Coverage Report

Created: 2026-07-13 21:32

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
be/src/storage/tablet/base_tablet.cpp
Line
Count
Source
1
// Licensed to the Apache Software Foundation (ASF) under one
2
// or more contributor license agreements.  See the NOTICE file
3
// distributed with this work for additional information
4
// regarding copyright ownership.  The ASF licenses this file
5
// to you under the Apache License, Version 2.0 (the
6
// "License"); you may not use this file except in compliance
7
// with the License.  You may obtain a copy of the License at
8
//
9
//   http://www.apache.org/licenses/LICENSE-2.0
10
//
11
// Unless required by applicable law or agreed to in writing,
12
// software distributed under the License is distributed on an
13
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
14
// KIND, either express or implied.  See the License for the
15
// specific language governing permissions and limitations
16
// under the License.
17
18
#include "storage/tablet/base_tablet.h"
19
20
#include <bthread/mutex.h>
21
#include <crc32c/crc32c.h>
22
#include <fmt/format.h>
23
#include <rapidjson/prettywriter.h>
24
25
#include <algorithm>
26
#include <cstdint>
27
#include <iterator>
28
#include <random>
29
#include <shared_mutex>
30
31
#include "cloud/cloud_tablet.h"
32
#include "cloud/config.h"
33
#include "common/cast_set.h"
34
#include "common/logging.h"
35
#include "common/metrics/doris_metrics.h"
36
#include "common/status.h"
37
#include "core/assert_cast.h"
38
#include "core/column/column_vector.h"
39
#include "core/data_type/data_type_factory.hpp"
40
#include "load/memtable/memtable.h"
41
#include "service/point_query_executor.h"
42
#include "storage/binlog.h"
43
#include "storage/compaction/cumulative_compaction_time_series_policy.h"
44
#include "storage/delete/calc_delete_bitmap_executor.h"
45
#include "storage/delete/delete_bitmap_calculator.h"
46
#include "storage/index/primary_key_index.h"
47
#include "storage/iterators.h"
48
#include "storage/partial_update_info.h"
49
#include "storage/rowid_conversion.h"
50
#include "storage/rowset/beta_rowset.h"
51
#include "storage/rowset/group_rowset_writer.h"
52
#include "storage/rowset/rowset.h"
53
#include "storage/rowset/rowset_factory.h"
54
#include "storage/rowset/rowset_fwd.h"
55
#include "storage/rowset/rowset_reader.h"
56
#include "storage/rowset/rowset_writer_context.h"
57
#include "storage/segment/column_reader.h"
58
#include "storage/tablet/tablet_fwd.h"
59
#include "storage/txn/txn_manager.h"
60
#include "util/bvar_helper.h"
61
#include "util/debug_points.h"
62
#include "util/jsonb/serialize.h"
63
64
namespace doris {
65
66
using namespace ErrorCode;
67
68
namespace {
69
70
bvar::LatencyRecorder g_tablet_commit_phase_update_delete_bitmap_latency(
71
        "doris_pk", "commit_phase_update_delete_bitmap");
72
bvar::LatencyRecorder g_tablet_lookup_rowkey_latency("doris_pk", "tablet_lookup_rowkey");
73
bvar::Adder<uint64_t> g_tablet_pk_not_found("doris_pk", "lookup_not_found");
74
bvar::PerSecond<bvar::Adder<uint64_t>> g_tablet_pk_not_found_per_second(
75
        "doris_pk", "lookup_not_found_per_second", &g_tablet_pk_not_found, 60);
76
bvar::LatencyRecorder g_tablet_update_delete_bitmap_latency("doris_pk", "update_delete_bitmap");
77
78
static bvar::Adder<size_t> g_total_tablet_num("doris_total_tablet_num");
79
80
Status _get_segment_column_iterator(const BetaRowsetSharedPtr& rowset, uint32_t segid,
81
                                    const TabletColumn& target_column,
82
                                    SegmentCacheHandle* segment_cache_handle,
83
                                    std::unique_ptr<segment_v2::ColumnIterator>* column_iterator,
84
                                    OlapReaderStatistics* stats,
85
0
                                    const io::IOContext* input_io_ctx = nullptr) {
86
0
    RETURN_IF_ERROR(SegmentLoader::instance()->load_segments(rowset, segment_cache_handle, true));
87
    // find segment
88
0
    auto it = std::find_if(
89
0
            segment_cache_handle->get_segments().begin(),
90
0
            segment_cache_handle->get_segments().end(),
91
0
            [&segid](const segment_v2::SegmentSharedPtr& seg) { return seg->id() == segid; });
92
0
    if (it == segment_cache_handle->get_segments().end()) {
93
0
        return Status::NotFound(fmt::format("rowset {} 's segemnt not found, seg_id {}",
94
0
                                            rowset->rowset_id().to_string(), segid));
95
0
    }
96
0
    segment_v2::SegmentSharedPtr segment = *it;
97
0
    StorageReadOptions opts;
98
0
    opts.stats = stats;
99
0
    if (input_io_ctx != nullptr) {
100
0
        opts.io_ctx = *input_io_ctx;
101
0
    }
102
0
    RETURN_IF_ERROR(segment->new_column_iterator(target_column, column_iterator, &opts));
103
0
    auto io_ctx = opts.io_ctx;
104
0
    io_ctx.reader_type = ReaderType::READER_QUERY;
105
0
    io_ctx.file_cache_stats = &stats->file_cache_stats;
106
0
    segment_v2::ColumnIteratorOptions opt {
107
0
            .use_page_cache = !config::disable_storage_page_cache,
108
0
            .file_reader = segment->file_reader().get(),
109
0
            .stats = stats,
110
0
            .io_ctx = io_ctx,
111
0
    };
112
0
    RETURN_IF_ERROR((*column_iterator)->init(opt));
113
0
    return Status::OK();
114
0
}
115
116
} // namespace
117
118
extern MetricPrototype METRIC_query_scan_bytes;
119
extern MetricPrototype METRIC_query_scan_rows;
120
extern MetricPrototype METRIC_query_scan_count;
121
DEFINE_COUNTER_METRIC_PROTOTYPE_2ARG(flush_bytes, MetricUnit::BYTES);
122
DEFINE_COUNTER_METRIC_PROTOTYPE_2ARG(flush_finish_count, MetricUnit::OPERATIONS);
123
124
768
BaseTablet::BaseTablet(TabletMetaSharedPtr tablet_meta) : _tablet_meta(std::move(tablet_meta)) {
125
768
    _metric_entity = DorisMetrics::instance()->metric_registry()->register_entity(
126
768
            fmt::format("Tablet.{}", tablet_id()), {{"tablet_id", std::to_string(tablet_id())}},
127
768
            MetricEntityType::kTablet);
128
768
    INT_COUNTER_METRIC_REGISTER(_metric_entity, query_scan_bytes);
129
768
    INT_COUNTER_METRIC_REGISTER(_metric_entity, query_scan_rows);
130
768
    INT_COUNTER_METRIC_REGISTER(_metric_entity, query_scan_count);
131
768
    INT_COUNTER_METRIC_REGISTER(_metric_entity, flush_bytes);
132
768
    INT_COUNTER_METRIC_REGISTER(_metric_entity, flush_finish_count);
133
134
    // construct _timestamped_versioned_tracker from rs and stale rs meta
135
768
    _timestamped_version_tracker.construct_versioned_tracker(_tablet_meta->all_rs_metas(),
136
768
                                                             _tablet_meta->all_stale_rs_metas());
137
768
    _row_binlog_version_tracker.construct_versioned_tracker(
138
768
            _tablet_meta->all_row_binlog_rs_metas());
139
140
    // if !_tablet_meta->all_rs_metas()[0]->tablet_schema(),
141
    // that mean the tablet_meta is still no upgrade to doris 1.2 versions.
142
    // Before doris 1.2 version, rowset metas don't have tablet schema.
143
    // And when upgrade to doris 1.2 version,
144
    // all rowset metas will be set the tablet schmea from tablet meta.
145
768
    if (_tablet_meta->all_rs_metas().empty() ||
146
768
        !_tablet_meta->all_rs_metas().begin()->second->tablet_schema()) {
147
698
        _max_version_schema = _tablet_meta->tablet_schema();
148
698
    } else {
149
70
        std::vector<RowsetMetaSharedPtr> rowset_metas(_tablet_meta->all_rs_metas().size());
150
70
        std::transform(_tablet_meta->all_rs_metas().begin(), _tablet_meta->all_rs_metas().end(),
151
3.17k
                       rowset_metas.begin(), [](const auto& it) { return it.second; });
152
70
        _max_version_schema = tablet_schema_with_merged_max_schema_version(rowset_metas);
153
70
    }
154
768
    DCHECK(_max_version_schema);
155
768
    g_total_tablet_num << 1;
156
768
}
157
158
768
BaseTablet::~BaseTablet() {
159
768
    DorisMetrics::instance()->metric_registry()->deregister_entity(_metric_entity);
160
768
    g_total_tablet_num << -1;
161
768
}
162
163
TabletSchemaSPtr BaseTablet::tablet_schema_with_merged_max_schema_version(
164
142
        const std::vector<RowsetMetaSharedPtr>& rowset_metas) {
165
142
    RowsetMetaSharedPtr max_schema_version_rs = *std::max_element(
166
3.21k
            rowset_metas.begin(), rowset_metas.end(), [](const auto& a, const auto& b) -> bool {
167
3.21k
                return !a->tablet_schema()
168
3.21k
                               ? true
169
3.21k
                               : (!b->tablet_schema()
170
3.21k
                                          ? false
171
3.21k
                                          : a->tablet_schema()->schema_version() <
172
3.21k
                                                    b->tablet_schema()->schema_version());
173
3.21k
            });
174
142
    return max_schema_version_rs->tablet_schema();
175
142
}
176
177
160
Status BaseTablet::set_tablet_state(TabletState state) {
178
160
    if (_tablet_meta->tablet_state() == TABLET_SHUTDOWN && state != TABLET_SHUTDOWN) {
179
0
        return Status::Error<META_INVALID_ARGUMENT>(
180
0
                "could not change tablet state from shutdown to {}", state);
181
0
    }
182
160
    _tablet_meta->set_tablet_state(state);
183
160
    return Status::OK();
184
160
}
185
186
0
void BaseTablet::update_max_version_schema(const TabletSchemaSPtr& tablet_schema) {
187
0
    std::lock_guard wrlock(_meta_lock);
188
    // Double Check for concurrent update
189
0
    if (!_max_version_schema ||
190
0
        tablet_schema->schema_version() > _max_version_schema->schema_version()) {
191
0
        _max_version_schema = tablet_schema;
192
0
    }
193
0
}
194
195
24
uint32_t BaseTablet::get_real_compaction_score() const {
196
24
    std::shared_lock l(_meta_lock);
197
24
    return get_real_compaction_score_unlocked();
198
24
}
199
200
142
uint32_t BaseTablet::get_real_compaction_score_unlocked() const {
201
142
    const auto& rs_metas = _tablet_meta->all_rs_metas();
202
5.24k
    return std::accumulate(rs_metas.begin(), rs_metas.end(), 0, [](uint32_t score, const auto& it) {
203
5.24k
        return score + it.second->get_compaction_score();
204
5.24k
    });
205
142
}
206
207
Status BaseTablet::capture_rs_readers_unlocked(const Versions& version_path,
208
0
                                               std::vector<RowSetSplits>* rs_splits) const {
209
0
    DCHECK(rs_splits != nullptr && rs_splits->empty());
210
0
    for (auto version : version_path) {
211
0
        auto it = _rs_version_map.find(version);
212
0
        if (it == _rs_version_map.end()) {
213
0
            VLOG_NOTICE << "fail to find Rowset in rs_version for version. tablet=" << tablet_id()
214
0
                        << ", version='" << version.first << "-" << version.second;
215
216
0
            it = _stale_rs_version_map.find(version);
217
0
            if (it == _stale_rs_version_map.end()) {
218
0
                return Status::Error<CAPTURE_ROWSET_READER_ERROR>(
219
0
                        "fail to find Rowset in stale_rs_version for version. tablet={}, "
220
0
                        "version={}-{}",
221
0
                        tablet_id(), version.first, version.second);
222
0
            }
223
0
        }
224
0
        RowsetReaderSharedPtr rs_reader;
225
0
        auto res = it->second->create_reader(&rs_reader);
226
0
        if (!res.ok()) {
227
0
            return Status::Error<CAPTURE_ROWSET_READER_ERROR>(
228
0
                    "failed to create reader for rowset:{}", it->second->rowset_id().to_string());
229
0
        }
230
0
        rs_splits->emplace_back(std::move(rs_reader));
231
0
    }
232
0
    return Status::OK();
233
0
}
234
235
// snapshot manager may call this api to check if version exists, so that
236
// the version maybe not exist
237
RowsetSharedPtr BaseTablet::get_rowset_by_version(const Version& version,
238
16
                                                  bool find_in_stale) const {
239
16
    auto iter = _rs_version_map.find(version);
240
16
    if (iter == _rs_version_map.end()) {
241
0
        if (find_in_stale) {
242
0
            return get_stale_rowset_by_version(version);
243
0
        }
244
0
        return nullptr;
245
0
    }
246
16
    return iter->second;
247
16
}
248
249
0
RowsetSharedPtr BaseTablet::get_stale_rowset_by_version(const Version& version) const {
250
0
    auto iter = _stale_rs_version_map.find(version);
251
0
    if (iter == _stale_rs_version_map.end()) {
252
0
        VLOG_NOTICE << "no rowset for version:" << version << ", tablet: " << tablet_id();
253
0
        return nullptr;
254
0
    }
255
0
    return iter->second;
256
0
}
257
258
0
RowsetSharedPtr BaseTablet::get_row_binlog_rowset_by_version(const Version& version) const {
259
0
    auto iter = _row_binlog_rs_version_map.find(version);
260
0
    if (iter == _row_binlog_rs_version_map.end()) {
261
0
        return nullptr;
262
0
    }
263
0
    return iter->second;
264
0
}
265
266
// Already under _meta_lock
267
42
RowsetSharedPtr BaseTablet::get_rowset_with_max_version() const {
268
42
    Version max_version = _tablet_meta->max_version();
269
42
    if (max_version.first == -1) {
270
0
        return nullptr;
271
0
    }
272
273
42
    auto iter = _rs_version_map.find(max_version);
274
42
    if (iter == _rs_version_map.end()) {
275
0
        DCHECK(false) << "invalid version:" << max_version;
276
0
        return nullptr;
277
0
    }
278
42
    return iter->second;
279
42
}
280
281
0
Status BaseTablet::get_all_rs_id(int64_t max_version, RowsetIdUnorderedSet* rowset_ids) const {
282
0
    std::shared_lock rlock(_meta_lock);
283
0
    return get_all_rs_id_unlocked(max_version, rowset_ids);
284
0
}
285
286
Status BaseTablet::get_all_rs_id_unlocked(int64_t max_version,
287
10
                                          RowsetIdUnorderedSet* rowset_ids) const {
288
    //  Ensure that the obtained versions of rowsets are continuous
289
10
    Version spec_version(0, max_version);
290
10
    Versions version_path;
291
10
    auto st = _timestamped_version_tracker.capture_consistent_versions(spec_version, &version_path);
292
10
    if (!st.ok()) [[unlikely]] {
293
0
        return st;
294
0
    }
295
296
12
    for (auto& ver : version_path) {
297
12
        if (ver.second == 1) {
298
            // [0-1] rowset is empty for each tablet, skip it
299
10
            continue;
300
10
        }
301
2
        auto it = _rs_version_map.find(ver);
302
2
        if (it == _rs_version_map.end()) {
303
0
            return Status::Error<CAPTURE_ROWSET_ERROR, false>(
304
0
                    "fail to find Rowset for version. tablet={}, version={}", tablet_id(),
305
0
                    ver.to_string());
306
0
        }
307
2
        rowset_ids->emplace(it->second->rowset_id());
308
2
    }
309
10
    return Status::OK();
310
10
}
311
312
0
Versions BaseTablet::get_missed_versions(int64_t spec_version) const {
313
0
    DCHECK(spec_version > 0) << "invalid spec_version: " << spec_version;
314
315
0
    Versions existing_versions;
316
0
    {
317
0
        std::shared_lock rdlock(_meta_lock);
318
0
        for (const auto& [ver, _] : _tablet_meta->all_rs_metas()) {
319
0
            existing_versions.emplace_back(ver);
320
0
        }
321
0
    }
322
0
    return calc_missed_versions(spec_version, std::move(existing_versions));
323
0
}
324
325
Versions BaseTablet::get_missed_versions_unlocked(int64_t spec_version,
326
3
                                                  bool capture_row_binlog) const {
327
3
    DCHECK(spec_version > 0) << "invalid spec_version: " << spec_version;
328
329
3
    Versions existing_versions;
330
3
    const auto& rs_metas = capture_row_binlog ? _tablet_meta->all_row_binlog_rs_metas()
331
3
                                              : _tablet_meta->all_rs_metas();
332
12
    for (const auto& [ver, _] : rs_metas) {
333
12
        existing_versions.emplace_back(ver);
334
12
    }
335
3
    return calc_missed_versions(spec_version, std::move(existing_versions));
336
3
}
337
338
2
void BaseTablet::_print_missed_versions(const Versions& missed_versions) const {
339
2
    std::stringstream ss;
340
2
    ss << tablet_id() << " has " << missed_versions.size() << " missed version:";
341
    // print at most 10 version
342
6
    for (int i = 0; i < 10 && i < missed_versions.size(); ++i) {
343
4
        ss << missed_versions[i] << ",";
344
4
    }
345
2
    LOG(WARNING) << ss.str();
346
2
}
347
348
2
bool BaseTablet::_reconstruct_version_tracker_if_necessary() {
349
2
    double data_orphan_vertex_ratio = _timestamped_version_tracker.get_orphan_vertex_ratio();
350
2
    double row_binlog_orphan_vertex_ratio = _row_binlog_version_tracker.get_orphan_vertex_ratio();
351
2
    if (data_orphan_vertex_ratio >= config::tablet_version_graph_orphan_vertex_ratio) {
352
2
        _timestamped_version_tracker.construct_versioned_tracker(
353
2
                _tablet_meta->all_rs_metas(), _tablet_meta->all_stale_rs_metas());
354
2
        return true;
355
2
    } else if (row_binlog_orphan_vertex_ratio >= config::tablet_version_graph_orphan_vertex_ratio) {
356
0
        _row_binlog_version_tracker.construct_versioned_tracker(
357
0
                _tablet_meta->all_row_binlog_rs_metas());
358
0
        return true;
359
0
    }
360
0
    return false;
361
2
}
362
363
// should use this method to get a copy of current tablet meta
364
// there are some rowset meta in local meta store and in in-memory tablet meta
365
// but not in tablet meta in local meta store
366
void BaseTablet::generate_tablet_meta_copy(TabletMeta& new_tablet_meta,
367
0
                                           bool cloud_get_rowset_meta) const {
368
0
    std::shared_lock rdlock(_meta_lock);
369
0
    generate_tablet_meta_copy_unlocked(new_tablet_meta, cloud_get_rowset_meta);
370
0
}
371
372
// this is a unlocked version of generate_tablet_meta_copy()
373
// some method already hold the _meta_lock before calling this,
374
// such as EngineCloneTask::_finish_clone -> tablet->revise_tablet_meta
375
void BaseTablet::generate_tablet_meta_copy_unlocked(TabletMeta& new_tablet_meta,
376
4
                                                    bool cloud_get_rowset_meta) const {
377
4
    TabletMetaPB tablet_meta_pb;
378
4
    _tablet_meta->to_meta_pb(&tablet_meta_pb, cloud_get_rowset_meta);
379
4
    new_tablet_meta.init_from_pb(tablet_meta_pb);
380
4
}
381
382
Status BaseTablet::calc_delete_bitmap_between_segments(
383
        TabletSchemaSPtr schema, RowsetId rowset_id,
384
        const std::vector<segment_v2::SegmentSharedPtr>& segments, DeleteBitmapPtr delete_bitmap,
385
0
        int64_t queue_time_us) {
386
0
    size_t const num_segments = segments.size();
387
0
    if (num_segments < 2) {
388
0
        return Status::OK();
389
0
    }
390
391
0
    OlapStopWatch watch;
392
0
    size_t seq_col_length = 0;
393
0
    if (schema->has_sequence_col()) {
394
0
        auto seq_col_idx = schema->sequence_col_idx();
395
0
        seq_col_length = schema->column(seq_col_idx).length() + 1;
396
0
    }
397
0
    size_t rowid_length = 0;
398
0
    if (!schema->cluster_key_uids().empty()) {
399
0
        rowid_length = PrimaryKeyIndexReader::ROW_ID_LENGTH;
400
0
    }
401
402
0
    MergeIndexDeleteBitmapCalculator calculator;
403
0
    RETURN_IF_ERROR(calculator.init(rowset_id, segments, seq_col_length, rowid_length));
404
405
0
    RETURN_IF_ERROR(calculator.calculate_all(delete_bitmap));
406
407
0
    delete_bitmap->add(
408
0
            {rowset_id, DeleteBitmap::INVALID_SEGMENT_ID, DeleteBitmap::TEMP_VERSION_COMMON},
409
0
            DeleteBitmap::ROWSET_SENTINEL_MARK);
410
0
    LOG(INFO) << fmt::format(
411
0
            "construct delete bitmap between segments, "
412
0
            "tablet: {}, rowset: {}, number of segments: {}, bitmap count: {}, bitmap cardinality: "
413
0
            "{}, queue_time_us: {}, cost {} (us)",
414
0
            tablet_id(), rowset_id.to_string(), num_segments,
415
0
            delete_bitmap->get_delete_bitmap_count(), delete_bitmap->cardinality(), queue_time_us,
416
0
            watch.get_elapse_time_us());
417
0
    return Status::OK();
418
0
}
419
420
std::vector<RowsetSharedPtr> BaseTablet::get_rowset_by_ids(
421
73
        const RowsetIdUnorderedSet* specified_rowset_ids) {
422
73
    std::vector<RowsetSharedPtr> rowsets;
423
73
    for (auto& rs : _rs_version_map) {
424
16
        if (!specified_rowset_ids ||
425
16
            specified_rowset_ids->find(rs.second->rowset_id()) != specified_rowset_ids->end()) {
426
1
            rowsets.push_back(rs.second);
427
1
        }
428
16
    }
429
430
73
    std::sort(rowsets.begin(), rowsets.end(), [](RowsetSharedPtr& lhs, RowsetSharedPtr& rhs) {
431
0
        return lhs->end_version() > rhs->end_version();
432
0
    });
433
73
    return rowsets;
434
73
}
435
436
Status BaseTablet::lookup_row_data(const Slice& encoded_key, const RowLocation& row_location,
437
                                   RowsetSharedPtr input_rowset, OlapReaderStatistics& stats,
438
                                   std::string& values, bool write_to_cache,
439
0
                                   const io::IOContext* io_ctx) {
440
0
    MonotonicStopWatch watch;
441
0
    size_t row_size = 1;
442
0
    watch.start();
443
0
    Defer _defer([&]() {
444
0
        LOG_EVERY_N(INFO, 500) << "get a single_row, cost(us):" << watch.elapsed_time() / 1000
445
0
                               << ", row_size:" << row_size;
446
0
    });
447
448
0
    BetaRowsetSharedPtr rowset = std::static_pointer_cast<BetaRowset>(input_rowset);
449
0
    CHECK(rowset);
450
0
    const TabletSchemaSPtr tablet_schema = rowset->tablet_schema();
451
0
    SegmentCacheHandle segment_cache_handle;
452
0
    std::unique_ptr<segment_v2::ColumnIterator> column_iterator;
453
0
    const auto& column = *DORIS_TRY(tablet_schema->column(BeConsts::ROW_STORE_COL));
454
0
    RETURN_IF_ERROR(_get_segment_column_iterator(rowset, row_location.segment_id, column,
455
0
                                                 &segment_cache_handle, &column_iterator, &stats,
456
0
                                                 io_ctx));
457
    // get and parse tuple row
458
0
    MutableColumnPtr column_ptr = ColumnString::create();
459
0
    std::vector<segment_v2::rowid_t> rowids {static_cast<segment_v2::rowid_t>(row_location.row_id)};
460
0
    RETURN_IF_ERROR(column_iterator->read_by_rowids(rowids.data(), 1, column_ptr));
461
0
    assert(column_ptr->size() == 1);
462
0
    auto* string_column = static_cast<ColumnString*>(column_ptr.get());
463
0
    StringRef value = string_column->get_data_at(0);
464
0
    values = value.to_string();
465
0
    if (write_to_cache) {
466
0
        RowCache::instance()->insert({tablet_id(), encoded_key}, Slice {value.data, value.size});
467
0
    }
468
0
    return Status::OK();
469
0
}
470
471
Status BaseTablet::lookup_row_key(const Slice& encoded_key, TabletSchema* latest_schema,
472
                                  bool with_seq_col,
473
                                  const std::vector<RowsetSharedPtr>& specified_rowsets,
474
                                  RowLocation* row_location, int64_t version,
475
                                  std::vector<std::unique_ptr<SegmentCacheHandle>>& segment_caches,
476
                                  RowsetSharedPtr* rowset, bool with_rowid,
477
                                  std::string* encoded_seq_value, OlapReaderStatistics* stats,
478
2
                                  DeleteBitmapPtr delete_bitmap) {
479
2
    SCOPED_BVAR_LATENCY(g_tablet_lookup_rowkey_latency);
480
2
    size_t seq_col_length = 0;
481
    // use the latest tablet schema to decide if the tablet has sequence column currently
482
2
    const TabletSchema* schema =
483
2
            (latest_schema == nullptr ? _tablet_meta->tablet_schema().get() : latest_schema);
484
2
    if (schema->has_sequence_col() && with_seq_col) {
485
2
        seq_col_length = schema->column(schema->sequence_col_idx()).length() + 1;
486
2
    }
487
2
    size_t rowid_length = 0;
488
2
    if (with_rowid && !schema->cluster_key_uids().empty()) {
489
0
        rowid_length = PrimaryKeyIndexReader::ROW_ID_LENGTH;
490
0
    }
491
2
    Slice key_without_seq =
492
2
            Slice(encoded_key.get_data(), encoded_key.get_size() - seq_col_length - rowid_length);
493
2
    RowLocation loc;
494
495
2
    auto tablet_delete_bitmap =
496
2
            delete_bitmap == nullptr ? _tablet_meta->delete_bitmap_ptr() : delete_bitmap;
497
2
    for (size_t i = 0; i < specified_rowsets.size(); i++) {
498
2
        const auto& rs = specified_rowsets[i];
499
2
        std::vector<KeyBoundsPB> segments_key_bounds;
500
2
        rs->rowset_meta()->get_segments_key_bounds(&segments_key_bounds);
501
2
        int num_segments = cast_set<int>(rs->num_segments());
502
        // MOW lookup requires per-segment bounds. Aggregation must be disabled
503
        // for MOW writers, but enforce at runtime too — indexing segments_key_bounds[j]
504
        // below would be out-of-bounds otherwise.
505
2
        if (UNLIKELY(rs->rowset_meta()->is_segments_key_bounds_aggregated() ||
506
2
                     static_cast<int>(segments_key_bounds.size()) != num_segments)) {
507
0
            return Status::InternalError(
508
0
                    "MOW lookup got rowset with inconsistent segments_key_bounds, rowset_id={}, "
509
0
                    "aggregated={}, bounds_size={}, num_segments={}",
510
0
                    rs->rowset_id().to_string(),
511
0
                    rs->rowset_meta()->is_segments_key_bounds_aggregated(),
512
0
                    segments_key_bounds.size(), num_segments);
513
0
        }
514
2
        std::vector<uint32_t> picked_segments;
515
4
        for (int j = num_segments - 1; j >= 0; j--) {
516
2
            if (_key_is_not_in_segment(key_without_seq, segments_key_bounds[j],
517
2
                                       rs->rowset_meta()->is_segments_key_bounds_truncated())) {
518
0
                continue;
519
0
            }
520
2
            picked_segments.emplace_back(j);
521
2
        }
522
2
        if (picked_segments.empty()) {
523
0
            continue;
524
0
        }
525
526
2
        if (UNLIKELY(segment_caches[i] == nullptr)) {
527
1
            segment_caches[i] = std::make_unique<SegmentCacheHandle>();
528
1
            RETURN_IF_ERROR(SegmentLoader::instance()->load_segments(
529
1
                    std::static_pointer_cast<BetaRowset>(rs), segment_caches[i].get(), true, true));
530
1
        }
531
2
        auto& segments = segment_caches[i]->get_segments();
532
2
        DCHECK_EQ(segments.size(), num_segments);
533
534
2
        for (auto id : picked_segments) {
535
2
            Status s = segments[id]->lookup_row_key(encoded_key, schema, with_seq_col, with_rowid,
536
2
                                                    &loc, stats, encoded_seq_value);
537
2
            if (s.is<KEY_NOT_FOUND>()) {
538
0
                continue;
539
0
            }
540
2
            if (!s.ok() && !s.is<KEY_ALREADY_EXISTS>()) {
541
0
                return s;
542
0
            }
543
2
            if (s.ok() && tablet_delete_bitmap->contains_agg_with_cache_if_eligible(
544
1
                                  {loc.rowset_id, loc.segment_id, version}, loc.row_id)) {
545
                // if has sequence col, we continue to compare the sequence_id of
546
                // all rowsets, util we find an existing key.
547
0
                if (schema->has_sequence_col()) {
548
0
                    continue;
549
0
                }
550
                // The key is deleted, we don't need to search for it any more.
551
0
                break;
552
0
            }
553
            // `st` is either OK or KEY_ALREADY_EXISTS now.
554
            // for partial update, even if the key is already exists, we still need to
555
            // read it's original values to keep all columns align.
556
2
            *row_location = loc;
557
2
            if (rowset) {
558
                // return it's rowset
559
2
                *rowset = rs;
560
2
            }
561
            // find it and return
562
2
            return s;
563
2
        }
564
2
    }
565
0
    g_tablet_pk_not_found << 1;
566
0
    return Status::Error<ErrorCode::KEY_NOT_FOUND>("can't find key in all rowsets");
567
2
}
568
569
// if user pass a token, then all calculation works will submit to a threadpool,
570
// user can get all delete bitmaps from that token.
571
// if `token` is nullptr, the calculation will run in local, and user can get the result
572
// delete bitmap from `delete_bitmap` directly.
573
Status BaseTablet::calc_delete_bitmap(const BaseTabletSPtr& tablet, RowsetSharedPtr rowset,
574
                                      const std::vector<segment_v2::SegmentSharedPtr>& segments,
575
                                      const std::vector<RowsetSharedPtr>& specified_rowsets,
576
                                      DeleteBitmapPtr delete_bitmap, int64_t end_version,
577
                                      CalcDeleteBitmapToken* token, RowsetWriter* rowset_writer,
578
69
                                      DeleteBitmapPtr tablet_delete_bitmap) {
579
69
    if (specified_rowsets.empty() || segments.empty()) {
580
68
        return Status::OK();
581
68
    }
582
583
1
    OlapStopWatch watch;
584
1
    for (const auto& segment : segments) {
585
1
        const auto& seg = segment;
586
1
        if (token != nullptr) {
587
1
            RETURN_IF_ERROR(token->submit(tablet, rowset, seg, specified_rowsets, end_version,
588
1
                                          delete_bitmap, rowset_writer, tablet_delete_bitmap));
589
1
        } else {
590
0
            RETURN_IF_ERROR(tablet->calc_segment_delete_bitmap(
591
0
                    rowset, segment, specified_rowsets, delete_bitmap, end_version, rowset_writer,
592
0
                    tablet_delete_bitmap));
593
0
        }
594
1
    }
595
596
1
    return Status::OK();
597
1
}
598
599
Status BaseTablet::calc_segment_delete_bitmap(RowsetSharedPtr rowset,
600
                                              const segment_v2::SegmentSharedPtr& seg,
601
                                              const std::vector<RowsetSharedPtr>& specified_rowsets,
602
                                              DeleteBitmapPtr delete_bitmap, int64_t end_version,
603
                                              RowsetWriter* rowset_writer,
604
                                              DeleteBitmapPtr tablet_delete_bitmap,
605
1
                                              int64_t queue_time_us) {
606
1
    OlapStopWatch watch;
607
1
    auto rowset_id = rowset->rowset_id();
608
1
    Version dummy_version(end_version + 1, end_version + 1);
609
1
    auto rowset_schema = rowset->tablet_schema();
610
611
1
    PartialUpdateInfo* partial_update_info =
612
1
            rowset_writer != nullptr ? rowset_writer->get_partial_update_info().get() : nullptr;
613
1
    bool is_partial_update = partial_update_info && partial_update_info->is_partial_update();
614
1
    bool need_rewrite_conflict = partial_update_info != nullptr;
615
    // `have_input_seq_column` is for fixed partial update only. For flexible partial update, we should use
616
    // the skip bitmap to determine wheather a row has specified the sequence column
617
1
    bool have_input_seq_column = false;
618
    // `rids_be_overwritten` is for flexible partial update only, it records row ids that is overwritten by
619
    // another row with higher seqeucne value
620
1
    std::set<uint32_t> rids_be_overwritten;
621
1
    if (is_partial_update) {
622
0
        if (partial_update_info->is_fixed_partial_update() && rowset_schema->has_sequence_col()) {
623
0
            std::vector<uint32_t> including_cids =
624
0
                    rowset_writer->get_partial_update_info()->update_cids;
625
0
            have_input_seq_column =
626
0
                    rowset_schema->has_sequence_col() &&
627
0
                    (std::find(including_cids.cbegin(), including_cids.cend(),
628
0
                               rowset_schema->sequence_col_idx()) != including_cids.cend());
629
0
        }
630
0
    }
631
632
1
    if (rowset_schema->num_variant_columns() > 0) {
633
        // During partial updates, the extracted columns of a variant should not be included in the rowset schema.
634
        // This is because the partial update for a variant needs to ignore the extracted columns.
635
        // Otherwise, the schema types in different rowsets might be inconsistent. When performing a partial update,
636
        // the complete variant is constructed by reading all the sub-columns of the variant.
637
0
        rowset_schema = rowset_schema->copy_without_variant_extracted_columns();
638
0
    }
639
    // use for partial update
640
1
    FixedReadPlan read_plan_ori;
641
1
    FixedReadPlan read_plan_update;
642
    // used to read sink-time LSN from input row-binlog rowset
643
1
    FixedReadPlan read_plan_lsn;
644
1
    RowsetSharedPtr row_binlog_rowset;
645
1
    if (rowset_writer != nullptr) {
646
0
        if (auto* group_writer = typeid_cast<GroupRowsetWriter*>(rowset_writer);
647
0
            group_writer != nullptr) {
648
0
            auto& binlog_ctx =
649
0
                    const_cast<RowsetWriterContext&>(group_writer->row_binlog_writer()->context());
650
0
            if (binlog_ctx.write_binlog_opt().enable) {
651
0
                row_binlog_rowset = binlog_ctx.write_binlog_opt()
652
0
                                            .write_binlog_config()
653
0
                                            .source.row_binlog_rowset;
654
0
                DCHECK(row_binlog_rowset != nullptr);
655
0
            }
656
0
        }
657
0
    }
658
1
    int64_t conflict_rows = 0;
659
1
    int64_t new_generated_rows = 0;
660
661
1
    std::map<RowsetId, RowsetSharedPtr> rsid_to_rowset;
662
1
    rsid_to_rowset[rowset_id] = rowset;
663
1
    Block block = rowset_schema->create_block();
664
1
    Block ordered_block = block.clone_empty();
665
1
    uint32_t pos = 0;
666
667
1
    RETURN_IF_ERROR(seg->load_pk_index_and_bf(nullptr)); // We need index blocks to iterate
668
1
    const auto* pk_idx = seg->get_primary_key_index();
669
1
    int64_t total = pk_idx->num_rows();
670
1
    uint32_t row_id = 0;
671
1
    int64_t remaining = total;
672
1
    bool exact_match = false;
673
1
    std::string last_key;
674
1
    int batch_size = 1024;
675
    // The data for each segment may be lookup multiple times. Creating a SegmentCacheHandle
676
    // will update the lru cache, and there will be obvious lock competition in multithreading
677
    // scenarios, so using a segment_caches to cache SegmentCacheHandle.
678
1
    std::vector<std::unique_ptr<SegmentCacheHandle>> segment_caches(specified_rowsets.size());
679
2
    while (remaining > 0) {
680
1
        std::unique_ptr<segment_v2::IndexedColumnIterator> iter;
681
1
        RETURN_IF_ERROR(pk_idx->new_iterator(&iter, nullptr));
682
683
1
        size_t num_to_read = std::min<int64_t>(batch_size, remaining);
684
1
        auto index_type = DataTypeFactory::instance().create_data_type(pk_idx->type(), 1, 0);
685
1
        auto index_column = index_type->create_column();
686
1
        Slice last_key_slice(last_key);
687
1
        RETURN_IF_ERROR(iter->seek_at_or_after(&last_key_slice, &exact_match));
688
1
        auto current_ordinal = iter->get_current_ordinal();
689
1
        DCHECK(total == remaining + current_ordinal)
690
0
                << "total: " << total << ", remaining: " << remaining
691
0
                << ", current_ordinal: " << current_ordinal;
692
693
1
        size_t num_read = num_to_read;
694
1
        RETURN_IF_ERROR(iter->next_batch(&num_read, index_column));
695
1
        DCHECK(num_to_read == num_read)
696
0
                << "num_to_read: " << num_to_read << ", num_read: " << num_read;
697
1
        last_key = index_column->get_data_at(num_read - 1).to_string();
698
699
        // exclude last_key, last_key will be read in next batch.
700
1
        if (num_read == batch_size && num_read != remaining) {
701
0
            num_read -= 1;
702
0
        }
703
3
        for (size_t i = 0; i < num_read; i++, row_id++) {
704
2
            Slice key = Slice(index_column->get_data_at(i).data, index_column->get_data_at(i).size);
705
2
            RowLocation loc;
706
            // calculate row id
707
2
            if (!_tablet_meta->tablet_schema()->cluster_key_uids().empty()) {
708
0
                size_t seq_col_length = 0;
709
0
                if (_tablet_meta->tablet_schema()->has_sequence_col()) {
710
0
                    seq_col_length =
711
0
                            _tablet_meta->tablet_schema()
712
0
                                    ->column(_tablet_meta->tablet_schema()->sequence_col_idx())
713
0
                                    .length() +
714
0
                            1;
715
0
                }
716
0
                size_t rowid_length = PrimaryKeyIndexReader::ROW_ID_LENGTH;
717
0
                Slice key_without_seq =
718
0
                        Slice(key.get_data(), key.get_size() - seq_col_length - rowid_length);
719
0
                Slice rowid_slice =
720
0
                        Slice(key.get_data() + key_without_seq.get_size() + seq_col_length + 1,
721
0
                              rowid_length - 1);
722
0
                const auto* rowid_coder = get_key_coder(FieldType::OLAP_FIELD_TYPE_UNSIGNED_INT);
723
0
                RETURN_IF_ERROR(rowid_coder->decode_ascending(&rowid_slice, rowid_length,
724
0
                                                              (uint8_t*)&row_id));
725
0
            }
726
            // same row in segments should be filtered
727
2
            if (delete_bitmap->contains({rowset_id, seg->id(), DeleteBitmap::TEMP_VERSION_COMMON},
728
2
                                        row_id)) {
729
0
                continue;
730
0
            }
731
732
2
            DBUG_EXECUTE_IF("BaseTablet::calc_segment_delete_bitmap.inject_err", {
733
2
                auto p = dp->param("percent", 0.01);
734
2
                std::mt19937 gen {std::random_device {}()};
735
2
                std::bernoulli_distribution inject_fault {p};
736
2
                if (inject_fault(gen)) {
737
2
                    return Status::InternalError(
738
2
                            "injection error in calc_segment_delete_bitmap, "
739
2
                            "tablet_id={}, rowset_id={}",
740
2
                            tablet_id(), rowset_id.to_string());
741
2
                }
742
2
            });
743
744
2
            RowsetSharedPtr rowset_find;
745
2
            Status st = Status::OK();
746
2
            if (tablet_delete_bitmap == nullptr) {
747
2
                st = lookup_row_key(key, rowset_schema.get(), true, specified_rowsets, &loc,
748
2
                                    dummy_version.first - 1, segment_caches, &rowset_find);
749
2
            } else {
750
0
                st = lookup_row_key(key, rowset_schema.get(), true, specified_rowsets, &loc,
751
0
                                    dummy_version.first - 1, segment_caches, &rowset_find, true,
752
0
                                    nullptr, nullptr, tablet_delete_bitmap);
753
0
            }
754
2
            bool expected_st = st.ok() || st.is<KEY_NOT_FOUND>() || st.is<KEY_ALREADY_EXISTS>();
755
            // It's a defensive DCHECK, we need to exclude some common errors to avoid core-dump
756
            // while stress test
757
2
            DCHECK(expected_st || st.is<MEM_LIMIT_EXCEEDED>())
758
0
                    << "unexpected error status while lookup_row_key:" << st;
759
2
            if (!expected_st) {
760
0
                return st;
761
0
            }
762
2
            if (st.is<KEY_NOT_FOUND>()) {
763
0
                continue;
764
0
            }
765
766
2
            ++conflict_rows;
767
2
            if (st.is<KEY_ALREADY_EXISTS>() &&
768
2
                (!is_partial_update ||
769
1
                 (partial_update_info->is_fixed_partial_update() && have_input_seq_column))) {
770
                // `st.is<KEY_ALREADY_EXISTS>()` means that there exists a row with the same key and larger value
771
                // in seqeunce column.
772
                // - If the current load is not a partial update, we just delete current row.
773
                // - Otherwise, it means that we are doing the alignment process in publish phase due to conflicts
774
                // during concurrent partial updates. And there exists another load which introduces a row with
775
                // the same keys and larger sequence column value published successfully after the commit phase
776
                // of the current load.
777
                //     - If the columns we update include sequence column, we should delete the current row becase the
778
                //       partial update on the current row has been `overwritten` by the previous one with larger sequence
779
                //       column value.
780
                //     - Otherwise, we should combine the values of the missing columns in the previous row and the values
781
                //       of the including columns in the current row into a new row.
782
1
                delete_bitmap->add({rowset_id, seg->id(), DeleteBitmap::TEMP_VERSION_COMMON},
783
1
                                   row_id);
784
1
                continue;
785
                // NOTE: for partial update which doesn't specify the sequence column, we can't use the sequence column value filled in flush phase
786
                // as its final value. Otherwise it may cause inconsistency between replicas.
787
1
            }
788
1
            if (need_rewrite_conflict) {
789
                // In publish version, record rows to be deleted for concurrent update
790
                // For example, if version 5 and 6 update a row, but version 6 only see
791
                // version 4 when write, and when publish version, version 5's value will
792
                // be marked as deleted and it's update is losed.
793
                // So here we should read version 5's columns and build a new row, which is
794
                // consists of version 6's update columns and version 5's origin columns
795
                // here we build 2 read plan for ori values and update values
796
797
                // - for fixed partial update, we should read update columns from current load's rowset
798
                // and read missing columns from previous rowsets to create the final block
799
                // - for flexible partial update, we should read all columns from current load's rowset
800
                // and read non sort key columns from previous rowsets to create the final block
801
                // - for upsert rewrite, we should read all columns from current load's rowset
802
                // So we only need to record rows to read for both mode partial update and upsert rewrite
803
0
                if (is_partial_update) {
804
0
                    read_plan_ori.prepare_to_read(loc, pos);
805
0
                }
806
0
                read_plan_update.prepare_to_read(RowLocation {rowset_id, seg->id(), row_id}, pos);
807
0
                if (row_binlog_rowset != nullptr) {
808
0
                    read_plan_lsn.prepare_to_read(
809
0
                            RowLocation {row_binlog_rowset->rowset_id(), seg->id(), row_id}, pos);
810
0
                }
811
812
                // For flexible partial update, we should use skip bitmap to determine wheather
813
                // a row has specified the sequence column. But skip bitmap should be read from the segment.
814
                // So we record these row ids and process and filter them in `generate_new_block_for_flexible_partial_update()`
815
0
                if (st.is<KEY_ALREADY_EXISTS>() &&
816
0
                    partial_update_info->is_flexible_partial_update()) {
817
0
                    rids_be_overwritten.insert(pos);
818
0
                }
819
820
0
                rsid_to_rowset[rowset_find->rowset_id()] = rowset_find;
821
0
                ++pos;
822
823
                // delete bitmap will be calculate when memtable flush and
824
                // publish. The two stages may see different versions.
825
                // When there is sequence column, the currently imported data
826
                // of rowset may be marked for deletion at memtablet flush or
827
                // publish because the seq column is smaller than the previous
828
                // rowset.
829
                // just set 0 as a unified temporary version number, and update to
830
                // the real version number later.
831
0
                delete_bitmap->add(
832
0
                        {loc.rowset_id, loc.segment_id, DeleteBitmap::TEMP_VERSION_COMMON},
833
0
                        loc.row_id);
834
0
                delete_bitmap->add({rowset_id, seg->id(), DeleteBitmap::TEMP_VERSION_COMMON},
835
0
                                   row_id);
836
0
                ++new_generated_rows;
837
0
                continue;
838
0
            }
839
            // when st = ok
840
1
            delete_bitmap->add({loc.rowset_id, loc.segment_id, DeleteBitmap::TEMP_VERSION_COMMON},
841
1
                               loc.row_id);
842
1
        }
843
1
        remaining -= num_read;
844
1
    }
845
    // DCHECK_EQ(total, row_id) << "segment total rows: " << total << " row_id:" << row_id;
846
847
1
    if (config::enable_merge_on_write_correctness_check) {
848
1
        RowsetIdUnorderedSet rowsetids;
849
1
        for (const auto& specified_rowset : specified_rowsets) {
850
1
            rowsetids.emplace(specified_rowset->rowset_id());
851
1
            VLOG_NOTICE << "[tabletID:" << tablet_id() << "]"
852
0
                        << "[add_sentinel_mark_to_delete_bitmap][end_version:" << end_version << "]"
853
0
                        << "add:" << specified_rowset->rowset_id();
854
1
        }
855
1
        add_sentinel_mark_to_delete_bitmap(delete_bitmap.get(), rowsetids);
856
1
    }
857
858
1
    if (pos > 0) {
859
0
        DCHECK(partial_update_info != nullptr);
860
0
        if (partial_update_info->is_flexible_partial_update()) {
861
0
            RETURN_IF_ERROR(generate_new_block_for_flexible_partial_update(
862
0
                    rowset_schema, partial_update_info, rids_be_overwritten, read_plan_ori,
863
0
                    read_plan_update, rsid_to_rowset, &block));
864
0
        } else {
865
0
            RETURN_IF_ERROR(generate_new_block_for_partial_update(
866
0
                    rowset_schema, partial_update_info, read_plan_ori, read_plan_update,
867
0
                    rsid_to_rowset, &block));
868
0
        }
869
870
        // read sink-time LSN from input row-binlog rowset, aligned with `block` rows.
871
0
        Block lsn_block;
872
0
        if (row_binlog_rowset != nullptr) {
873
0
            auto row_binlog_schema = row_binlog_rowset->tablet_schema();
874
0
            std::vector<uint32_t> lsn_cids = {
875
0
                    static_cast<uint32_t>(row_binlog_schema->binlog_lsn_col_idx())};
876
0
            lsn_block = row_binlog_schema->create_block_by_cids(lsn_cids);
877
0
            std::map<RowsetId, RowsetSharedPtr> rsid_to_row_binlog {
878
0
                    {row_binlog_rowset->rowset_id(), row_binlog_rowset}};
879
0
            std::map<uint32_t, uint32_t> read_index;
880
0
            RETURN_IF_ERROR(read_plan_lsn.read_columns_by_plan(*row_binlog_schema, lsn_cids,
881
0
                                                               rsid_to_row_binlog, lsn_block,
882
0
                                                               &read_index, false));
883
0
        }
884
885
0
        std::vector<uint32_t> sort_perm;
886
0
        RETURN_IF_ERROR(sort_block(block, ordered_block, &sort_perm));
887
0
        auto segment_id = rowset_writer->allocate_segment_id();
888
889
        // Publish-phase partial update may flush transient segments to a GroupRowsetWriter.
890
        // For row-binlog writing, RowBinlogSegmentWriter requires `seg_id -> lsn_ids` to be
891
        // registered before the segment writer is constructed.
892
0
        if (auto* group_writer = typeid_cast<GroupRowsetWriter*>(rowset_writer);
893
0
            group_writer != nullptr) {
894
0
            auto binlog_writer = group_writer->row_binlog_writer();
895
0
            auto& binlog_ctx = const_cast<RowsetWriterContext&>(binlog_writer->context());
896
0
            if (binlog_ctx.write_binlog_opt().enable) {
897
0
                const auto& src =
898
0
                        assert_cast<const ColumnInt64&>(*lsn_block.get_by_position(0).column);
899
0
                auto lsn_ids = std::make_shared<std::vector<int64_t>>();
900
0
                lsn_ids->reserve(sort_perm.size());
901
0
                for (auto p : sort_perm) {
902
0
                    lsn_ids->emplace_back(src.get_data()[p]);
903
0
                }
904
0
                binlog_ctx.write_binlog_opt().write_binlog_config().insert_seg_lsn(
905
0
                        segment_id, std::move(lsn_ids));
906
0
            }
907
0
        }
908
0
        RETURN_IF_ERROR(rowset_writer->flush_single_block(&ordered_block, segment_id));
909
0
        auto cost_us = watch.get_elapse_time_us();
910
0
        if (config::enable_mow_verbose_log || cost_us > 10 * 1000 || queue_time_us > 10 * 1000) {
911
0
            LOG(INFO) << "calc segment delete bitmap for "
912
0
                      << partial_update_info->partial_update_mode_str()
913
0
                      << ", tablet: " << tablet_id() << " rowset: " << rowset_id
914
0
                      << " seg_id: " << seg->id() << " dummy_version: " << end_version + 1
915
0
                      << " rows: " << seg->num_rows() << " conflict rows: " << conflict_rows
916
0
                      << " new generated rows: " << new_generated_rows
917
0
                      << " bitmap num: " << delete_bitmap->get_delete_bitmap_count()
918
0
                      << " bitmap cardinality: " << delete_bitmap->cardinality()
919
0
                      << " queue_time_us: " << queue_time_us << ", cost: " << cost_us << "(us)";
920
0
        }
921
0
        return Status::OK();
922
0
    }
923
1
    auto cost_us = watch.get_elapse_time_us();
924
1
    if (config::enable_mow_verbose_log || cost_us > 10 * 1000 || queue_time_us > 10 * 1000) {
925
0
        LOG(INFO) << "calc segment delete bitmap, tablet: " << tablet_id()
926
0
                  << " rowset: " << rowset_id << " seg_id: " << seg->id()
927
0
                  << " dummy_version: " << end_version + 1 << " rows: " << seg->num_rows()
928
0
                  << " conflict rows: " << conflict_rows
929
0
                  << " bitmap num: " << delete_bitmap->get_delete_bitmap_count()
930
0
                  << " bitmap cardinality: " << delete_bitmap->cardinality()
931
0
                  << " queue_time_us: " << queue_time_us << ", cost: " << cost_us << "(us)";
932
0
    }
933
1
    return Status::OK();
934
1
}
935
936
Status BaseTablet::sort_block(Block& in_block, Block& output_block,
937
0
                              std::vector<uint32_t>* permutation) {
938
0
    ScopedMutableBlock scoped_input_block(&in_block);
939
0
    auto& mutable_input_block = scoped_input_block.mutable_block();
940
0
    ScopedMutableBlock scoped_output_block(&output_block);
941
0
    auto& mutable_output_block = scoped_output_block.mutable_block();
942
943
0
    std::shared_ptr<RowInBlockComparator> vec_row_comparator =
944
0
            std::make_shared<RowInBlockComparator>(_tablet_meta->tablet_schema());
945
0
    vec_row_comparator->set_block(&mutable_input_block);
946
947
0
    std::vector<std::unique_ptr<RowInBlock>> row_in_blocks;
948
0
    const auto input_rows = mutable_input_block.rows();
949
0
    DCHECK(input_rows <= std::numeric_limits<int>::max());
950
0
    row_in_blocks.reserve(input_rows);
951
0
    for (size_t i = 0; i < input_rows; ++i) {
952
0
        row_in_blocks.emplace_back(std::make_unique<RowInBlock>(i));
953
0
    }
954
0
    std::sort(row_in_blocks.begin(), row_in_blocks.end(),
955
0
              [&](const std::unique_ptr<RowInBlock>& l,
956
0
                  const std::unique_ptr<RowInBlock>& r) -> bool {
957
0
                  auto value = (*vec_row_comparator)(l.get(), r.get());
958
0
                  DCHECK(value != 0) << "value equel when sort block, l_pos: " << l->_row_pos
959
0
                                     << " r_pos: " << r->_row_pos;
960
0
                  return value < 0;
961
0
              });
962
0
    std::vector<uint32_t> row_pos_vec;
963
0
    row_pos_vec.reserve(input_rows);
964
0
    for (auto& block : row_in_blocks) {
965
0
        row_pos_vec.emplace_back(block->_row_pos);
966
0
    }
967
0
    scoped_input_block.restore();
968
0
    RETURN_IF_ERROR(mutable_output_block.add_rows(&in_block, row_pos_vec.data(),
969
0
                                                  row_pos_vec.data() + input_rows));
970
0
    if (permutation != nullptr) {
971
0
        *permutation = std::move(row_pos_vec);
972
0
    }
973
0
    return Status::OK();
974
0
}
975
976
// fetch value by row column
977
Status BaseTablet::fetch_value_through_row_column(RowsetSharedPtr input_rowset,
978
                                                  const TabletSchema& tablet_schema, uint32_t segid,
979
                                                  const std::vector<uint32_t>& rowids,
980
0
                                                  const std::vector<uint32_t>& cids, Block& block) {
981
0
    MonotonicStopWatch watch;
982
0
    watch.start();
983
0
    Defer _defer([&]() {
984
0
        LOG_EVERY_N(INFO, 500) << "fetch_value_by_rowids, cost(us):" << watch.elapsed_time() / 1000
985
0
                               << ", row_batch_size:" << rowids.size();
986
0
    });
987
988
0
    BetaRowsetSharedPtr rowset = std::static_pointer_cast<BetaRowset>(input_rowset);
989
0
    CHECK(rowset);
990
0
    CHECK(tablet_schema.has_row_store_for_all_columns());
991
0
    SegmentCacheHandle segment_cache_handle;
992
0
    std::unique_ptr<segment_v2::ColumnIterator> column_iterator;
993
0
    OlapReaderStatistics stats;
994
0
    const auto& column = *DORIS_TRY(tablet_schema.column(BeConsts::ROW_STORE_COL));
995
0
    RETURN_IF_ERROR(_get_segment_column_iterator(rowset, segid, column, &segment_cache_handle,
996
0
                                                 &column_iterator, &stats));
997
    // get and parse tuple row
998
0
    MutableColumnPtr column_ptr = ColumnString::create();
999
0
    RETURN_IF_ERROR(column_iterator->read_by_rowids(rowids.data(), rowids.size(), column_ptr));
1000
0
    assert(column_ptr->size() == rowids.size());
1001
0
    auto* string_column = static_cast<ColumnString*>(column_ptr.get());
1002
0
    DataTypeSerDeSPtrs serdes;
1003
0
    serdes.resize(cids.size());
1004
0
    std::unordered_map<uint32_t, uint32_t> col_uid_to_idx;
1005
0
    std::vector<std::string> default_values;
1006
0
    default_values.resize(cids.size());
1007
0
    for (int i = 0; i < cids.size(); ++i) {
1008
0
        const TabletColumn& tablet_column = tablet_schema.column(cids[i]);
1009
0
        DataTypePtr type = DataTypeFactory::instance().create_data_type(tablet_column);
1010
0
        col_uid_to_idx[tablet_column.unique_id()] = i;
1011
0
        default_values[i] = tablet_column.default_value();
1012
0
        serdes[i] = type->get_serde();
1013
0
    }
1014
0
    RETURN_IF_ERROR(JsonbSerializeUtil::jsonb_to_block(serdes, *string_column, col_uid_to_idx,
1015
0
                                                       block, default_values, {}));
1016
0
    return Status::OK();
1017
0
}
1018
1019
Status BaseTablet::fetch_value_by_rowids(RowsetSharedPtr input_rowset, uint32_t segid,
1020
                                         const std::vector<uint32_t>& rowids,
1021
0
                                         const TabletColumn& tablet_column, MutableColumnPtr& dst) {
1022
0
    MonotonicStopWatch watch;
1023
0
    watch.start();
1024
0
    Defer _defer([&]() {
1025
0
        LOG_EVERY_N(INFO, 500) << "fetch_value_by_rowids, cost(us):" << watch.elapsed_time() / 1000
1026
0
                               << ", row_batch_size:" << rowids.size();
1027
0
    });
1028
1029
    // read row data
1030
0
    BetaRowsetSharedPtr rowset = std::static_pointer_cast<BetaRowset>(input_rowset);
1031
0
    CHECK(rowset);
1032
0
    SegmentCacheHandle segment_cache_handle;
1033
0
    std::unique_ptr<segment_v2::ColumnIterator> column_iterator;
1034
0
    OlapReaderStatistics stats;
1035
0
    RETURN_IF_ERROR(_get_segment_column_iterator(rowset, segid, tablet_column,
1036
0
                                                 &segment_cache_handle, &column_iterator, &stats));
1037
0
    RETURN_IF_ERROR(column_iterator->read_by_rowids(rowids.data(), rowids.size(), dst));
1038
0
    return Status::OK();
1039
0
}
1040
1041
const signed char* BaseTablet::get_delete_sign_column_data(const Block& block,
1042
0
                                                           size_t rows_at_least) {
1043
0
    if (int pos = block.get_position_by_name(DELETE_SIGN); pos != -1) {
1044
0
        const ColumnWithTypeAndName& delete_sign_column = block.get_by_position(pos);
1045
0
        const auto& delete_sign_col = assert_cast<const ColumnInt8&>(*(delete_sign_column.column));
1046
0
        if (delete_sign_col.size() >= rows_at_least) {
1047
0
            return delete_sign_col.get_data().data();
1048
0
        }
1049
0
    }
1050
0
    return nullptr;
1051
0
};
1052
1053
Status BaseTablet::generate_default_value_block(const TabletSchema& schema,
1054
                                                const std::vector<uint32_t>& cids,
1055
                                                const std::vector<std::string>& default_values,
1056
                                                const Block& ref_block,
1057
0
                                                Block& default_value_block) {
1058
0
    auto mutable_default_value_columns_guard = default_value_block.mutate_columns_scoped();
1059
0
    auto& mutable_default_value_columns = mutable_default_value_columns_guard.mutable_columns();
1060
0
    for (auto i = 0; i < cids.size(); ++i) {
1061
0
        const auto& column = schema.column(cids[i]);
1062
0
        if (column.has_default_value()) {
1063
0
            const auto& default_value = default_values[i];
1064
0
            StringRef str(default_value);
1065
0
            RETURN_IF_ERROR(ref_block.get_by_position(i).type->get_serde()->default_from_string(
1066
0
                    str, *mutable_default_value_columns[i]));
1067
0
        }
1068
0
    }
1069
0
    return Status::OK();
1070
0
}
1071
1072
Status BaseTablet::generate_new_block_for_partial_update(
1073
        TabletSchemaSPtr rowset_schema, const PartialUpdateInfo* partial_update_info,
1074
        const FixedReadPlan& read_plan_ori, const FixedReadPlan& read_plan_update,
1075
0
        const std::map<RowsetId, RowsetSharedPtr>& rsid_to_rowset, Block* output_block) {
1076
    // do partial update related works
1077
    // 1. read columns by read plan
1078
    // 2. generate new block
1079
    // 3. write a new segment and modify rowset meta
1080
    // 4. mark current keys deleted
1081
0
    CHECK(output_block);
1082
0
    auto full_mutable_columns_guard = output_block->mutate_columns_scoped();
1083
0
    auto& full_mutable_columns = full_mutable_columns_guard.mutable_columns();
1084
0
    const auto& missing_cids = partial_update_info->missing_cids;
1085
0
    const auto& update_cids = partial_update_info->update_cids;
1086
0
    auto old_block = rowset_schema->create_block_by_cids(missing_cids);
1087
0
    auto update_block = rowset_schema->create_block_by_cids(update_cids);
1088
1089
0
    bool have_input_seq_column = false;
1090
0
    if (rowset_schema->has_sequence_col()) {
1091
0
        have_input_seq_column =
1092
0
                (std::find(update_cids.cbegin(), update_cids.cend(),
1093
0
                           rowset_schema->sequence_col_idx()) != update_cids.cend());
1094
0
    }
1095
1096
    // rowid in the final block(start from 0, increase continuously) -> rowid to read in update_block
1097
0
    std::map<uint32_t, uint32_t> read_index_update;
1098
1099
    // read current rowset first, if a row in the current rowset has delete sign mark
1100
    // we don't need to read values from old block
1101
0
    RETURN_IF_ERROR(read_plan_update.read_columns_by_plan(
1102
0
            *rowset_schema, update_cids, rsid_to_rowset, update_block, &read_index_update, false));
1103
0
    size_t update_rows = read_index_update.size();
1104
0
    for (auto i = 0; i < update_cids.size(); ++i) {
1105
0
        for (auto idx = 0; idx < update_rows; ++idx) {
1106
0
            full_mutable_columns[update_cids[i]]->insert_from(
1107
0
                    *update_block.get_by_position(i).column, read_index_update[idx]);
1108
0
        }
1109
0
    }
1110
1111
    // if there is sequence column in the table, we need to read the sequence column,
1112
    // otherwise it may cause the merge-on-read based compaction policy to produce incorrect results
1113
0
    const auto* __restrict new_block_delete_signs =
1114
0
            rowset_schema->has_sequence_col()
1115
0
                    ? nullptr
1116
0
                    : get_delete_sign_column_data(update_block, update_rows);
1117
1118
    // 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
1119
0
    std::map<uint32_t, uint32_t> read_index_old;
1120
0
    RETURN_IF_ERROR(read_plan_ori.read_columns_by_plan(*rowset_schema, missing_cids, rsid_to_rowset,
1121
0
                                                       old_block, &read_index_old, true,
1122
0
                                                       new_block_delete_signs));
1123
0
    size_t old_rows = read_index_old.size();
1124
0
    const auto* __restrict old_block_delete_signs =
1125
0
            get_delete_sign_column_data(old_block, old_rows);
1126
0
    DCHECK(old_block_delete_signs != nullptr);
1127
    // build default value block
1128
0
    auto default_value_block = old_block.clone_empty();
1129
0
    RETURN_IF_ERROR(BaseTablet::generate_default_value_block(*rowset_schema, missing_cids,
1130
0
                                                             partial_update_info->default_values,
1131
0
                                                             old_block, default_value_block));
1132
1133
0
    CHECK(update_rows >= old_rows);
1134
1135
    // build full block
1136
0
    for (auto i = 0; i < missing_cids.size(); ++i) {
1137
0
        const auto& rs_column = rowset_schema->column(missing_cids[i]);
1138
0
        auto& mutable_column = full_mutable_columns[missing_cids[i]];
1139
0
        for (auto idx = 0; idx < update_rows; ++idx) {
1140
            // There are two cases we don't need to read values from old data:
1141
            //     1. if the conflicting new row's delete sign is marked, which means the value columns
1142
            //     of the row will not be read. So we don't need to read the missing values from the previous rows.
1143
            //     2. if the conflicting old row's delete sign is marked, which means that the key is not exist now,
1144
            //     we should not read old values from the deleted data, and should use default value instead.
1145
            //     NOTE: since now we are in the publishing phase, all data is commited
1146
            //         before, even the `strict_mode` is true (which requires partial update
1147
            //         load job can't insert new keys), this "new" key MUST be written into
1148
            //         the new generated segment file.
1149
0
            bool new_row_delete_sign =
1150
0
                    (new_block_delete_signs != nullptr && new_block_delete_signs[idx]);
1151
0
            if (new_row_delete_sign) {
1152
0
                mutable_column->insert_default();
1153
0
            } else {
1154
0
                bool use_default = false;
1155
0
                bool old_row_delete_sign = (old_block_delete_signs != nullptr &&
1156
0
                                            old_block_delete_signs[read_index_old.at(idx)] != 0);
1157
0
                if (old_row_delete_sign) {
1158
0
                    if (!rowset_schema->has_sequence_col()) {
1159
0
                        use_default = true;
1160
0
                    } else if (have_input_seq_column || !rs_column.is_seqeunce_col()) {
1161
                        // to keep the sequence column value not decreasing, we should read values of seq column
1162
                        // from old rows even if the old row is deleted when the input don't specify the sequence column, otherwise
1163
                        // it may cause the merge-on-read based compaction to produce incorrect results
1164
0
                        use_default = true;
1165
0
                    }
1166
0
                }
1167
1168
0
                if (use_default) {
1169
0
                    if (rs_column.has_default_value()) {
1170
0
                        mutable_column->insert_from(*default_value_block.get_by_position(i).column,
1171
0
                                                    0);
1172
0
                    } else if (rs_column.is_nullable()) {
1173
0
                        assert_cast<ColumnNullable*, TypeCheckOnRelease::DISABLE>(
1174
0
                                mutable_column.get())
1175
0
                                ->insert_default();
1176
0
                    } else {
1177
0
                        mutable_column->insert_default();
1178
0
                    }
1179
0
                } else {
1180
0
                    mutable_column->insert_from(*old_block.get_by_position(i).column,
1181
0
                                                read_index_old[idx]);
1182
0
                }
1183
0
            }
1184
0
        }
1185
0
    }
1186
0
    full_mutable_columns_guard.restore();
1187
0
    VLOG_DEBUG << "full block when publish: " << output_block->dump_data();
1188
0
    return Status::OK();
1189
0
}
1190
1191
static void fill_cell_for_flexible_partial_update(
1192
        std::map<uint32_t, uint32_t>& read_index_old,
1193
        std::map<uint32_t, uint32_t>& read_index_update, const TabletSchemaSPtr& rowset_schema,
1194
        const PartialUpdateInfo* partial_update_info, const TabletColumn& tablet_column,
1195
        std::size_t idx, MutableColumnPtr& new_col, const IColumn& default_value_col,
1196
        const IColumn& old_value_col, const IColumn& cur_col, bool skipped,
1197
0
        bool row_has_sequence_col, const signed char* delete_sign_column_data) {
1198
0
    if (skipped) {
1199
0
        bool use_default = false;
1200
0
        bool old_row_delete_sign =
1201
0
                (delete_sign_column_data != nullptr &&
1202
0
                 delete_sign_column_data[read_index_old[cast_set<uint32_t>(idx)]] != 0);
1203
0
        if (old_row_delete_sign) {
1204
0
            if (!rowset_schema->has_sequence_col()) {
1205
0
                use_default = true;
1206
0
            } else if (row_has_sequence_col || (!tablet_column.is_seqeunce_col() &&
1207
0
                                                (tablet_column.unique_id() !=
1208
0
                                                 partial_update_info->sequence_map_col_uid()))) {
1209
                // to keep the sequence column value not decreasing, we should read values of seq column(and seq map column)
1210
                // from old rows even if the old row is deleted when the input don't specify the sequence column, otherwise
1211
                // it may cause the merge-on-read based compaction to produce incorrect results
1212
0
                use_default = true;
1213
0
            }
1214
0
        }
1215
0
        if (use_default) {
1216
0
            if (tablet_column.has_default_value()) {
1217
0
                new_col->insert_from(default_value_col, 0);
1218
0
            } else if (tablet_column.is_nullable()) {
1219
0
                assert_cast<ColumnNullable*, TypeCheckOnRelease::DISABLE>(new_col.get())
1220
0
                        ->insert_many_defaults(1);
1221
0
            } else if (tablet_column.is_auto_increment()) {
1222
                // For auto-increment column, its default value(generated value) is filled in current block in flush phase
1223
                // when the load doesn't specify the auto-increment column
1224
                //     - if the previous conflicting row is deleted, we should use the value in current block as its final value
1225
                //     - if the previous conflicting row is an insert, we should use the value in old block as its final value to
1226
                //       keep consistency between replicas
1227
0
                new_col->insert_from(cur_col, read_index_update[cast_set<uint32_t>(idx)]);
1228
0
            } else {
1229
0
                new_col->insert_default();
1230
0
            }
1231
0
        } else {
1232
0
            new_col->insert_from(old_value_col, read_index_old[cast_set<uint32_t>(idx)]);
1233
0
        }
1234
0
    } else {
1235
0
        new_col->insert_from(cur_col, read_index_update[cast_set<uint32_t>(idx)]);
1236
0
    }
1237
0
}
1238
1239
Status BaseTablet::generate_new_block_for_flexible_partial_update(
1240
        TabletSchemaSPtr rowset_schema, const PartialUpdateInfo* partial_update_info,
1241
        std::set<uint32_t>& rids_be_overwritten, const FixedReadPlan& read_plan_ori,
1242
        const FixedReadPlan& read_plan_update,
1243
0
        const std::map<RowsetId, RowsetSharedPtr>& rsid_to_rowset, Block* output_block) {
1244
0
    CHECK(output_block);
1245
1246
0
    int32_t seq_col_unique_id = -1;
1247
0
    if (rowset_schema->has_sequence_col()) {
1248
0
        seq_col_unique_id = rowset_schema->column(rowset_schema->sequence_col_idx()).unique_id();
1249
0
    }
1250
0
    const auto& non_sort_key_cids = partial_update_info->missing_cids;
1251
0
    std::vector<uint32_t> all_cids(rowset_schema->num_columns());
1252
0
    std::iota(all_cids.begin(), all_cids.end(), 0);
1253
0
    auto old_block = rowset_schema->create_block_by_cids(non_sort_key_cids);
1254
0
    auto update_block = rowset_schema->create_block_by_cids(all_cids);
1255
1256
    // rowid in the final block(start from 0, increase continuously) -> rowid to read in update_block
1257
0
    std::map<uint32_t, uint32_t> read_index_update;
1258
1259
    // 1. read the current rowset first, if a row in the current rowset has delete sign mark
1260
    // we don't need to read values from old block for that row
1261
0
    RETURN_IF_ERROR(read_plan_update.read_columns_by_plan(*rowset_schema, all_cids, rsid_to_rowset,
1262
0
                                                          update_block, &read_index_update, true));
1263
0
    size_t update_rows = read_index_update.size();
1264
1265
    // TODO(bobhan1): add the delete sign optimazation here
1266
    // // if there is sequence column in the table, we need to read the sequence column,
1267
    // // otherwise it may cause the merge-on-read based compaction policy to produce incorrect results
1268
    // const auto* __restrict new_block_delete_signs =
1269
    //         rowset_schema->has_sequence_col()
1270
    //                 ? nullptr
1271
    //                 : get_delete_sign_column_data(update_block, update_rows);
1272
1273
    // 2. read previous rowsets
1274
    // 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
1275
0
    std::map<uint32_t, uint32_t> read_index_old;
1276
0
    RETURN_IF_ERROR(read_plan_ori.read_columns_by_plan(
1277
0
            *rowset_schema, non_sort_key_cids, rsid_to_rowset, old_block, &read_index_old, true));
1278
0
    size_t old_rows = read_index_old.size();
1279
0
    DCHECK(update_rows == old_rows);
1280
0
    const auto* __restrict old_block_delete_signs =
1281
0
            get_delete_sign_column_data(old_block, old_rows);
1282
0
    DCHECK(old_block_delete_signs != nullptr);
1283
1284
    // 3. build default value block
1285
0
    auto default_value_block = old_block.clone_empty();
1286
0
    RETURN_IF_ERROR(BaseTablet::generate_default_value_block(*rowset_schema, non_sort_key_cids,
1287
0
                                                             partial_update_info->default_values,
1288
0
                                                             old_block, default_value_block));
1289
1290
    // 4. build the final block
1291
0
    auto full_mutable_columns_guard = output_block->mutate_columns_scoped();
1292
0
    auto& full_mutable_columns = full_mutable_columns_guard.mutable_columns();
1293
0
    DCHECK(rowset_schema->has_skip_bitmap_col());
1294
0
    auto skip_bitmap_col_idx = rowset_schema->skip_bitmap_col_idx();
1295
0
    const std::vector<BitmapValue>* skip_bitmaps =
1296
0
            &(assert_cast<const ColumnBitmap*, TypeCheckOnRelease::DISABLE>(
1297
0
                      update_block.get_by_position(skip_bitmap_col_idx).column->get_ptr().get())
1298
0
                      ->get_data());
1299
1300
0
    if (rowset_schema->has_sequence_col() && !rids_be_overwritten.empty()) {
1301
        // If the row specifies the sequence column, we should delete the current row becase the
1302
        // flexible partial update on the current row has been `overwritten` by the previous one with larger sequence
1303
        // column value.
1304
0
        for (auto it = rids_be_overwritten.begin(); it != rids_be_overwritten.end();) {
1305
0
            auto rid = *it;
1306
0
            if (!skip_bitmaps->at(rid).contains(seq_col_unique_id)) {
1307
0
                ++it;
1308
0
            } else {
1309
0
                it = rids_be_overwritten.erase(it);
1310
0
            }
1311
0
        }
1312
0
    }
1313
1314
0
    for (std::size_t cid {0}; cid < rowset_schema->num_columns(); cid++) {
1315
0
        MutableColumnPtr& new_col = full_mutable_columns[cid];
1316
0
        const IColumn& cur_col = *update_block.get_by_position(cid).column;
1317
0
        const auto& rs_column = rowset_schema->column(cid);
1318
0
        auto col_uid = rs_column.unique_id();
1319
0
        for (auto idx = 0; idx < update_rows; ++idx) {
1320
0
            if (cid < rowset_schema->num_key_columns()) {
1321
0
                new_col->insert_from(cur_col, read_index_update[idx]);
1322
0
            } else {
1323
0
                const IColumn& default_value_col =
1324
0
                        *default_value_block.get_by_position(cid - rowset_schema->num_key_columns())
1325
0
                                 .column;
1326
0
                const IColumn& old_value_col =
1327
0
                        *old_block.get_by_position(cid - rowset_schema->num_key_columns()).column;
1328
0
                if (rids_be_overwritten.contains(idx)) {
1329
0
                    new_col->insert_from(old_value_col, read_index_old[idx]);
1330
0
                } else {
1331
0
                    fill_cell_for_flexible_partial_update(
1332
0
                            read_index_old, read_index_update, rowset_schema, partial_update_info,
1333
0
                            rs_column, idx, new_col, default_value_col, old_value_col, cur_col,
1334
0
                            skip_bitmaps->at(idx).contains(col_uid),
1335
0
                            rowset_schema->has_sequence_col()
1336
0
                                    ? !skip_bitmaps->at(idx).contains(seq_col_unique_id)
1337
0
                                    : false,
1338
0
                            old_block_delete_signs);
1339
0
                }
1340
0
            }
1341
0
        }
1342
0
        DCHECK_EQ(full_mutable_columns[cid]->size(), update_rows);
1343
0
    }
1344
1345
0
    full_mutable_columns_guard.restore();
1346
0
    VLOG_DEBUG << "full block when publish: " << output_block->dump_data();
1347
0
    return Status::OK();
1348
0
}
1349
1350
Status BaseTablet::commit_phase_update_delete_bitmap(
1351
        const BaseTabletSPtr& tablet, const RowsetSharedPtr& rowset,
1352
        RowsetIdUnorderedSet& pre_rowset_ids, DeleteBitmapPtr delete_bitmap,
1353
        const std::vector<segment_v2::SegmentSharedPtr>& segments, int64_t txn_id,
1354
3
        CalcDeleteBitmapToken* token, RowsetWriter* rowset_writer) {
1355
3
    DBUG_EXECUTE_IF("BaseTablet::commit_phase_update_delete_bitmap.enable_spin_wait", {
1356
3
        auto tok = dp->param<std::string>("token", "invalid_token");
1357
3
        while (DebugPoints::instance()->is_enable(
1358
3
                "BaseTablet::commit_phase_update_delete_bitmap.block")) {
1359
3
            auto block_dp = DebugPoints::instance()->get_debug_point(
1360
3
                    "BaseTablet::commit_phase_update_delete_bitmap.block");
1361
3
            if (block_dp) {
1362
3
                auto pass_token = block_dp->param<std::string>("pass_token", "");
1363
3
                if (pass_token == tok) {
1364
3
                    break;
1365
3
                }
1366
3
            }
1367
3
            std::this_thread::sleep_for(std::chrono::milliseconds(50));
1368
3
        }
1369
3
    });
1370
3
    SCOPED_BVAR_LATENCY(g_tablet_commit_phase_update_delete_bitmap_latency);
1371
3
    RowsetIdUnorderedSet cur_rowset_ids;
1372
3
    RowsetIdUnorderedSet rowset_ids_to_add;
1373
3
    RowsetIdUnorderedSet rowset_ids_to_del;
1374
3
    int64_t cur_version;
1375
1376
3
    std::vector<RowsetSharedPtr> specified_rowsets;
1377
3
    {
1378
        // to prevent seeing intermediate state of a tablet
1379
3
        std::unique_lock<bthread::Mutex> sync_lock;
1380
3
        if (config::is_cloud_mode()) {
1381
0
            sync_lock = std::unique_lock<bthread::Mutex>(
1382
0
                    std::static_pointer_cast<CloudTablet>(tablet)->get_sync_meta_lock());
1383
0
        }
1384
3
        std::shared_lock meta_rlock(tablet->_meta_lock);
1385
3
        if (tablet->tablet_state() == TABLET_NOTREADY) {
1386
            // tablet is under alter process. The delete bitmap will be calculated after conversion.
1387
0
            LOG(INFO) << "tablet is under alter process, delete bitmap will be calculated later, "
1388
0
                         "tablet_id: "
1389
0
                      << tablet->tablet_id() << " txn_id: " << txn_id;
1390
0
            return Status::OK();
1391
0
        }
1392
3
        cur_version = tablet->max_version_unlocked();
1393
3
        RETURN_IF_ERROR(tablet->get_all_rs_id_unlocked(cur_version, &cur_rowset_ids));
1394
3
        _rowset_ids_difference(cur_rowset_ids, pre_rowset_ids, &rowset_ids_to_add,
1395
3
                               &rowset_ids_to_del);
1396
3
        specified_rowsets = tablet->get_rowset_by_ids(&rowset_ids_to_add);
1397
3
    }
1398
0
    for (const auto& to_del : rowset_ids_to_del) {
1399
0
        delete_bitmap->remove({to_del, 0, 0}, {to_del, UINT32_MAX, INT64_MAX});
1400
0
    }
1401
1402
3
    RETURN_IF_ERROR(calc_delete_bitmap(tablet, rowset, segments, specified_rowsets, delete_bitmap,
1403
3
                                       cur_version, token, rowset_writer));
1404
3
    size_t total_rows = std::accumulate(
1405
3
            segments.begin(), segments.end(), 0,
1406
3
            [](size_t sum, const segment_v2::SegmentSharedPtr& s) { return sum += s->num_rows(); });
1407
3
    LOG(INFO) << "[Before Commit] construct delete bitmap tablet: " << tablet->tablet_id()
1408
3
              << ", rowset_ids to add: " << rowset_ids_to_add.size()
1409
3
              << ", rowset_ids to del: " << rowset_ids_to_del.size()
1410
3
              << ", cur max_version: " << cur_version << ", transaction_id: " << txn_id
1411
3
              << ", total rows: " << total_rows;
1412
3
    pre_rowset_ids = cur_rowset_ids;
1413
3
    return Status::OK();
1414
3
}
1415
1416
void BaseTablet::add_sentinel_mark_to_delete_bitmap(DeleteBitmap* delete_bitmap,
1417
5
                                                    const RowsetIdUnorderedSet& rowsetids) {
1418
5
    for (const auto& rowsetid : rowsetids) {
1419
5
        delete_bitmap->add(
1420
5
                {rowsetid, DeleteBitmap::INVALID_SEGMENT_ID, DeleteBitmap::TEMP_VERSION_COMMON},
1421
5
                DeleteBitmap::ROWSET_SENTINEL_MARK);
1422
5
    }
1423
5
}
1424
1425
void BaseTablet::_rowset_ids_difference(const RowsetIdUnorderedSet& cur,
1426
                                        const RowsetIdUnorderedSet& pre,
1427
                                        RowsetIdUnorderedSet* to_add,
1428
6
                                        RowsetIdUnorderedSet* to_del) {
1429
6
    for (const auto& id : cur) {
1430
2
        if (pre.find(id) == pre.end()) {
1431
1
            to_add->insert(id);
1432
1
        }
1433
2
    }
1434
6
    for (const auto& id : pre) {
1435
1
        if (cur.find(id) == cur.end()) {
1436
0
            to_del->insert(id);
1437
0
        }
1438
1
    }
1439
6
}
1440
1441
Status BaseTablet::check_delete_bitmap_correctness(DeleteBitmapPtr delete_bitmap,
1442
                                                   int64_t max_version, int64_t txn_id,
1443
                                                   const RowsetIdUnorderedSet& rowset_ids,
1444
6
                                                   std::vector<RowsetSharedPtr>* rowsets) {
1445
6
    RowsetIdUnorderedSet missing_ids;
1446
6
    for (const auto& rowsetid : rowset_ids) {
1447
2
        if (!delete_bitmap->delete_bitmap.contains({rowsetid, DeleteBitmap::INVALID_SEGMENT_ID,
1448
2
                                                    DeleteBitmap::TEMP_VERSION_COMMON})) {
1449
0
            missing_ids.insert(rowsetid);
1450
0
        }
1451
2
    }
1452
1453
6
    if (!missing_ids.empty()) {
1454
0
        LOG(WARNING) << "[txn_id:" << txn_id << "][tablet_id:" << tablet_id()
1455
0
                     << "][max_version: " << max_version
1456
0
                     << "] check delete bitmap correctness failed!";
1457
0
        rapidjson::Document root;
1458
0
        root.SetObject();
1459
0
        rapidjson::Document required_rowsets_arr;
1460
0
        required_rowsets_arr.SetArray();
1461
0
        rapidjson::Document missing_rowsets_arr;
1462
0
        missing_rowsets_arr.SetArray();
1463
1464
0
        if (rowsets != nullptr) {
1465
0
            for (const auto& rowset : *rowsets) {
1466
0
                rapidjson::Value value;
1467
0
                std::string version_str = rowset->get_rowset_info_str();
1468
0
                value.SetString(version_str.c_str(),
1469
0
                                cast_set<rapidjson::SizeType>(version_str.length()),
1470
0
                                required_rowsets_arr.GetAllocator());
1471
0
                required_rowsets_arr.PushBack(value, required_rowsets_arr.GetAllocator());
1472
0
            }
1473
0
        } else {
1474
0
            std::vector<RowsetSharedPtr> tablet_rowsets;
1475
0
            {
1476
0
                std::shared_lock meta_rlock(_meta_lock);
1477
0
                tablet_rowsets = get_rowset_by_ids(&rowset_ids);
1478
0
            }
1479
0
            for (const auto& rowset : tablet_rowsets) {
1480
0
                rapidjson::Value value;
1481
0
                std::string version_str = rowset->get_rowset_info_str();
1482
0
                value.SetString(version_str.c_str(),
1483
0
                                cast_set<rapidjson::SizeType>(version_str.length()),
1484
0
                                required_rowsets_arr.GetAllocator());
1485
0
                required_rowsets_arr.PushBack(value, required_rowsets_arr.GetAllocator());
1486
0
            }
1487
0
        }
1488
0
        for (const auto& missing_rowset_id : missing_ids) {
1489
0
            rapidjson::Value miss_value;
1490
0
            std::string rowset_id_str = missing_rowset_id.to_string();
1491
0
            miss_value.SetString(rowset_id_str.c_str(),
1492
0
                                 cast_set<rapidjson::SizeType>(rowset_id_str.length()),
1493
0
                                 missing_rowsets_arr.GetAllocator());
1494
0
            missing_rowsets_arr.PushBack(miss_value, missing_rowsets_arr.GetAllocator());
1495
0
        }
1496
1497
0
        root.AddMember("required_rowsets", required_rowsets_arr, root.GetAllocator());
1498
0
        root.AddMember("missing_rowsets", missing_rowsets_arr, root.GetAllocator());
1499
0
        rapidjson::StringBuffer strbuf;
1500
0
        rapidjson::PrettyWriter<rapidjson::StringBuffer> writer(strbuf);
1501
0
        root.Accept(writer);
1502
0
        std::string rowset_status_string = std::string(strbuf.GetString());
1503
0
        LOG_EVERY_SECOND(WARNING) << rowset_status_string;
1504
        // let it crash if correctness check failed in Debug mode
1505
0
        DCHECK(false) << "delete bitmap correctness check failed in publish phase!";
1506
0
        return Status::InternalError("check delete bitmap failed!");
1507
0
    }
1508
6
    return Status::OK();
1509
6
}
1510
1511
Status BaseTablet::update_delete_bitmap(const BaseTabletSPtr& self, TabletTxnInfo* txn_info,
1512
                                        int64_t txn_id, int64_t txn_expiration,
1513
3
                                        DeleteBitmapPtr tablet_delete_bitmap) {
1514
3
    SCOPED_BVAR_LATENCY(g_tablet_update_delete_bitmap_latency);
1515
3
    RowsetIdUnorderedSet cur_rowset_ids;
1516
3
    RowsetIdUnorderedSet rowset_ids_to_add;
1517
3
    RowsetIdUnorderedSet rowset_ids_to_del;
1518
3
    RowsetSharedPtr rowset = txn_info->rowset;
1519
3
    RowsetSharedPtr row_binlog_rowset;
1520
3
    bool build_row_binlog = false;
1521
3
    int64_t cur_version = rowset->start_version();
1522
3
    std::unique_ptr<RowsetWriter> transient_rs_writer;
1523
3
    DeleteBitmapPtr delete_bitmap = txn_info->delete_bitmap;
1524
3
    bool is_partial_update =
1525
3
            txn_info->partial_update_info && txn_info->partial_update_info->is_partial_update();
1526
3
    for (const auto& rs : txn_info->attach_rowsets) {
1527
0
        if (rs != nullptr && rs->rowset_meta() != nullptr && rs->rowset_meta()->is_row_binlog()) {
1528
0
            row_binlog_rowset = rs;
1529
0
            build_row_binlog = is_partial_update ||
1530
0
                               self->tablet_meta()->binlog_config().need_historical_value();
1531
0
            break;
1532
0
        }
1533
0
    }
1534
1535
    // rewrite conflict only when partial update or need before
1536
3
    if (is_partial_update || build_row_binlog) {
1537
0
        if (txn_info->partial_update_info == nullptr) {
1538
0
            txn_info->partial_update_info = std::make_shared<PartialUpdateInfo>();
1539
0
        }
1540
0
        if (txn_info->partial_update_info->partial_update_mode == UniqueKeyUpdateModePB::UPSERT) {
1541
0
            txn_info->partial_update_info->partial_update_input_columns.clear();
1542
0
            txn_info->partial_update_info->missing_cids.clear();
1543
0
            txn_info->partial_update_info->default_values.clear();
1544
0
            auto& update_cids = txn_info->partial_update_info->update_cids;
1545
0
            update_cids.resize(rowset->tablet_schema()->num_columns());
1546
0
            std::iota(update_cids.begin(), update_cids.end(), 0);
1547
0
        }
1548
1549
0
        DCHECK(txn_info->partial_update_info != nullptr);
1550
1551
0
        transient_rs_writer = DORIS_TRY(self->create_transient_rowset_writer(
1552
0
                *rowset, txn_info->partial_update_info, txn_expiration));
1553
0
        DBUG_EXECUTE_IF("BaseTablet::update_delete_bitmap.after.create_transient_rs_writer",
1554
0
                        DBUG_BLOCK);
1555
        // Partial update or upsert rewrite might generate new segments when there is conflicts while publish, and mark
1556
        // the same key in original segments as delete.
1557
        // When the new segment flush fails or the rowset build fails, the deletion marker for the
1558
        // duplicate key of the original segment should not remain in `txn_info->delete_bitmap`,
1559
        // so we need to make a copy of `txn_info->delete_bitmap` and make changes on it.
1560
0
        delete_bitmap = std::make_shared<DeleteBitmap>(*(txn_info->delete_bitmap));
1561
0
    }
1562
1563
3
    OlapStopWatch watch;
1564
3
    std::vector<segment_v2::SegmentSharedPtr> segments;
1565
3
    RETURN_IF_ERROR(std::dynamic_pointer_cast<BetaRowset>(rowset)->load_segments(&segments));
1566
3
    auto t1 = watch.get_elapse_time_us();
1567
1568
3
    int64_t next_visible_version = txn_info->is_txn_load ? txn_info->next_visible_version
1569
3
                                                         : txn_info->rowset->start_version();
1570
3
    {
1571
3
        std::shared_lock meta_rlock(self->_meta_lock);
1572
        // tablet is under alter process. The delete bitmap will be calculated after conversion.
1573
3
        if (self->tablet_state() == TABLET_NOTREADY) {
1574
0
            LOG(INFO) << "tablet is under alter process, update delete bitmap later, tablet_id="
1575
0
                      << self->tablet_id();
1576
0
            return Status::OK();
1577
0
        }
1578
3
        RETURN_IF_ERROR(self->get_all_rs_id_unlocked(next_visible_version - 1, &cur_rowset_ids));
1579
3
    }
1580
3
    auto t2 = watch.get_elapse_time_us();
1581
1582
3
    _rowset_ids_difference(cur_rowset_ids, txn_info->rowset_ids, &rowset_ids_to_add,
1583
3
                           &rowset_ids_to_del);
1584
3
    for (const auto& to_del : rowset_ids_to_del) {
1585
0
        delete_bitmap->remove({to_del, 0, 0}, {to_del, UINT32_MAX, INT64_MAX});
1586
0
    }
1587
1588
3
    std::vector<RowsetSharedPtr> specified_rowsets;
1589
3
    {
1590
3
        std::shared_lock meta_rlock(self->_meta_lock);
1591
3
        specified_rowsets = self->get_rowset_by_ids(&rowset_ids_to_add);
1592
3
    }
1593
3
    if (txn_info->is_txn_load) {
1594
0
        for (auto invisible_rowset : txn_info->invisible_rowsets) {
1595
0
            specified_rowsets.emplace_back(invisible_rowset);
1596
0
        }
1597
0
        std::sort(specified_rowsets.begin(), specified_rowsets.end(),
1598
0
                  [](RowsetSharedPtr& lhs, RowsetSharedPtr& rhs) {
1599
0
                      return lhs->end_version() > rhs->end_version();
1600
0
                  });
1601
0
    }
1602
3
    auto t3 = watch.get_elapse_time_us();
1603
1604
    // If a rowset is produced by compaction before the commit phase of the partial update load
1605
    // and is not included in txn_info->rowset_ids, we can skip the alignment process of that rowset
1606
    // because data remains the same before and after compaction. But we still need to calculate the
1607
    // the delete bitmap for that rowset.
1608
3
    std::vector<RowsetSharedPtr> rowsets_skip_alignment;
1609
3
    if (is_partial_update) {
1610
0
        int64_t max_version_in_flush_phase =
1611
0
                txn_info->partial_update_info->max_version_in_flush_phase;
1612
0
        DCHECK(max_version_in_flush_phase != -1);
1613
0
        std::vector<RowsetSharedPtr> remained_rowsets;
1614
0
        for (const auto& specified_rowset : specified_rowsets) {
1615
0
            if (specified_rowset->end_version() <= max_version_in_flush_phase &&
1616
0
                specified_rowset->produced_by_compaction()) {
1617
0
                rowsets_skip_alignment.emplace_back(specified_rowset);
1618
0
            } else {
1619
0
                remained_rowsets.emplace_back(specified_rowset);
1620
0
            }
1621
0
        }
1622
0
        if (!rowsets_skip_alignment.empty()) {
1623
0
            specified_rowsets = std::move(remained_rowsets);
1624
0
        }
1625
0
    }
1626
1627
3
    DBUG_EXECUTE_IF("BaseTablet::update_delete_bitmap.enable_spin_wait", {
1628
3
        auto token = dp->param<std::string>("token", "invalid_token");
1629
3
        while (DebugPoints::instance()->is_enable("BaseTablet::update_delete_bitmap.block")) {
1630
3
            auto block_dp = DebugPoints::instance()->get_debug_point(
1631
3
                    "BaseTablet::update_delete_bitmap.block");
1632
3
            if (block_dp) {
1633
3
                auto wait_token = block_dp->param<std::string>("wait_token", "");
1634
3
                LOG(INFO) << "BaseTablet::update_delete_bitmap.enable_spin_wait, wait_token: "
1635
3
                          << wait_token << ", token: " << token;
1636
3
                if (wait_token != token) {
1637
3
                    break;
1638
3
                }
1639
3
            }
1640
3
            std::this_thread::sleep_for(std::chrono::milliseconds(50));
1641
3
        }
1642
3
    });
1643
1644
3
    if (!rowsets_skip_alignment.empty()) {
1645
0
        auto token = self->calc_delete_bitmap_executor()->create_token();
1646
        // set rowset_writer to nullptr to skip the alignment process
1647
0
        RETURN_IF_ERROR(calc_delete_bitmap(self, rowset, segments, rowsets_skip_alignment,
1648
0
                                           delete_bitmap, cur_version - 1, token.get(), nullptr,
1649
0
                                           tablet_delete_bitmap));
1650
0
        RETURN_IF_ERROR(token->wait());
1651
0
    }
1652
1653
    // Publish-phase partial update or upsert rewrite may generate transient segments. If row binlog is enabled for
1654
    // this txn (row binlog rowset attached), we need to build row binlog segments together with
1655
    // transient data segments:
1656
    // 1) create a transient row binlog writer that appends to the same version row binlog rowset
1657
    // 2) pre-allocate per-row LSNs for each transient segment and register them in binlog options
1658
    // 3) wrap both writers into a GroupRowsetWriter so calc_delete_bitmap writes to both.
1659
3
    if (build_row_binlog) {
1660
0
        DCHECK(transient_rs_writer != nullptr);
1661
1662
        // Create transient row binlog writer for publish-phase segment appending.
1663
0
        auto transient_row_binlog_writer = DORIS_TRY(self->create_transient_rowset_writer(
1664
0
                *row_binlog_rowset, txn_info->partial_update_info, txn_expiration));
1665
1666
        // Prepare source MOW context for historical row retrieval in binlog writer.
1667
0
        auto& data_ctx = const_cast<RowsetWriterContext&>(transient_rs_writer->context());
1668
0
        data_ctx.mow_context = std::make_shared<MowContext>(
1669
0
                cur_version - 1, txn_id, std::make_shared<RowsetIdUnorderedSet>(),
1670
0
                specified_rowsets, nullptr);
1671
1672
0
        auto& binlog_ctx = const_cast<RowsetWriterContext&>(transient_row_binlog_writer->context());
1673
0
        auto& cfg = binlog_ctx.write_binlog_opt().write_binlog_config();
1674
0
        cfg.source.tablet_schema = data_ctx.tablet_schema;
1675
0
        cfg.source.partial_update_info = data_ctx.partial_update_info;
1676
0
        cfg.source.mow_context = data_ctx.mow_context;
1677
0
        cfg.source.is_transient_rowset_writer = data_ctx.is_transient_rowset_writer;
1678
0
        cfg.source.source_write_type = data_ctx.write_type;
1679
0
        cfg.source.row_binlog_rowset = row_binlog_rowset;
1680
1681
        // Wrap two transient writers into a group writer for dual flush/build.
1682
0
        RowsetWriterSharedPtr data_writer_sp(std::move(transient_rs_writer));
1683
0
        RowsetWriterSharedPtr row_binlog_writer_sp(std::move(transient_row_binlog_writer));
1684
0
        std::unique_ptr<GroupRowsetWriter> group_writer;
1685
0
        RETURN_IF_ERROR(RowsetFactory::create_empty_group_rowset_writer(&group_writer));
1686
0
        group_writer->set_data_writer(data_writer_sp);
1687
0
        group_writer->set_row_binlog_writer(row_binlog_writer_sp);
1688
0
        transient_rs_writer = std::move(group_writer);
1689
0
    }
1690
1691
    // When there is only one segment, it will be calculated in the current thread.
1692
    // Otherwise, it will be submitted to the thread pool for calculation.
1693
3
    if (segments.size() <= 1) {
1694
3
        RETURN_IF_ERROR(calc_delete_bitmap(self, rowset, segments, specified_rowsets, delete_bitmap,
1695
3
                                           cur_version - 1, nullptr, transient_rs_writer.get(),
1696
3
                                           tablet_delete_bitmap));
1697
1698
3
    } else {
1699
0
        auto token = self->calc_delete_bitmap_executor()->create_token();
1700
0
        RETURN_IF_ERROR(calc_delete_bitmap(self, rowset, segments, specified_rowsets, delete_bitmap,
1701
0
                                           cur_version - 1, token.get(), transient_rs_writer.get(),
1702
0
                                           tablet_delete_bitmap));
1703
0
        RETURN_IF_ERROR(token->wait());
1704
0
    }
1705
1706
3
    std::stringstream ss;
1707
3
    ss << "cost(us): (load segments: " << t1 << ", get all rsid: " << t2 - t1
1708
3
       << ", get rowsets: " << t3 - t2
1709
3
       << ", calc delete bitmap: " << watch.get_elapse_time_us() - t3 << ")";
1710
1711
3
    if (config::enable_merge_on_write_correctness_check && rowset->num_rows() != 0) {
1712
        // only do correctness check if the rowset has at least one row written
1713
        // check if all the rowset has ROWSET_SENTINEL_MARK
1714
3
        auto st = self->check_delete_bitmap_correctness(delete_bitmap, cur_version - 1, -1,
1715
3
                                                        cur_rowset_ids, &specified_rowsets);
1716
3
        if (!st.ok()) {
1717
0
            LOG(WARNING) << fmt::format("delete bitmap correctness check failed in publish phase!");
1718
0
        }
1719
3
    }
1720
1721
3
    if (transient_rs_writer) {
1722
0
        auto t4 = watch.get_elapse_time_us();
1723
0
        DBUG_EXECUTE_IF("Tablet.update_delete_bitmap.partial_update_write_rowset_fail", {
1724
0
            if (rand() % 100 < (100 * dp->param("percent", 0.5))) {
1725
0
                LOG_WARNING("Tablet.update_delete_bitmap.partial_update_write_rowset random failed")
1726
0
                        .tag("txn_id", txn_id);
1727
0
                return Status::InternalError(
1728
0
                        "debug update_delete_bitmap partial update write rowset random failed");
1729
0
            }
1730
0
        });
1731
        // build rowset writer and merge transient rowset
1732
0
        RETURN_IF_ERROR(transient_rs_writer->flush());
1733
0
        RowsetSharedPtr transient_rowset;
1734
0
        RowsetSharedPtr transient_row_binlog;
1735
0
        if (build_row_binlog) {
1736
0
            auto* group_rowset_writer = typeid_cast<GroupRowsetWriter*>(transient_rs_writer.get());
1737
0
            DCHECK(group_rowset_writer != nullptr);
1738
0
            std::vector<RowsetSharedPtr> waited_build_rowsets;
1739
0
            RETURN_IF_ERROR(group_rowset_writer->build_rowsets(waited_build_rowsets));
1740
0
            transient_rowset = waited_build_rowsets.at(0);
1741
0
            transient_row_binlog = waited_build_rowsets.at(1);
1742
0
        } else {
1743
0
            RETURN_IF_ERROR(transient_rs_writer->build(transient_rowset));
1744
0
        }
1745
0
        auto old_segments = rowset->num_segments();
1746
0
        rowset->merge_rowset_meta(*transient_rowset->rowset_meta());
1747
0
        auto new_segments = rowset->num_segments();
1748
0
        ss << ", " << txn_info->partial_update_info->partial_update_mode_str()
1749
0
           << " flush rowset (old segment num: " << old_segments
1750
0
           << ", new segment num: " << new_segments << ")"
1751
0
           << ", cost:" << watch.get_elapse_time_us() - t4 << "(us)";
1752
1753
0
        if (build_row_binlog) {
1754
0
            DCHECK(row_binlog_rowset != nullptr);
1755
0
            old_segments = row_binlog_rowset->num_segments();
1756
0
            row_binlog_rowset->merge_rowset_meta(*transient_row_binlog->rowset_meta());
1757
0
            new_segments = row_binlog_rowset->num_segments();
1758
0
            ss << ", " << txn_info->partial_update_info->partial_update_mode_str()
1759
0
               << " flush binlog<row> (old segment num: " << old_segments
1760
0
               << ", new segment num: " << new_segments << ")";
1761
1762
0
            SegmentLoader::instance()->erase_segments(row_binlog_rowset->rowset_id(),
1763
0
                                                      row_binlog_rowset->num_segments());
1764
0
        }
1765
1766
        // update the shared_ptr to new bitmap, which is consistent with current rowset.
1767
0
        txn_info->delete_bitmap = delete_bitmap;
1768
        // erase segment cache cause we will add a segment to rowset
1769
0
        SegmentLoader::instance()->erase_segments(rowset->rowset_id(), rowset->num_segments());
1770
0
    }
1771
1772
3
    size_t total_rows = std::accumulate(
1773
3
            segments.begin(), segments.end(), 0,
1774
3
            [](size_t sum, const segment_v2::SegmentSharedPtr& s) { return sum += s->num_rows(); });
1775
3
    auto t5 = watch.get_elapse_time_us();
1776
3
    int64_t lock_id = txn_info->is_txn_load ? txn_info->lock_id : -1;
1777
3
    RETURN_IF_ERROR(self->save_delete_bitmap(txn_info, txn_id, delete_bitmap,
1778
3
                                             transient_rs_writer.get(), cur_rowset_ids, lock_id,
1779
3
                                             next_visible_version));
1780
1781
    // defensive check, check that the delete bitmap cache we wrote is correct
1782
3
    RETURN_IF_ERROR(self->check_delete_bitmap_cache(txn_id, delete_bitmap.get()));
1783
1784
3
    LOG(INFO) << "[Publish] construct delete bitmap tablet: " << self->tablet_id()
1785
3
              << ", rowset_ids to add: "
1786
3
              << (specified_rowsets.size() + rowsets_skip_alignment.size())
1787
3
              << ", rowset_ids to del: " << rowset_ids_to_del.size()
1788
3
              << ", cur version: " << cur_version << ", transaction_id: " << txn_id << ","
1789
3
              << ss.str() << " , total rows: " << total_rows
1790
3
              << ", update delete_bitmap cost: " << watch.get_elapse_time_us() - t5 << "(us)";
1791
3
    return Status::OK();
1792
3
}
1793
1794
void BaseTablet::calc_compaction_output_rowset_delete_bitmap(
1795
        const std::vector<RowsetSharedPtr>& input_rowsets, const RowIdConversion& rowid_conversion,
1796
        uint64_t start_version, uint64_t end_version, std::set<RowLocation>* missed_rows,
1797
        std::map<RowsetSharedPtr, std::list<std::pair<RowLocation, RowLocation>>>* location_map,
1798
0
        const DeleteBitmap& input_delete_bitmap, DeleteBitmap* output_rowset_delete_bitmap) {
1799
0
    RowLocation src;
1800
0
    RowLocation dst;
1801
0
    for (auto& rowset : input_rowsets) {
1802
0
        src.rowset_id = rowset->rowset_id();
1803
0
        for (uint32_t seg_id = 0; seg_id < rowset->num_segments(); ++seg_id) {
1804
0
            src.segment_id = seg_id;
1805
0
            DeleteBitmap subset_map(tablet_id());
1806
0
            input_delete_bitmap.subset({rowset->rowset_id(), seg_id, start_version},
1807
0
                                       {rowset->rowset_id(), seg_id, end_version}, &subset_map);
1808
            // traverse all versions and convert rowid
1809
0
            for (auto iter = subset_map.delete_bitmap.begin();
1810
0
                 iter != subset_map.delete_bitmap.end(); ++iter) {
1811
0
                auto cur_version = std::get<2>(iter->first);
1812
0
                for (auto index = iter->second.begin(); index != iter->second.end(); ++index) {
1813
0
                    src.row_id = *index;
1814
0
                    if (rowid_conversion.get(src, &dst) != 0) {
1815
0
                        VLOG_CRITICAL << "Can't find rowid, may be deleted by the delete_handler, "
1816
0
                                      << " src loaction: |" << src.rowset_id << "|"
1817
0
                                      << src.segment_id << "|" << src.row_id
1818
0
                                      << " version: " << cur_version;
1819
0
                        if (missed_rows) {
1820
0
                            missed_rows->insert(src);
1821
0
                        }
1822
0
                        continue;
1823
0
                    }
1824
0
                    VLOG_DEBUG << "calc_compaction_output_rowset_delete_bitmap dst location: |"
1825
0
                               << dst.rowset_id << "|" << dst.segment_id << "|" << dst.row_id
1826
0
                               << " src location: |" << src.rowset_id << "|" << src.segment_id
1827
0
                               << "|" << src.row_id << " start version: " << start_version
1828
0
                               << "end version" << end_version;
1829
0
                    if (location_map) {
1830
0
                        (*location_map)[rowset].emplace_back(src, dst);
1831
0
                    }
1832
0
                    output_rowset_delete_bitmap->add({dst.rowset_id, dst.segment_id, cur_version},
1833
0
                                                     dst.row_id);
1834
0
                }
1835
0
            }
1836
0
        }
1837
0
    }
1838
0
}
1839
1840
Status BaseTablet::check_rowid_conversion(
1841
        RowsetSharedPtr dst_rowset,
1842
        const std::map<RowsetSharedPtr, std::list<std::pair<RowLocation, RowLocation>>>&
1843
0
                location_map) {
1844
0
    if (location_map.empty()) {
1845
0
        VLOG_DEBUG << "check_rowid_conversion, location_map is empty";
1846
0
        return Status::OK();
1847
0
    }
1848
0
    std::vector<segment_v2::SegmentSharedPtr> dst_segments;
1849
1850
0
    RETURN_IF_ERROR(
1851
0
            std::dynamic_pointer_cast<BetaRowset>(dst_rowset)->load_segments(&dst_segments));
1852
0
    std::unordered_map<RowsetId, std::vector<segment_v2::SegmentSharedPtr>> input_rowsets_segment;
1853
1854
0
    VLOG_DEBUG << "check_rowid_conversion, dst_segments size: " << dst_segments.size();
1855
0
    for (auto [src_rowset, locations] : location_map) {
1856
0
        std::vector<segment_v2::SegmentSharedPtr>& segments =
1857
0
                input_rowsets_segment[src_rowset->rowset_id()];
1858
0
        if (segments.empty()) {
1859
0
            RETURN_IF_ERROR(
1860
0
                    std::dynamic_pointer_cast<BetaRowset>(src_rowset)->load_segments(&segments));
1861
0
        }
1862
0
        for (auto& [src, dst] : locations) {
1863
0
            std::string src_key;
1864
0
            std::string dst_key;
1865
0
            Status s = segments[src.segment_id]->read_key_by_rowid(src.row_id, &src_key);
1866
0
            if (UNLIKELY(s.is<NOT_IMPLEMENTED_ERROR>())) {
1867
0
                LOG(INFO) << "primary key index of old version does not "
1868
0
                             "support reading key by rowid";
1869
0
                break;
1870
0
            }
1871
0
            if (UNLIKELY(!s)) {
1872
0
                LOG(WARNING) << "failed to get src key: |" << src.rowset_id << "|" << src.segment_id
1873
0
                             << "|" << src.row_id << " status: " << s;
1874
0
                DCHECK(false);
1875
0
                return s;
1876
0
            }
1877
1878
0
            s = dst_segments[dst.segment_id]->read_key_by_rowid(dst.row_id, &dst_key);
1879
0
            if (UNLIKELY(!s)) {
1880
0
                LOG(WARNING) << "failed to get dst key: |" << dst.rowset_id << "|" << dst.segment_id
1881
0
                             << "|" << dst.row_id << " status: " << s;
1882
0
                DCHECK(false);
1883
0
                return s;
1884
0
            }
1885
1886
0
            VLOG_DEBUG << "check_rowid_conversion, src: |" << src.rowset_id << "|" << src.segment_id
1887
0
                       << "|" << src.row_id << "|" << src_key << " dst: |" << dst.rowset_id << "|"
1888
0
                       << dst.segment_id << "|" << dst.row_id << "|" << dst_key;
1889
0
            if (UNLIKELY(src_key.compare(dst_key) != 0)) {
1890
0
                LOG(WARNING) << "failed to check key, src key: |" << src.rowset_id << "|"
1891
0
                             << src.segment_id << "|" << src.row_id << "|" << src_key
1892
0
                             << " dst key: |" << dst.rowset_id << "|" << dst.segment_id << "|"
1893
0
                             << dst.row_id << "|" << dst_key;
1894
0
                DCHECK(false);
1895
0
                return Status::InternalError("failed to check rowid conversion");
1896
0
            }
1897
0
        }
1898
0
    }
1899
0
    return Status::OK();
1900
0
}
1901
1902
// The caller should hold _rowset_update_lock and _meta_lock lock.
1903
Status BaseTablet::update_delete_bitmap_without_lock(
1904
        const BaseTabletSPtr& self, const RowsetSharedPtr& rowset,
1905
0
        const std::vector<RowsetSharedPtr>* specified_base_rowsets) {
1906
0
    DBUG_EXECUTE_IF("BaseTablet.update_delete_bitmap_without_lock.random_failed", {
1907
0
        auto rnd = rand() % 100;
1908
0
        auto percent = dp->param("percent", 0.1);
1909
0
        if (rnd < (100 * percent)) {
1910
0
            LOG(WARNING) << "BaseTablet.update_delete_bitmap_without_lock.random_failed";
1911
0
            return Status::InternalError(
1912
0
                    "debug tablet update delete bitmap without lock random failed");
1913
0
        } else {
1914
0
            LOG(INFO) << "BaseTablet.update_delete_bitmap_without_lock.random_failed not "
1915
0
                         "triggered"
1916
0
                      << ", rnd:" << rnd << ", percent: " << percent;
1917
0
        }
1918
0
    });
1919
0
    int64_t cur_version = rowset->start_version();
1920
0
    std::vector<segment_v2::SegmentSharedPtr> segments;
1921
0
    RETURN_IF_ERROR(std::dynamic_pointer_cast<BetaRowset>(rowset)->load_segments(&segments));
1922
1923
    // If this rowset does not have a segment, there is no need for an update.
1924
0
    if (segments.empty()) {
1925
0
        LOG(INFO) << "[Schema Change or Clone] skip to construct delete bitmap tablet: "
1926
0
                  << self->tablet_id() << " cur max_version: " << cur_version;
1927
0
        return Status::OK();
1928
0
    }
1929
1930
    // calculate delete bitmap between segments if necessary.
1931
0
    DeleteBitmapPtr delete_bitmap = std::make_shared<DeleteBitmap>(self->tablet_id());
1932
0
    RETURN_IF_ERROR(self->calc_delete_bitmap_between_segments(
1933
0
            rowset->tablet_schema(), rowset->rowset_id(), segments, delete_bitmap));
1934
1935
    // get all base rowsets to calculate on
1936
0
    std::vector<RowsetSharedPtr> specified_rowsets;
1937
0
    RowsetIdUnorderedSet cur_rowset_ids;
1938
0
    if (specified_base_rowsets == nullptr) {
1939
0
        RETURN_IF_ERROR(self->get_all_rs_id_unlocked(cur_version - 1, &cur_rowset_ids));
1940
0
        specified_rowsets = self->get_rowset_by_ids(&cur_rowset_ids);
1941
0
    } else {
1942
0
        specified_rowsets = *specified_base_rowsets;
1943
0
    }
1944
1945
0
    OlapStopWatch watch;
1946
0
    auto token = self->calc_delete_bitmap_executor()->create_token();
1947
0
    RETURN_IF_ERROR(calc_delete_bitmap(self, rowset, segments, specified_rowsets, delete_bitmap,
1948
0
                                       cur_version - 1, token.get()));
1949
0
    RETURN_IF_ERROR(token->wait());
1950
0
    size_t total_rows = std::accumulate(
1951
0
            segments.begin(), segments.end(), 0,
1952
0
            [](size_t sum, const segment_v2::SegmentSharedPtr& s) { return sum += s->num_rows(); });
1953
0
    LOG(INFO) << "[Schema Change or Clone] construct delete bitmap tablet: " << self->tablet_id()
1954
0
              << ", rowset_ids: " << cur_rowset_ids.size() << ", cur max_version: " << cur_version
1955
0
              << ", transaction_id: " << -1 << ", cost: " << watch.get_elapse_time_us()
1956
0
              << "(us), total rows: " << total_rows;
1957
0
    if (config::enable_merge_on_write_correctness_check) {
1958
        // check if all the rowset has ROWSET_SENTINEL_MARK
1959
0
        auto st = self->check_delete_bitmap_correctness(delete_bitmap, cur_version - 1, -1,
1960
0
                                                        cur_rowset_ids, &specified_rowsets);
1961
0
        if (!st.ok()) {
1962
0
            LOG(WARNING) << fmt::format("delete bitmap correctness check failed in publish phase!");
1963
0
        }
1964
0
        delete_bitmap->remove_sentinel_marks();
1965
0
    }
1966
0
    for (auto& iter : delete_bitmap->delete_bitmap) {
1967
0
        self->_tablet_meta->delete_bitmap().merge(
1968
0
                {std::get<0>(iter.first), std::get<1>(iter.first), cur_version}, iter.second);
1969
0
    }
1970
1971
0
    return Status::OK();
1972
0
}
1973
1974
void BaseTablet::agg_delete_bitmap_for_stale_rowsets(
1975
0
        Version version, DeleteBitmapKeyRanges& remove_delete_bitmap_key_ranges) {
1976
0
    if (!config::enable_agg_and_remove_pre_rowsets_delete_bitmap) {
1977
0
        return;
1978
0
    }
1979
0
    if (!(keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write())) {
1980
0
        return;
1981
0
    }
1982
0
    int64_t start_version = version.first;
1983
0
    int64_t end_version = version.second;
1984
0
    if (start_version == end_version) {
1985
0
        return;
1986
0
    }
1987
0
    DCHECK(start_version < end_version)
1988
0
            << ". start_version: " << start_version << ", end_version: " << end_version;
1989
    // get pre rowsets
1990
0
    std::vector<RowsetSharedPtr> pre_rowsets {};
1991
0
    {
1992
0
        std::shared_lock rdlock(_meta_lock);
1993
0
        for (const auto& it2 : _rs_version_map) {
1994
0
            if (it2.first.second < start_version) {
1995
0
                pre_rowsets.emplace_back(it2.second);
1996
0
            }
1997
0
        }
1998
0
    }
1999
0
    std::sort(pre_rowsets.begin(), pre_rowsets.end(), Rowset::comparator);
2000
    // do agg for pre rowsets
2001
0
    DeleteBitmapPtr new_delete_bitmap = std::make_shared<DeleteBitmap>(tablet_id());
2002
0
    for (auto& rowset : pre_rowsets) {
2003
0
        for (uint32_t seg_id = 0; seg_id < rowset->num_segments(); ++seg_id) {
2004
0
            auto d = tablet_meta()->delete_bitmap().get_agg_without_cache(
2005
0
                    {rowset->rowset_id(), seg_id, end_version}, start_version);
2006
0
            if (d->isEmpty()) {
2007
0
                continue;
2008
0
            }
2009
0
            VLOG_DEBUG << "agg delete bitmap for tablet_id=" << tablet_id()
2010
0
                       << ", rowset_id=" << rowset->rowset_id() << ", seg_id=" << seg_id
2011
0
                       << ", rowset_version=" << rowset->version().to_string()
2012
0
                       << ". compaction start_version=" << start_version
2013
0
                       << ", end_version=" << end_version << ", delete_bitmap=" << d->cardinality();
2014
0
            DeleteBitmap::BitmapKey start_key {rowset->rowset_id(), seg_id, start_version};
2015
0
            DeleteBitmap::BitmapKey end_key {rowset->rowset_id(), seg_id, end_version};
2016
0
            new_delete_bitmap->set(end_key, *d);
2017
0
            remove_delete_bitmap_key_ranges.emplace_back(start_key, end_key);
2018
0
        }
2019
0
    }
2020
0
    DBUG_EXECUTE_IF("BaseTablet.agg_delete_bitmap_for_stale_rowsets.merge_delete_bitmap.block",
2021
0
                    DBUG_BLOCK);
2022
0
    tablet_meta()->delete_bitmap().merge(*new_delete_bitmap);
2023
0
}
2024
2025
void BaseTablet::check_agg_delete_bitmap_for_stale_rowsets(int64_t& useless_rowset_count,
2026
0
                                                           int64_t& useless_rowset_version_count) {
2027
0
    std::map<RowsetId, Version> rowset_ids;
2028
0
    std::set<int64_t> end_versions;
2029
0
    traverse_rowsets(
2030
0
            [&rowset_ids, &end_versions](const RowsetSharedPtr& rs) {
2031
0
                rowset_ids[rs->rowset_id()] = rs->version();
2032
0
                end_versions.emplace(rs->end_version());
2033
0
            },
2034
0
            true);
2035
2036
0
    std::set<RowsetId> useless_rowsets;
2037
0
    std::map<RowsetId, std::vector<int64_t>> useless_rowset_versions;
2038
0
    {
2039
0
        _tablet_meta->delete_bitmap().traverse_rowset_and_version(
2040
                // 0: rowset and rowset with version exists
2041
                // -1: rowset does not exist
2042
                // -2: find next <rowset, version>
2043
                //     rowset exist, rowset with version does not exist
2044
                //     sequence table
2045
0
                [&](const RowsetId& rowset_id, int64_t version) {
2046
0
                    auto rowset_it = rowset_ids.find(rowset_id);
2047
0
                    if (rowset_it == rowset_ids.end()) {
2048
0
                        useless_rowsets.emplace(rowset_id);
2049
0
                        return -1;
2050
0
                    }
2051
0
                    if (end_versions.find(version) == end_versions.end()) {
2052
0
                        if (tablet_schema()->has_sequence_col()) {
2053
0
                            auto rowset_version = rowset_it->second;
2054
0
                            if (version >= rowset_version.first &&
2055
0
                                version <= rowset_version.second) {
2056
0
                                return -2;
2057
0
                            }
2058
0
                        }
2059
0
                        if (useless_rowset_versions.find(rowset_id) ==
2060
0
                            useless_rowset_versions.end()) {
2061
0
                            useless_rowset_versions[rowset_id] = {};
2062
0
                        }
2063
0
                        useless_rowset_versions[rowset_id].emplace_back(version);
2064
0
                        return -2;
2065
0
                    }
2066
0
                    return 0;
2067
0
                });
2068
0
    }
2069
0
    useless_rowset_count = useless_rowsets.size();
2070
0
    useless_rowset_version_count = useless_rowset_versions.size();
2071
0
    if (!useless_rowsets.empty() || !useless_rowset_versions.empty()) {
2072
0
        std::stringstream ss;
2073
0
        if (!useless_rowsets.empty()) {
2074
0
            ss << "useless rowsets: {";
2075
0
            for (auto it = useless_rowsets.begin(); it != useless_rowsets.end(); ++it) {
2076
0
                if (it != useless_rowsets.begin()) {
2077
0
                    ss << ", ";
2078
0
                }
2079
0
                ss << it->to_string();
2080
0
            }
2081
0
            ss << "}. ";
2082
0
        }
2083
0
        if (!useless_rowset_versions.empty()) {
2084
0
            ss << "useless rowset versions: {";
2085
0
            for (auto iter = useless_rowset_versions.begin(); iter != useless_rowset_versions.end();
2086
0
                 ++iter) {
2087
0
                if (iter != useless_rowset_versions.begin()) {
2088
0
                    ss << ", ";
2089
0
                }
2090
0
                ss << iter->first.to_string() << ": [";
2091
                // some versions are continuous, such as [8, 9, 10, 11, 13, 17, 18]
2092
                // print as [8-11, 13, 17-18]
2093
0
                int64_t last_start_version = -1;
2094
0
                int64_t last_end_version = -1;
2095
0
                for (int64_t version : iter->second) {
2096
0
                    if (last_start_version == -1) {
2097
0
                        last_start_version = version;
2098
0
                        last_end_version = version;
2099
0
                        continue;
2100
0
                    }
2101
0
                    if (last_end_version + 1 == version) {
2102
0
                        last_end_version = version;
2103
0
                    } else {
2104
0
                        if (last_start_version == last_end_version) {
2105
0
                            ss << last_start_version << ", ";
2106
0
                        } else {
2107
0
                            ss << last_start_version << "-" << last_end_version << ", ";
2108
0
                        }
2109
0
                        last_start_version = version;
2110
0
                        last_end_version = version;
2111
0
                    }
2112
0
                }
2113
0
                if (last_start_version == last_end_version) {
2114
0
                    ss << last_start_version;
2115
0
                } else {
2116
0
                    ss << last_start_version << "-" << last_end_version;
2117
0
                }
2118
2119
0
                ss << "]";
2120
0
            }
2121
0
            ss << "}.";
2122
0
        }
2123
0
        LOG(WARNING) << "failed check_agg_delete_bitmap_for_stale_rowsets for tablet_id="
2124
0
                     << tablet_id() << ". " << ss.str();
2125
0
    } else {
2126
0
        LOG(INFO) << "succeed check_agg_delete_bitmap_for_stale_rowsets for tablet_id="
2127
0
                  << tablet_id();
2128
0
    }
2129
0
}
2130
2131
0
RowsetSharedPtr BaseTablet::get_rowset(const RowsetId& rowset_id) {
2132
0
    std::shared_lock rdlock(_meta_lock);
2133
0
    for (auto& version_rowset : _rs_version_map) {
2134
0
        if (version_rowset.second->rowset_id() == rowset_id) {
2135
0
            return version_rowset.second;
2136
0
        }
2137
0
    }
2138
0
    for (auto& stale_version_rowset : _stale_rs_version_map) {
2139
0
        if (stale_version_rowset.second->rowset_id() == rowset_id) {
2140
0
            return stale_version_rowset.second;
2141
0
        }
2142
0
    }
2143
0
    return nullptr;
2144
0
}
2145
2146
1
std::vector<RowsetSharedPtr> BaseTablet::get_snapshot_rowset(bool include_stale_rowset) const {
2147
1
    std::shared_lock rdlock(_meta_lock);
2148
1
    std::vector<RowsetSharedPtr> rowsets;
2149
1
    std::transform(_rs_version_map.cbegin(), _rs_version_map.cend(), std::back_inserter(rowsets),
2150
28
                   [](auto& kv) { return kv.second; });
2151
1
    if (include_stale_rowset) {
2152
0
        std::transform(_stale_rs_version_map.cbegin(), _stale_rs_version_map.cend(),
2153
0
                       std::back_inserter(rowsets), [](auto& kv) { return kv.second; });
2154
0
    }
2155
1
    return rowsets;
2156
1
}
2157
2158
void BaseTablet::calc_consecutive_empty_rowsets(
2159
        std::vector<RowsetSharedPtr>* empty_rowsets,
2160
4
        const std::vector<RowsetSharedPtr>& candidate_rowsets, int64_t limit) {
2161
4
    int len = cast_set<int>(candidate_rowsets.size());
2162
12
    for (int i = 0; i < len - 1; ++i) {
2163
9
        auto rowset = candidate_rowsets[i];
2164
9
        auto next_rowset = candidate_rowsets[i + 1];
2165
2166
        // identify two consecutive rowsets that are empty
2167
9
        if (rowset->num_segments() == 0 && next_rowset->num_segments() == 0 &&
2168
9
            !rowset->rowset_meta()->has_delete_predicate() &&
2169
9
            !next_rowset->rowset_meta()->has_delete_predicate() &&
2170
9
            rowset->end_version() == next_rowset->start_version() - 1) {
2171
1
            empty_rowsets->emplace_back(rowset);
2172
1
            empty_rowsets->emplace_back(next_rowset);
2173
1
            rowset = next_rowset;
2174
1
            int next_index = i + 2;
2175
2176
            // keep searching for consecutive empty rowsets
2177
6
            while (next_index < len && candidate_rowsets[next_index]->num_segments() == 0 &&
2178
6
                   !candidate_rowsets[next_index]->rowset_meta()->has_delete_predicate() &&
2179
6
                   rowset->end_version() == candidate_rowsets[next_index]->start_version() - 1) {
2180
5
                empty_rowsets->emplace_back(candidate_rowsets[next_index]);
2181
5
                rowset = candidate_rowsets[next_index++];
2182
5
            }
2183
            // if the number of consecutive empty rowset reach the limit,
2184
            // and there are still rowsets following them
2185
1
            if (empty_rowsets->size() >= limit && next_index < len) {
2186
1
                return;
2187
1
            } else {
2188
                // current rowset is not empty, start searching from that rowset in the next
2189
0
                i = next_index - 1;
2190
0
                empty_rowsets->clear();
2191
0
            }
2192
1
        }
2193
9
    }
2194
4
}
2195
2196
Status BaseTablet::calc_file_crc(uint32_t* crc_value, int64_t start_version, int64_t end_version,
2197
0
                                 uint32_t* rowset_count, int64_t* file_count) {
2198
0
    Version v(start_version, end_version);
2199
0
    std::vector<RowsetSharedPtr> rowsets;
2200
0
    traverse_rowsets([&rowsets, &v](const auto& rs) {
2201
        // get all rowsets
2202
0
        if (v.contains(rs->version())) {
2203
0
            rowsets.emplace_back(rs);
2204
0
        }
2205
0
    });
2206
0
    std::sort(rowsets.begin(), rowsets.end(), Rowset::comparator);
2207
0
    *rowset_count = cast_set<uint32_t>(rowsets.size());
2208
2209
0
    *crc_value = 0;
2210
0
    *file_count = 0;
2211
0
    for (const auto& rs : rowsets) {
2212
0
        uint32_t rs_crc_value = 0;
2213
0
        int64_t rs_file_count = 0;
2214
0
        auto rowset = std::static_pointer_cast<BetaRowset>(rs);
2215
0
        auto st = rowset->calc_file_crc(&rs_crc_value, &rs_file_count);
2216
0
        if (!st.ok()) {
2217
0
            return st;
2218
0
        }
2219
        // crc_value is calculated based on the crc_value of each rowset.
2220
0
        *crc_value = crc32c::Extend(*crc_value, reinterpret_cast<const uint8_t*>(&rs_crc_value),
2221
0
                                    sizeof(rs_crc_value));
2222
0
        *file_count += rs_file_count;
2223
0
    }
2224
0
    return Status::OK();
2225
0
}
2226
2227
0
Status BaseTablet::show_nested_index_file(std::string* json_meta) {
2228
0
    Version v(0, max_version_unlocked());
2229
0
    std::vector<RowsetSharedPtr> rowsets;
2230
0
    traverse_rowsets([&rowsets, &v](const auto& rs) {
2231
        // get all rowsets
2232
0
        if (v.contains(rs->version())) {
2233
0
            rowsets.emplace_back(rs);
2234
0
        }
2235
0
    });
2236
0
    std::sort(rowsets.begin(), rowsets.end(), Rowset::comparator);
2237
2238
0
    rapidjson::Document doc;
2239
0
    doc.SetObject();
2240
0
    rapidjson::Document::AllocatorType& allocator = doc.GetAllocator();
2241
0
    rapidjson::Value tabletIdValue(tablet_id());
2242
0
    doc.AddMember("tablet_id", tabletIdValue, allocator);
2243
2244
0
    rapidjson::Value rowsets_value(rapidjson::kArrayType);
2245
2246
0
    for (const auto& rs : rowsets) {
2247
0
        rapidjson::Value rowset_value(rapidjson::kObjectType);
2248
2249
0
        auto rowset = std::static_pointer_cast<BetaRowset>(rs);
2250
0
        RETURN_IF_ERROR(rowset->show_nested_index_file(&rowset_value, allocator));
2251
0
        rowsets_value.PushBack(rowset_value, allocator);
2252
0
    }
2253
0
    doc.AddMember("rowsets", rowsets_value, allocator);
2254
2255
0
    rapidjson::StringBuffer buffer;
2256
0
    rapidjson::PrettyWriter<rapidjson::StringBuffer> writer(buffer);
2257
0
    doc.Accept(writer);
2258
0
    *json_meta = std::string(buffer.GetString());
2259
2260
0
    return Status::OK();
2261
0
}
2262
2263
void BaseTablet::get_base_rowset_delete_bitmap_count(
2264
        uint64_t* max_base_rowset_delete_bitmap_score,
2265
0
        int64_t* max_base_rowset_delete_bitmap_score_tablet_id) {
2266
0
    std::vector<RowsetSharedPtr> rowsets_;
2267
0
    std::string base_rowset_id_str;
2268
0
    {
2269
0
        std::shared_lock rowset_ldlock(this->get_header_lock());
2270
0
        for (const auto& it : _rs_version_map) {
2271
0
            rowsets_.emplace_back(it.second);
2272
0
        }
2273
0
    }
2274
0
    std::sort(rowsets_.begin(), rowsets_.end(), Rowset::comparator);
2275
0
    if (!rowsets_.empty()) {
2276
0
        bool base_found = false;
2277
0
        for (auto& rowset : rowsets_) {
2278
0
            if (rowset->start_version() > 2) {
2279
0
                break;
2280
0
            }
2281
0
            base_found = true;
2282
0
            uint64_t base_rowset_delete_bitmap_count =
2283
0
                    this->tablet_meta()->delete_bitmap().get_count_with_range(
2284
0
                            {rowset->rowset_id(), 0, 0},
2285
0
                            {rowset->rowset_id(), UINT32_MAX, UINT64_MAX});
2286
0
            if (base_rowset_delete_bitmap_count > *max_base_rowset_delete_bitmap_score) {
2287
0
                *max_base_rowset_delete_bitmap_score = base_rowset_delete_bitmap_count;
2288
0
                *max_base_rowset_delete_bitmap_score_tablet_id = this->tablet_id();
2289
0
            }
2290
0
        }
2291
0
        if (!base_found) {
2292
0
            LOG(WARNING) << "can not found base rowset for tablet " << tablet_id();
2293
0
        }
2294
0
    }
2295
0
}
2296
2297
386
void TabletReadSource::fill_delete_predicates() {
2298
386
    DCHECK_EQ(delete_predicates.size(), 0);
2299
386
    auto delete_pred_view =
2300
1.33k
            rs_splits | std::views::transform([](auto&& split) {
2301
1.33k
                return split.rs_reader->rowset()->rowset_meta();
2302
1.33k
            }) |
2303
1.18k
            std::views::filter([](const auto& rs_meta) { return rs_meta->has_delete_predicate(); });
2304
386
    delete_predicates = {delete_pred_view.begin(), delete_pred_view.end()};
2305
386
}
2306
2307
57
int32_t BaseTablet::max_version_config() {
2308
57
    int32_t max_version = tablet_meta()->compaction_policy() == CUMULATIVE_TIME_SERIES_POLICY
2309
57
                                  ? std::max(config::time_series_max_tablet_version_num,
2310
0
                                             config::max_tablet_version_num)
2311
57
                                  : config::max_tablet_version_num;
2312
57
    return max_version;
2313
57
}
2314
2315
0
void BaseTablet::prefill_dbm_agg_cache(const RowsetSharedPtr& rowset, int64_t version) {
2316
0
    for (std::size_t i = 0; i < rowset->num_segments(); i++) {
2317
0
        tablet_meta()->delete_bitmap().get_agg({rowset->rowset_id(), i, version});
2318
0
    }
2319
0
}
2320
2321
30
void BaseTablet::prefill_dbm_agg_cache_after_compaction(const RowsetSharedPtr& output_rowset) {
2322
30
    if (keys_type() == KeysType::UNIQUE_KEYS && enable_unique_key_merge_on_write() &&
2323
30
        (config::enable_prefill_output_dbm_agg_cache_after_compaction ||
2324
0
         config::enable_prefill_all_dbm_agg_cache_after_compaction)) {
2325
0
        int64_t cur_max_version {-1};
2326
0
        {
2327
0
            std::shared_lock rlock(get_header_lock());
2328
0
            cur_max_version = max_version_unlocked();
2329
0
        }
2330
0
        if (config::enable_prefill_all_dbm_agg_cache_after_compaction) {
2331
0
            traverse_rowsets(
2332
0
                    [&](const RowsetSharedPtr& rs) { prefill_dbm_agg_cache(rs, cur_max_version); },
2333
0
                    false);
2334
0
        } else if (config::enable_prefill_output_dbm_agg_cache_after_compaction) {
2335
0
            prefill_dbm_agg_cache(output_rowset, cur_max_version);
2336
0
        }
2337
0
    }
2338
30
}
2339
2340
bool BaseTablet::_key_is_not_in_segment(Slice key, const KeyBoundsPB& segment_key_bounds,
2341
2
                                        bool is_segments_key_bounds_truncated) {
2342
2
    Slice maybe_truncated_min_key {segment_key_bounds.min_key()};
2343
2
    Slice maybe_truncated_max_key {segment_key_bounds.max_key()};
2344
2
    bool res1 = Slice::lhs_is_strictly_less_than_rhs(key, false, maybe_truncated_min_key,
2345
2
                                                     is_segments_key_bounds_truncated);
2346
2
    bool res2 = Slice::lhs_is_strictly_less_than_rhs(maybe_truncated_max_key,
2347
2
                                                     is_segments_key_bounds_truncated, key, false);
2348
2
    return res1 || res2;
2349
2
}
2350
2351
} // namespace doris