Coverage Report

Created: 2026-01-13 00:27

/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 <bthread/mutex.h>
21
#include <fmt/format.h>
22
#include <rapidjson/prettywriter.h>
23
24
#include <algorithm>
25
#include <random>
26
#include <shared_mutex>
27
28
#include "cloud/cloud_tablet.h"
29
#include "cloud/config.h"
30
#include "common/logging.h"
31
#include "common/status.h"
32
#include "olap/calc_delete_bitmap_executor.h"
33
#include "olap/cumulative_compaction_time_series_policy.h"
34
#include "olap/delete_bitmap_calculator.h"
35
#include "olap/iterators.h"
36
#include "olap/memtable.h"
37
#include "olap/partial_update_info.h"
38
#include "olap/primary_key_index.h"
39
#include "olap/rowid_conversion.h"
40
#include "olap/rowset/beta_rowset.h"
41
#include "olap/rowset/rowset.h"
42
#include "olap/rowset/rowset_fwd.h"
43
#include "olap/rowset/rowset_reader.h"
44
#include "olap/tablet_fwd.h"
45
#include "olap/txn_manager.h"
46
#include "service/point_query_executor.h"
47
#include "util/bvar_helper.h"
48
#include "util/crc32c.h"
49
#include "util/debug_points.h"
50
#include "util/doris_metrics.h"
51
#include "util/key_util.h"
52
#include "vec/common/assert_cast.h"
53
#include "vec/common/schema_util.h"
54
#include "vec/data_types/data_type_factory.hpp"
55
#include "vec/jsonb/serialize.h"
56
57
namespace doris {
58
using namespace ErrorCode;
59
60
namespace {
61
62
bvar::LatencyRecorder g_tablet_commit_phase_update_delete_bitmap_latency(
63
        "doris_pk", "commit_phase_update_delete_bitmap");
64
bvar::LatencyRecorder g_tablet_lookup_rowkey_latency("doris_pk", "tablet_lookup_rowkey");
65
bvar::Adder<uint64_t> g_tablet_pk_not_found("doris_pk", "lookup_not_found");
66
bvar::PerSecond<bvar::Adder<uint64_t>> g_tablet_pk_not_found_per_second(
67
        "doris_pk", "lookup_not_found_per_second", &g_tablet_pk_not_found, 60);
68
bvar::LatencyRecorder g_tablet_update_delete_bitmap_latency("doris_pk", "update_delete_bitmap");
69
70
static bvar::Adder<size_t> g_total_tablet_num("doris_total_tablet_num");
71
72
Status _get_segment_column_iterator(const BetaRowsetSharedPtr& rowset, uint32_t segid,
73
                                    const TabletColumn& target_column,
74
                                    SegmentCacheHandle* segment_cache_handle,
75
                                    std::unique_ptr<segment_v2::ColumnIterator>* column_iterator,
76
0
                                    OlapReaderStatistics* stats) {
77
0
    RETURN_IF_ERROR(SegmentLoader::instance()->load_segments(rowset, segment_cache_handle, true));
78
    // find segment
79
0
    auto it = std::find_if(
80
0
            segment_cache_handle->get_segments().begin(),
81
0
            segment_cache_handle->get_segments().end(),
82
0
            [&segid](const segment_v2::SegmentSharedPtr& seg) { return seg->id() == segid; });
83
0
    if (it == segment_cache_handle->get_segments().end()) {
84
0
        return Status::NotFound(fmt::format("rowset {} 's segemnt not found, seg_id {}",
85
0
                                            rowset->rowset_id().to_string(), segid));
86
0
    }
87
0
    segment_v2::SegmentSharedPtr segment = *it;
88
0
    StorageReadOptions opts;
89
0
    opts.stats = stats;
90
0
    RETURN_IF_ERROR(segment->new_column_iterator(target_column, column_iterator, &opts));
91
0
    segment_v2::ColumnIteratorOptions opt {
92
0
            .use_page_cache = !config::disable_storage_page_cache,
93
0
            .file_reader = segment->file_reader().get(),
94
0
            .stats = stats,
95
0
            .io_ctx = io::IOContext {.reader_type = ReaderType::READER_QUERY,
96
0
                                     .file_cache_stats = &stats->file_cache_stats},
97
0
    };
98
0
    RETURN_IF_ERROR((*column_iterator)->init(opt));
99
0
    return Status::OK();
100
0
}
101
102
} // namespace
103
104
extern MetricPrototype METRIC_query_scan_bytes;
105
extern MetricPrototype METRIC_query_scan_rows;
106
extern MetricPrototype METRIC_query_scan_count;
107
DEFINE_COUNTER_METRIC_PROTOTYPE_2ARG(flush_bytes, MetricUnit::BYTES);
108
DEFINE_COUNTER_METRIC_PROTOTYPE_2ARG(flush_finish_count, MetricUnit::OPERATIONS);
109
110
556
BaseTablet::BaseTablet(TabletMetaSharedPtr tablet_meta) : _tablet_meta(std::move(tablet_meta)) {
111
556
    _metric_entity = DorisMetrics::instance()->metric_registry()->register_entity(
112
556
            fmt::format("Tablet.{}", tablet_id()), {{"tablet_id", std::to_string(tablet_id())}},
113
556
            MetricEntityType::kTablet);
114
556
    INT_COUNTER_METRIC_REGISTER(_metric_entity, query_scan_bytes);
115
556
    INT_COUNTER_METRIC_REGISTER(_metric_entity, query_scan_rows);
116
556
    INT_COUNTER_METRIC_REGISTER(_metric_entity, query_scan_count);
117
556
    INT_COUNTER_METRIC_REGISTER(_metric_entity, flush_bytes);
118
556
    INT_COUNTER_METRIC_REGISTER(_metric_entity, flush_finish_count);
119
120
    // construct _timestamped_versioned_tracker from rs and stale rs meta
121
556
    _timestamped_version_tracker.construct_versioned_tracker(_tablet_meta->all_rs_metas(),
122
556
                                                             _tablet_meta->all_stale_rs_metas());
123
124
    // if !_tablet_meta->all_rs_metas()[0]->tablet_schema(),
125
    // that mean the tablet_meta is still no upgrade to doris 1.2 versions.
126
    // Before doris 1.2 version, rowset metas don't have tablet schema.
127
    // And when upgrade to doris 1.2 version,
128
    // all rowset metas will be set the tablet schmea from tablet meta.
129
556
    if (_tablet_meta->all_rs_metas().empty() || !_tablet_meta->all_rs_metas()[0]->tablet_schema()) {
130
519
        _max_version_schema = _tablet_meta->tablet_schema();
131
519
    } else {
132
37
        _max_version_schema =
133
37
                tablet_schema_with_merged_max_schema_version(_tablet_meta->all_rs_metas());
134
37
    }
135
556
    DCHECK(_max_version_schema);
136
556
    g_total_tablet_num << 1;
137
556
}
138
139
556
BaseTablet::~BaseTablet() {
140
556
    DorisMetrics::instance()->metric_registry()->deregister_entity(_metric_entity);
141
556
    g_total_tablet_num << -1;
142
556
}
143
144
TabletSchemaSPtr BaseTablet::tablet_schema_with_merged_max_schema_version(
145
76
        const std::vector<RowsetMetaSharedPtr>& rowset_metas) {
146
76
    RowsetMetaSharedPtr max_schema_version_rs = *std::max_element(
147
76
            rowset_metas.begin(), rowset_metas.end(),
148
245
            [](const RowsetMetaSharedPtr& a, const RowsetMetaSharedPtr& b) {
149
245
                return !a->tablet_schema()
150
245
                               ? true
151
245
                               : (!b->tablet_schema()
152
245
                                          ? false
153
245
                                          : a->tablet_schema()->schema_version() <
154
245
                                                    b->tablet_schema()->schema_version());
155
245
            });
156
76
    TabletSchemaSPtr target_schema = max_schema_version_rs->tablet_schema();
157
76
    if (target_schema->num_variant_columns() > 0) {
158
        // For variant columns tablet schema need to be the merged wide tablet schema
159
0
        std::vector<TabletSchemaSPtr> schemas;
160
0
        std::transform(rowset_metas.begin(), rowset_metas.end(), std::back_inserter(schemas),
161
0
                       [](const RowsetMetaSharedPtr& rs_meta) { return rs_meta->tablet_schema(); });
162
0
        static_cast<void>(
163
0
                vectorized::schema_util::get_least_common_schema(schemas, nullptr, target_schema));
164
0
        VLOG_DEBUG << "dump schema: " << target_schema->dump_full_schema();
165
0
    }
166
76
    return target_schema;
167
76
}
168
169
255
Status BaseTablet::set_tablet_state(TabletState state) {
170
255
    if (_tablet_meta->tablet_state() == TABLET_SHUTDOWN && state != TABLET_SHUTDOWN) {
171
0
        return Status::Error<META_INVALID_ARGUMENT>(
172
0
                "could not change tablet state from shutdown to {}", state);
173
0
    }
174
255
    _tablet_meta->set_tablet_state(state);
175
255
    return Status::OK();
176
255
}
177
178
0
void BaseTablet::update_max_version_schema(const TabletSchemaSPtr& tablet_schema) {
179
0
    std::lock_guard wrlock(_meta_lock);
180
    // Double Check for concurrent update
181
0
    if (!_max_version_schema ||
182
0
        tablet_schema->schema_version() > _max_version_schema->schema_version()) {
183
0
        _max_version_schema = tablet_schema;
184
0
    }
185
0
}
186
187
0
Status BaseTablet::update_by_least_common_schema(const TabletSchemaSPtr& update_schema) {
188
0
    std::lock_guard wrlock(_meta_lock);
189
0
    CHECK(_max_version_schema->schema_version() >= update_schema->schema_version());
190
0
    TabletSchemaSPtr final_schema;
191
0
    bool check_column_size = true;
192
0
    VLOG_DEBUG << "dump _max_version_schema: " << _max_version_schema->dump_full_schema();
193
0
    VLOG_DEBUG << "dump update_schema: " << update_schema->dump_full_schema();
194
0
    RETURN_IF_ERROR(vectorized::schema_util::get_least_common_schema(
195
0
            {_max_version_schema, update_schema}, _max_version_schema, final_schema,
196
0
            check_column_size));
197
0
    _max_version_schema = final_schema;
198
0
    VLOG_DEBUG << "dump updated tablet schema: " << final_schema->dump_full_schema();
199
0
    return Status::OK();
200
0
}
201
202
0
uint32_t BaseTablet::get_real_compaction_score() const {
203
0
    std::shared_lock l(_meta_lock);
204
0
    const auto& rs_metas = _tablet_meta->all_rs_metas();
205
0
    return std::accumulate(rs_metas.begin(), rs_metas.end(), 0,
206
0
                           [](uint32_t score, const RowsetMetaSharedPtr& rs_meta) {
207
0
                               return score + rs_meta->get_compaction_score();
208
0
                           });
209
0
}
210
211
Status BaseTablet::capture_rs_readers_unlocked(const Versions& version_path,
212
0
                                               std::vector<RowSetSplits>* rs_splits) const {
213
0
    DCHECK(rs_splits != nullptr && rs_splits->empty());
214
0
    for (auto version : version_path) {
215
0
        auto it = _rs_version_map.find(version);
216
0
        if (it == _rs_version_map.end()) {
217
0
            VLOG_NOTICE << "fail to find Rowset in rs_version for version. tablet=" << tablet_id()
218
0
                        << ", version='" << version.first << "-" << version.second;
219
220
0
            it = _stale_rs_version_map.find(version);
221
0
            if (it == _stale_rs_version_map.end()) {
222
0
                return Status::Error<CAPTURE_ROWSET_READER_ERROR>(
223
0
                        "fail to find Rowset in stale_rs_version for version. tablet={}, "
224
0
                        "version={}-{}",
225
0
                        tablet_id(), version.first, version.second);
226
0
            }
227
0
        }
228
0
        RowsetReaderSharedPtr rs_reader;
229
0
        auto res = it->second->create_reader(&rs_reader);
230
0
        if (!res.ok()) {
231
0
            return Status::Error<CAPTURE_ROWSET_READER_ERROR>(
232
0
                    "failed to create reader for rowset:{}", it->second->rowset_id().to_string());
233
0
        }
234
0
        rs_splits->emplace_back(std::move(rs_reader));
235
0
    }
236
0
    return Status::OK();
237
0
}
238
239
// snapshot manager may call this api to check if version exists, so that
240
// the version maybe not exist
241
RowsetSharedPtr BaseTablet::get_rowset_by_version(const Version& version,
242
1
                                                  bool find_in_stale) const {
243
1
    auto iter = _rs_version_map.find(version);
244
1
    if (iter == _rs_version_map.end()) {
245
0
        if (find_in_stale) {
246
0
            return get_stale_rowset_by_version(version);
247
0
        }
248
0
        return nullptr;
249
0
    }
250
1
    return iter->second;
251
1
}
252
253
0
RowsetSharedPtr BaseTablet::get_stale_rowset_by_version(const Version& version) const {
254
0
    auto iter = _stale_rs_version_map.find(version);
255
0
    if (iter == _stale_rs_version_map.end()) {
256
0
        VLOG_NOTICE << "no rowset for version:" << version << ", tablet: " << tablet_id();
257
0
        return nullptr;
258
0
    }
259
0
    return iter->second;
260
0
}
261
262
// Already under _meta_lock
263
42
RowsetSharedPtr BaseTablet::get_rowset_with_max_version() const {
264
42
    Version max_version = _tablet_meta->max_version();
265
42
    if (max_version.first == -1) {
266
0
        return nullptr;
267
0
    }
268
269
42
    auto iter = _rs_version_map.find(max_version);
270
42
    if (iter == _rs_version_map.end()) {
271
0
        DCHECK(false) << "invalid version:" << max_version;
272
0
        return nullptr;
273
0
    }
274
42
    return iter->second;
275
42
}
276
277
0
Status BaseTablet::get_all_rs_id(int64_t max_version, RowsetIdUnorderedSet* rowset_ids) const {
278
0
    std::shared_lock rlock(_meta_lock);
279
0
    return get_all_rs_id_unlocked(max_version, rowset_ids);
280
0
}
281
282
Status BaseTablet::get_all_rs_id_unlocked(int64_t max_version,
283
9
                                          RowsetIdUnorderedSet* rowset_ids) const {
284
    //  Ensure that the obtained versions of rowsets are continuous
285
9
    Version spec_version(0, max_version);
286
9
    Versions version_path;
287
9
    auto st = _timestamped_version_tracker.capture_consistent_versions(spec_version, &version_path);
288
9
    if (!st.ok()) [[unlikely]] {
289
0
        return st;
290
0
    }
291
292
11
    for (auto& ver : version_path) {
293
11
        if (ver.second == 1) {
294
            // [0-1] rowset is empty for each tablet, skip it
295
9
            continue;
296
9
        }
297
2
        auto it = _rs_version_map.find(ver);
298
2
        if (it == _rs_version_map.end()) {
299
0
            return Status::Error<CAPTURE_ROWSET_ERROR, false>(
300
0
                    "fail to find Rowset for version. tablet={}, version={}", tablet_id(),
301
0
                    ver.to_string());
302
0
        }
303
2
        rowset_ids->emplace(it->second->rowset_id());
304
2
    }
305
9
    return Status::OK();
306
9
}
307
308
0
Versions BaseTablet::get_missed_versions(int64_t spec_version) const {
309
0
    DCHECK(spec_version > 0) << "invalid spec_version: " << spec_version;
310
311
0
    Versions existing_versions;
312
0
    {
313
0
        std::shared_lock rdlock(_meta_lock);
314
0
        for (const auto& rs : _tablet_meta->all_rs_metas()) {
315
0
            existing_versions.emplace_back(rs->version());
316
0
        }
317
0
    }
318
0
    return calc_missed_versions(spec_version, std::move(existing_versions));
319
0
}
320
321
2
Versions BaseTablet::get_missed_versions_unlocked(int64_t spec_version) const {
322
2
    DCHECK(spec_version > 0) << "invalid spec_version: " << spec_version;
323
324
2
    Versions existing_versions;
325
8
    for (const auto& rs : _tablet_meta->all_rs_metas()) {
326
8
        existing_versions.emplace_back(rs->version());
327
8
    }
328
2
    return calc_missed_versions(spec_version, std::move(existing_versions));
329
2
}
330
331
1
void BaseTablet::_print_missed_versions(const Versions& missed_versions) const {
332
1
    std::stringstream ss;
333
1
    ss << tablet_id() << " has " << missed_versions.size() << " missed version:";
334
    // print at most 10 version
335
3
    for (int i = 0; i < 10 && i < missed_versions.size(); ++i) {
336
2
        ss << missed_versions[i] << ",";
337
2
    }
338
1
    LOG(WARNING) << ss.str();
339
1
}
340
341
1
bool BaseTablet::_reconstruct_version_tracker_if_necessary() {
342
1
    double orphan_vertex_ratio = _timestamped_version_tracker.get_orphan_vertex_ratio();
343
1
    if (orphan_vertex_ratio >= config::tablet_version_graph_orphan_vertex_ratio) {
344
1
        _timestamped_version_tracker.construct_versioned_tracker(
345
1
                _tablet_meta->all_rs_metas(), _tablet_meta->all_stale_rs_metas());
346
1
        return true;
347
1
    }
348
0
    return false;
349
1
}
350
351
// should use this method to get a copy of current tablet meta
352
// there are some rowset meta in local meta store and in in-memory tablet meta
353
// but not in tablet meta in local meta store
354
0
void BaseTablet::generate_tablet_meta_copy(TabletMeta& new_tablet_meta) const {
355
0
    TabletMetaPB tablet_meta_pb;
356
0
    {
357
0
        std::shared_lock rdlock(_meta_lock);
358
0
        _tablet_meta->to_meta_pb(&tablet_meta_pb);
359
0
    }
360
0
    generate_tablet_meta_copy_unlocked(new_tablet_meta);
361
0
}
362
363
// this is a unlocked version of generate_tablet_meta_copy()
364
// some method already hold the _meta_lock before calling this,
365
// such as EngineCloneTask::_finish_clone -> tablet->revise_tablet_meta
366
4
void BaseTablet::generate_tablet_meta_copy_unlocked(TabletMeta& new_tablet_meta) const {
367
4
    TabletMetaPB tablet_meta_pb;
368
4
    _tablet_meta->to_meta_pb(&tablet_meta_pb);
369
4
    new_tablet_meta.init_from_pb(tablet_meta_pb);
370
4
}
371
372
Status BaseTablet::calc_delete_bitmap_between_segments(
373
        TabletSchemaSPtr schema, const RowsetId& rowset_id,
374
0
        const std::vector<segment_v2::SegmentSharedPtr>& segments, DeleteBitmapPtr delete_bitmap) {
375
0
    size_t const num_segments = segments.size();
376
0
    if (num_segments < 2) {
377
0
        return Status::OK();
378
0
    }
379
380
0
    OlapStopWatch watch;
381
0
    size_t seq_col_length = 0;
382
0
    if (schema->has_sequence_col()) {
383
0
        auto seq_col_idx = schema->sequence_col_idx();
384
0
        seq_col_length = schema->column(seq_col_idx).length() + 1;
385
0
    }
386
0
    size_t rowid_length = 0;
387
0
    if (!schema->cluster_key_idxes().empty()) {
388
0
        rowid_length = PrimaryKeyIndexReader::ROW_ID_LENGTH;
389
0
    }
390
391
0
    MergeIndexDeleteBitmapCalculator calculator;
392
0
    RETURN_IF_ERROR(calculator.init(rowset_id, segments, seq_col_length, rowid_length));
393
394
0
    RETURN_IF_ERROR(calculator.calculate_all(delete_bitmap));
395
396
0
    delete_bitmap->add(
397
0
            {rowset_id, DeleteBitmap::INVALID_SEGMENT_ID, DeleteBitmap::TEMP_VERSION_COMMON},
398
0
            DeleteBitmap::ROWSET_SENTINEL_MARK);
399
0
    LOG(INFO) << fmt::format(
400
0
            "construct delete bitmap between segments, "
401
0
            "tablet: {}, rowset: {}, number of segments: {}, bitmap count: {}, bitmap cardinality: "
402
0
            "{}, cost {} (us)",
403
0
            tablet_id(), rowset_id.to_string(), num_segments,
404
0
            delete_bitmap->get_delete_bitmap_count(), delete_bitmap->cardinality(),
405
0
            watch.get_elapse_time_us());
406
0
    return Status::OK();
407
0
}
408
409
std::vector<RowsetSharedPtr> BaseTablet::get_rowset_by_ids(
410
72
        const RowsetIdUnorderedSet* specified_rowset_ids) {
411
72
    std::vector<RowsetSharedPtr> rowsets;
412
72
    for (auto& rs : _rs_version_map) {
413
15
        if (!specified_rowset_ids ||
414
15
            specified_rowset_ids->find(rs.second->rowset_id()) != specified_rowset_ids->end()) {
415
1
            rowsets.push_back(rs.second);
416
1
        }
417
15
    }
418
419
72
    std::sort(rowsets.begin(), rowsets.end(), [](RowsetSharedPtr& lhs, RowsetSharedPtr& rhs) {
420
0
        return lhs->end_version() > rhs->end_version();
421
0
    });
422
72
    return rowsets;
423
72
}
424
425
Status BaseTablet::lookup_row_data(const Slice& encoded_key, const RowLocation& row_location,
426
                                   RowsetSharedPtr input_rowset, const TupleDescriptor* desc,
427
                                   OlapReaderStatistics& stats, std::string& values,
428
0
                                   bool write_to_cache) {
429
0
    MonotonicStopWatch watch;
430
0
    size_t row_size = 1;
431
0
    watch.start();
432
0
    Defer _defer([&]() {
433
0
        LOG_EVERY_N(INFO, 500) << "get a single_row, cost(us):" << watch.elapsed_time() / 1000
434
0
                               << ", row_size:" << row_size;
435
0
    });
436
437
0
    BetaRowsetSharedPtr rowset = std::static_pointer_cast<BetaRowset>(input_rowset);
438
0
    CHECK(rowset);
439
0
    const TabletSchemaSPtr tablet_schema = rowset->tablet_schema();
440
0
    SegmentCacheHandle segment_cache_handle;
441
0
    std::unique_ptr<segment_v2::ColumnIterator> column_iterator;
442
0
    const auto& column = *DORIS_TRY(tablet_schema->column(BeConsts::ROW_STORE_COL));
443
0
    RETURN_IF_ERROR(_get_segment_column_iterator(rowset, row_location.segment_id, column,
444
0
                                                 &segment_cache_handle, &column_iterator, &stats));
445
    // get and parse tuple row
446
0
    vectorized::MutableColumnPtr column_ptr = vectorized::ColumnString::create();
447
0
    std::vector<segment_v2::rowid_t> rowids {static_cast<segment_v2::rowid_t>(row_location.row_id)};
448
0
    RETURN_IF_ERROR(column_iterator->read_by_rowids(rowids.data(), 1, column_ptr));
449
0
    assert(column_ptr->size() == 1);
450
0
    auto* string_column = static_cast<vectorized::ColumnString*>(column_ptr.get());
451
0
    StringRef value = string_column->get_data_at(0);
452
0
    values = value.to_string();
453
0
    if (write_to_cache) {
454
0
        StringRef value = string_column->get_data_at(0);
455
0
        RowCache::instance()->insert({tablet_id(), encoded_key}, Slice {value.data, value.size});
456
0
    }
457
0
    return Status::OK();
458
0
}
459
460
Status BaseTablet::lookup_row_key(const Slice& encoded_key, TabletSchema* latest_schema,
461
                                  bool with_seq_col,
462
                                  const std::vector<RowsetSharedPtr>& specified_rowsets,
463
                                  RowLocation* row_location, uint32_t version,
464
                                  std::vector<std::unique_ptr<SegmentCacheHandle>>& segment_caches,
465
                                  RowsetSharedPtr* rowset, bool with_rowid,
466
2
                                  OlapReaderStatistics* stats) {
467
2
    SCOPED_BVAR_LATENCY(g_tablet_lookup_rowkey_latency);
468
2
    size_t seq_col_length = 0;
469
    // use the latest tablet schema to decide if the tablet has sequence column currently
470
2
    const TabletSchema* schema =
471
2
            (latest_schema == nullptr ? _tablet_meta->tablet_schema().get() : latest_schema);
472
2
    if (schema->has_sequence_col() && with_seq_col) {
473
2
        seq_col_length = schema->column(schema->sequence_col_idx()).length() + 1;
474
2
    }
475
2
    size_t rowid_length = 0;
476
2
    if (with_rowid && !schema->cluster_key_idxes().empty()) {
477
0
        rowid_length = PrimaryKeyIndexReader::ROW_ID_LENGTH;
478
0
    }
479
2
    Slice key_without_seq =
480
2
            Slice(encoded_key.get_data(), encoded_key.get_size() - seq_col_length - rowid_length);
481
2
    RowLocation loc;
482
483
2
    for (size_t i = 0; i < specified_rowsets.size(); i++) {
484
2
        const auto& rs = specified_rowsets[i];
485
2
        std::vector<KeyBoundsPB> segments_key_bounds;
486
2
        rs->rowset_meta()->get_segments_key_bounds(&segments_key_bounds);
487
2
        int num_segments = static_cast<int>(rs->num_segments());
488
2
        DCHECK_EQ(segments_key_bounds.size(), num_segments);
489
2
        std::vector<uint32_t> picked_segments;
490
4
        for (int j = num_segments - 1; j >= 0; j--) {
491
            // If mow table has cluster keys, the key bounds is short keys, not primary keys
492
            // use PrimaryKeyIndexMetaPB in primary key index?
493
2
            if (schema->cluster_key_idxes().empty()) {
494
2
                if (key_is_not_in_segment(key_without_seq, segments_key_bounds[j],
495
2
                                          rs->rowset_meta()->is_segments_key_bounds_truncated())) {
496
0
                    continue;
497
0
                }
498
2
            }
499
2
            picked_segments.emplace_back(j);
500
2
        }
501
2
        if (picked_segments.empty()) {
502
0
            continue;
503
0
        }
504
505
2
        if (UNLIKELY(segment_caches[i] == nullptr)) {
506
1
            segment_caches[i] = std::make_unique<SegmentCacheHandle>();
507
1
            RETURN_IF_ERROR(SegmentLoader::instance()->load_segments(
508
1
                    std::static_pointer_cast<BetaRowset>(rs), segment_caches[i].get(), true, true));
509
1
        }
510
2
        auto& segments = segment_caches[i]->get_segments();
511
2
        DCHECK_EQ(segments.size(), num_segments);
512
513
2
        for (auto id : picked_segments) {
514
2
            Status s = segments[id]->lookup_row_key(encoded_key, schema, with_seq_col, with_rowid,
515
2
                                                    &loc, stats);
516
2
            if (s.is<KEY_NOT_FOUND>()) {
517
0
                continue;
518
0
            }
519
2
            if (!s.ok() && !s.is<KEY_ALREADY_EXISTS>()) {
520
0
                return s;
521
0
            }
522
2
            if (s.ok() && _tablet_meta->delete_bitmap()->contains_agg_without_cache(
523
1
                                  {loc.rowset_id, loc.segment_id, version}, loc.row_id)) {
524
                // if has sequence col, we continue to compare the sequence_id of
525
                // all rowsets, util we find an existing key.
526
0
                if (schema->has_sequence_col()) {
527
0
                    continue;
528
0
                }
529
                // The key is deleted, we don't need to search for it any more.
530
0
                break;
531
0
            }
532
            // `st` is either OK or KEY_ALREADY_EXISTS now.
533
            // for partial update, even if the key is already exists, we still need to
534
            // read it's original values to keep all columns align.
535
2
            *row_location = loc;
536
2
            if (rowset) {
537
                // return it's rowset
538
2
                *rowset = rs;
539
2
            }
540
            // find it and return
541
2
            return s;
542
2
        }
543
2
    }
544
0
    g_tablet_pk_not_found << 1;
545
0
    return Status::Error<ErrorCode::KEY_NOT_FOUND>("can't find key in all rowsets");
546
2
}
547
548
// if user pass a token, then all calculation works will submit to a threadpool,
549
// user can get all delete bitmaps from that token.
550
// if `token` is nullptr, the calculation will run in local, and user can get the result
551
// delete bitmap from `delete_bitmap` directly.
552
Status BaseTablet::calc_delete_bitmap(const BaseTabletSPtr& tablet, RowsetSharedPtr rowset,
553
                                      const std::vector<segment_v2::SegmentSharedPtr>& segments,
554
                                      const std::vector<RowsetSharedPtr>& specified_rowsets,
555
                                      DeleteBitmapPtr delete_bitmap, int64_t end_version,
556
69
                                      CalcDeleteBitmapToken* token, RowsetWriter* rowset_writer) {
557
69
    if (specified_rowsets.empty() || segments.empty()) {
558
68
        return Status::OK();
559
68
    }
560
561
1
    OlapStopWatch watch;
562
1
    for (const auto& segment : segments) {
563
1
        const auto& seg = segment;
564
1
        if (token != nullptr) {
565
1
            RETURN_IF_ERROR(token->submit(tablet, rowset, seg, specified_rowsets, end_version,
566
1
                                          delete_bitmap, rowset_writer));
567
1
        } else {
568
0
            RETURN_IF_ERROR(tablet->calc_segment_delete_bitmap(
569
0
                    rowset, segment, specified_rowsets, delete_bitmap, end_version, rowset_writer));
570
0
        }
571
1
    }
572
573
1
    return Status::OK();
574
1
}
575
576
Status BaseTablet::calc_segment_delete_bitmap(RowsetSharedPtr rowset,
577
                                              const segment_v2::SegmentSharedPtr& seg,
578
                                              const std::vector<RowsetSharedPtr>& specified_rowsets,
579
                                              DeleteBitmapPtr delete_bitmap, int64_t end_version,
580
1
                                              RowsetWriter* rowset_writer) {
581
1
    OlapStopWatch watch;
582
1
    auto rowset_id = rowset->rowset_id();
583
1
    Version dummy_version(end_version + 1, end_version + 1);
584
1
    auto rowset_schema = rowset->tablet_schema();
585
1
    bool is_partial_update = rowset_writer && rowset_writer->is_partial_update();
586
1
    bool have_input_seq_column = false;
587
1
    if (is_partial_update && rowset_schema->has_sequence_col()) {
588
0
        std::vector<uint32_t> including_cids =
589
0
                rowset_writer->get_partial_update_info()->update_cids;
590
0
        have_input_seq_column =
591
0
                rowset_schema->has_sequence_col() &&
592
0
                (std::find(including_cids.cbegin(), including_cids.cend(),
593
0
                           rowset_schema->sequence_col_idx()) != including_cids.cend());
594
0
    }
595
1
    if (rowset_schema->num_variant_columns() > 0) {
596
        // During partial updates, the extracted columns of a variant should not be included in the rowset schema.
597
        // This is because the partial update for a variant needs to ignore the extracted columns.
598
        // Otherwise, the schema types in different rowsets might be inconsistent. When performing a partial update,
599
        // the complete variant is constructed by reading all the sub-columns of the variant.
600
0
        rowset_schema = rowset_schema->copy_without_variant_extracted_columns();
601
0
    }
602
    // use for partial update
603
1
    PartialUpdateReadPlan read_plan_ori;
604
1
    PartialUpdateReadPlan read_plan_update;
605
1
    int64_t conflict_rows = 0;
606
1
    int64_t new_generated_rows = 0;
607
608
1
    std::map<RowsetId, RowsetSharedPtr> rsid_to_rowset;
609
1
    rsid_to_rowset[rowset_id] = rowset;
610
1
    vectorized::Block block = rowset_schema->create_block();
611
1
    vectorized::Block ordered_block = block.clone_empty();
612
1
    uint32_t pos = 0;
613
614
1
    RETURN_IF_ERROR(seg->load_pk_index_and_bf(nullptr)); // We need index blocks to iterate
615
1
    const auto* pk_idx = seg->get_primary_key_index();
616
1
    int total = pk_idx->num_rows();
617
1
    uint32_t row_id = 0;
618
1
    int32_t remaining = total;
619
1
    bool exact_match = false;
620
1
    std::string last_key;
621
1
    int batch_size = 1024;
622
    // The data for each segment may be lookup multiple times. Creating a SegmentCacheHandle
623
    // will update the lru cache, and there will be obvious lock competition in multithreading
624
    // scenarios, so using a segment_caches to cache SegmentCacheHandle.
625
1
    std::vector<std::unique_ptr<SegmentCacheHandle>> segment_caches(specified_rowsets.size());
626
2
    while (remaining > 0) {
627
1
        std::unique_ptr<segment_v2::IndexedColumnIterator> iter;
628
1
        RETURN_IF_ERROR(pk_idx->new_iterator(&iter, nullptr));
629
630
1
        size_t num_to_read = std::min(batch_size, remaining);
631
1
        auto index_type = vectorized::DataTypeFactory::instance().create_data_type(
632
1
                pk_idx->type_info()->type(), 1, 0);
633
1
        auto index_column = index_type->create_column();
634
1
        Slice last_key_slice(last_key);
635
1
        RETURN_IF_ERROR(iter->seek_at_or_after(&last_key_slice, &exact_match));
636
1
        auto current_ordinal = iter->get_current_ordinal();
637
1
        DCHECK(total == remaining + current_ordinal)
638
0
                << "total: " << total << ", remaining: " << remaining
639
0
                << ", current_ordinal: " << current_ordinal;
640
641
1
        size_t num_read = num_to_read;
642
1
        RETURN_IF_ERROR(iter->next_batch(&num_read, index_column));
643
1
        DCHECK(num_to_read == num_read)
644
0
                << "num_to_read: " << num_to_read << ", num_read: " << num_read;
645
1
        last_key = index_column->get_data_at(num_read - 1).to_string();
646
647
        // exclude last_key, last_key will be read in next batch.
648
1
        if (num_read == batch_size && num_read != remaining) {
649
0
            num_read -= 1;
650
0
        }
651
3
        for (size_t i = 0; i < num_read; i++, row_id++) {
652
2
            Slice key = Slice(index_column->get_data_at(i).data, index_column->get_data_at(i).size);
653
2
            RowLocation loc;
654
            // calculate row id
655
2
            if (!_tablet_meta->tablet_schema()->cluster_key_idxes().empty()) {
656
0
                size_t seq_col_length = 0;
657
0
                if (_tablet_meta->tablet_schema()->has_sequence_col()) {
658
0
                    seq_col_length =
659
0
                            _tablet_meta->tablet_schema()
660
0
                                    ->column(_tablet_meta->tablet_schema()->sequence_col_idx())
661
0
                                    .length() +
662
0
                            1;
663
0
                }
664
0
                size_t rowid_length = PrimaryKeyIndexReader::ROW_ID_LENGTH;
665
0
                Slice key_without_seq =
666
0
                        Slice(key.get_data(), key.get_size() - seq_col_length - rowid_length);
667
0
                Slice rowid_slice =
668
0
                        Slice(key.get_data() + key_without_seq.get_size() + seq_col_length + 1,
669
0
                              rowid_length - 1);
670
0
                const auto* type_info =
671
0
                        get_scalar_type_info<FieldType::OLAP_FIELD_TYPE_UNSIGNED_INT>();
672
0
                const auto* rowid_coder = get_key_coder(type_info->type());
673
0
                RETURN_IF_ERROR(rowid_coder->decode_ascending(&rowid_slice, rowid_length,
674
0
                                                              (uint8_t*)&row_id));
675
0
            }
676
            // same row in segments should be filtered
677
2
            if (delete_bitmap->contains({rowset_id, seg->id(), DeleteBitmap::TEMP_VERSION_COMMON},
678
2
                                        row_id)) {
679
0
                continue;
680
0
            }
681
682
2
            DBUG_EXECUTE_IF("BaseTablet::calc_segment_delete_bitmap.inject_err", {
683
2
                auto p = dp->param("percent", 0.01);
684
2
                std::mt19937 gen {std::random_device {}()};
685
2
                std::bernoulli_distribution inject_fault {p};
686
2
                if (inject_fault(gen)) {
687
2
                    return Status::InternalError(
688
2
                            "injection error in calc_segment_delete_bitmap, "
689
2
                            "tablet_id={}, rowset_id={}",
690
2
                            tablet_id(), rowset_id.to_string());
691
2
                }
692
2
            });
693
694
2
            RowsetSharedPtr rowset_find;
695
2
            auto st = lookup_row_key(key, rowset_schema.get(), true, specified_rowsets, &loc,
696
2
                                     dummy_version.first - 1, segment_caches, &rowset_find);
697
2
            bool expected_st = st.ok() || st.is<KEY_NOT_FOUND>() || st.is<KEY_ALREADY_EXISTS>();
698
            // It's a defensive DCHECK, we need to exclude some common errors to avoid core-dump
699
            // while stress test
700
2
            DCHECK(expected_st || st.is<MEM_LIMIT_EXCEEDED>())
701
0
                    << "unexpected error status while lookup_row_key:" << st;
702
2
            if (!expected_st) {
703
0
                return st;
704
0
            }
705
2
            if (st.is<KEY_NOT_FOUND>()) {
706
0
                continue;
707
0
            }
708
709
2
            if (st.is<KEY_ALREADY_EXISTS>() && (!is_partial_update || have_input_seq_column)) {
710
                // `st.is<KEY_ALREADY_EXISTS>()` means that there exists a row with the same key and larger value
711
                // in seqeunce column.
712
                // - If the current load is not a partial update, we just delete current row.
713
                // - Otherwise, it means that we are doing the alignment process in publish phase due to conflicts
714
                // during concurrent partial updates. And there exists another load which introduces a row with
715
                // the same keys and larger sequence column value published successfully after the commit phase
716
                // of the current load.
717
                //     - If the columns we update include sequence column, we should delete the current row becase the
718
                //       partial update on the current row has been `overwritten` by the previous one with larger sequence
719
                //       column value.
720
                //     - Otherwise, we should combine the values of the missing columns in the previous row and the values
721
                //       of the including columns in the current row into a new row.
722
1
                delete_bitmap->add({rowset_id, seg->id(), DeleteBitmap::TEMP_VERSION_COMMON},
723
1
                                   row_id);
724
1
                ++conflict_rows;
725
1
                continue;
726
                // NOTE: for partial update which doesn't specify the sequence column, we can't use the sequence column value filled in flush phase
727
                // as its final value. Otherwise it may cause inconsistency between replicas.
728
1
            }
729
1
            if (is_partial_update && rowset_writer != nullptr) {
730
                // In publish version, record rows to be deleted for concurrent update
731
                // For example, if version 5 and 6 update a row, but version 6 only see
732
                // version 4 when write, and when publish version, version 5's value will
733
                // be marked as deleted and it's update is losed.
734
                // So here we should read version 5's columns and build a new row, which is
735
                // consists of version 6's update columns and version 5's origin columns
736
                // here we build 2 read plan for ori values and update values
737
0
                read_plan_ori.prepare_to_read(loc, pos);
738
0
                read_plan_update.prepare_to_read(RowLocation {rowset_id, seg->id(), row_id}, pos);
739
0
                rsid_to_rowset[rowset_find->rowset_id()] = rowset_find;
740
0
                ++pos;
741
                // delete bitmap will be calculate when memtable flush and
742
                // publish. The two stages may see different versions.
743
                // When there is sequence column, the currently imported data
744
                // of rowset may be marked for deletion at memtablet flush or
745
                // publish because the seq column is smaller than the previous
746
                // rowset.
747
                // just set 0 as a unified temporary version number, and update to
748
                // the real version number later.
749
0
                delete_bitmap->add(
750
0
                        {loc.rowset_id, loc.segment_id, DeleteBitmap::TEMP_VERSION_COMMON},
751
0
                        loc.row_id);
752
0
                delete_bitmap->add({rowset_id, seg->id(), DeleteBitmap::TEMP_VERSION_COMMON},
753
0
                                   row_id);
754
0
                ++conflict_rows;
755
0
                ++new_generated_rows;
756
0
                continue;
757
0
            }
758
            // when st = ok
759
1
            delete_bitmap->add({loc.rowset_id, loc.segment_id, DeleteBitmap::TEMP_VERSION_COMMON},
760
1
                               loc.row_id);
761
1
            ++conflict_rows;
762
1
        }
763
1
        remaining -= num_read;
764
1
    }
765
    // DCHECK_EQ(total, row_id) << "segment total rows: " << total << " row_id:" << row_id;
766
767
1
    if (config::enable_merge_on_write_correctness_check) {
768
1
        RowsetIdUnorderedSet rowsetids;
769
1
        for (const auto& rowset : specified_rowsets) {
770
1
            rowsetids.emplace(rowset->rowset_id());
771
1
            VLOG_NOTICE << "[tabletID:" << tablet_id() << "]"
772
0
                        << "[add_sentinel_mark_to_delete_bitmap][end_version:" << end_version << "]"
773
0
                        << "add:" << rowset->rowset_id();
774
1
        }
775
1
        add_sentinel_mark_to_delete_bitmap(delete_bitmap.get(), rowsetids);
776
1
    }
777
778
1
    if (pos > 0) {
779
0
        auto partial_update_info = rowset_writer->get_partial_update_info();
780
0
        DCHECK(partial_update_info);
781
0
        RETURN_IF_ERROR(generate_new_block_for_partial_update(
782
0
                rowset_schema, partial_update_info.get(), read_plan_ori, read_plan_update,
783
0
                rsid_to_rowset, &block));
784
0
        RETURN_IF_ERROR(sort_block(block, ordered_block));
785
0
        RETURN_IF_ERROR(rowset_writer->flush_single_block(&ordered_block));
786
0
        auto cost_us = watch.get_elapse_time_us();
787
0
        if (config::enable_mow_verbose_log || cost_us > 10 * 1000) {
788
0
            LOG(INFO) << "calc segment delete bitmap for partial update, tablet: " << tablet_id()
789
0
                      << " rowset: " << rowset_id << " seg_id: " << seg->id()
790
0
                      << " dummy_version: " << end_version + 1 << " rows: " << seg->num_rows()
791
0
                      << " conflict rows: " << conflict_rows
792
0
                      << " new generated rows: " << new_generated_rows
793
0
                      << " bitmap num: " << delete_bitmap->get_delete_bitmap_count()
794
0
                      << " bitmap cardinality: " << delete_bitmap->cardinality()
795
0
                      << " cost: " << cost_us << "(us)";
796
0
        }
797
0
        return Status::OK();
798
0
    }
799
1
    auto cost_us = watch.get_elapse_time_us();
800
1
    if (config::enable_mow_verbose_log || cost_us > 10 * 1000) {
801
0
        LOG(INFO) << "calc segment delete bitmap, tablet: " << tablet_id()
802
0
                  << " rowset: " << rowset_id << " seg_id: " << seg->id()
803
0
                  << " dummy_version: " << end_version + 1 << " rows: " << seg->num_rows()
804
0
                  << " conflict rows: " << conflict_rows
805
0
                  << " bitmap num: " << delete_bitmap->get_delete_bitmap_count()
806
0
                  << " bitmap cardinality: " << delete_bitmap->cardinality() << " cost: " << cost_us
807
0
                  << "(us)";
808
0
    }
809
1
    return Status::OK();
810
1
}
811
812
0
Status BaseTablet::sort_block(vectorized::Block& in_block, vectorized::Block& output_block) {
813
0
    vectorized::MutableBlock mutable_input_block =
814
0
            vectorized::MutableBlock::build_mutable_block(&in_block);
815
0
    vectorized::MutableBlock mutable_output_block =
816
0
            vectorized::MutableBlock::build_mutable_block(&output_block);
817
818
0
    std::shared_ptr<RowInBlockComparator> vec_row_comparator =
819
0
            std::make_shared<RowInBlockComparator>(_tablet_meta->tablet_schema());
820
0
    vec_row_comparator->set_block(&mutable_input_block);
821
822
0
    std::vector<std::unique_ptr<RowInBlock>> row_in_blocks;
823
0
    DCHECK(in_block.rows() <= std::numeric_limits<int>::max());
824
0
    row_in_blocks.reserve(in_block.rows());
825
0
    for (size_t i = 0; i < in_block.rows(); ++i) {
826
0
        row_in_blocks.emplace_back(std::make_unique<RowInBlock>(i));
827
0
    }
828
0
    std::sort(row_in_blocks.begin(), row_in_blocks.end(),
829
0
              [&](const std::unique_ptr<RowInBlock>& l,
830
0
                  const std::unique_ptr<RowInBlock>& r) -> bool {
831
0
                  auto value = (*vec_row_comparator)(l.get(), r.get());
832
0
                  DCHECK(value != 0) << "value equel when sort block, l_pos: " << l->_row_pos
833
0
                                     << " r_pos: " << r->_row_pos;
834
0
                  return value < 0;
835
0
              });
836
0
    std::vector<uint32_t> row_pos_vec;
837
0
    row_pos_vec.reserve(in_block.rows());
838
0
    for (auto& block : row_in_blocks) {
839
0
        row_pos_vec.emplace_back(block->_row_pos);
840
0
    }
841
0
    return mutable_output_block.add_rows(&in_block, row_pos_vec.data(),
842
0
                                         row_pos_vec.data() + in_block.rows());
843
0
}
844
845
// fetch value by row column
846
Status BaseTablet::fetch_value_through_row_column(RowsetSharedPtr input_rowset,
847
                                                  const TabletSchema& tablet_schema, uint32_t segid,
848
                                                  const std::vector<uint32_t>& rowids,
849
                                                  const std::vector<uint32_t>& cids,
850
0
                                                  vectorized::Block& block) {
851
0
    MonotonicStopWatch watch;
852
0
    watch.start();
853
0
    Defer _defer([&]() {
854
0
        LOG_EVERY_N(INFO, 500) << "fetch_value_by_rowids, cost(us):" << watch.elapsed_time() / 1000
855
0
                               << ", row_batch_size:" << rowids.size();
856
0
    });
857
858
0
    BetaRowsetSharedPtr rowset = std::static_pointer_cast<BetaRowset>(input_rowset);
859
0
    CHECK(rowset);
860
0
    CHECK(tablet_schema.has_row_store_for_all_columns());
861
0
    SegmentCacheHandle segment_cache_handle;
862
0
    std::unique_ptr<segment_v2::ColumnIterator> column_iterator;
863
0
    OlapReaderStatistics stats;
864
0
    const auto& column = *DORIS_TRY(tablet_schema.column(BeConsts::ROW_STORE_COL));
865
0
    RETURN_IF_ERROR(_get_segment_column_iterator(rowset, segid, column, &segment_cache_handle,
866
0
                                                 &column_iterator, &stats));
867
    // get and parse tuple row
868
0
    vectorized::MutableColumnPtr column_ptr = vectorized::ColumnString::create();
869
0
    RETURN_IF_ERROR(column_iterator->read_by_rowids(rowids.data(), rowids.size(), column_ptr));
870
0
    assert(column_ptr->size() == rowids.size());
871
0
    auto* string_column = static_cast<vectorized::ColumnString*>(column_ptr.get());
872
0
    vectorized::DataTypeSerDeSPtrs serdes;
873
0
    serdes.resize(cids.size());
874
0
    std::unordered_map<uint32_t, uint32_t> col_uid_to_idx;
875
0
    std::vector<std::string> default_values;
876
0
    default_values.resize(cids.size());
877
0
    for (int i = 0; i < cids.size(); ++i) {
878
0
        const TabletColumn& column = tablet_schema.column(cids[i]);
879
0
        vectorized::DataTypePtr type =
880
0
                vectorized::DataTypeFactory::instance().create_data_type(column);
881
0
        col_uid_to_idx[column.unique_id()] = i;
882
0
        default_values[i] = column.default_value();
883
0
        serdes[i] = type->get_serde();
884
0
    }
885
0
    vectorized::JsonbSerializeUtil::jsonb_to_block(serdes, *string_column, col_uid_to_idx, block,
886
0
                                                   default_values, {});
887
0
    return Status::OK();
888
0
}
889
890
Status BaseTablet::fetch_value_by_rowids(RowsetSharedPtr input_rowset, uint32_t segid,
891
                                         const std::vector<uint32_t>& rowids,
892
                                         const TabletColumn& tablet_column,
893
0
                                         vectorized::MutableColumnPtr& dst) {
894
0
    MonotonicStopWatch watch;
895
0
    watch.start();
896
0
    Defer _defer([&]() {
897
0
        LOG_EVERY_N(INFO, 500) << "fetch_value_by_rowids, cost(us):" << watch.elapsed_time() / 1000
898
0
                               << ", row_batch_size:" << rowids.size();
899
0
    });
900
901
    // read row data
902
0
    BetaRowsetSharedPtr rowset = std::static_pointer_cast<BetaRowset>(input_rowset);
903
0
    CHECK(rowset);
904
0
    SegmentCacheHandle segment_cache_handle;
905
0
    std::unique_ptr<segment_v2::ColumnIterator> column_iterator;
906
0
    OlapReaderStatistics stats;
907
0
    RETURN_IF_ERROR(_get_segment_column_iterator(rowset, segid, tablet_column,
908
0
                                                 &segment_cache_handle, &column_iterator, &stats));
909
0
    RETURN_IF_ERROR(column_iterator->read_by_rowids(rowids.data(), rowids.size(), dst));
910
0
    return Status::OK();
911
0
}
912
913
const signed char* BaseTablet::get_delete_sign_column_data(vectorized::Block& block,
914
0
                                                           size_t rows_at_least) {
915
0
    if (const vectorized::ColumnWithTypeAndName* delete_sign_column =
916
0
                block.try_get_by_name(DELETE_SIGN);
917
0
        delete_sign_column != nullptr) {
918
0
        const auto& delete_sign_col =
919
0
                reinterpret_cast<const vectorized::ColumnInt8&>(*(delete_sign_column->column));
920
0
        if (delete_sign_col.size() >= rows_at_least) {
921
0
            return delete_sign_col.get_data().data();
922
0
        }
923
0
    }
924
0
    return nullptr;
925
0
};
926
927
Status BaseTablet::generate_default_value_block(const TabletSchema& schema,
928
                                                const std::vector<uint32_t>& cids,
929
                                                const std::vector<std::string>& default_values,
930
                                                const vectorized::Block& ref_block,
931
0
                                                vectorized::Block& default_value_block) {
932
0
    auto mutable_default_value_columns = default_value_block.mutate_columns();
933
0
    for (auto i = 0; i < cids.size(); ++i) {
934
0
        const auto& column = schema.column(cids[i]);
935
0
        if (column.has_default_value()) {
936
0
            const auto& default_value = default_values[i];
937
0
            vectorized::ReadBuffer rb(const_cast<char*>(default_value.c_str()),
938
0
                                      default_value.size());
939
0
            RETURN_IF_ERROR(ref_block.get_by_position(i).type->from_string(
940
0
                    rb, mutable_default_value_columns[i].get()));
941
0
        }
942
0
    }
943
0
    default_value_block.set_columns(std::move(mutable_default_value_columns));
944
0
    return Status::OK();
945
0
}
946
947
Status BaseTablet::generate_new_block_for_partial_update(
948
        TabletSchemaSPtr rowset_schema, const PartialUpdateInfo* partial_update_info,
949
        const PartialUpdateReadPlan& read_plan_ori, const PartialUpdateReadPlan& read_plan_update,
950
        const std::map<RowsetId, RowsetSharedPtr>& rsid_to_rowset,
951
0
        vectorized::Block* output_block) {
952
    // do partial update related works
953
    // 1. read columns by read plan
954
    // 2. generate new block
955
    // 3. write a new segment and modify rowset meta
956
    // 4. mark current keys deleted
957
0
    CHECK(output_block);
958
0
    auto full_mutable_columns = output_block->mutate_columns();
959
0
    const auto& missing_cids = partial_update_info->missing_cids;
960
0
    const auto& update_cids = partial_update_info->update_cids;
961
0
    auto old_block = rowset_schema->create_block_by_cids(missing_cids);
962
0
    auto update_block = rowset_schema->create_block_by_cids(update_cids);
963
964
    // rowid in the final block(start from 0, increase continuously) -> rowid to read in update_block
965
0
    std::map<uint32_t, uint32_t> read_index_update;
966
967
    // read current rowset first, if a row in the current rowset has delete sign mark
968
    // we don't need to read values from old block
969
0
    RETURN_IF_ERROR(read_plan_update.read_columns_by_plan(
970
0
            *rowset_schema, update_cids, rsid_to_rowset, update_block, &read_index_update, false));
971
0
    size_t update_rows = read_index_update.size();
972
0
    for (auto i = 0; i < update_cids.size(); ++i) {
973
0
        for (auto idx = 0; idx < update_rows; ++idx) {
974
0
            full_mutable_columns[update_cids[i]]->insert_from(
975
0
                    *update_block.get_columns_with_type_and_name()[i].column.get(),
976
0
                    read_index_update[idx]);
977
0
        }
978
0
    }
979
980
    // if there is sequence column in the table, we need to read the sequence column,
981
    // otherwise it may cause the merge-on-read based compaction policy to produce incorrect results
982
0
    const auto* __restrict new_block_delete_signs =
983
0
            rowset_schema->has_sequence_col()
984
0
                    ? nullptr
985
0
                    : get_delete_sign_column_data(update_block, update_rows);
986
987
    // 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
988
0
    std::map<uint32_t, uint32_t> read_index_old;
989
0
    RETURN_IF_ERROR(read_plan_ori.read_columns_by_plan(*rowset_schema, missing_cids, rsid_to_rowset,
990
0
                                                       old_block, &read_index_old, true,
991
0
                                                       new_block_delete_signs));
992
0
    size_t old_rows = read_index_old.size();
993
0
    const auto* __restrict old_block_delete_signs =
994
0
            get_delete_sign_column_data(old_block, old_rows);
995
0
    DCHECK(old_block_delete_signs != nullptr);
996
    // build default value block
997
0
    auto default_value_block = old_block.clone_empty();
998
0
    RETURN_IF_ERROR(BaseTablet::generate_default_value_block(*rowset_schema, missing_cids,
999
0
                                                             partial_update_info->default_values,
1000
0
                                                             old_block, default_value_block));
1001
0
    auto mutable_default_value_columns = default_value_block.mutate_columns();
1002
1003
0
    CHECK(update_rows >= old_rows);
1004
1005
    // build full block
1006
0
    for (auto i = 0; i < missing_cids.size(); ++i) {
1007
0
        const auto& rs_column = rowset_schema->column(missing_cids[i]);
1008
0
        auto& mutable_column = full_mutable_columns[missing_cids[i]];
1009
0
        for (auto idx = 0; idx < update_rows; ++idx) {
1010
            // There are two cases we don't need to read values from old data:
1011
            //     1. if the conflicting new row's delete sign is marked, which means the value columns
1012
            //     of the row will not be read. So we don't need to read the missing values from the previous rows.
1013
            //     2. if the conflicting old row's delete sign is marked, which means that the key is not exist now,
1014
            //     we should not read old values from the deleted data, and should use default value instead.
1015
            //     NOTE: since now we are in the publishing phase, all data is commited
1016
            //         before, even the `strict_mode` is true (which requires partial update
1017
            //         load job can't insert new keys), this "new" key MUST be written into
1018
            //         the new generated segment file.
1019
0
            if (new_block_delete_signs != nullptr && new_block_delete_signs[idx]) {
1020
0
                mutable_column->insert_default();
1021
0
            } else if (old_block_delete_signs != nullptr &&
1022
0
                       old_block_delete_signs[read_index_old[idx]] != 0) {
1023
0
                if (rs_column.has_default_value()) {
1024
0
                    mutable_column->insert_from(*mutable_default_value_columns[i].get(), 0);
1025
0
                } else if (rs_column.is_nullable()) {
1026
0
                    assert_cast<vectorized::ColumnNullable*, TypeCheckOnRelease::DISABLE>(
1027
0
                            mutable_column.get())
1028
0
                            ->insert_null_elements(1);
1029
0
                } else {
1030
0
                    mutable_column->insert(rs_column.get_vec_type()->get_default());
1031
0
                }
1032
0
            } else {
1033
0
                mutable_column->insert_from(
1034
0
                        *old_block.get_columns_with_type_and_name()[i].column.get(),
1035
0
                        read_index_old[idx]);
1036
0
            }
1037
0
        }
1038
0
    }
1039
0
    output_block->set_columns(std::move(full_mutable_columns));
1040
0
    VLOG_DEBUG << "full block when publish: " << output_block->dump_data();
1041
0
    return Status::OK();
1042
0
}
1043
1044
Status BaseTablet::commit_phase_update_delete_bitmap(
1045
        const BaseTabletSPtr& tablet, const RowsetSharedPtr& rowset,
1046
        RowsetIdUnorderedSet& pre_rowset_ids, DeleteBitmapPtr delete_bitmap,
1047
        const std::vector<segment_v2::SegmentSharedPtr>& segments, int64_t txn_id,
1048
3
        CalcDeleteBitmapToken* token, RowsetWriter* rowset_writer) {
1049
3
    SCOPED_BVAR_LATENCY(g_tablet_commit_phase_update_delete_bitmap_latency);
1050
3
    RowsetIdUnorderedSet cur_rowset_ids;
1051
3
    RowsetIdUnorderedSet rowset_ids_to_add;
1052
3
    RowsetIdUnorderedSet rowset_ids_to_del;
1053
3
    int64_t cur_version;
1054
1055
3
    std::vector<RowsetSharedPtr> specified_rowsets;
1056
3
    {
1057
        // to prevent seeing intermediate state of a tablet
1058
3
        std::unique_lock<bthread::Mutex> sync_lock;
1059
3
        if (config::is_cloud_mode()) {
1060
0
            sync_lock = std::unique_lock<bthread::Mutex>(
1061
0
                    std::static_pointer_cast<CloudTablet>(tablet)->get_sync_meta_lock());
1062
0
        }
1063
3
        std::shared_lock meta_rlock(tablet->_meta_lock);
1064
3
        if (tablet->tablet_state() == TABLET_NOTREADY) {
1065
            // tablet is under alter process. The delete bitmap will be calculated after conversion.
1066
0
            LOG(INFO) << "tablet is under alter process, delete bitmap will be calculated later, "
1067
0
                         "tablet_id: "
1068
0
                      << tablet->tablet_id() << " txn_id: " << txn_id;
1069
0
            return Status::OK();
1070
0
        }
1071
3
        cur_version = tablet->max_version_unlocked();
1072
3
        RETURN_IF_ERROR(tablet->get_all_rs_id_unlocked(cur_version, &cur_rowset_ids));
1073
3
        _rowset_ids_difference(cur_rowset_ids, pre_rowset_ids, &rowset_ids_to_add,
1074
3
                               &rowset_ids_to_del);
1075
3
        specified_rowsets = tablet->get_rowset_by_ids(&rowset_ids_to_add);
1076
3
    }
1077
0
    for (const auto& to_del : rowset_ids_to_del) {
1078
0
        delete_bitmap->remove({to_del, 0, 0}, {to_del, UINT32_MAX, INT64_MAX});
1079
0
    }
1080
1081
3
    RETURN_IF_ERROR(calc_delete_bitmap(tablet, rowset, segments, specified_rowsets, delete_bitmap,
1082
3
                                       cur_version, token, rowset_writer));
1083
3
    size_t total_rows = std::accumulate(
1084
3
            segments.begin(), segments.end(), 0,
1085
3
            [](size_t sum, const segment_v2::SegmentSharedPtr& s) { return sum += s->num_rows(); });
1086
3
    LOG(INFO) << "[Before Commit] construct delete bitmap tablet: " << tablet->tablet_id()
1087
3
              << ", rowset_ids to add: " << rowset_ids_to_add.size()
1088
3
              << ", rowset_ids to del: " << rowset_ids_to_del.size()
1089
3
              << ", cur max_version: " << cur_version << ", transaction_id: " << txn_id
1090
3
              << ", total rows: " << total_rows;
1091
3
    pre_rowset_ids = cur_rowset_ids;
1092
3
    return Status::OK();
1093
3
}
1094
1095
void BaseTablet::add_sentinel_mark_to_delete_bitmap(DeleteBitmap* delete_bitmap,
1096
5
                                                    const RowsetIdUnorderedSet& rowsetids) {
1097
5
    for (const auto& rowsetid : rowsetids) {
1098
5
        delete_bitmap->add(
1099
5
                {rowsetid, DeleteBitmap::INVALID_SEGMENT_ID, DeleteBitmap::TEMP_VERSION_COMMON},
1100
5
                DeleteBitmap::ROWSET_SENTINEL_MARK);
1101
5
    }
1102
5
}
1103
1104
void BaseTablet::_rowset_ids_difference(const RowsetIdUnorderedSet& cur,
1105
                                        const RowsetIdUnorderedSet& pre,
1106
                                        RowsetIdUnorderedSet* to_add,
1107
6
                                        RowsetIdUnorderedSet* to_del) {
1108
6
    for (const auto& id : cur) {
1109
2
        if (pre.find(id) == pre.end()) {
1110
1
            to_add->insert(id);
1111
1
        }
1112
2
    }
1113
6
    for (const auto& id : pre) {
1114
1
        if (cur.find(id) == cur.end()) {
1115
0
            to_del->insert(id);
1116
0
        }
1117
1
    }
1118
6
}
1119
1120
Status BaseTablet::check_delete_bitmap_correctness(DeleteBitmapPtr delete_bitmap,
1121
                                                   int64_t max_version, int64_t txn_id,
1122
                                                   const RowsetIdUnorderedSet& rowset_ids,
1123
6
                                                   std::vector<RowsetSharedPtr>* rowsets) {
1124
6
    RowsetIdUnorderedSet missing_ids;
1125
6
    for (const auto& rowsetid : rowset_ids) {
1126
2
        if (!delete_bitmap->delete_bitmap.contains({rowsetid, DeleteBitmap::INVALID_SEGMENT_ID,
1127
2
                                                    DeleteBitmap::TEMP_VERSION_COMMON})) {
1128
0
            missing_ids.insert(rowsetid);
1129
0
        }
1130
2
    }
1131
1132
6
    if (!missing_ids.empty()) {
1133
0
        LOG(WARNING) << "[txn_id:" << txn_id << "][tablet_id:" << tablet_id()
1134
0
                     << "][max_version: " << max_version
1135
0
                     << "] check delete bitmap correctness failed!";
1136
0
        rapidjson::Document root;
1137
0
        root.SetObject();
1138
0
        rapidjson::Document required_rowsets_arr;
1139
0
        required_rowsets_arr.SetArray();
1140
0
        rapidjson::Document missing_rowsets_arr;
1141
0
        missing_rowsets_arr.SetArray();
1142
1143
0
        if (rowsets != nullptr) {
1144
0
            for (const auto& rowset : *rowsets) {
1145
0
                rapidjson::Value value;
1146
0
                std::string version_str = rowset->get_rowset_info_str();
1147
0
                value.SetString(version_str.c_str(), version_str.length(),
1148
0
                                required_rowsets_arr.GetAllocator());
1149
0
                required_rowsets_arr.PushBack(value, required_rowsets_arr.GetAllocator());
1150
0
            }
1151
0
        } else {
1152
0
            std::vector<RowsetSharedPtr> rowsets;
1153
0
            {
1154
0
                std::shared_lock meta_rlock(_meta_lock);
1155
0
                rowsets = get_rowset_by_ids(&rowset_ids);
1156
0
            }
1157
0
            for (const auto& rowset : rowsets) {
1158
0
                rapidjson::Value value;
1159
0
                std::string version_str = rowset->get_rowset_info_str();
1160
0
                value.SetString(version_str.c_str(), version_str.length(),
1161
0
                                required_rowsets_arr.GetAllocator());
1162
0
                required_rowsets_arr.PushBack(value, required_rowsets_arr.GetAllocator());
1163
0
            }
1164
0
        }
1165
0
        for (const auto& missing_rowset_id : missing_ids) {
1166
0
            rapidjson::Value miss_value;
1167
0
            std::string rowset_id_str = missing_rowset_id.to_string();
1168
0
            miss_value.SetString(rowset_id_str.c_str(), rowset_id_str.length(),
1169
0
                                 missing_rowsets_arr.GetAllocator());
1170
0
            missing_rowsets_arr.PushBack(miss_value, missing_rowsets_arr.GetAllocator());
1171
0
        }
1172
1173
0
        root.AddMember("required_rowsets", required_rowsets_arr, root.GetAllocator());
1174
0
        root.AddMember("missing_rowsets", missing_rowsets_arr, root.GetAllocator());
1175
0
        rapidjson::StringBuffer strbuf;
1176
0
        rapidjson::PrettyWriter<rapidjson::StringBuffer> writer(strbuf);
1177
0
        root.Accept(writer);
1178
0
        std::string rowset_status_string = std::string(strbuf.GetString());
1179
0
        LOG_EVERY_SECOND(WARNING) << rowset_status_string;
1180
        // let it crash if correctness check failed in Debug mode
1181
0
        DCHECK(false) << "delete bitmap correctness check failed in publish phase!";
1182
0
        return Status::InternalError("check delete bitmap failed!");
1183
0
    }
1184
6
    return Status::OK();
1185
6
}
1186
1187
Status BaseTablet::update_delete_bitmap(const BaseTabletSPtr& self, TabletTxnInfo* txn_info,
1188
3
                                        int64_t txn_id, int64_t txn_expiration) {
1189
3
    SCOPED_BVAR_LATENCY(g_tablet_update_delete_bitmap_latency);
1190
3
    RowsetIdUnorderedSet cur_rowset_ids;
1191
3
    RowsetIdUnorderedSet rowset_ids_to_add;
1192
3
    RowsetIdUnorderedSet rowset_ids_to_del;
1193
3
    RowsetSharedPtr rowset = txn_info->rowset;
1194
3
    int64_t cur_version = rowset->start_version();
1195
1196
3
    std::unique_ptr<RowsetWriter> transient_rs_writer;
1197
3
    DeleteBitmapPtr delete_bitmap = txn_info->delete_bitmap;
1198
3
    bool is_partial_update =
1199
3
            txn_info->partial_update_info && txn_info->partial_update_info->is_partial_update;
1200
3
    if (is_partial_update) {
1201
0
        transient_rs_writer = DORIS_TRY(self->create_transient_rowset_writer(
1202
0
                *rowset, txn_info->partial_update_info, txn_expiration));
1203
0
        DBUG_EXECUTE_IF("BaseTablet::update_delete_bitmap.after.create_transient_rs_writer",
1204
0
                        DBUG_BLOCK);
1205
        // Partial update might generate new segments when there is conflicts while publish, and mark
1206
        // the same key in original segments as delete.
1207
        // When the new segment flush fails or the rowset build fails, the deletion marker for the
1208
        // duplicate key of the original segment should not remain in `txn_info->delete_bitmap`,
1209
        // so we need to make a copy of `txn_info->delete_bitmap` and make changes on it.
1210
0
        delete_bitmap = std::make_shared<DeleteBitmap>(*(txn_info->delete_bitmap));
1211
0
    }
1212
1213
3
    OlapStopWatch watch;
1214
3
    std::vector<segment_v2::SegmentSharedPtr> segments;
1215
3
    RETURN_IF_ERROR(std::dynamic_pointer_cast<BetaRowset>(rowset)->load_segments(&segments));
1216
3
    auto t1 = watch.get_elapse_time_us();
1217
1218
3
    {
1219
3
        std::shared_lock meta_rlock(self->_meta_lock);
1220
        // tablet is under alter process. The delete bitmap will be calculated after conversion.
1221
3
        if (self->tablet_state() == TABLET_NOTREADY) {
1222
0
            LOG(INFO) << "tablet is under alter process, update delete bitmap later, tablet_id="
1223
0
                      << self->tablet_id();
1224
0
            return Status::OK();
1225
0
        }
1226
3
        RETURN_IF_ERROR(self->get_all_rs_id_unlocked(cur_version - 1, &cur_rowset_ids));
1227
3
    }
1228
3
    auto t2 = watch.get_elapse_time_us();
1229
1230
3
    _rowset_ids_difference(cur_rowset_ids, txn_info->rowset_ids, &rowset_ids_to_add,
1231
3
                           &rowset_ids_to_del);
1232
3
    for (const auto& to_del : rowset_ids_to_del) {
1233
0
        delete_bitmap->remove({to_del, 0, 0}, {to_del, UINT32_MAX, INT64_MAX});
1234
0
    }
1235
1236
3
    std::vector<RowsetSharedPtr> specified_rowsets;
1237
3
    {
1238
3
        std::shared_lock meta_rlock(self->_meta_lock);
1239
3
        specified_rowsets = self->get_rowset_by_ids(&rowset_ids_to_add);
1240
3
    }
1241
3
    auto t3 = watch.get_elapse_time_us();
1242
1243
    // If a rowset is produced by compaction before the commit phase of the partial update load
1244
    // and is not included in txn_info->rowset_ids, we can skip the alignment process of that rowset
1245
    // because data remains the same before and after compaction. But we still need to calculate the
1246
    // the delete bitmap for that rowset.
1247
3
    std::vector<RowsetSharedPtr> rowsets_skip_alignment;
1248
3
    if (is_partial_update) {
1249
0
        int64_t max_version_in_flush_phase =
1250
0
                txn_info->partial_update_info->max_version_in_flush_phase;
1251
0
        DCHECK(max_version_in_flush_phase != -1);
1252
0
        std::vector<RowsetSharedPtr> remained_rowsets;
1253
0
        for (const auto& rowset : specified_rowsets) {
1254
0
            if (rowset->end_version() <= max_version_in_flush_phase &&
1255
0
                rowset->produced_by_compaction()) {
1256
0
                rowsets_skip_alignment.emplace_back(rowset);
1257
0
            } else {
1258
0
                remained_rowsets.emplace_back(rowset);
1259
0
            }
1260
0
        }
1261
0
        if (!rowsets_skip_alignment.empty()) {
1262
0
            specified_rowsets = std::move(remained_rowsets);
1263
0
        }
1264
0
    }
1265
1266
3
    DBUG_EXECUTE_IF("BaseTablet::update_delete_bitmap.enable_spin_wait", {
1267
3
        auto token = dp->param<std::string>("token", "invalid_token");
1268
3
        while (DebugPoints::instance()->is_enable("BaseTablet::update_delete_bitmap.block")) {
1269
3
            auto block_dp = DebugPoints::instance()->get_debug_point(
1270
3
                    "BaseTablet::update_delete_bitmap.block");
1271
3
            if (block_dp) {
1272
3
                auto wait_token = block_dp->param<std::string>("wait_token", "");
1273
3
                LOG(INFO) << "BaseTablet::update_delete_bitmap.enable_spin_wait, wait_token: "
1274
3
                          << wait_token << ", token: " << token;
1275
3
                if (wait_token != token) {
1276
3
                    break;
1277
3
                }
1278
3
            }
1279
3
            std::this_thread::sleep_for(std::chrono::milliseconds(50));
1280
3
        }
1281
3
    });
1282
1283
3
    if (!rowsets_skip_alignment.empty()) {
1284
0
        auto token = self->calc_delete_bitmap_executor()->create_token();
1285
        // set rowset_writer to nullptr to skip the alignment process
1286
0
        RETURN_IF_ERROR(calc_delete_bitmap(self, rowset, segments, rowsets_skip_alignment,
1287
0
                                           delete_bitmap, cur_version - 1, token.get(), nullptr));
1288
0
        RETURN_IF_ERROR(token->wait());
1289
0
    }
1290
1291
    // When there is only one segment, it will be calculated in the current thread.
1292
    // Otherwise, it will be submitted to the thread pool for calculation.
1293
3
    if (segments.size() <= 1) {
1294
3
        RETURN_IF_ERROR(calc_delete_bitmap(self, rowset, segments, specified_rowsets, delete_bitmap,
1295
3
                                           cur_version - 1, nullptr, transient_rs_writer.get()));
1296
1297
3
    } else {
1298
0
        auto token = self->calc_delete_bitmap_executor()->create_token();
1299
0
        RETURN_IF_ERROR(calc_delete_bitmap(self, rowset, segments, specified_rowsets, delete_bitmap,
1300
0
                                           cur_version - 1, token.get(),
1301
0
                                           transient_rs_writer.get()));
1302
0
        RETURN_IF_ERROR(token->wait());
1303
0
    }
1304
1305
3
    std::stringstream ss;
1306
3
    ss << "cost(us): (load segments: " << t1 << ", get all rsid: " << t2 - t1
1307
3
       << ", get rowsets: " << t3 - t2
1308
3
       << ", calc delete bitmap: " << watch.get_elapse_time_us() - t3 << ")";
1309
1310
3
    if (config::enable_merge_on_write_correctness_check && rowset->num_rows() != 0) {
1311
        // only do correctness check if the rowset has at least one row written
1312
        // check if all the rowset has ROWSET_SENTINEL_MARK
1313
3
        auto st = self->check_delete_bitmap_correctness(delete_bitmap, cur_version - 1, -1,
1314
3
                                                        cur_rowset_ids, &specified_rowsets);
1315
3
        if (!st.ok()) {
1316
0
            LOG(WARNING) << fmt::format("delete bitmap correctness check failed in publish phase!");
1317
0
        }
1318
3
    }
1319
1320
3
    if (transient_rs_writer) {
1321
0
        auto t4 = watch.get_elapse_time_us();
1322
0
        DBUG_EXECUTE_IF("Tablet.update_delete_bitmap.partial_update_write_rowset_fail", {
1323
0
            if (rand() % 100 < (100 * dp->param("percent", 0.5))) {
1324
0
                LOG_WARNING("Tablet.update_delete_bitmap.partial_update_write_rowset random failed")
1325
0
                        .tag("txn_id", txn_id);
1326
0
                return Status::InternalError(
1327
0
                        "debug update_delete_bitmap partial update write rowset random failed");
1328
0
            }
1329
0
        });
1330
        // build rowset writer and merge transient rowset
1331
0
        RETURN_IF_ERROR(transient_rs_writer->flush());
1332
0
        RowsetSharedPtr transient_rowset;
1333
0
        RETURN_IF_ERROR(transient_rs_writer->build(transient_rowset));
1334
0
        auto old_segments = rowset->num_segments();
1335
0
        rowset->merge_rowset_meta(*transient_rowset->rowset_meta());
1336
0
        auto new_segments = rowset->num_segments();
1337
0
        ss << ", partial update flush rowset (old segment num: " << old_segments
1338
0
           << ", new segment num: " << new_segments << ")"
1339
0
           << ", cost:" << watch.get_elapse_time_us() - t4 << "(us)";
1340
1341
        // update the shared_ptr to new bitmap, which is consistent with current rowset.
1342
0
        txn_info->delete_bitmap = delete_bitmap;
1343
        // erase segment cache cause we will add a segment to rowset
1344
0
        SegmentLoader::instance()->erase_segments(rowset->rowset_id(), rowset->num_segments());
1345
0
    }
1346
1347
3
    size_t total_rows = std::accumulate(
1348
3
            segments.begin(), segments.end(), 0,
1349
3
            [](size_t sum, const segment_v2::SegmentSharedPtr& s) { return sum += s->num_rows(); });
1350
3
    auto t5 = watch.get_elapse_time_us();
1351
3
    RETURN_IF_ERROR(self->save_delete_bitmap(txn_info, txn_id, delete_bitmap,
1352
3
                                             transient_rs_writer.get(), cur_rowset_ids,
1353
3
                                             cur_version));
1354
1355
    // defensive check, check that the delete bitmap cache we wrote is correct
1356
3
    RETURN_IF_ERROR(self->check_delete_bitmap_cache(txn_id, delete_bitmap.get()));
1357
1358
3
    LOG(INFO) << "[Publish] construct delete bitmap tablet: " << self->tablet_id()
1359
3
              << ", rowset_ids to add: "
1360
3
              << (specified_rowsets.size() + rowsets_skip_alignment.size())
1361
3
              << ", rowset_ids to del: " << rowset_ids_to_del.size()
1362
3
              << ", cur version: " << cur_version << ", transaction_id: " << txn_id << ","
1363
3
              << ss.str() << " , total rows: " << total_rows
1364
3
              << ", update delete_bitmap cost: " << watch.get_elapse_time_us() - t5 << "(us)";
1365
3
    return Status::OK();
1366
3
}
1367
1368
void BaseTablet::calc_compaction_output_rowset_delete_bitmap(
1369
        const std::vector<RowsetSharedPtr>& input_rowsets, const RowIdConversion& rowid_conversion,
1370
        uint64_t start_version, uint64_t end_version, std::set<RowLocation>* missed_rows,
1371
        std::map<RowsetSharedPtr, std::list<std::pair<RowLocation, RowLocation>>>* location_map,
1372
0
        const DeleteBitmap& input_delete_bitmap, DeleteBitmap* output_rowset_delete_bitmap) {
1373
0
    RowLocation src;
1374
0
    RowLocation dst;
1375
0
    for (auto& rowset : input_rowsets) {
1376
0
        src.rowset_id = rowset->rowset_id();
1377
0
        for (uint32_t seg_id = 0; seg_id < rowset->num_segments(); ++seg_id) {
1378
0
            src.segment_id = seg_id;
1379
0
            DeleteBitmap subset_map(tablet_id());
1380
0
            input_delete_bitmap.subset({rowset->rowset_id(), seg_id, start_version},
1381
0
                                       {rowset->rowset_id(), seg_id, end_version}, &subset_map);
1382
            // traverse all versions and convert rowid
1383
0
            for (auto iter = subset_map.delete_bitmap.begin();
1384
0
                 iter != subset_map.delete_bitmap.end(); ++iter) {
1385
0
                auto cur_version = std::get<2>(iter->first);
1386
0
                for (auto index = iter->second.begin(); index != iter->second.end(); ++index) {
1387
0
                    src.row_id = *index;
1388
0
                    if (rowid_conversion.get(src, &dst) != 0) {
1389
0
                        VLOG_CRITICAL << "Can't find rowid, may be deleted by the delete_handler, "
1390
0
                                      << " src loaction: |" << src.rowset_id << "|"
1391
0
                                      << src.segment_id << "|" << src.row_id
1392
0
                                      << " version: " << cur_version;
1393
0
                        if (missed_rows) {
1394
0
                            missed_rows->insert(src);
1395
0
                        }
1396
0
                        continue;
1397
0
                    }
1398
0
                    VLOG_DEBUG << "calc_compaction_output_rowset_delete_bitmap dst location: |"
1399
0
                               << dst.rowset_id << "|" << dst.segment_id << "|" << dst.row_id
1400
0
                               << " src location: |" << src.rowset_id << "|" << src.segment_id
1401
0
                               << "|" << src.row_id << " start version: " << start_version
1402
0
                               << "end version" << end_version;
1403
0
                    if (location_map) {
1404
0
                        (*location_map)[rowset].emplace_back(src, dst);
1405
0
                    }
1406
0
                    output_rowset_delete_bitmap->add({dst.rowset_id, dst.segment_id, cur_version},
1407
0
                                                     dst.row_id);
1408
0
                }
1409
0
            }
1410
0
        }
1411
0
    }
1412
0
}
1413
1414
Status BaseTablet::check_rowid_conversion(
1415
        RowsetSharedPtr dst_rowset,
1416
        const std::map<RowsetSharedPtr, std::list<std::pair<RowLocation, RowLocation>>>&
1417
0
                location_map) {
1418
0
    if (location_map.empty()) {
1419
0
        VLOG_DEBUG << "check_rowid_conversion, location_map is empty";
1420
0
        return Status::OK();
1421
0
    }
1422
0
    std::vector<segment_v2::SegmentSharedPtr> dst_segments;
1423
1424
0
    RETURN_IF_ERROR(
1425
0
            std::dynamic_pointer_cast<BetaRowset>(dst_rowset)->load_segments(&dst_segments));
1426
0
    std::unordered_map<RowsetId, std::vector<segment_v2::SegmentSharedPtr>> input_rowsets_segment;
1427
1428
0
    VLOG_DEBUG << "check_rowid_conversion, dst_segments size: " << dst_segments.size();
1429
0
    for (auto [src_rowset, locations] : location_map) {
1430
0
        std::vector<segment_v2::SegmentSharedPtr>& segments =
1431
0
                input_rowsets_segment[src_rowset->rowset_id()];
1432
0
        if (segments.empty()) {
1433
0
            RETURN_IF_ERROR(
1434
0
                    std::dynamic_pointer_cast<BetaRowset>(src_rowset)->load_segments(&segments));
1435
0
        }
1436
0
        for (auto& [src, dst] : locations) {
1437
0
            std::string src_key;
1438
0
            std::string dst_key;
1439
0
            Status s = segments[src.segment_id]->read_key_by_rowid(src.row_id, &src_key);
1440
0
            if (UNLIKELY(s.is<NOT_IMPLEMENTED_ERROR>())) {
1441
0
                LOG(INFO) << "primary key index of old version does not "
1442
0
                             "support reading key by rowid";
1443
0
                break;
1444
0
            }
1445
0
            if (UNLIKELY(!s)) {
1446
0
                LOG(WARNING) << "failed to get src key: |" << src.rowset_id << "|" << src.segment_id
1447
0
                             << "|" << src.row_id << " status: " << s;
1448
0
                DCHECK(false);
1449
0
                return s;
1450
0
            }
1451
1452
0
            s = dst_segments[dst.segment_id]->read_key_by_rowid(dst.row_id, &dst_key);
1453
0
            if (UNLIKELY(!s)) {
1454
0
                LOG(WARNING) << "failed to get dst key: |" << dst.rowset_id << "|" << dst.segment_id
1455
0
                             << "|" << dst.row_id << " status: " << s;
1456
0
                DCHECK(false);
1457
0
                return s;
1458
0
            }
1459
1460
0
            VLOG_DEBUG << "check_rowid_conversion, src: |" << src.rowset_id << "|" << src.segment_id
1461
0
                       << "|" << src.row_id << "|" << src_key << " dst: |" << dst.rowset_id << "|"
1462
0
                       << dst.segment_id << "|" << dst.row_id << "|" << dst_key;
1463
0
            if (UNLIKELY(src_key.compare(dst_key) != 0)) {
1464
0
                LOG(WARNING) << "failed to check key, src key: |" << src.rowset_id << "|"
1465
0
                             << src.segment_id << "|" << src.row_id << "|" << src_key
1466
0
                             << " dst key: |" << dst.rowset_id << "|" << dst.segment_id << "|"
1467
0
                             << dst.row_id << "|" << dst_key;
1468
0
                DCHECK(false);
1469
0
                return Status::InternalError("failed to check rowid conversion");
1470
0
            }
1471
0
        }
1472
0
    }
1473
0
    return Status::OK();
1474
0
}
1475
1476
// The caller should hold _rowset_update_lock and _meta_lock lock.
1477
Status BaseTablet::update_delete_bitmap_without_lock(
1478
        const BaseTabletSPtr& self, const RowsetSharedPtr& rowset,
1479
0
        const std::vector<RowsetSharedPtr>* specified_base_rowsets) {
1480
0
    DBUG_EXECUTE_IF("BaseTablet.update_delete_bitmap_without_lock.random_failed", {
1481
0
        auto rnd = rand() % 100;
1482
0
        auto percent = dp->param("percent", 0.1);
1483
0
        if (rnd < (100 * percent)) {
1484
0
            LOG(WARNING) << "BaseTablet.update_delete_bitmap_without_lock.random_failed";
1485
0
            return Status::InternalError(
1486
0
                    "debug tablet update delete bitmap without lock random failed");
1487
0
        } else {
1488
0
            LOG(INFO) << "BaseTablet.update_delete_bitmap_without_lock.random_failed not "
1489
0
                         "triggered"
1490
0
                      << ", rnd:" << rnd << ", percent: " << percent;
1491
0
        }
1492
0
    });
1493
0
    int64_t cur_version = rowset->start_version();
1494
0
    std::vector<segment_v2::SegmentSharedPtr> segments;
1495
0
    RETURN_IF_ERROR(std::dynamic_pointer_cast<BetaRowset>(rowset)->load_segments(&segments));
1496
1497
    // If this rowset does not have a segment, there is no need for an update.
1498
0
    if (segments.empty()) {
1499
0
        LOG(INFO) << "[Schema Change or Clone] skip to construct delete bitmap tablet: "
1500
0
                  << self->tablet_id() << " cur max_version: " << cur_version;
1501
0
        return Status::OK();
1502
0
    }
1503
1504
    // calculate delete bitmap between segments if necessary.
1505
0
    DeleteBitmapPtr delete_bitmap = std::make_shared<DeleteBitmap>(self->tablet_id());
1506
0
    RETURN_IF_ERROR(self->calc_delete_bitmap_between_segments(
1507
0
            rowset->tablet_schema(), rowset->rowset_id(), segments, delete_bitmap));
1508
1509
    // get all base rowsets to calculate on
1510
0
    std::vector<RowsetSharedPtr> specified_rowsets;
1511
0
    RowsetIdUnorderedSet cur_rowset_ids;
1512
0
    if (specified_base_rowsets == nullptr) {
1513
0
        RETURN_IF_ERROR(self->get_all_rs_id_unlocked(cur_version - 1, &cur_rowset_ids));
1514
0
        specified_rowsets = self->get_rowset_by_ids(&cur_rowset_ids);
1515
0
    } else {
1516
0
        specified_rowsets = *specified_base_rowsets;
1517
0
    }
1518
1519
0
    OlapStopWatch watch;
1520
0
    auto token = self->calc_delete_bitmap_executor()->create_token();
1521
0
    RETURN_IF_ERROR(calc_delete_bitmap(self, rowset, segments, specified_rowsets, delete_bitmap,
1522
0
                                       cur_version - 1, token.get()));
1523
0
    RETURN_IF_ERROR(token->wait());
1524
0
    size_t total_rows = std::accumulate(
1525
0
            segments.begin(), segments.end(), 0,
1526
0
            [](size_t sum, const segment_v2::SegmentSharedPtr& s) { return sum += s->num_rows(); });
1527
0
    LOG(INFO) << "[Schema Change or Clone] construct delete bitmap tablet: " << self->tablet_id()
1528
0
              << ", rowset_ids: " << cur_rowset_ids.size() << ", cur max_version: " << cur_version
1529
0
              << ", transaction_id: " << -1 << ", cost: " << watch.get_elapse_time_us()
1530
0
              << "(us), total rows: " << total_rows;
1531
0
    if (config::enable_merge_on_write_correctness_check) {
1532
        // check if all the rowset has ROWSET_SENTINEL_MARK
1533
0
        auto st = self->check_delete_bitmap_correctness(delete_bitmap, cur_version - 1, -1,
1534
0
                                                        cur_rowset_ids, &specified_rowsets);
1535
0
        if (!st.ok()) {
1536
0
            LOG(WARNING) << fmt::format("delete bitmap correctness check failed in publish phase!");
1537
0
        }
1538
0
        delete_bitmap->remove_sentinel_marks();
1539
0
    }
1540
0
    for (auto& iter : delete_bitmap->delete_bitmap) {
1541
0
        self->_tablet_meta->delete_bitmap()->merge(
1542
0
                {std::get<0>(iter.first), std::get<1>(iter.first), cur_version}, iter.second);
1543
0
    }
1544
1545
0
    return Status::OK();
1546
0
}
1547
1548
0
RowsetSharedPtr BaseTablet::get_rowset(const RowsetId& rowset_id) {
1549
0
    std::shared_lock rdlock(_meta_lock);
1550
0
    for (auto& version_rowset : _rs_version_map) {
1551
0
        if (version_rowset.second->rowset_id() == rowset_id) {
1552
0
            return version_rowset.second;
1553
0
        }
1554
0
    }
1555
0
    for (auto& stale_version_rowset : _stale_rs_version_map) {
1556
0
        if (stale_version_rowset.second->rowset_id() == rowset_id) {
1557
0
            return stale_version_rowset.second;
1558
0
        }
1559
0
    }
1560
0
    return nullptr;
1561
0
}
1562
1563
0
std::vector<RowsetSharedPtr> BaseTablet::get_snapshot_rowset(bool include_stale_rowset) const {
1564
0
    std::shared_lock rdlock(_meta_lock);
1565
0
    std::vector<RowsetSharedPtr> rowsets;
1566
0
    std::transform(_rs_version_map.cbegin(), _rs_version_map.cend(), std::back_inserter(rowsets),
1567
0
                   [](auto& kv) { return kv.second; });
1568
0
    if (include_stale_rowset) {
1569
0
        std::transform(_stale_rs_version_map.cbegin(), _stale_rs_version_map.cend(),
1570
0
                       std::back_inserter(rowsets), [](auto& kv) { return kv.second; });
1571
0
    }
1572
0
    return rowsets;
1573
0
}
1574
1575
void BaseTablet::calc_consecutive_empty_rowsets(
1576
        std::vector<RowsetSharedPtr>* empty_rowsets,
1577
4
        const std::vector<RowsetSharedPtr>& candidate_rowsets, int64_t limit) {
1578
4
    int len = static_cast<int>(candidate_rowsets.size());
1579
12
    for (int i = 0; i < len - 1; ++i) {
1580
9
        auto rowset = candidate_rowsets[i];
1581
9
        auto next_rowset = candidate_rowsets[i + 1];
1582
1583
        // identify two consecutive rowsets that are empty
1584
9
        if (rowset->num_segments() == 0 && next_rowset->num_segments() == 0 &&
1585
9
            !rowset->rowset_meta()->has_delete_predicate() &&
1586
9
            !next_rowset->rowset_meta()->has_delete_predicate() &&
1587
9
            rowset->end_version() == next_rowset->start_version() - 1) {
1588
1
            empty_rowsets->emplace_back(rowset);
1589
1
            empty_rowsets->emplace_back(next_rowset);
1590
1
            rowset = next_rowset;
1591
1
            int next_index = i + 2;
1592
1593
            // keep searching for consecutive empty rowsets
1594
6
            while (next_index < len && candidate_rowsets[next_index]->num_segments() == 0 &&
1595
6
                   !candidate_rowsets[next_index]->rowset_meta()->has_delete_predicate() &&
1596
6
                   rowset->end_version() == candidate_rowsets[next_index]->start_version() - 1) {
1597
5
                empty_rowsets->emplace_back(candidate_rowsets[next_index]);
1598
5
                rowset = candidate_rowsets[next_index++];
1599
5
            }
1600
            // if the number of consecutive empty rowset reach the limit,
1601
            // and there are still rowsets following them
1602
1
            if (empty_rowsets->size() >= limit && next_index < len) {
1603
1
                return;
1604
1
            } else {
1605
                // current rowset is not empty, start searching from that rowset in the next
1606
0
                i = next_index - 1;
1607
0
                empty_rowsets->clear();
1608
0
            }
1609
1
        }
1610
9
    }
1611
4
}
1612
1613
Status BaseTablet::calc_file_crc(uint32_t* crc_value, int64_t start_version, int64_t end_version,
1614
0
                                 int32_t* rowset_count, int64_t* file_count) {
1615
0
    Version v(start_version, end_version);
1616
0
    std::vector<RowsetSharedPtr> rowsets;
1617
0
    traverse_rowsets([&rowsets, &v](const auto& rs) {
1618
        // get all rowsets
1619
0
        if (v.contains(rs->version())) {
1620
0
            rowsets.emplace_back(rs);
1621
0
        }
1622
0
    });
1623
0
    std::sort(rowsets.begin(), rowsets.end(), Rowset::comparator);
1624
0
    *rowset_count = rowsets.size();
1625
1626
0
    *crc_value = 0;
1627
0
    *file_count = 0;
1628
0
    for (const auto& rs : rowsets) {
1629
0
        uint32_t rs_crc_value = 0;
1630
0
        int64_t rs_file_count = 0;
1631
0
        auto rowset = std::static_pointer_cast<BetaRowset>(rs);
1632
0
        auto st = rowset->calc_file_crc(&rs_crc_value, &rs_file_count);
1633
0
        if (!st.ok()) {
1634
0
            return st;
1635
0
        }
1636
        // crc_value is calculated based on the crc_value of each rowset.
1637
0
        *crc_value = crc32c::Extend(*crc_value, reinterpret_cast<const char*>(&rs_crc_value),
1638
0
                                    sizeof(rs_crc_value));
1639
0
        *file_count += rs_file_count;
1640
0
    }
1641
0
    return Status::OK();
1642
0
}
1643
1644
0
Status BaseTablet::show_nested_index_file(std::string* json_meta) {
1645
0
    Version v(0, max_version_unlocked());
1646
0
    std::vector<RowsetSharedPtr> rowsets;
1647
0
    traverse_rowsets([&rowsets, &v](const auto& rs) {
1648
        // get all rowsets
1649
0
        if (v.contains(rs->version())) {
1650
0
            rowsets.emplace_back(rs);
1651
0
        }
1652
0
    });
1653
0
    std::sort(rowsets.begin(), rowsets.end(), Rowset::comparator);
1654
1655
0
    rapidjson::Document doc;
1656
0
    doc.SetObject();
1657
0
    rapidjson::Document::AllocatorType& allocator = doc.GetAllocator();
1658
0
    rapidjson::Value tabletIdValue(tablet_id());
1659
0
    doc.AddMember("tablet_id", tabletIdValue, allocator);
1660
1661
0
    rapidjson::Value rowsets_value(rapidjson::kArrayType);
1662
1663
0
    for (const auto& rs : rowsets) {
1664
0
        rapidjson::Value rowset_value(rapidjson::kObjectType);
1665
1666
0
        auto rowset = std::static_pointer_cast<BetaRowset>(rs);
1667
0
        RETURN_IF_ERROR(rowset->show_nested_index_file(&rowset_value, allocator));
1668
0
        rowsets_value.PushBack(rowset_value, allocator);
1669
0
    }
1670
0
    doc.AddMember("rowsets", rowsets_value, allocator);
1671
1672
0
    rapidjson::StringBuffer buffer;
1673
0
    rapidjson::PrettyWriter<rapidjson::StringBuffer> writer(buffer);
1674
0
    doc.Accept(writer);
1675
0
    *json_meta = std::string(buffer.GetString());
1676
1677
0
    return Status::OK();
1678
0
}
1679
1680
void BaseTablet::get_base_rowset_delete_bitmap_count(
1681
        uint64_t* max_base_rowset_delete_bitmap_score,
1682
0
        int64_t* max_base_rowset_delete_bitmap_score_tablet_id) {
1683
0
    std::vector<RowsetSharedPtr> rowsets_;
1684
0
    std::string base_rowset_id_str;
1685
0
    {
1686
0
        std::shared_lock rowset_ldlock(this->get_header_lock());
1687
0
        for (const auto& it : _rs_version_map) {
1688
0
            rowsets_.emplace_back(it.second);
1689
0
        }
1690
0
    }
1691
0
    std::sort(rowsets_.begin(), rowsets_.end(), Rowset::comparator);
1692
0
    if (!rowsets_.empty()) {
1693
0
        bool base_found = false;
1694
0
        for (auto& rowset : rowsets_) {
1695
0
            if (rowset->start_version() > 2) {
1696
0
                break;
1697
0
            }
1698
0
            base_found = true;
1699
0
            uint64_t base_rowset_delete_bitmap_count =
1700
0
                    this->tablet_meta()->delete_bitmap()->get_count_with_range(
1701
0
                            {rowset->rowset_id(), 0, 0},
1702
0
                            {rowset->rowset_id(), UINT32_MAX, UINT64_MAX});
1703
0
            if (base_rowset_delete_bitmap_count > *max_base_rowset_delete_bitmap_score) {
1704
0
                *max_base_rowset_delete_bitmap_score = base_rowset_delete_bitmap_count;
1705
0
                *max_base_rowset_delete_bitmap_score_tablet_id = this->tablet_id();
1706
0
            }
1707
0
        }
1708
0
        if (!base_found) {
1709
0
            LOG(WARNING) << "can not found base rowset for tablet " << tablet_id();
1710
0
        }
1711
0
    }
1712
0
}
1713
1714
331
void TabletReadSource::fill_delete_predicates() {
1715
331
    DCHECK_EQ(delete_predicates.size(), 0);
1716
331
    auto delete_pred_view =
1717
1.20k
            rs_splits | std::views::transform([](auto&& split) {
1718
1.20k
                return split.rs_reader->rowset()->rowset_meta();
1719
1.20k
            }) |
1720
1.05k
            std::views::filter([](const auto& rs_meta) { return rs_meta->has_delete_predicate(); });
1721
331
    delete_predicates = {delete_pred_view.begin(), delete_pred_view.end()};
1722
331
}
1723
1724
32
int32_t BaseTablet::max_version_config() {
1725
32
    int32_t max_version = tablet_meta()->compaction_policy() == CUMULATIVE_TIME_SERIES_POLICY
1726
32
                                  ? std::max(config::time_series_max_tablet_version_num,
1727
0
                                             config::max_tablet_version_num)
1728
32
                                  : config::max_tablet_version_num;
1729
32
    return max_version;
1730
32
}
1731
1732
} // namespace doris