0) { |
1536 | 0 | VLOG_CRITICAL << "Can't find rowid, may be deleted by the delete_handler, " |
1537 | 0 | << " src loaction: |" << src.rowset_id << "|" |
1538 | 0 | << src.segment_id << "|" << src.row_id |
1539 | 0 | << " version: " << cur_version; |
1540 | 0 | if (missed_rows) { |
1541 | 0 | missed_rows->insert(src); |
1542 | 0 | } |
1543 | 0 | continue; |
1544 | 0 | } |
1545 | 0 | VLOG_DEBUG << "calc_compaction_output_rowset_delete_bitmap dst location: |" |
1546 | 0 | << dst.rowset_id << "|" << dst.segment_id << "|" << dst.row_id |
1547 | 0 | << " src location: |" << src.rowset_id << "|" << src.segment_id |
1548 | 0 | << "|" << src.row_id << " start version: " << start_version |
1549 | 0 | << "end version" << end_version; |
1550 | 0 | if (location_map) { |
1551 | 0 | (*location_map)[rowset].emplace_back(src, dst); |
1552 | 0 | } |
1553 | 0 | output_rowset_delete_bitmap->add({dst.rowset_id, dst.segment_id, cur_version}, |
1554 | 0 | dst.row_id); |
1555 | 0 | } |
1556 | 0 | } |
1557 | 0 | } |
1558 | 0 | } |
1559 | 0 | } |
1560 | | |
1561 | | Status BaseTablet::check_rowid_conversion( |
1562 | | RowsetSharedPtr dst_rowset, |
1563 | | const std::map<RowsetSharedPtr, std::list<std::pair<RowLocation, RowLocation>>>& |
1564 | 0 | location_map) { |
1565 | 0 | if (location_map.empty()) { |
1566 | 0 | VLOG_DEBUG << "check_rowid_conversion, location_map is empty"; |
1567 | 0 | return Status::OK(); |
1568 | 0 | } |
1569 | 0 | if (!tablet_schema()->cluster_key_idxes().empty()) { |
1570 | 0 | VLOG_DEBUG << "skip check_rowid_conversion for mow tables with cluster keys"; |
1571 | 0 | return Status::OK(); |
1572 | 0 | } |
1573 | 0 | std::vector<segment_v2::SegmentSharedPtr> dst_segments; |
1574 | |
|
1575 | 0 | RETURN_IF_ERROR( |
1576 | 0 | std::dynamic_pointer_cast<BetaRowset>(dst_rowset)->load_segments(&dst_segments)); |
1577 | 0 | std::unordered_map<RowsetId, std::vector<segment_v2::SegmentSharedPtr>> input_rowsets_segment; |
1578 | |
|
1579 | 0 | VLOG_DEBUG << "check_rowid_conversion, dst_segments size: " << dst_segments.size(); |
1580 | 0 | for (auto [src_rowset, locations] : location_map) { |
1581 | 0 | std::vector<segment_v2::SegmentSharedPtr>& segments = |
1582 | 0 | input_rowsets_segment[src_rowset->rowset_id()]; |
1583 | 0 | if (segments.empty()) { |
1584 | 0 | RETURN_IF_ERROR( |
1585 | 0 | std::dynamic_pointer_cast<BetaRowset>(src_rowset)->load_segments(&segments)); |
1586 | 0 | } |
1587 | 0 | for (auto& [src, dst] : locations) { |
1588 | 0 | std::string src_key; |
1589 | 0 | std::string dst_key; |
1590 | 0 | Status s = segments[src.segment_id]->read_key_by_rowid(src.row_id, &src_key); |
1591 | 0 | if (UNLIKELY(s.is<NOT_IMPLEMENTED_ERROR>())) { |
1592 | 0 | LOG(INFO) << "primary key index of old version does not " |
1593 | 0 | "support reading key by rowid"; |
1594 | 0 | break; |
1595 | 0 | } |
1596 | 0 | if (UNLIKELY(!s)) { |
1597 | 0 | LOG(WARNING) << "failed to get src key: |" << src.rowset_id << "|" << src.segment_id |
1598 | 0 | << "|" << src.row_id << " status: " << s; |
1599 | 0 | DCHECK(false); |
1600 | 0 | return s; |
1601 | 0 | } |
1602 | | |
1603 | 0 | s = dst_segments[dst.segment_id]->read_key_by_rowid(dst.row_id, &dst_key); |
1604 | 0 | if (UNLIKELY(!s)) { |
1605 | 0 | LOG(WARNING) << "failed to get dst key: |" << dst.rowset_id << "|" << dst.segment_id |
1606 | 0 | << "|" << dst.row_id << " status: " << s; |
1607 | 0 | DCHECK(false); |
1608 | 0 | return s; |
1609 | 0 | } |
1610 | | |
1611 | 0 | VLOG_DEBUG << "check_rowid_conversion, src: |" << src.rowset_id << "|" << src.segment_id |
1612 | 0 | << "|" << src.row_id << "|" << src_key << " dst: |" << dst.rowset_id << "|" |
1613 | 0 | << dst.segment_id << "|" << dst.row_id << "|" << dst_key; |
1614 | 0 | if (UNLIKELY(src_key.compare(dst_key) != 0)) { |
1615 | 0 | LOG(WARNING) << "failed to check key, src key: |" << src.rowset_id << "|" |
1616 | 0 | << src.segment_id << "|" << src.row_id << "|" << src_key |
1617 | 0 | << " dst key: |" << dst.rowset_id << "|" << dst.segment_id << "|" |
1618 | 0 | << dst.row_id << "|" << dst_key; |
1619 | 0 | DCHECK(false); |
1620 | 0 | return Status::InternalError("failed to check rowid conversion"); |
1621 | 0 | } |
1622 | 0 | } |
1623 | 0 | } |
1624 | 0 | return Status::OK(); |
1625 | 0 | } |
1626 | | |
1627 | | // The caller should hold _rowset_update_lock and _meta_lock lock. |
1628 | | Status BaseTablet::update_delete_bitmap_without_lock( |
1629 | | const BaseTabletSPtr& self, const RowsetSharedPtr& rowset, |
1630 | 0 | const std::vector<RowsetSharedPtr>* specified_base_rowsets) { |
1631 | 0 | DBUG_EXECUTE_IF("BaseTablet.update_delete_bitmap_without_lock.random_failed", { |
1632 | 0 | auto rnd = rand() % 100; |
1633 | 0 | auto percent = dp->param("percent", 0.1); |
1634 | 0 | if (rnd < (100 * percent)) { |
1635 | 0 | LOG(WARNING) << "BaseTablet.update_delete_bitmap_without_lock.random_failed"; |
1636 | 0 | return Status::InternalError( |
1637 | 0 | "debug tablet update delete bitmap without lock random failed"); |
1638 | 0 | } else { |
1639 | 0 | LOG(INFO) << "BaseTablet.update_delete_bitmap_without_lock.random_failed not " |
1640 | 0 | "triggered" |
1641 | 0 | << ", rnd:" << rnd << ", percent: " << percent; |
1642 | 0 | } |
1643 | 0 | }); |
1644 | 0 | int64_t cur_version = rowset->start_version(); |
1645 | 0 | std::vector<segment_v2::SegmentSharedPtr> segments; |
1646 | 0 | RETURN_IF_ERROR(std::dynamic_pointer_cast<BetaRowset>(rowset)->load_segments(&segments)); |
1647 | | |
1648 | | // If this rowset does not have a segment, there is no need for an update. |
1649 | 0 | if (segments.empty()) { |
1650 | 0 | LOG(INFO) << "[Schema Change or Clone] skip to construct delete bitmap tablet: " |
1651 | 0 | << self->tablet_id() << " cur max_version: " << cur_version; |
1652 | 0 | return Status::OK(); |
1653 | 0 | } |
1654 | | |
1655 | | // calculate delete bitmap between segments if necessary. |
1656 | 0 | DeleteBitmapPtr delete_bitmap = std::make_shared<DeleteBitmap>(self->tablet_id()); |
1657 | 0 | RETURN_IF_ERROR(self->calc_delete_bitmap_between_segments(rowset, segments, delete_bitmap)); |
1658 | | |
1659 | | // get all base rowsets to calculate on |
1660 | 0 | std::vector<RowsetSharedPtr> specified_rowsets; |
1661 | 0 | RowsetIdUnorderedSet cur_rowset_ids; |
1662 | 0 | if (specified_base_rowsets == nullptr) { |
1663 | 0 | RETURN_IF_ERROR(self->get_all_rs_id_unlocked(cur_version - 1, &cur_rowset_ids)); |
1664 | 0 | specified_rowsets = self->get_rowset_by_ids(&cur_rowset_ids); |
1665 | 0 | } else { |
1666 | 0 | specified_rowsets = *specified_base_rowsets; |
1667 | 0 | } |
1668 | | |
1669 | 0 | OlapStopWatch watch; |
1670 | 0 | auto token = self->calc_delete_bitmap_executor()->create_token(); |
1671 | 0 | RETURN_IF_ERROR(calc_delete_bitmap(self, rowset, segments, specified_rowsets, delete_bitmap, |
1672 | 0 | cur_version - 1, token.get())); |
1673 | 0 | RETURN_IF_ERROR(token->wait()); |
1674 | 0 | size_t total_rows = std::accumulate( |
1675 | 0 | segments.begin(), segments.end(), 0, |
1676 | 0 | [](size_t sum, const segment_v2::SegmentSharedPtr& s) { return sum += s->num_rows(); }); |
1677 | 0 | LOG(INFO) << "[Schema Change or Clone] construct delete bitmap tablet: " << self->tablet_id() |
1678 | 0 | << ", rowset_ids: " << cur_rowset_ids.size() << ", cur max_version: " << cur_version |
1679 | 0 | << ", transaction_id: " << -1 << ", cost: " << watch.get_elapse_time_us() |
1680 | 0 | << "(us), total rows: " << total_rows; |
1681 | 0 | if (config::enable_merge_on_write_correctness_check) { |
1682 | | // check if all the rowset has ROWSET_SENTINEL_MARK |
1683 | 0 | auto st = self->check_delete_bitmap_correctness(delete_bitmap, cur_version - 1, -1, |
1684 | 0 | cur_rowset_ids, &specified_rowsets); |
1685 | 0 | if (!st.ok()) { |
1686 | 0 | LOG(WARNING) << fmt::format("delete bitmap correctness check failed in publish phase!"); |
1687 | 0 | } |
1688 | 0 | delete_bitmap->remove_sentinel_marks(); |
1689 | 0 | } |
1690 | 0 | for (auto& iter : delete_bitmap->delete_bitmap) { |
1691 | 0 | self->_tablet_meta->delete_bitmap().merge( |
1692 | 0 | {std::get<0>(iter.first), std::get<1>(iter.first), cur_version}, iter.second); |
1693 | 0 | } |
1694 | |
|
1695 | 0 | return Status::OK(); |
1696 | 0 | } |
1697 | | |
1698 | 0 | RowsetSharedPtr BaseTablet::get_rowset(const RowsetId& rowset_id) { |
1699 | 0 | std::shared_lock rdlock(_meta_lock); |
1700 | 0 | for (auto& version_rowset : _rs_version_map) { |
1701 | 0 | if (version_rowset.second->rowset_id() == rowset_id) { |
1702 | 0 | return version_rowset.second; |
1703 | 0 | } |
1704 | 0 | } |
1705 | 0 | for (auto& stale_version_rowset : _stale_rs_version_map) { |
1706 | 0 | if (stale_version_rowset.second->rowset_id() == rowset_id) { |
1707 | 0 | return stale_version_rowset.second; |
1708 | 0 | } |
1709 | 0 | } |
1710 | 0 | return nullptr; |
1711 | 0 | } |
1712 | | |
1713 | 1 | std::vector<RowsetSharedPtr> BaseTablet::get_snapshot_rowset(bool include_stale_rowset) const { |
1714 | 1 | std::shared_lock rdlock(_meta_lock); |
1715 | 1 | std::vector<RowsetSharedPtr> rowsets; |
1716 | 1 | std::transform(_rs_version_map.cbegin(), _rs_version_map.cend(), std::back_inserter(rowsets), |
1717 | 28 | [](auto& kv) { return kv.second; }); |
1718 | 1 | if (include_stale_rowset) { |
1719 | 0 | std::transform(_stale_rs_version_map.cbegin(), _stale_rs_version_map.cend(), |
1720 | 0 | std::back_inserter(rowsets), [](auto& kv) { return kv.second; }); |
1721 | 0 | } |
1722 | 1 | return rowsets; |
1723 | 1 | } |
1724 | | |
1725 | | void BaseTablet::calc_consecutive_empty_rowsets( |
1726 | | std::vector<RowsetSharedPtr>* empty_rowsets, |
1727 | 3 | const std::vector<RowsetSharedPtr>& candidate_rowsets, int limit) { |
1728 | 3 | int len = candidate_rowsets.size(); |
1729 | 7 | for (int i = 0; i < len - 1; ++i) { |
1730 | 5 | auto rowset = candidate_rowsets[i]; |
1731 | 5 | auto next_rowset = candidate_rowsets[i + 1]; |
1732 | | |
1733 | | // identify two consecutive rowsets that are empty |
1734 | 5 | if (rowset->num_segments() == 0 && next_rowset->num_segments() == 0 && |
1735 | 5 | !rowset->rowset_meta()->has_delete_predicate() && |
1736 | 5 | !next_rowset->rowset_meta()->has_delete_predicate() && |
1737 | 5 | rowset->end_version() == next_rowset->start_version() - 1) { |
1738 | 1 | empty_rowsets->emplace_back(rowset); |
1739 | 1 | empty_rowsets->emplace_back(next_rowset); |
1740 | 1 | rowset = next_rowset; |
1741 | 1 | int next_index = i + 2; |
1742 | | |
1743 | | // keep searching for consecutive empty rowsets |
1744 | 6 | while (next_index < len && candidate_rowsets[next_index]->num_segments() == 0 && |
1745 | 6 | !candidate_rowsets[next_index]->rowset_meta()->has_delete_predicate() && |
1746 | 6 | rowset->end_version() == candidate_rowsets[next_index]->start_version() - 1) { |
1747 | 5 | empty_rowsets->emplace_back(candidate_rowsets[next_index]); |
1748 | 5 | rowset = candidate_rowsets[next_index++]; |
1749 | 5 | } |
1750 | | // if the number of consecutive empty rowset reach the limit, |
1751 | | // and there are still rowsets following them |
1752 | 1 | if (empty_rowsets->size() >= limit && next_index < len) { |
1753 | 1 | return; |
1754 | 1 | } else { |
1755 | | // current rowset is not empty, start searching from that rowset in the next |
1756 | 0 | i = next_index - 1; |
1757 | 0 | empty_rowsets->clear(); |
1758 | 0 | } |
1759 | 1 | } |
1760 | 5 | } |
1761 | 3 | } |
1762 | | |
1763 | | Status BaseTablet::calc_file_crc(uint32_t* crc_value, int64_t start_version, int64_t end_version, |
1764 | 0 | int32_t* rowset_count, int64_t* file_count) { |
1765 | 0 | Version v(start_version, end_version); |
1766 | 0 | std::vector<RowsetSharedPtr> rowsets; |
1767 | 0 | traverse_rowsets([&rowsets, &v](const auto& rs) { |
1768 | | // get all rowsets |
1769 | 0 | if (v.contains(rs->version())) { |
1770 | 0 | rowsets.emplace_back(rs); |
1771 | 0 | } |
1772 | 0 | }); |
1773 | 0 | std::sort(rowsets.begin(), rowsets.end(), Rowset::comparator); |
1774 | 0 | *rowset_count = rowsets.size(); |
1775 | |
|
1776 | 0 | *crc_value = 0; |
1777 | 0 | *file_count = 0; |
1778 | 0 | for (const auto& rs : rowsets) { |
1779 | 0 | uint32_t rs_crc_value = 0; |
1780 | 0 | int64_t rs_file_count = 0; |
1781 | 0 | auto rowset = std::static_pointer_cast<BetaRowset>(rs); |
1782 | 0 | auto st = rowset->calc_file_crc(&rs_crc_value, &rs_file_count); |
1783 | 0 | if (!st.ok()) { |
1784 | 0 | return st; |
1785 | 0 | } |
1786 | | // crc_value is calculated based on the crc_value of each rowset. |
1787 | 0 | *crc_value = crc32c::Extend(*crc_value, reinterpret_cast<const char*>(&rs_crc_value), |
1788 | 0 | sizeof(rs_crc_value)); |
1789 | 0 | *file_count += rs_file_count; |
1790 | 0 | } |
1791 | 0 | return Status::OK(); |
1792 | 0 | } |
1793 | | |
1794 | 0 | Status BaseTablet::show_nested_index_file(std::string* json_meta) { |
1795 | 0 | Version v(0, max_version_unlocked()); |
1796 | 0 | std::vector<RowsetSharedPtr> rowsets; |
1797 | 0 | traverse_rowsets([&rowsets, &v](const auto& rs) { |
1798 | | // get all rowsets |
1799 | 0 | if (v.contains(rs->version())) { |
1800 | 0 | rowsets.emplace_back(rs); |
1801 | 0 | } |
1802 | 0 | }); |
1803 | 0 | std::sort(rowsets.begin(), rowsets.end(), Rowset::comparator); |
1804 | |
|
1805 | 0 | rapidjson::Document doc; |
1806 | 0 | doc.SetObject(); |
1807 | 0 | rapidjson::Document::AllocatorType& allocator = doc.GetAllocator(); |
1808 | 0 | rapidjson::Value tabletIdValue(tablet_id()); |
1809 | 0 | doc.AddMember("tablet_id", tabletIdValue, allocator); |
1810 | |
|
1811 | 0 | rapidjson::Value rowsets_value(rapidjson::kArrayType); |
1812 | |
|
1813 | 0 | for (const auto& rs : rowsets) { |
1814 | 0 | rapidjson::Value rowset_value(rapidjson::kObjectType); |
1815 | |
|
1816 | 0 | auto rowset = std::static_pointer_cast<BetaRowset>(rs); |
1817 | 0 | RETURN_IF_ERROR(rowset->show_nested_index_file(&rowset_value, allocator)); |
1818 | 0 | rowsets_value.PushBack(rowset_value, allocator); |
1819 | 0 | } |
1820 | 0 | doc.AddMember("rowsets", rowsets_value, allocator); |
1821 | |
|
1822 | 0 | rapidjson::StringBuffer buffer; |
1823 | 0 | rapidjson::PrettyWriter<rapidjson::StringBuffer> writer(buffer); |
1824 | 0 | doc.Accept(writer); |
1825 | 0 | *json_meta = std::string(buffer.GetString()); |
1826 | |
|
1827 | 0 | return Status::OK(); |
1828 | 0 | } |
1829 | | |
1830 | | } // namespace doris |