Coverage Report

Created: 2025-09-10 18:13

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
/root/doris/be/src/olap/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 "olap/base_tablet.h"
19
20
#include <bthread/mutex.h>
21
#include <fmt/format.h>
22
#include <rapidjson/prettywriter.h>
23
24
#include <algorithm>
25
#include <cstdint>
26
#include <iterator>
27
#include <random>
28
#include <shared_mutex>
29
30
#include "cloud/cloud_tablet.h"
31
#include "cloud/config.h"
32
#include "common/cast_set.h"
33
#include "common/logging.h"
34
#include "common/status.h"
35
#include "olap/calc_delete_bitmap_executor.h"
36
#include "olap/cumulative_compaction_time_series_policy.h"
37
#include "olap/delete_bitmap_calculator.h"
38
#include "olap/iterators.h"
39
#include "olap/memtable.h"
40
#include "olap/partial_update_info.h"
41
#include "olap/primary_key_index.h"
42
#include "olap/rowid_conversion.h"
43
#include "olap/rowset/beta_rowset.h"
44
#include "olap/rowset/rowset.h"
45
#include "olap/rowset/rowset_fwd.h"
46
#include "olap/rowset/rowset_reader.h"
47
#include "olap/tablet_fwd.h"
48
#include "olap/txn_manager.h"
49
#include "service/point_query_executor.h"
50
#include "util/bvar_helper.h"
51
#include "util/crc32c.h"
52
#include "util/debug_points.h"
53
#include "util/doris_metrics.h"
54
#include "util/key_util.h"
55
#include "vec/common/assert_cast.h"
56
#include "vec/common/schema_util.h"
57
#include "vec/data_types/data_type_factory.hpp"
58
#include "vec/jsonb/serialize.h"
59
60
namespace doris {
61
#include "common/compile_check_begin.h"
62
63
using namespace ErrorCode;
64
65
namespace {
66
67
bvar::LatencyRecorder g_tablet_commit_phase_update_delete_bitmap_latency(
68
        "doris_pk", "commit_phase_update_delete_bitmap");
69
bvar::LatencyRecorder g_tablet_lookup_rowkey_latency("doris_pk", "tablet_lookup_rowkey");
70
bvar::Adder<uint64_t> g_tablet_pk_not_found("doris_pk", "lookup_not_found");
71
bvar::PerSecond<bvar::Adder<uint64_t>> g_tablet_pk_not_found_per_second(
72
        "doris_pk", "lookup_not_found_per_second", &g_tablet_pk_not_found, 60);
73
bvar::LatencyRecorder g_tablet_update_delete_bitmap_latency("doris_pk", "update_delete_bitmap");
74
75
static bvar::Adder<size_t> g_total_tablet_num("doris_total_tablet_num");
76
77
Status _get_segment_column_iterator(const BetaRowsetSharedPtr& rowset, uint32_t segid,
78
                                    const TabletColumn& target_column,
79
                                    SegmentCacheHandle* segment_cache_handle,
80
                                    std::unique_ptr<segment_v2::ColumnIterator>* column_iterator,
81
0
                                    OlapReaderStatistics* stats) {
82
0
    RETURN_IF_ERROR(SegmentLoader::instance()->load_segments(rowset, segment_cache_handle, true));
83
    // find segment
84
0
    auto it = std::find_if(
85
0
            segment_cache_handle->get_segments().begin(),
86
0
            segment_cache_handle->get_segments().end(),
87
0
            [&segid](const segment_v2::SegmentSharedPtr& seg) { return seg->id() == segid; });
88
0
    if (it == segment_cache_handle->get_segments().end()) {
89
0
        return Status::NotFound(fmt::format("rowset {} 's segemnt not found, seg_id {}",
90
0
                                            rowset->rowset_id().to_string(), segid));
91
0
    }
92
0
    segment_v2::SegmentSharedPtr segment = *it;
93
0
    StorageReadOptions opts;
94
0
    opts.stats = stats;
95
0
    RETURN_IF_ERROR(segment->new_column_iterator(target_column, column_iterator, &opts));
96
0
    segment_v2::ColumnIteratorOptions opt {
97
0
            .use_page_cache = !config::disable_storage_page_cache,
98
0
            .file_reader = segment->file_reader().get(),
99
0
            .stats = stats,
100
0
            .io_ctx = io::IOContext {.reader_type = ReaderType::READER_QUERY,
101
0
                                     .file_cache_stats = &stats->file_cache_stats},
102
0
    };
103
0
    RETURN_IF_ERROR((*column_iterator)->init(opt));
104
0
    return Status::OK();
105
0
}
106
107
} // namespace
108
109
extern MetricPrototype METRIC_query_scan_bytes;
110
extern MetricPrototype METRIC_query_scan_rows;
111
extern MetricPrototype METRIC_query_scan_count;
112
DEFINE_COUNTER_METRIC_PROTOTYPE_2ARG(flush_bytes, MetricUnit::BYTES);
113
DEFINE_COUNTER_METRIC_PROTOTYPE_2ARG(flush_finish_count, MetricUnit::OPERATIONS);
114
115
608
BaseTablet::BaseTablet(TabletMetaSharedPtr tablet_meta) : _tablet_meta(std::move(tablet_meta)) {
116
608
    _metric_entity = DorisMetrics::instance()->metric_registry()->register_entity(
117
608
            fmt::format("Tablet.{}", tablet_id()), {{"tablet_id", std::to_string(tablet_id())}},
118
608
            MetricEntityType::kTablet);
119
608
    INT_COUNTER_METRIC_REGISTER(_metric_entity, query_scan_bytes);
120
608
    INT_COUNTER_METRIC_REGISTER(_metric_entity, query_scan_rows);
121
608
    INT_COUNTER_METRIC_REGISTER(_metric_entity, query_scan_count);
122
608
    INT_COUNTER_METRIC_REGISTER(_metric_entity, flush_bytes);
123
608
    INT_COUNTER_METRIC_REGISTER(_metric_entity, flush_finish_count);
124
125
    // construct _timestamped_versioned_tracker from rs and stale rs meta
126
608
    _timestamped_version_tracker.construct_versioned_tracker(_tablet_meta->all_rs_metas(),
127
608
                                                             _tablet_meta->all_stale_rs_metas());
128
129
    // if !_tablet_meta->all_rs_metas()[0]->tablet_schema(),
130
    // that mean the tablet_meta is still no upgrade to doris 1.2 versions.
131
    // Before doris 1.2 version, rowset metas don't have tablet schema.
132
    // And when upgrade to doris 1.2 version,
133
    // all rowset metas will be set the tablet schmea from tablet meta.
134
608
    if (_tablet_meta->all_rs_metas().empty() || !_tablet_meta->all_rs_metas()[0]->tablet_schema()) {
135
571
        _max_version_schema = _tablet_meta->tablet_schema();
136
571
    } else {
137
37
        _max_version_schema =
138
37
                tablet_schema_with_merged_max_schema_version(_tablet_meta->all_rs_metas());
139
37
    }
140
608
    DCHECK(_max_version_schema);
141
608
    g_total_tablet_num << 1;
142
608
}
143
144
608
BaseTablet::~BaseTablet() {
145
608
    DorisMetrics::instance()->metric_registry()->deregister_entity(_metric_entity);
146
608
    g_total_tablet_num << -1;
147
608
}
148
149
TabletSchemaSPtr BaseTablet::tablet_schema_with_merged_max_schema_version(
150
79
        const std::vector<RowsetMetaSharedPtr>& rowset_metas) {
151
79
    RowsetMetaSharedPtr max_schema_version_rs = *std::max_element(
152
79
            rowset_metas.begin(), rowset_metas.end(),
153
248
            [](const RowsetMetaSharedPtr& a, const RowsetMetaSharedPtr& b) {
154
248
                return !a->tablet_schema()
155
248
                               ? true
156
248
                               : (!b->tablet_schema()
157
248
                                          ? false
158
248
                                          : a->tablet_schema()->schema_version() <
159
248
                                                    b->tablet_schema()->schema_version());
160
248
            });
161
79
    return max_schema_version_rs->tablet_schema();
162
79
}
163
164
257
Status BaseTablet::set_tablet_state(TabletState state) {
165
257
    if (_tablet_meta->tablet_state() == TABLET_SHUTDOWN && state != TABLET_SHUTDOWN) {
166
0
        return Status::Error<META_INVALID_ARGUMENT>(
167
0
                "could not change tablet state from shutdown to {}", state);
168
0
    }
169
257
    _tablet_meta->set_tablet_state(state);
170
257
    return Status::OK();
171
257
}
172
173
0
void BaseTablet::update_max_version_schema(const TabletSchemaSPtr& tablet_schema) {
174
0
    std::lock_guard wrlock(_meta_lock);
175
    // Double Check for concurrent update
176
0
    if (!_max_version_schema ||
177
0
        tablet_schema->schema_version() > _max_version_schema->schema_version()) {
178
0
        _max_version_schema = tablet_schema;
179
0
    }
180
0
}
181
182
233
uint32_t BaseTablet::get_real_compaction_score() const {
183
233
    std::shared_lock l(_meta_lock);
184
233
    const auto& rs_metas = _tablet_meta->all_rs_metas();
185
233
    return std::accumulate(rs_metas.begin(), rs_metas.end(), 0,
186
10.7k
                           [](uint32_t score, const RowsetMetaSharedPtr& rs_meta) {
187
10.7k
                               return score + rs_meta->get_compaction_score();
188
10.7k
                           });
189
233
}
190
191
Status BaseTablet::capture_rs_readers_unlocked(const Versions& version_path,
192
1
                                               std::vector<RowSetSplits>* rs_splits) const {
193
1
    DCHECK(rs_splits != nullptr && rs_splits->empty());
194
1
    for (auto version : version_path) {
195
1
        auto it = _rs_version_map.find(version);
196
1
        if (it == _rs_version_map.end()) {
197
0
            VLOG_NOTICE << "fail to find Rowset in rs_version for version. tablet=" << tablet_id()
198
0
                        << ", version='" << version.first << "-" << version.second;
199
200
0
            it = _stale_rs_version_map.find(version);
201
0
            if (it == _stale_rs_version_map.end()) {
202
0
                return Status::Error<CAPTURE_ROWSET_READER_ERROR>(
203
0
                        "fail to find Rowset in stale_rs_version for version. tablet={}, "
204
0
                        "version={}-{}",
205
0
                        tablet_id(), version.first, version.second);
206
0
            }
207
0
        }
208
1
        RowsetReaderSharedPtr rs_reader;
209
1
        auto res = it->second->create_reader(&rs_reader);
210
1
        if (!res.ok()) {
211
0
            return Status::Error<CAPTURE_ROWSET_READER_ERROR>(
212
0
                    "failed to create reader for rowset:{}", it->second->rowset_id().to_string());
213
0
        }
214
1
        rs_splits->emplace_back(std::move(rs_reader));
215
1
    }
216
1
    return Status::OK();
217
1
}
218
219
// snapshot manager may call this api to check if version exists, so that
220
// the version maybe not exist
221
RowsetSharedPtr BaseTablet::get_rowset_by_version(const Version& version,
222
12
                                                  bool find_in_stale) const {
223
12
    auto iter = _rs_version_map.find(version);
224
12
    if (iter == _rs_version_map.end()) {
225
0
        if (find_in_stale) {
226
0
            return get_stale_rowset_by_version(version);
227
0
        }
228
0
        return nullptr;
229
0
    }
230
12
    return iter->second;
231
12
}
232
233
0
RowsetSharedPtr BaseTablet::get_stale_rowset_by_version(const Version& version) const {
234
0
    auto iter = _stale_rs_version_map.find(version);
235
0
    if (iter == _stale_rs_version_map.end()) {
236
0
        VLOG_NOTICE << "no rowset for version:" << version << ", tablet: " << tablet_id();
237
0
        return nullptr;
238
0
    }
239
0
    return iter->second;
240
0
}
241
242
// Already under _meta_lock
243
42
RowsetSharedPtr BaseTablet::get_rowset_with_max_version() const {
244
42
    Version max_version = _tablet_meta->max_version();
245
42
    if (max_version.first == -1) {
246
0
        return nullptr;
247
0
    }
248
249
42
    auto iter = _rs_version_map.find(max_version);
250
42
    if (iter == _rs_version_map.end()) {
251
0
        DCHECK(false) << "invalid version:" << max_version;
252
0
        return nullptr;
253
0
    }
254
42
    return iter->second;
255
42
}
256
257
0
Status BaseTablet::get_all_rs_id(int64_t max_version, RowsetIdUnorderedSet* rowset_ids) const {
258
0
    std::shared_lock rlock(_meta_lock);
259
0
    return get_all_rs_id_unlocked(max_version, rowset_ids);
260
0
}
261
262
Status BaseTablet::get_all_rs_id_unlocked(int64_t max_version,
263
9
                                          RowsetIdUnorderedSet* rowset_ids) const {
264
    //  Ensure that the obtained versions of rowsets are continuous
265
9
    Version spec_version(0, max_version);
266
9
    Versions version_path;
267
9
    auto st = _timestamped_version_tracker.capture_consistent_versions(spec_version, &version_path);
268
9
    if (!st.ok()) [[unlikely]] {
269
0
        return st;
270
0
    }
271
272
11
    for (auto& ver : version_path) {
273
11
        if (ver.second == 1) {
274
            // [0-1] rowset is empty for each tablet, skip it
275
9
            continue;
276
9
        }
277
2
        auto it = _rs_version_map.find(ver);
278
2
        if (it == _rs_version_map.end()) {
279
0
            return Status::Error<CAPTURE_ROWSET_ERROR, false>(
280
0
                    "fail to find Rowset for version. tablet={}, version={}", tablet_id(),
281
0
                    ver.to_string());
282
0
        }
283
2
        rowset_ids->emplace(it->second->rowset_id());
284
2
    }
285
9
    return Status::OK();
286
9
}
287
288
0
Versions BaseTablet::get_missed_versions(int64_t spec_version) const {
289
0
    DCHECK(spec_version > 0) << "invalid spec_version: " << spec_version;
290
291
0
    Versions existing_versions;
292
0
    {
293
0
        std::shared_lock rdlock(_meta_lock);
294
0
        for (const auto& rs : _tablet_meta->all_rs_metas()) {
295
0
            existing_versions.emplace_back(rs->version());
296
0
        }
297
0
    }
298
0
    return calc_missed_versions(spec_version, std::move(existing_versions));
299
0
}
300
301
2
Versions BaseTablet::get_missed_versions_unlocked(int64_t spec_version) const {
302
2
    DCHECK(spec_version > 0) << "invalid spec_version: " << spec_version;
303
304
2
    Versions existing_versions;
305
8
    for (const auto& rs : _tablet_meta->all_rs_metas()) {
306
8
        existing_versions.emplace_back(rs->version());
307
8
    }
308
2
    return calc_missed_versions(spec_version, std::move(existing_versions));
309
2
}
310
311
1
void BaseTablet::_print_missed_versions(const Versions& missed_versions) const {
312
1
    std::stringstream ss;
313
1
    ss << tablet_id() << " has " << missed_versions.size() << " missed version:";
314
    // print at most 10 version
315
3
    for (int i = 0; i < 10 && i < missed_versions.size(); ++i) {
316
2
        ss << missed_versions[i] << ",";
317
2
    }
318
1
    LOG(WARNING) << ss.str();
319
1
}
320
321
1
bool BaseTablet::_reconstruct_version_tracker_if_necessary() {
322
1
    double orphan_vertex_ratio = _timestamped_version_tracker.get_orphan_vertex_ratio();
323
1
    if (orphan_vertex_ratio >= config::tablet_version_graph_orphan_vertex_ratio) {
324
1
        _timestamped_version_tracker.construct_versioned_tracker(
325
1
                _tablet_meta->all_rs_metas(), _tablet_meta->all_stale_rs_metas());
326
1
        return true;
327
1
    }
328
0
    return false;
329
1
}
330
331
// should use this method to get a copy of current tablet meta
332
// there are some rowset meta in local meta store and in in-memory tablet meta
333
// but not in tablet meta in local meta store
334
0
void BaseTablet::generate_tablet_meta_copy(TabletMeta& new_tablet_meta) const {
335
0
    TabletMetaPB tablet_meta_pb;
336
0
    {
337
0
        std::shared_lock rdlock(_meta_lock);
338
0
        _tablet_meta->to_meta_pb(&tablet_meta_pb);
339
0
    }
340
0
    generate_tablet_meta_copy_unlocked(new_tablet_meta);
341
0
}
342
343
// this is a unlocked version of generate_tablet_meta_copy()
344
// some method already hold the _meta_lock before calling this,
345
// such as EngineCloneTask::_finish_clone -> tablet->revise_tablet_meta
346
4
void BaseTablet::generate_tablet_meta_copy_unlocked(TabletMeta& new_tablet_meta) const {
347
4
    TabletMetaPB tablet_meta_pb;
348
4
    _tablet_meta->to_meta_pb(&tablet_meta_pb);
349
4
    new_tablet_meta.init_from_pb(tablet_meta_pb);
350
4
}
351
352
Status BaseTablet::calc_delete_bitmap_between_segments(
353
        TabletSchemaSPtr schema, RowsetId rowset_id,
354
0
        const std::vector<segment_v2::SegmentSharedPtr>& segments, DeleteBitmapPtr delete_bitmap) {
355
0
    size_t const num_segments = segments.size();
356
0
    if (num_segments < 2) {
357
0
        return Status::OK();
358
0
    }
359
360
0
    OlapStopWatch watch;
361
0
    size_t seq_col_length = 0;
362
0
    if (schema->has_sequence_col()) {
363
0
        auto seq_col_idx = schema->sequence_col_idx();
364
0
        seq_col_length = schema->column(seq_col_idx).length() + 1;
365
0
    }
366
0
    size_t rowid_length = 0;
367
0
    if (!schema->cluster_key_uids().empty()) {
368
0
        rowid_length = PrimaryKeyIndexReader::ROW_ID_LENGTH;
369
0
    }
370
371
0
    MergeIndexDeleteBitmapCalculator calculator;
372
0
    RETURN_IF_ERROR(calculator.init(rowset_id, segments, seq_col_length, rowid_length));
373
374
0
    RETURN_IF_ERROR(calculator.calculate_all(delete_bitmap));
375
376
0
    delete_bitmap->add(
377
0
            {rowset_id, DeleteBitmap::INVALID_SEGMENT_ID, DeleteBitmap::TEMP_VERSION_COMMON},
378
0
            DeleteBitmap::ROWSET_SENTINEL_MARK);
379
0
    LOG(INFO) << fmt::format(
380
0
            "construct delete bitmap between segments, "
381
0
            "tablet: {}, rowset: {}, number of segments: {}, bitmap count: {}, bitmap cardinality: "
382
0
            "{}, cost {} (us)",
383
0
            tablet_id(), rowset_id.to_string(), num_segments,
384
0
            delete_bitmap->get_delete_bitmap_count(), delete_bitmap->cardinality(),
385
0
            watch.get_elapse_time_us());
386
0
    return Status::OK();
387
0
}
388
389
std::vector<RowsetSharedPtr> BaseTablet::get_rowset_by_ids(
390
72
        const RowsetIdUnorderedSet* specified_rowset_ids) {
391
72
    std::vector<RowsetSharedPtr> rowsets;
392
72
    for (auto& rs : _rs_version_map) {
393
15
        if (!specified_rowset_ids ||
394
15
            specified_rowset_ids->find(rs.second->rowset_id()) != specified_rowset_ids->end()) {
395
1
            rowsets.push_back(rs.second);
396
1
        }
397
15
    }
398
399
72
    std::sort(rowsets.begin(), rowsets.end(), [](RowsetSharedPtr& lhs, RowsetSharedPtr& rhs) {
400
0
        return lhs->end_version() > rhs->end_version();
401
0
    });
402
72
    return rowsets;
403
72
}
404
405
Status BaseTablet::lookup_row_data(const Slice& encoded_key, const RowLocation& row_location,
406
                                   RowsetSharedPtr input_rowset, OlapReaderStatistics& stats,
407
0
                                   std::string& values, bool write_to_cache) {
408
0
    MonotonicStopWatch watch;
409
0
    size_t row_size = 1;
410
0
    watch.start();
411
0
    Defer _defer([&]() {
412
0
        LOG_EVERY_N(INFO, 500) << "get a single_row, cost(us):" << watch.elapsed_time() / 1000
413
0
                               << ", row_size:" << row_size;
414
0
    });
415
416
0
    BetaRowsetSharedPtr rowset = std::static_pointer_cast<BetaRowset>(input_rowset);
417
0
    CHECK(rowset);
418
0
    const TabletSchemaSPtr tablet_schema = rowset->tablet_schema();
419
0
    SegmentCacheHandle segment_cache_handle;
420
0
    std::unique_ptr<segment_v2::ColumnIterator> column_iterator;
421
0
    const auto& column = *DORIS_TRY(tablet_schema->column(BeConsts::ROW_STORE_COL));
422
0
    RETURN_IF_ERROR(_get_segment_column_iterator(rowset, row_location.segment_id, column,
423
0
                                                 &segment_cache_handle, &column_iterator, &stats));
424
    // get and parse tuple row
425
0
    vectorized::MutableColumnPtr column_ptr = vectorized::ColumnString::create();
426
0
    std::vector<segment_v2::rowid_t> rowids {static_cast<segment_v2::rowid_t>(row_location.row_id)};
427
0
    RETURN_IF_ERROR(column_iterator->read_by_rowids(rowids.data(), 1, column_ptr));
428
0
    assert(column_ptr->size() == 1);
429
0
    auto* string_column = static_cast<vectorized::ColumnString*>(column_ptr.get());
430
0
    StringRef value = string_column->get_data_at(0);
431
0
    values = value.to_string();
432
0
    if (write_to_cache) {
433
0
        RowCache::instance()->insert({tablet_id(), encoded_key}, Slice {value.data, value.size});
434
0
    }
435
0
    return Status::OK();
436
0
}
437
438
Status BaseTablet::lookup_row_key(const Slice& encoded_key, TabletSchema* latest_schema,
439
                                  bool with_seq_col,
440
                                  const std::vector<RowsetSharedPtr>& specified_rowsets,
441
                                  RowLocation* row_location, int64_t version,
442
                                  std::vector<std::unique_ptr<SegmentCacheHandle>>& segment_caches,
443
                                  RowsetSharedPtr* rowset, bool with_rowid,
444
                                  std::string* encoded_seq_value, OlapReaderStatistics* stats,
445
2
                                  DeleteBitmapPtr delete_bitmap) {
446
2
    SCOPED_BVAR_LATENCY(g_tablet_lookup_rowkey_latency);
447
2
    size_t seq_col_length = 0;
448
    // use the latest tablet schema to decide if the tablet has sequence column currently
449
2
    const TabletSchema* schema =
450
2
            (latest_schema == nullptr ? _tablet_meta->tablet_schema().get() : latest_schema);
451
2
    if (schema->has_sequence_col() && with_seq_col) {
452
2
        seq_col_length = schema->column(schema->sequence_col_idx()).length() + 1;
453
2
    }
454
2
    size_t rowid_length = 0;
455
2
    if (with_rowid && !schema->cluster_key_uids().empty()) {
456
0
        rowid_length = PrimaryKeyIndexReader::ROW_ID_LENGTH;
457
0
    }
458
2
    Slice key_without_seq =
459
2
            Slice(encoded_key.get_data(), encoded_key.get_size() - seq_col_length - rowid_length);
460
2
    RowLocation loc;
461
462
2
    auto tablet_delete_bitmap =
463
2
            delete_bitmap == nullptr ? _tablet_meta->delete_bitmap_ptr() : delete_bitmap;
464
2
    for (size_t i = 0; i < specified_rowsets.size(); i++) {
465
2
        const auto& rs = specified_rowsets[i];
466
2
        std::vector<KeyBoundsPB> segments_key_bounds;
467
2
        rs->rowset_meta()->get_segments_key_bounds(&segments_key_bounds);
468
2
        int num_segments = cast_set<int>(rs->num_segments());
469
2
        DCHECK_EQ(segments_key_bounds.size(), num_segments);
470
2
        std::vector<uint32_t> picked_segments;
471
4
        for (int j = num_segments - 1; j >= 0; j--) {
472
2
            if (key_is_not_in_segment(key_without_seq, segments_key_bounds[j],
473
2
                                      rs->rowset_meta()->is_segments_key_bounds_truncated())) {
474
0
                continue;
475
0
            }
476
2
            picked_segments.emplace_back(j);
477
2
        }
478
2
        if (picked_segments.empty()) {
479
0
            continue;
480
0
        }
481
482
2
        if (UNLIKELY(segment_caches[i] == nullptr)) {
483
1
            segment_caches[i] = std::make_unique<SegmentCacheHandle>();
484
1
            RETURN_IF_ERROR(SegmentLoader::instance()->load_segments(
485
1
                    std::static_pointer_cast<BetaRowset>(rs), segment_caches[i].get(), true, true));
486
1
        }
487
2
        auto& segments = segment_caches[i]->get_segments();
488
2
        DCHECK_EQ(segments.size(), num_segments);
489
490
2
        for (auto id : picked_segments) {
491
2
            Status s = segments[id]->lookup_row_key(encoded_key, schema, with_seq_col, with_rowid,
492
2
                                                    &loc, stats, encoded_seq_value);
493
2
            if (s.is<KEY_NOT_FOUND>()) {
494
0
                continue;
495
0
            }
496
2
            if (!s.ok() && !s.is<KEY_ALREADY_EXISTS>()) {
497
0
                return s;
498
0
            }
499
2
            if (s.ok() && tablet_delete_bitmap->contains_agg_with_cache_if_eligible(
500
1
                                  {loc.rowset_id, loc.segment_id, version}, loc.row_id)) {
501
                // if has sequence col, we continue to compare the sequence_id of
502
                // all rowsets, util we find an existing key.
503
0
                if (schema->has_sequence_col()) {
504
0
                    continue;
505
0
                }
506
                // The key is deleted, we don't need to search for it any more.
507
0
                break;
508
0
            }
509
            // `st` is either OK or KEY_ALREADY_EXISTS now.
510
            // for partial update, even if the key is already exists, we still need to
511
            // read it's original values to keep all columns align.
512
2
            *row_location = loc;
513
2
            if (rowset) {
514
                // return it's rowset
515
2
                *rowset = rs;
516
2
            }
517
            // find it and return
518
2
            return s;
519
2
        }
520
2
    }
521
0
    g_tablet_pk_not_found << 1;
522
0
    return Status::Error<ErrorCode::KEY_NOT_FOUND>("can't find key in all rowsets");
523
2
}
524
525
// if user pass a token, then all calculation works will submit to a threadpool,
526
// user can get all delete bitmaps from that token.
527
// if `token` is nullptr, the calculation will run in local, and user can get the result
528
// delete bitmap from `delete_bitmap` directly.
529
Status BaseTablet::calc_delete_bitmap(
530
        const BaseTabletSPtr& tablet, RowsetSharedPtr rowset,
531
        const std::vector<segment_v2::SegmentSharedPtr>& segments,
532
        const std::vector<RowsetSharedPtr>& specified_rowsets, DeleteBitmapPtr delete_bitmap,
533
        int64_t end_version, CalcDeleteBitmapToken* token, RowsetWriter* rowset_writer,
534
        DeleteBitmapPtr tablet_delete_bitmap,
535
69
        std::function<void(segment_v2::SegmentSharedPtr, Status)> callback) {
536
69
    if (specified_rowsets.empty() || segments.empty()) {
537
68
        return Status::OK();
538
68
    }
539
540
1
    OlapStopWatch watch;
541
1
    for (const auto& segment : segments) {
542
1
        const auto& seg = segment;
543
1
        if (token != nullptr) {
544
1
            RETURN_IF_ERROR(token->submit(tablet, rowset, seg, specified_rowsets, end_version,
545
1
                                          delete_bitmap, rowset_writer, tablet_delete_bitmap,
546
1
                                          callback));
547
1
        } else {
548
0
            RETURN_IF_ERROR(tablet->calc_segment_delete_bitmap(
549
0
                    rowset, segment, specified_rowsets, delete_bitmap, end_version, rowset_writer,
550
0
                    tablet_delete_bitmap));
551
0
        }
552
1
    }
553
554
1
    return Status::OK();
555
1
}
556
557
Status BaseTablet::calc_segment_delete_bitmap(RowsetSharedPtr rowset,
558
                                              const segment_v2::SegmentSharedPtr& seg,
559
                                              const std::vector<RowsetSharedPtr>& specified_rowsets,
560
                                              DeleteBitmapPtr delete_bitmap, int64_t end_version,
561
                                              RowsetWriter* rowset_writer,
562
1
                                              DeleteBitmapPtr tablet_delete_bitmap) {
563
1
    OlapStopWatch watch;
564
1
    auto rowset_id = rowset->rowset_id();
565
1
    Version dummy_version(end_version + 1, end_version + 1);
566
1
    auto rowset_schema = rowset->tablet_schema();
567
568
1
    PartialUpdateInfo* partial_update_info {nullptr};
569
1
    bool is_partial_update = rowset_writer && rowset_writer->is_partial_update();
570
    // `have_input_seq_column` is for fixed partial update only. For flexible partial update, we should use
571
    // the skip bitmap to determine wheather a row has specified the sequence column
572
1
    bool have_input_seq_column = false;
573
    // `rids_be_overwritten` is for flexible partial update only, it records row ids that is overwritten by
574
    // another row with higher seqeucne value
575
1
    std::set<uint32_t> rids_be_overwritten;
576
1
    if (is_partial_update) {
577
0
        partial_update_info = rowset_writer->get_partial_update_info().get();
578
0
        if (partial_update_info->is_fixed_partial_update() && rowset_schema->has_sequence_col()) {
579
0
            std::vector<uint32_t> including_cids =
580
0
                    rowset_writer->get_partial_update_info()->update_cids;
581
0
            have_input_seq_column =
582
0
                    rowset_schema->has_sequence_col() &&
583
0
                    (std::find(including_cids.cbegin(), including_cids.cend(),
584
0
                               rowset_schema->sequence_col_idx()) != including_cids.cend());
585
0
        }
586
0
    }
587
588
1
    DBUG_EXECUTE_IF("BaseTablet::calc_segment_delete_bitmap.sleep", {
589
1
        auto target_tablet_id = dp->param<int64_t>("tablet_id", -1);
590
1
        auto sleep = dp->param<int64_t>("sleep", 10);
591
1
        if (target_tablet_id == tablet_id()) {
592
1
            std::this_thread::sleep_for(std::chrono::seconds(sleep));
593
1
        }
594
1
    });
595
596
1
    if (rowset_schema->num_variant_columns() > 0) {
597
        // During partial updates, the extracted columns of a variant should not be included in the rowset schema.
598
        // This is because the partial update for a variant needs to ignore the extracted columns.
599
        // Otherwise, the schema types in different rowsets might be inconsistent. When performing a partial update,
600
        // the complete variant is constructed by reading all the sub-columns of the variant.
601
0
        rowset_schema = rowset_schema->copy_without_variant_extracted_columns();
602
0
    }
603
    // use for partial update
604
1
    FixedReadPlan read_plan_ori;
605
1
    FixedReadPlan read_plan_update;
606
1
    int64_t conflict_rows = 0;
607
1
    int64_t new_generated_rows = 0;
608
609
1
    std::map<RowsetId, RowsetSharedPtr> rsid_to_rowset;
610
1
    rsid_to_rowset[rowset_id] = rowset;
611
1
    vectorized::Block block = rowset_schema->create_block();
612
1
    vectorized::Block ordered_block = block.clone_empty();
613
1
    uint32_t pos = 0;
614
615
1
    RETURN_IF_ERROR(seg->load_pk_index_and_bf(nullptr)); // We need index blocks to iterate
616
1
    const auto* pk_idx = seg->get_primary_key_index();
617
1
    int64_t total = pk_idx->num_rows();
618
1
    uint32_t row_id = 0;
619
1
    int64_t remaining = total;
620
1
    bool exact_match = false;
621
1
    std::string last_key;
622
1
    int batch_size = 1024;
623
    // The data for each segment may be lookup multiple times. Creating a SegmentCacheHandle
624
    // will update the lru cache, and there will be obvious lock competition in multithreading
625
    // scenarios, so using a segment_caches to cache SegmentCacheHandle.
626
1
    std::vector<std::unique_ptr<SegmentCacheHandle>> segment_caches(specified_rowsets.size());
627
2
    while (remaining > 0) {
628
1
        std::unique_ptr<segment_v2::IndexedColumnIterator> iter;
629
1
        RETURN_IF_ERROR(pk_idx->new_iterator(&iter, nullptr));
630
631
1
        size_t num_to_read = std::min<int64_t>(batch_size, remaining);
632
1
        auto index_type = vectorized::DataTypeFactory::instance().create_data_type(
633
1
                pk_idx->type_info()->type(), 1, 0);
634
1
        auto index_column = index_type->create_column();
635
1
        Slice last_key_slice(last_key);
636
1
        RETURN_IF_ERROR(iter->seek_at_or_after(&last_key_slice, &exact_match));
637
1
        auto current_ordinal = iter->get_current_ordinal();
638
1
        DCHECK(total == remaining + current_ordinal)
639
0
                << "total: " << total << ", remaining: " << remaining
640
0
                << ", current_ordinal: " << current_ordinal;
641
642
1
        size_t num_read = num_to_read;
643
1
        RETURN_IF_ERROR(iter->next_batch(&num_read, index_column));
644
1
        DCHECK(num_to_read == num_read)
645
0
                << "num_to_read: " << num_to_read << ", num_read: " << num_read;
646
1
        last_key = index_column->get_data_at(num_read - 1).to_string();
647
648
        // exclude last_key, last_key will be read in next batch.
649
1
        if (num_read == batch_size && num_read != remaining) {
650
0
            num_read -= 1;
651
0
        }
652
3
        for (size_t i = 0; i < num_read; i++, row_id++) {
653
2
            Slice key = Slice(index_column->get_data_at(i).data, index_column->get_data_at(i).size);
654
2
            RowLocation loc;
655
            // calculate row id
656
2
            if (!_tablet_meta->tablet_schema()->cluster_key_uids().empty()) {
657
0
                size_t seq_col_length = 0;
658
0
                if (_tablet_meta->tablet_schema()->has_sequence_col()) {
659
0
                    seq_col_length =
660
0
                            _tablet_meta->tablet_schema()
661
0
                                    ->column(_tablet_meta->tablet_schema()->sequence_col_idx())
662
0
                                    .length() +
663
0
                            1;
664
0
                }
665
0
                size_t rowid_length = PrimaryKeyIndexReader::ROW_ID_LENGTH;
666
0
                Slice key_without_seq =
667
0
                        Slice(key.get_data(), key.get_size() - seq_col_length - rowid_length);
668
0
                Slice rowid_slice =
669
0
                        Slice(key.get_data() + key_without_seq.get_size() + seq_col_length + 1,
670
0
                              rowid_length - 1);
671
0
                const auto* type_info =
672
0
                        get_scalar_type_info<FieldType::OLAP_FIELD_TYPE_UNSIGNED_INT>();
673
0
                const auto* rowid_coder = get_key_coder(type_info->type());
674
0
                RETURN_IF_ERROR(rowid_coder->decode_ascending(&rowid_slice, rowid_length,
675
0
                                                              (uint8_t*)&row_id));
676
0
            }
677
            // same row in segments should be filtered
678
2
            if (delete_bitmap->contains({rowset_id, seg->id(), DeleteBitmap::TEMP_VERSION_COMMON},
679
2
                                        row_id)) {
680
0
                continue;
681
0
            }
682
683
2
            DBUG_EXECUTE_IF("BaseTablet::calc_segment_delete_bitmap.inject_err", {
684
2
                auto p = dp->param("percent", 0.01);
685
2
                std::mt19937 gen {std::random_device {}()};
686
2
                std::bernoulli_distribution inject_fault {p};
687
2
                if (inject_fault(gen)) {
688
2
                    return Status::InternalError(
689
2
                            "injection error in calc_segment_delete_bitmap, "
690
2
                            "tablet_id={}, rowset_id={}",
691
2
                            tablet_id(), rowset_id.to_string());
692
2
                }
693
2
            });
694
695
2
            RowsetSharedPtr rowset_find;
696
2
            Status st = Status::OK();
697
2
            if (tablet_delete_bitmap == nullptr) {
698
2
                st = lookup_row_key(key, rowset_schema.get(), true, specified_rowsets, &loc,
699
2
                                    dummy_version.first - 1, segment_caches, &rowset_find);
700
2
            } else {
701
0
                st = lookup_row_key(key, rowset_schema.get(), true, specified_rowsets, &loc,
702
0
                                    dummy_version.first - 1, segment_caches, &rowset_find, true,
703
0
                                    nullptr, nullptr, tablet_delete_bitmap);
704
0
            }
705
2
            bool expected_st = st.ok() || st.is<KEY_NOT_FOUND>() || st.is<KEY_ALREADY_EXISTS>();
706
            // It's a defensive DCHECK, we need to exclude some common errors to avoid core-dump
707
            // while stress test
708
2
            DCHECK(expected_st || st.is<MEM_LIMIT_EXCEEDED>())
709
0
                    << "unexpected error status while lookup_row_key:" << st;
710
2
            if (!expected_st) {
711
0
                return st;
712
0
            }
713
2
            if (st.is<KEY_NOT_FOUND>()) {
714
0
                continue;
715
0
            }
716
717
2
            ++conflict_rows;
718
2
            if (st.is<KEY_ALREADY_EXISTS>() &&
719
2
                (!is_partial_update ||
720
1
                 (partial_update_info->is_fixed_partial_update() && have_input_seq_column))) {
721
                // `st.is<KEY_ALREADY_EXISTS>()` means that there exists a row with the same key and larger value
722
                // in seqeunce column.
723
                // - If the current load is not a partial update, we just delete current row.
724
                // - Otherwise, it means that we are doing the alignment process in publish phase due to conflicts
725
                // during concurrent partial updates. And there exists another load which introduces a row with
726
                // the same keys and larger sequence column value published successfully after the commit phase
727
                // of the current load.
728
                //     - If the columns we update include sequence column, we should delete the current row becase the
729
                //       partial update on the current row has been `overwritten` by the previous one with larger sequence
730
                //       column value.
731
                //     - Otherwise, we should combine the values of the missing columns in the previous row and the values
732
                //       of the including columns in the current row into a new row.
733
1
                delete_bitmap->add({rowset_id, seg->id(), DeleteBitmap::TEMP_VERSION_COMMON},
734
1
                                   row_id);
735
1
                continue;
736
                // NOTE: for partial update which doesn't specify the sequence column, we can't use the sequence column value filled in flush phase
737
                // as its final value. Otherwise it may cause inconsistency between replicas.
738
1
            }
739
1
            if (is_partial_update && rowset_writer != nullptr) {
740
                // In publish version, record rows to be deleted for concurrent update
741
                // For example, if version 5 and 6 update a row, but version 6 only see
742
                // version 4 when write, and when publish version, version 5's value will
743
                // be marked as deleted and it's update is losed.
744
                // So here we should read version 5's columns and build a new row, which is
745
                // consists of version 6's update columns and version 5's origin columns
746
                // here we build 2 read plan for ori values and update values
747
748
                // - for fixed partial update, we should read update columns from current load's rowset
749
                // and read missing columns from previous rowsets to create the final block
750
                // - for flexible partial update, we should read all columns from current load's rowset
751
                // and read non sort key columns from previous rowsets to create the final block
752
                // So we only need to record rows to read for both mode partial update
753
0
                read_plan_ori.prepare_to_read(loc, pos);
754
0
                read_plan_update.prepare_to_read(RowLocation {rowset_id, seg->id(), row_id}, pos);
755
756
                // For flexible partial update, we should use skip bitmap to determine wheather
757
                // a row has specified the sequence column. But skip bitmap should be read from the segment.
758
                // So we record these row ids and process and filter them in `generate_new_block_for_flexible_partial_update()`
759
0
                if (st.is<KEY_ALREADY_EXISTS>() &&
760
0
                    partial_update_info->is_flexible_partial_update()) {
761
0
                    rids_be_overwritten.insert(pos);
762
0
                }
763
764
0
                rsid_to_rowset[rowset_find->rowset_id()] = rowset_find;
765
0
                ++pos;
766
767
                // delete bitmap will be calculate when memtable flush and
768
                // publish. The two stages may see different versions.
769
                // When there is sequence column, the currently imported data
770
                // of rowset may be marked for deletion at memtablet flush or
771
                // publish because the seq column is smaller than the previous
772
                // rowset.
773
                // just set 0 as a unified temporary version number, and update to
774
                // the real version number later.
775
0
                delete_bitmap->add(
776
0
                        {loc.rowset_id, loc.segment_id, DeleteBitmap::TEMP_VERSION_COMMON},
777
0
                        loc.row_id);
778
0
                delete_bitmap->add({rowset_id, seg->id(), DeleteBitmap::TEMP_VERSION_COMMON},
779
0
                                   row_id);
780
0
                ++new_generated_rows;
781
0
                continue;
782
0
            }
783
            // when st = ok
784
1
            delete_bitmap->add({loc.rowset_id, loc.segment_id, DeleteBitmap::TEMP_VERSION_COMMON},
785
1
                               loc.row_id);
786
1
        }
787
1
        remaining -= num_read;
788
1
    }
789
    // DCHECK_EQ(total, row_id) << "segment total rows: " << total << " row_id:" << row_id;
790
791
1
    if (config::enable_merge_on_write_correctness_check) {
792
1
        RowsetIdUnorderedSet rowsetids;
793
1
        for (const auto& specified_rowset : specified_rowsets) {
794
1
            rowsetids.emplace(specified_rowset->rowset_id());
795
1
            VLOG_NOTICE << "[tabletID:" << tablet_id() << "]"
796
0
                        << "[add_sentinel_mark_to_delete_bitmap][end_version:" << end_version << "]"
797
0
                        << "add:" << specified_rowset->rowset_id();
798
1
        }
799
1
        add_sentinel_mark_to_delete_bitmap(delete_bitmap.get(), rowsetids);
800
1
    }
801
802
1
    if (pos > 0) {
803
0
        DCHECK(partial_update_info);
804
0
        if (partial_update_info->is_fixed_partial_update()) {
805
0
            RETURN_IF_ERROR(generate_new_block_for_partial_update(
806
0
                    rowset_schema, partial_update_info, read_plan_ori, read_plan_update,
807
0
                    rsid_to_rowset, &block));
808
0
        } else {
809
0
            RETURN_IF_ERROR(generate_new_block_for_flexible_partial_update(
810
0
                    rowset_schema, partial_update_info, rids_be_overwritten, read_plan_ori,
811
0
                    read_plan_update, rsid_to_rowset, &block));
812
0
        }
813
0
        RETURN_IF_ERROR(sort_block(block, ordered_block));
814
0
        RETURN_IF_ERROR(rowset_writer->flush_single_block(&ordered_block));
815
0
        auto cost_us = watch.get_elapse_time_us();
816
0
        if (config::enable_mow_verbose_log || cost_us > 10 * 1000) {
817
0
            LOG(INFO) << "calc segment delete bitmap for "
818
0
                      << partial_update_info->partial_update_mode_str()
819
0
                      << ", tablet: " << tablet_id() << " rowset: " << rowset_id
820
0
                      << " seg_id: " << seg->id() << " dummy_version: " << end_version + 1
821
0
                      << " rows: " << seg->num_rows() << " conflict rows: " << conflict_rows
822
0
                      << " new generated rows: " << new_generated_rows
823
0
                      << " bitmap num: " << delete_bitmap->get_delete_bitmap_count()
824
0
                      << " bitmap cardinality: " << delete_bitmap->cardinality()
825
0
                      << " cost: " << cost_us << "(us)";
826
0
        }
827
0
        return Status::OK();
828
0
    }
829
1
    auto cost_us = watch.get_elapse_time_us();
830
1
    if (config::enable_mow_verbose_log || cost_us > 10 * 1000) {
831
0
        LOG(INFO) << "calc segment delete bitmap, tablet: " << tablet_id()
832
0
                  << " rowset: " << rowset_id << " seg_id: " << seg->id()
833
0
                  << " dummy_version: " << end_version + 1 << " rows: " << seg->num_rows()
834
0
                  << " conflict rows: " << conflict_rows
835
0
                  << " bitmap num: " << delete_bitmap->get_delete_bitmap_count()
836
0
                  << " bitmap cardinality: " << delete_bitmap->cardinality() << " cost: " << cost_us
837
0
                  << "(us)";
838
0
    }
839
1
    return Status::OK();
840
1
}
841
842
0
Status BaseTablet::sort_block(vectorized::Block& in_block, vectorized::Block& output_block) {
843
0
    vectorized::MutableBlock mutable_input_block =
844
0
            vectorized::MutableBlock::build_mutable_block(&in_block);
845
0
    vectorized::MutableBlock mutable_output_block =
846
0
            vectorized::MutableBlock::build_mutable_block(&output_block);
847
848
0
    std::shared_ptr<RowInBlockComparator> vec_row_comparator =
849
0
            std::make_shared<RowInBlockComparator>(_tablet_meta->tablet_schema());
850
0
    vec_row_comparator->set_block(&mutable_input_block);
851
852
0
    std::vector<std::unique_ptr<RowInBlock>> row_in_blocks;
853
0
    DCHECK(in_block.rows() <= std::numeric_limits<int>::max());
854
0
    row_in_blocks.reserve(in_block.rows());
855
0
    for (size_t i = 0; i < in_block.rows(); ++i) {
856
0
        row_in_blocks.emplace_back(std::make_unique<RowInBlock>(i));
857
0
    }
858
0
    std::sort(row_in_blocks.begin(), row_in_blocks.end(),
859
0
              [&](const std::unique_ptr<RowInBlock>& l,
860
0
                  const std::unique_ptr<RowInBlock>& r) -> bool {
861
0
                  auto value = (*vec_row_comparator)(l.get(), r.get());
862
0
                  DCHECK(value != 0) << "value equel when sort block, l_pos: " << l->_row_pos
863
0
                                     << " r_pos: " << r->_row_pos;
864
0
                  return value < 0;
865
0
              });
866
0
    std::vector<uint32_t> row_pos_vec;
867
0
    row_pos_vec.reserve(in_block.rows());
868
0
    for (auto& block : row_in_blocks) {
869
0
        row_pos_vec.emplace_back(block->_row_pos);
870
0
    }
871
0
    return mutable_output_block.add_rows(&in_block, row_pos_vec.data(),
872
0
                                         row_pos_vec.data() + in_block.rows());
873
0
}
874
875
// fetch value by row column
876
Status BaseTablet::fetch_value_through_row_column(RowsetSharedPtr input_rowset,
877
                                                  const TabletSchema& tablet_schema, uint32_t segid,
878
                                                  const std::vector<uint32_t>& rowids,
879
                                                  const std::vector<uint32_t>& cids,
880
0
                                                  vectorized::Block& block) {
881
0
    MonotonicStopWatch watch;
882
0
    watch.start();
883
0
    Defer _defer([&]() {
884
0
        LOG_EVERY_N(INFO, 500) << "fetch_value_by_rowids, cost(us):" << watch.elapsed_time() / 1000
885
0
                               << ", row_batch_size:" << rowids.size();
886
0
    });
887
888
0
    BetaRowsetSharedPtr rowset = std::static_pointer_cast<BetaRowset>(input_rowset);
889
0
    CHECK(rowset);
890
0
    CHECK(tablet_schema.has_row_store_for_all_columns());
891
0
    SegmentCacheHandle segment_cache_handle;
892
0
    std::unique_ptr<segment_v2::ColumnIterator> column_iterator;
893
0
    OlapReaderStatistics stats;
894
0
    const auto& column = *DORIS_TRY(tablet_schema.column(BeConsts::ROW_STORE_COL));
895
0
    RETURN_IF_ERROR(_get_segment_column_iterator(rowset, segid, column, &segment_cache_handle,
896
0
                                                 &column_iterator, &stats));
897
    // get and parse tuple row
898
0
    vectorized::MutableColumnPtr column_ptr = vectorized::ColumnString::create();
899
0
    RETURN_IF_ERROR(column_iterator->read_by_rowids(rowids.data(), rowids.size(), column_ptr));
900
0
    assert(column_ptr->size() == rowids.size());
901
0
    auto* string_column = static_cast<vectorized::ColumnString*>(column_ptr.get());
902
0
    vectorized::DataTypeSerDeSPtrs serdes;
903
0
    serdes.resize(cids.size());
904
0
    std::unordered_map<uint32_t, uint32_t> col_uid_to_idx;
905
0
    std::vector<std::string> default_values;
906
0
    default_values.resize(cids.size());
907
0
    for (int i = 0; i < cids.size(); ++i) {
908
0
        const TabletColumn& tablet_column = tablet_schema.column(cids[i]);
909
0
        vectorized::DataTypePtr type =
910
0
                vectorized::DataTypeFactory::instance().create_data_type(tablet_column);
911
0
        col_uid_to_idx[tablet_column.unique_id()] = i;
912
0
        default_values[i] = tablet_column.default_value();
913
0
        serdes[i] = type->get_serde();
914
0
    }
915
0
    RETURN_IF_ERROR(vectorized::JsonbSerializeUtil::jsonb_to_block(
916
0
            serdes, *string_column, col_uid_to_idx, block, default_values, {}));
917
0
    return Status::OK();
918
0
}
919
920
Status BaseTablet::fetch_value_by_rowids(RowsetSharedPtr input_rowset, uint32_t segid,
921
                                         const std::vector<uint32_t>& rowids,
922
                                         const TabletColumn& tablet_column,
923
0
                                         vectorized::MutableColumnPtr& dst) {
924
0
    MonotonicStopWatch watch;
925
0
    watch.start();
926
0
    Defer _defer([&]() {
927
0
        LOG_EVERY_N(INFO, 500) << "fetch_value_by_rowids, cost(us):" << watch.elapsed_time() / 1000
928
0
                               << ", row_batch_size:" << rowids.size();
929
0
    });
930
931
    // read row data
932
0
    BetaRowsetSharedPtr rowset = std::static_pointer_cast<BetaRowset>(input_rowset);
933
0
    CHECK(rowset);
934
0
    SegmentCacheHandle segment_cache_handle;
935
0
    std::unique_ptr<segment_v2::ColumnIterator> column_iterator;
936
0
    OlapReaderStatistics stats;
937
0
    RETURN_IF_ERROR(_get_segment_column_iterator(rowset, segid, tablet_column,
938
0
                                                 &segment_cache_handle, &column_iterator, &stats));
939
0
    RETURN_IF_ERROR(column_iterator->read_by_rowids(rowids.data(), rowids.size(), dst));
940
0
    return Status::OK();
941
0
}
942
943
const signed char* BaseTablet::get_delete_sign_column_data(const vectorized::Block& block,
944
0
                                                           size_t rows_at_least) {
945
0
    if (const vectorized::ColumnWithTypeAndName* delete_sign_column =
946
0
                block.try_get_by_name(DELETE_SIGN);
947
0
        delete_sign_column != nullptr) {
948
0
        const auto& delete_sign_col =
949
0
                reinterpret_cast<const vectorized::ColumnInt8&>(*(delete_sign_column->column));
950
0
        if (delete_sign_col.size() >= rows_at_least) {
951
0
            return delete_sign_col.get_data().data();
952
0
        }
953
0
    }
954
0
    return nullptr;
955
0
};
956
957
Status BaseTablet::generate_default_value_block(const TabletSchema& schema,
958
                                                const std::vector<uint32_t>& cids,
959
                                                const std::vector<std::string>& default_values,
960
                                                const vectorized::Block& ref_block,
961
0
                                                vectorized::Block& default_value_block) {
962
0
    auto mutable_default_value_columns = default_value_block.mutate_columns();
963
0
    for (auto i = 0; i < cids.size(); ++i) {
964
0
        const auto& column = schema.column(cids[i]);
965
0
        if (column.has_default_value()) {
966
0
            const auto& default_value = default_values[i];
967
0
            StringRef str(default_value);
968
0
            RETURN_IF_ERROR(ref_block.get_by_position(i).type->get_serde()->default_from_string(
969
0
                    str, *mutable_default_value_columns[i]));
970
0
        }
971
0
    }
972
0
    default_value_block.set_columns(std::move(mutable_default_value_columns));
973
0
    return Status::OK();
974
0
}
975
976
Status BaseTablet::generate_new_block_for_partial_update(
977
        TabletSchemaSPtr rowset_schema, const PartialUpdateInfo* partial_update_info,
978
        const FixedReadPlan& read_plan_ori, const FixedReadPlan& read_plan_update,
979
        const std::map<RowsetId, RowsetSharedPtr>& rsid_to_rowset,
980
0
        vectorized::Block* output_block) {
981
    // do partial update related works
982
    // 1. read columns by read plan
983
    // 2. generate new block
984
    // 3. write a new segment and modify rowset meta
985
    // 4. mark current keys deleted
986
0
    CHECK(output_block);
987
0
    auto full_mutable_columns = output_block->mutate_columns();
988
0
    const auto& missing_cids = partial_update_info->missing_cids;
989
0
    const auto& update_cids = partial_update_info->update_cids;
990
0
    auto old_block = rowset_schema->create_block_by_cids(missing_cids);
991
0
    auto update_block = rowset_schema->create_block_by_cids(update_cids);
992
993
0
    bool have_input_seq_column = false;
994
0
    if (rowset_schema->has_sequence_col()) {
995
0
        have_input_seq_column =
996
0
                (std::find(update_cids.cbegin(), update_cids.cend(),
997
0
                           rowset_schema->sequence_col_idx()) != update_cids.cend());
998
0
    }
999
1000
    // rowid in the final block(start from 0, increase continuously) -> rowid to read in update_block
1001
0
    std::map<uint32_t, uint32_t> read_index_update;
1002
1003
    // read current rowset first, if a row in the current rowset has delete sign mark
1004
    // we don't need to read values from old block
1005
0
    RETURN_IF_ERROR(read_plan_update.read_columns_by_plan(
1006
0
            *rowset_schema, update_cids, rsid_to_rowset, update_block, &read_index_update, false));
1007
0
    size_t update_rows = read_index_update.size();
1008
0
    for (auto i = 0; i < update_cids.size(); ++i) {
1009
0
        for (auto idx = 0; idx < update_rows; ++idx) {
1010
0
            full_mutable_columns[update_cids[i]]->insert_from(
1011
0
                    *update_block.get_by_position(i).column, read_index_update[idx]);
1012
0
        }
1013
0
    }
1014
1015
    // if there is sequence column in the table, we need to read the sequence column,
1016
    // otherwise it may cause the merge-on-read based compaction policy to produce incorrect results
1017
0
    const auto* __restrict new_block_delete_signs =
1018
0
            rowset_schema->has_sequence_col()
1019
0
                    ? nullptr
1020
0
                    : get_delete_sign_column_data(update_block, update_rows);
1021
1022
    // 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
1023
0
    std::map<uint32_t, uint32_t> read_index_old;
1024
0
    RETURN_IF_ERROR(read_plan_ori.read_columns_by_plan(*rowset_schema, missing_cids, rsid_to_rowset,
1025
0
                                                       old_block, &read_index_old, true,
1026
0
                                                       new_block_delete_signs));
1027
0
    size_t old_rows = read_index_old.size();
1028
0
    const auto* __restrict old_block_delete_signs =
1029
0
            get_delete_sign_column_data(old_block, old_rows);
1030
0
    DCHECK(old_block_delete_signs != nullptr);
1031
    // build default value block
1032
0
    auto default_value_block = old_block.clone_empty();
1033
0
    RETURN_IF_ERROR(BaseTablet::generate_default_value_block(*rowset_schema, missing_cids,
1034
0
                                                             partial_update_info->default_values,
1035
0
                                                             old_block, default_value_block));
1036
1037
0
    CHECK(update_rows >= old_rows);
1038
1039
    // build full block
1040
0
    for (auto i = 0; i < missing_cids.size(); ++i) {
1041
0
        const auto& rs_column = rowset_schema->column(missing_cids[i]);
1042
0
        auto& mutable_column = full_mutable_columns[missing_cids[i]];
1043
0
        for (auto idx = 0; idx < update_rows; ++idx) {
1044
            // There are two cases we don't need to read values from old data:
1045
            //     1. if the conflicting new row's delete sign is marked, which means the value columns
1046
            //     of the row will not be read. So we don't need to read the missing values from the previous rows.
1047
            //     2. if the conflicting old row's delete sign is marked, which means that the key is not exist now,
1048
            //     we should not read old values from the deleted data, and should use default value instead.
1049
            //     NOTE: since now we are in the publishing phase, all data is commited
1050
            //         before, even the `strict_mode` is true (which requires partial update
1051
            //         load job can't insert new keys), this "new" key MUST be written into
1052
            //         the new generated segment file.
1053
0
            bool new_row_delete_sign =
1054
0
                    (new_block_delete_signs != nullptr && new_block_delete_signs[idx]);
1055
0
            if (new_row_delete_sign) {
1056
0
                mutable_column->insert_default();
1057
0
            } else {
1058
0
                bool use_default = false;
1059
0
                bool old_row_delete_sign = (old_block_delete_signs != nullptr &&
1060
0
                                            old_block_delete_signs[read_index_old.at(idx)] != 0);
1061
0
                if (old_row_delete_sign) {
1062
0
                    if (!rowset_schema->has_sequence_col()) {
1063
0
                        use_default = true;
1064
0
                    } else if (have_input_seq_column || !rs_column.is_seqeunce_col()) {
1065
                        // to keep the sequence column value not decreasing, we should read values of seq column
1066
                        // from old rows even if the old row is deleted when the input don't specify the sequence column, otherwise
1067
                        // it may cause the merge-on-read based compaction to produce incorrect results
1068
0
                        use_default = true;
1069
0
                    }
1070
0
                }
1071
1072
0
                if (use_default) {
1073
0
                    if (rs_column.has_default_value()) {
1074
0
                        mutable_column->insert_from(*default_value_block.get_by_position(i).column,
1075
0
                                                    0);
1076
0
                    } else if (rs_column.is_nullable()) {
1077
0
                        assert_cast<vectorized::ColumnNullable*, TypeCheckOnRelease::DISABLE>(
1078
0
                                mutable_column.get())
1079
0
                                ->insert_default();
1080
0
                    } else {
1081
0
                        mutable_column->insert(rs_column.get_vec_type()->get_default());
1082
0
                    }
1083
0
                } else {
1084
0
                    mutable_column->insert_from(*old_block.get_by_position(i).column,
1085
0
                                                read_index_old[idx]);
1086
0
                }
1087
0
            }
1088
0
        }
1089
0
    }
1090
0
    output_block->set_columns(std::move(full_mutable_columns));
1091
0
    VLOG_DEBUG << "full block when publish: " << output_block->dump_data();
1092
0
    return Status::OK();
1093
0
}
1094
1095
Status BaseTablet::generate_new_block_for_flexible_partial_update(
1096
        TabletSchemaSPtr rowset_schema, const PartialUpdateInfo* partial_update_info,
1097
        std::set<uint32_t>& rids_be_overwritten, const FixedReadPlan& read_plan_ori,
1098
        const FixedReadPlan& read_plan_update,
1099
        const std::map<RowsetId, RowsetSharedPtr>& rsid_to_rowset,
1100
0
        vectorized::Block* output_block) {
1101
0
    CHECK(output_block);
1102
1103
0
    int32_t seq_col_unique_id = -1;
1104
0
    if (rowset_schema->has_sequence_col()) {
1105
0
        seq_col_unique_id = rowset_schema->column(rowset_schema->sequence_col_idx()).unique_id();
1106
0
    }
1107
0
    const auto& non_sort_key_cids = partial_update_info->missing_cids;
1108
0
    std::vector<uint32_t> all_cids(rowset_schema->num_columns());
1109
0
    std::iota(all_cids.begin(), all_cids.end(), 0);
1110
0
    auto old_block = rowset_schema->create_block_by_cids(non_sort_key_cids);
1111
0
    auto update_block = rowset_schema->create_block_by_cids(all_cids);
1112
1113
    // rowid in the final block(start from 0, increase continuously) -> rowid to read in update_block
1114
0
    std::map<uint32_t, uint32_t> read_index_update;
1115
1116
    // 1. read the current rowset first, if a row in the current rowset has delete sign mark
1117
    // we don't need to read values from old block for that row
1118
0
    RETURN_IF_ERROR(read_plan_update.read_columns_by_plan(*rowset_schema, all_cids, rsid_to_rowset,
1119
0
                                                          update_block, &read_index_update, true));
1120
0
    size_t update_rows = read_index_update.size();
1121
1122
    // TODO(bobhan1): add the delete sign optimazation here
1123
    // // if there is sequence column in the table, we need to read the sequence column,
1124
    // // otherwise it may cause the merge-on-read based compaction policy to produce incorrect results
1125
    // const auto* __restrict new_block_delete_signs =
1126
    //         rowset_schema->has_sequence_col()
1127
    //                 ? nullptr
1128
    //                 : get_delete_sign_column_data(update_block, update_rows);
1129
1130
    // 2. read previous rowsets
1131
    // 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
1132
0
    std::map<uint32_t, uint32_t> read_index_old;
1133
0
    RETURN_IF_ERROR(read_plan_ori.read_columns_by_plan(
1134
0
            *rowset_schema, non_sort_key_cids, rsid_to_rowset, old_block, &read_index_old, true));
1135
0
    size_t old_rows = read_index_old.size();
1136
0
    DCHECK(update_rows == old_rows);
1137
0
    const auto* __restrict old_block_delete_signs =
1138
0
            get_delete_sign_column_data(old_block, old_rows);
1139
0
    DCHECK(old_block_delete_signs != nullptr);
1140
1141
    // 3. build default value block
1142
0
    auto default_value_block = old_block.clone_empty();
1143
0
    RETURN_IF_ERROR(BaseTablet::generate_default_value_block(*rowset_schema, non_sort_key_cids,
1144
0
                                                             partial_update_info->default_values,
1145
0
                                                             old_block, default_value_block));
1146
1147
    // 4. build the final block
1148
0
    auto full_mutable_columns = output_block->mutate_columns();
1149
0
    DCHECK(rowset_schema->has_skip_bitmap_col());
1150
0
    auto skip_bitmap_col_idx = rowset_schema->skip_bitmap_col_idx();
1151
0
    const std::vector<BitmapValue>* skip_bitmaps =
1152
0
            &(assert_cast<const vectorized::ColumnBitmap*, TypeCheckOnRelease::DISABLE>(
1153
0
                      update_block.get_by_position(skip_bitmap_col_idx).column->get_ptr().get())
1154
0
                      ->get_data());
1155
1156
0
    if (rowset_schema->has_sequence_col() && !rids_be_overwritten.empty()) {
1157
        // If the row specifies the sequence column, we should delete the current row becase the
1158
        // flexible partial update on the current row has been `overwritten` by the previous one with larger sequence
1159
        // column value.
1160
0
        for (auto it = rids_be_overwritten.begin(); it != rids_be_overwritten.end();) {
1161
0
            auto rid = *it;
1162
0
            if (!skip_bitmaps->at(rid).contains(seq_col_unique_id)) {
1163
0
                ++it;
1164
0
            } else {
1165
0
                it = rids_be_overwritten.erase(it);
1166
0
            }
1167
0
        }
1168
0
    }
1169
1170
0
    auto fill_one_cell = [&read_index_old, &read_index_update, &rowset_schema, partial_update_info](
1171
0
                                 const TabletColumn& tablet_column, std::size_t idx,
1172
0
                                 vectorized::MutableColumnPtr& new_col,
1173
0
                                 const vectorized::IColumn& default_value_col,
1174
0
                                 const vectorized::IColumn& old_value_col,
1175
0
                                 const vectorized::IColumn& cur_col, bool skipped,
1176
0
                                 bool row_has_sequence_col,
1177
0
                                 const signed char* delete_sign_column_data) {
1178
0
        if (skipped) {
1179
0
            bool use_default = false;
1180
0
            bool old_row_delete_sign =
1181
0
                    (delete_sign_column_data != nullptr &&
1182
0
                     delete_sign_column_data[read_index_old[cast_set<uint32_t>(idx)]] != 0);
1183
0
            if (old_row_delete_sign) {
1184
0
                if (!rowset_schema->has_sequence_col()) {
1185
0
                    use_default = true;
1186
0
                } else if (row_has_sequence_col ||
1187
0
                           (!tablet_column.is_seqeunce_col() &&
1188
0
                            (tablet_column.unique_id() !=
1189
0
                             partial_update_info->sequence_map_col_uid()))) {
1190
                    // to keep the sequence column value not decreasing, we should read values of seq column(and seq map column)
1191
                    // from old rows even if the old row is deleted when the input don't specify the sequence column, otherwise
1192
                    // it may cause the merge-on-read based compaction to produce incorrect results
1193
0
                    use_default = true;
1194
0
                }
1195
0
            }
1196
0
            if (use_default) {
1197
0
                if (tablet_column.has_default_value()) {
1198
0
                    new_col->insert_from(default_value_col, 0);
1199
0
                } else if (tablet_column.is_nullable()) {
1200
0
                    assert_cast<vectorized::ColumnNullable*, TypeCheckOnRelease::DISABLE>(
1201
0
                            new_col.get())
1202
0
                            ->insert_many_defaults(1);
1203
0
                } else if (tablet_column.is_auto_increment()) {
1204
                    // For auto-increment column, its default value(generated value) is filled in current block in flush phase
1205
                    // when the load doesn't specify the auto-increment column
1206
                    //     - if the previous conflicting row is deleted, we should use the value in current block as its final value
1207
                    //     - if the previous conflicting row is an insert, we should use the value in old block as its final value to
1208
                    //       keep consistency between replicas
1209
0
                    new_col->insert_from(cur_col, read_index_update[cast_set<uint32_t>(idx)]);
1210
0
                } else {
1211
0
                    new_col->insert(tablet_column.get_vec_type()->get_default());
1212
0
                }
1213
0
            } else {
1214
0
                new_col->insert_from(old_value_col, read_index_old[cast_set<uint32_t>(idx)]);
1215
0
            }
1216
0
        } else {
1217
0
            new_col->insert_from(cur_col, read_index_update[cast_set<uint32_t>(idx)]);
1218
0
        }
1219
0
    };
1220
1221
0
    for (std::size_t cid {0}; cid < rowset_schema->num_columns(); cid++) {
1222
0
        vectorized::MutableColumnPtr& new_col = full_mutable_columns[cid];
1223
0
        const vectorized::IColumn& cur_col = *update_block.get_by_position(cid).column;
1224
0
        const auto& rs_column = rowset_schema->column(cid);
1225
0
        auto col_uid = rs_column.unique_id();
1226
0
        for (auto idx = 0; idx < update_rows; ++idx) {
1227
0
            if (cid < rowset_schema->num_key_columns()) {
1228
0
                new_col->insert_from(cur_col, read_index_update[idx]);
1229
0
            } else {
1230
0
                const vectorized::IColumn& default_value_col =
1231
0
                        *default_value_block.get_by_position(cid - rowset_schema->num_key_columns())
1232
0
                                 .column;
1233
0
                const vectorized::IColumn& old_value_col =
1234
0
                        *old_block.get_by_position(cid - rowset_schema->num_key_columns()).column;
1235
0
                if (rids_be_overwritten.contains(idx)) {
1236
0
                    new_col->insert_from(old_value_col, read_index_old[idx]);
1237
0
                } else {
1238
0
                    fill_one_cell(rs_column, idx, new_col, default_value_col, old_value_col,
1239
0
                                  cur_col, skip_bitmaps->at(idx).contains(col_uid),
1240
0
                                  rowset_schema->has_sequence_col()
1241
0
                                          ? !skip_bitmaps->at(idx).contains(seq_col_unique_id)
1242
0
                                          : false,
1243
0
                                  old_block_delete_signs);
1244
0
                }
1245
0
            }
1246
0
        }
1247
0
        DCHECK_EQ(full_mutable_columns[cid]->size(), update_rows);
1248
0
    }
1249
1250
0
    output_block->set_columns(std::move(full_mutable_columns));
1251
0
    VLOG_DEBUG << "full block when publish: " << output_block->dump_data();
1252
0
    return Status::OK();
1253
0
}
1254
1255
Status BaseTablet::commit_phase_update_delete_bitmap(
1256
        const BaseTabletSPtr& tablet, const RowsetSharedPtr& rowset,
1257
        RowsetIdUnorderedSet& pre_rowset_ids, DeleteBitmapPtr delete_bitmap,
1258
        const std::vector<segment_v2::SegmentSharedPtr>& segments, int64_t txn_id,
1259
3
        CalcDeleteBitmapToken* token, RowsetWriter* rowset_writer) {
1260
3
    DBUG_EXECUTE_IF("BaseTablet::commit_phase_update_delete_bitmap.enable_spin_wait", {
1261
3
        auto tok = dp->param<std::string>("token", "invalid_token");
1262
3
        while (DebugPoints::instance()->is_enable(
1263
3
                "BaseTablet::commit_phase_update_delete_bitmap.block")) {
1264
3
            auto block_dp = DebugPoints::instance()->get_debug_point(
1265
3
                    "BaseTablet::commit_phase_update_delete_bitmap.block");
1266
3
            if (block_dp) {
1267
3
                auto pass_token = block_dp->param<std::string>("pass_token", "");
1268
3
                if (pass_token == tok) {
1269
3
                    break;
1270
3
                }
1271
3
            }
1272
3
            std::this_thread::sleep_for(std::chrono::milliseconds(50));
1273
3
        }
1274
3
    });
1275
3
    SCOPED_BVAR_LATENCY(g_tablet_commit_phase_update_delete_bitmap_latency);
1276
3
    RowsetIdUnorderedSet cur_rowset_ids;
1277
3
    RowsetIdUnorderedSet rowset_ids_to_add;
1278
3
    RowsetIdUnorderedSet rowset_ids_to_del;
1279
3
    int64_t cur_version;
1280
1281
3
    std::vector<RowsetSharedPtr> specified_rowsets;
1282
3
    {
1283
        // to prevent seeing intermediate state of a tablet
1284
3
        std::unique_lock<bthread::Mutex> sync_lock;
1285
3
        if (config::is_cloud_mode()) {
1286
0
            sync_lock = std::unique_lock<bthread::Mutex>(
1287
0
                    std::static_pointer_cast<CloudTablet>(tablet)->get_sync_meta_lock());
1288
0
        }
1289
3
        std::shared_lock meta_rlock(tablet->_meta_lock);
1290
3
        if (tablet->tablet_state() == TABLET_NOTREADY) {
1291
            // tablet is under alter process. The delete bitmap will be calculated after conversion.
1292
0
            LOG(INFO) << "tablet is under alter process, delete bitmap will be calculated later, "
1293
0
                         "tablet_id: "
1294
0
                      << tablet->tablet_id() << " txn_id: " << txn_id;
1295
0
            return Status::OK();
1296
0
        }
1297
3
        cur_version = tablet->max_version_unlocked();
1298
3
        RETURN_IF_ERROR(tablet->get_all_rs_id_unlocked(cur_version, &cur_rowset_ids));
1299
3
        _rowset_ids_difference(cur_rowset_ids, pre_rowset_ids, &rowset_ids_to_add,
1300
3
                               &rowset_ids_to_del);
1301
3
        specified_rowsets = tablet->get_rowset_by_ids(&rowset_ids_to_add);
1302
3
    }
1303
0
    for (const auto& to_del : rowset_ids_to_del) {
1304
0
        delete_bitmap->remove({to_del, 0, 0}, {to_del, UINT32_MAX, INT64_MAX});
1305
0
    }
1306
1307
3
    RETURN_IF_ERROR(calc_delete_bitmap(tablet, rowset, segments, specified_rowsets, delete_bitmap,
1308
3
                                       cur_version, token, rowset_writer));
1309
3
    size_t total_rows = std::accumulate(
1310
3
            segments.begin(), segments.end(), 0,
1311
3
            [](size_t sum, const segment_v2::SegmentSharedPtr& s) { return sum += s->num_rows(); });
1312
3
    LOG(INFO) << "[Before Commit] construct delete bitmap tablet: " << tablet->tablet_id()
1313
3
              << ", rowset_ids to add: " << rowset_ids_to_add.size()
1314
3
              << ", rowset_ids to del: " << rowset_ids_to_del.size()
1315
3
              << ", cur max_version: " << cur_version << ", transaction_id: " << txn_id
1316
3
              << ", total rows: " << total_rows;
1317
3
    pre_rowset_ids = cur_rowset_ids;
1318
3
    return Status::OK();
1319
3
}
1320
1321
void BaseTablet::add_sentinel_mark_to_delete_bitmap(DeleteBitmap* delete_bitmap,
1322
5
                                                    const RowsetIdUnorderedSet& rowsetids) {
1323
5
    for (const auto& rowsetid : rowsetids) {
1324
5
        delete_bitmap->add(
1325
5
                {rowsetid, DeleteBitmap::INVALID_SEGMENT_ID, DeleteBitmap::TEMP_VERSION_COMMON},
1326
5
                DeleteBitmap::ROWSET_SENTINEL_MARK);
1327
5
    }
1328
5
}
1329
1330
void BaseTablet::_rowset_ids_difference(const RowsetIdUnorderedSet& cur,
1331
                                        const RowsetIdUnorderedSet& pre,
1332
                                        RowsetIdUnorderedSet* to_add,
1333
6
                                        RowsetIdUnorderedSet* to_del) {
1334
6
    for (const auto& id : cur) {
1335
2
        if (pre.find(id) == pre.end()) {
1336
1
            to_add->insert(id);
1337
1
        }
1338
2
    }
1339
6
    for (const auto& id : pre) {
1340
1
        if (cur.find(id) == cur.end()) {
1341
0
            to_del->insert(id);
1342
0
        }
1343
1
    }
1344
6
}
1345
1346
Status BaseTablet::_capture_consistent_rowsets_unlocked(
1347
4
        const std::vector<Version>& version_path, std::vector<RowsetSharedPtr>* rowsets) const {
1348
4
    DCHECK(rowsets != nullptr);
1349
4
    rowsets->reserve(version_path.size());
1350
8
    for (const auto& version : version_path) {
1351
8
        bool is_find = false;
1352
8
        do {
1353
8
            auto it = _rs_version_map.find(version);
1354
8
            if (it != _rs_version_map.end()) {
1355
8
                is_find = true;
1356
8
                rowsets->push_back(it->second);
1357
8
                break;
1358
8
            }
1359
1360
0
            auto it_expired = _stale_rs_version_map.find(version);
1361
0
            if (it_expired != _stale_rs_version_map.end()) {
1362
0
                is_find = true;
1363
0
                rowsets->push_back(it_expired->second);
1364
0
                break;
1365
0
            }
1366
0
        } while (false);
1367
1368
8
        if (!is_find) {
1369
0
            return Status::Error<CAPTURE_ROWSET_ERROR>(
1370
0
                    "fail to find Rowset for version. tablet={}, version={}", tablet_id(),
1371
0
                    version.to_string());
1372
0
        }
1373
8
    }
1374
4
    return Status::OK();
1375
4
}
1376
1377
Status BaseTablet::check_delete_bitmap_correctness(DeleteBitmapPtr delete_bitmap,
1378
                                                   int64_t max_version, int64_t txn_id,
1379
                                                   const RowsetIdUnorderedSet& rowset_ids,
1380
6
                                                   std::vector<RowsetSharedPtr>* rowsets) {
1381
6
    RowsetIdUnorderedSet missing_ids;
1382
6
    for (const auto& rowsetid : rowset_ids) {
1383
2
        if (!delete_bitmap->delete_bitmap.contains({rowsetid, DeleteBitmap::INVALID_SEGMENT_ID,
1384
2
                                                    DeleteBitmap::TEMP_VERSION_COMMON})) {
1385
0
            missing_ids.insert(rowsetid);
1386
0
        }
1387
2
    }
1388
1389
6
    if (!missing_ids.empty()) {
1390
0
        LOG(WARNING) << "[txn_id:" << txn_id << "][tablet_id:" << tablet_id()
1391
0
                     << "][max_version: " << max_version
1392
0
                     << "] check delete bitmap correctness failed!";
1393
0
        rapidjson::Document root;
1394
0
        root.SetObject();
1395
0
        rapidjson::Document required_rowsets_arr;
1396
0
        required_rowsets_arr.SetArray();
1397
0
        rapidjson::Document missing_rowsets_arr;
1398
0
        missing_rowsets_arr.SetArray();
1399
1400
0
        if (rowsets != nullptr) {
1401
0
            for (const auto& rowset : *rowsets) {
1402
0
                rapidjson::Value value;
1403
0
                std::string version_str = rowset->get_rowset_info_str();
1404
0
                value.SetString(version_str.c_str(),
1405
0
                                cast_set<rapidjson::SizeType>(version_str.length()),
1406
0
                                required_rowsets_arr.GetAllocator());
1407
0
                required_rowsets_arr.PushBack(value, required_rowsets_arr.GetAllocator());
1408
0
            }
1409
0
        } else {
1410
0
            std::vector<RowsetSharedPtr> tablet_rowsets;
1411
0
            {
1412
0
                std::shared_lock meta_rlock(_meta_lock);
1413
0
                tablet_rowsets = get_rowset_by_ids(&rowset_ids);
1414
0
            }
1415
0
            for (const auto& rowset : tablet_rowsets) {
1416
0
                rapidjson::Value value;
1417
0
                std::string version_str = rowset->get_rowset_info_str();
1418
0
                value.SetString(version_str.c_str(),
1419
0
                                cast_set<rapidjson::SizeType>(version_str.length()),
1420
0
                                required_rowsets_arr.GetAllocator());
1421
0
                required_rowsets_arr.PushBack(value, required_rowsets_arr.GetAllocator());
1422
0
            }
1423
0
        }
1424
0
        for (const auto& missing_rowset_id : missing_ids) {
1425
0
            rapidjson::Value miss_value;
1426
0
            std::string rowset_id_str = missing_rowset_id.to_string();
1427
0
            miss_value.SetString(rowset_id_str.c_str(),
1428
0
                                 cast_set<rapidjson::SizeType>(rowset_id_str.length()),
1429
0
                                 missing_rowsets_arr.GetAllocator());
1430
0
            missing_rowsets_arr.PushBack(miss_value, missing_rowsets_arr.GetAllocator());
1431
0
        }
1432
1433
0
        root.AddMember("required_rowsets", required_rowsets_arr, root.GetAllocator());
1434
0
        root.AddMember("missing_rowsets", missing_rowsets_arr, root.GetAllocator());
1435
0
        rapidjson::StringBuffer strbuf;
1436
0
        rapidjson::PrettyWriter<rapidjson::StringBuffer> writer(strbuf);
1437
0
        root.Accept(writer);
1438
0
        std::string rowset_status_string = std::string(strbuf.GetString());
1439
0
        LOG_EVERY_SECOND(WARNING) << rowset_status_string;
1440
        // let it crash if correctness check failed in Debug mode
1441
0
        DCHECK(false) << "delete bitmap correctness check failed in publish phase!";
1442
0
        return Status::InternalError("check delete bitmap failed!");
1443
0
    }
1444
6
    return Status::OK();
1445
6
}
1446
1447
Status BaseTablet::update_delete_bitmap(const BaseTabletSPtr& self, TabletTxnInfo* txn_info,
1448
                                        int64_t txn_id, int64_t txn_expiration,
1449
3
                                        DeleteBitmapPtr tablet_delete_bitmap) {
1450
3
    SCOPED_BVAR_LATENCY(g_tablet_update_delete_bitmap_latency);
1451
3
    RowsetIdUnorderedSet cur_rowset_ids;
1452
3
    RowsetIdUnorderedSet rowset_ids_to_add;
1453
3
    RowsetIdUnorderedSet rowset_ids_to_del;
1454
3
    RowsetSharedPtr rowset = txn_info->rowset;
1455
3
    int64_t cur_version = rowset->start_version();
1456
1457
3
    std::unique_ptr<RowsetWriter> transient_rs_writer;
1458
3
    DeleteBitmapPtr delete_bitmap = txn_info->delete_bitmap;
1459
3
    bool is_partial_update =
1460
3
            txn_info->partial_update_info && txn_info->partial_update_info->is_partial_update();
1461
3
    if (is_partial_update) {
1462
0
        transient_rs_writer = DORIS_TRY(self->create_transient_rowset_writer(
1463
0
                *rowset, txn_info->partial_update_info, txn_expiration));
1464
0
        DBUG_EXECUTE_IF("BaseTablet::update_delete_bitmap.after.create_transient_rs_writer",
1465
0
                        DBUG_BLOCK);
1466
        // Partial update might generate new segments when there is conflicts while publish, and mark
1467
        // the same key in original segments as delete.
1468
        // When the new segment flush fails or the rowset build fails, the deletion marker for the
1469
        // duplicate key of the original segment should not remain in `txn_info->delete_bitmap`,
1470
        // so we need to make a copy of `txn_info->delete_bitmap` and make changes on it.
1471
0
        delete_bitmap = std::make_shared<DeleteBitmap>(*(txn_info->delete_bitmap));
1472
0
    }
1473
1474
3
    OlapStopWatch watch;
1475
3
    std::vector<segment_v2::SegmentSharedPtr> segments;
1476
3
    RETURN_IF_ERROR(std::dynamic_pointer_cast<BetaRowset>(rowset)->load_segments(&segments));
1477
3
    auto t1 = watch.get_elapse_time_us();
1478
1479
3
    int64_t next_visible_version = txn_info->is_txn_load ? txn_info->next_visible_version
1480
3
                                                         : txn_info->rowset->start_version();
1481
3
    {
1482
3
        std::shared_lock meta_rlock(self->_meta_lock);
1483
        // tablet is under alter process. The delete bitmap will be calculated after conversion.
1484
3
        if (self->tablet_state() == TABLET_NOTREADY) {
1485
0
            LOG(INFO) << "tablet is under alter process, update delete bitmap later, tablet_id="
1486
0
                      << self->tablet_id();
1487
0
            return Status::OK();
1488
0
        }
1489
3
        RETURN_IF_ERROR(self->get_all_rs_id_unlocked(next_visible_version - 1, &cur_rowset_ids));
1490
3
    }
1491
3
    auto t2 = watch.get_elapse_time_us();
1492
1493
3
    _rowset_ids_difference(cur_rowset_ids, txn_info->rowset_ids, &rowset_ids_to_add,
1494
3
                           &rowset_ids_to_del);
1495
3
    for (const auto& to_del : rowset_ids_to_del) {
1496
0
        delete_bitmap->remove({to_del, 0, 0}, {to_del, UINT32_MAX, INT64_MAX});
1497
0
    }
1498
1499
3
    std::vector<RowsetSharedPtr> specified_rowsets;
1500
3
    {
1501
3
        std::shared_lock meta_rlock(self->_meta_lock);
1502
3
        specified_rowsets = self->get_rowset_by_ids(&rowset_ids_to_add);
1503
3
    }
1504
3
    if (txn_info->is_txn_load) {
1505
0
        for (auto invisible_rowset : txn_info->invisible_rowsets) {
1506
0
            specified_rowsets.emplace_back(invisible_rowset);
1507
0
        }
1508
0
        std::sort(specified_rowsets.begin(), specified_rowsets.end(),
1509
0
                  [](RowsetSharedPtr& lhs, RowsetSharedPtr& rhs) {
1510
0
                      return lhs->end_version() > rhs->end_version();
1511
0
                  });
1512
0
    }
1513
3
    auto t3 = watch.get_elapse_time_us();
1514
1515
    // If a rowset is produced by compaction before the commit phase of the partial update load
1516
    // and is not included in txn_info->rowset_ids, we can skip the alignment process of that rowset
1517
    // because data remains the same before and after compaction. But we still need to calculate the
1518
    // the delete bitmap for that rowset.
1519
3
    std::vector<RowsetSharedPtr> rowsets_skip_alignment;
1520
3
    if (is_partial_update) {
1521
0
        int64_t max_version_in_flush_phase =
1522
0
                txn_info->partial_update_info->max_version_in_flush_phase;
1523
0
        DCHECK(max_version_in_flush_phase != -1);
1524
0
        std::vector<RowsetSharedPtr> remained_rowsets;
1525
0
        for (const auto& specified_rowset : specified_rowsets) {
1526
0
            if (specified_rowset->end_version() <= max_version_in_flush_phase &&
1527
0
                specified_rowset->produced_by_compaction()) {
1528
0
                rowsets_skip_alignment.emplace_back(specified_rowset);
1529
0
            } else {
1530
0
                remained_rowsets.emplace_back(specified_rowset);
1531
0
            }
1532
0
        }
1533
0
        if (!rowsets_skip_alignment.empty()) {
1534
0
            specified_rowsets = std::move(remained_rowsets);
1535
0
        }
1536
0
    }
1537
1538
3
    DBUG_EXECUTE_IF("BaseTablet::update_delete_bitmap.enable_spin_wait", {
1539
3
        auto token = dp->param<std::string>("token", "invalid_token");
1540
3
        while (DebugPoints::instance()->is_enable("BaseTablet::update_delete_bitmap.block")) {
1541
3
            auto block_dp = DebugPoints::instance()->get_debug_point(
1542
3
                    "BaseTablet::update_delete_bitmap.block");
1543
3
            if (block_dp) {
1544
3
                auto wait_token = block_dp->param<std::string>("wait_token", "");
1545
3
                if (wait_token != token) {
1546
3
                    break;
1547
3
                }
1548
3
            }
1549
3
            std::this_thread::sleep_for(std::chrono::milliseconds(50));
1550
3
        }
1551
3
    });
1552
1553
3
    if (!rowsets_skip_alignment.empty()) {
1554
0
        auto token = self->calc_delete_bitmap_executor()->create_token();
1555
        // set rowset_writer to nullptr to skip the alignment process
1556
0
        RETURN_IF_ERROR(calc_delete_bitmap(self, rowset, segments, rowsets_skip_alignment,
1557
0
                                           delete_bitmap, cur_version - 1, token.get(), nullptr,
1558
0
                                           tablet_delete_bitmap));
1559
0
        RETURN_IF_ERROR(token->wait());
1560
0
    }
1561
1562
    // When there is only one segment, it will be calculated in the current thread.
1563
    // Otherwise, it will be submitted to the thread pool for calculation.
1564
3
    if (segments.size() <= 1) {
1565
3
        RETURN_IF_ERROR(calc_delete_bitmap(self, rowset, segments, specified_rowsets, delete_bitmap,
1566
3
                                           cur_version - 1, nullptr, transient_rs_writer.get(),
1567
3
                                           tablet_delete_bitmap));
1568
1569
3
    } else {
1570
0
        auto token = self->calc_delete_bitmap_executor()->create_token();
1571
0
        RETURN_IF_ERROR(calc_delete_bitmap(self, rowset, segments, specified_rowsets, delete_bitmap,
1572
0
                                           cur_version - 1, token.get(), transient_rs_writer.get(),
1573
0
                                           tablet_delete_bitmap));
1574
0
        RETURN_IF_ERROR(token->wait());
1575
0
    }
1576
1577
3
    std::stringstream ss;
1578
3
    ss << "cost(us): (load segments: " << t1 << ", get all rsid: " << t2 - t1
1579
3
       << ", get rowsets: " << t3 - t2
1580
3
       << ", calc delete bitmap: " << watch.get_elapse_time_us() - t3 << ")";
1581
1582
3
    if (config::enable_merge_on_write_correctness_check && rowset->num_rows() != 0) {
1583
        // only do correctness check if the rowset has at least one row written
1584
        // check if all the rowset has ROWSET_SENTINEL_MARK
1585
3
        auto st = self->check_delete_bitmap_correctness(delete_bitmap, cur_version - 1, -1,
1586
3
                                                        cur_rowset_ids, &specified_rowsets);
1587
3
        if (!st.ok()) {
1588
0
            LOG(WARNING) << fmt::format("delete bitmap correctness check failed in publish phase!");
1589
0
        }
1590
3
    }
1591
1592
3
    if (transient_rs_writer) {
1593
0
        auto t4 = watch.get_elapse_time_us();
1594
0
        DBUG_EXECUTE_IF("Tablet.update_delete_bitmap.partial_update_write_rowset_fail", {
1595
0
            if (rand() % 100 < (100 * dp->param("percent", 0.5))) {
1596
0
                LOG_WARNING("Tablet.update_delete_bitmap.partial_update_write_rowset random failed")
1597
0
                        .tag("txn_id", txn_id);
1598
0
                return Status::InternalError(
1599
0
                        "debug update_delete_bitmap partial update write rowset random failed");
1600
0
            }
1601
0
        });
1602
        // build rowset writer and merge transient rowset
1603
0
        RETURN_IF_ERROR(transient_rs_writer->flush());
1604
0
        RowsetSharedPtr transient_rowset;
1605
0
        RETURN_IF_ERROR(transient_rs_writer->build(transient_rowset));
1606
0
        auto old_segments = rowset->num_segments();
1607
0
        rowset->merge_rowset_meta(*transient_rowset->rowset_meta());
1608
0
        auto new_segments = rowset->num_segments();
1609
0
        ss << ", " << txn_info->partial_update_info->partial_update_mode_str()
1610
0
           << " flush rowset (old segment num: " << old_segments
1611
0
           << ", new segment num: " << new_segments << ")"
1612
0
           << ", cost:" << watch.get_elapse_time_us() - t4 << "(us)";
1613
1614
        // update the shared_ptr to new bitmap, which is consistent with current rowset.
1615
0
        txn_info->delete_bitmap = delete_bitmap;
1616
        // erase segment cache cause we will add a segment to rowset
1617
0
        SegmentLoader::instance()->erase_segments(rowset->rowset_id(), rowset->num_segments());
1618
0
    }
1619
1620
3
    size_t total_rows = std::accumulate(
1621
3
            segments.begin(), segments.end(), 0,
1622
3
            [](size_t sum, const segment_v2::SegmentSharedPtr& s) { return sum += s->num_rows(); });
1623
3
    auto t5 = watch.get_elapse_time_us();
1624
3
    int64_t lock_id = txn_info->is_txn_load ? txn_info->lock_id : -1;
1625
3
    RETURN_IF_ERROR(self->save_delete_bitmap(txn_info, txn_id, delete_bitmap,
1626
3
                                             transient_rs_writer.get(), cur_rowset_ids, lock_id,
1627
3
                                             next_visible_version));
1628
1629
    // defensive check, check that the delete bitmap cache we wrote is correct
1630
3
    RETURN_IF_ERROR(self->check_delete_bitmap_cache(txn_id, delete_bitmap.get()));
1631
1632
3
    LOG(INFO) << "[Publish] construct delete bitmap tablet: " << self->tablet_id()
1633
3
              << ", rowset_ids to add: "
1634
3
              << (specified_rowsets.size() + rowsets_skip_alignment.size())
1635
3
              << ", rowset_ids to del: " << rowset_ids_to_del.size()
1636
3
              << ", cur version: " << cur_version << ", transaction_id: " << txn_id << ","
1637
3
              << ss.str() << " , total rows: " << total_rows
1638
3
              << ", update delete_bitmap cost: " << watch.get_elapse_time_us() - t5 << "(us)";
1639
3
    return Status::OK();
1640
3
}
1641
1642
void BaseTablet::calc_compaction_output_rowset_delete_bitmap(
1643
        const std::vector<RowsetSharedPtr>& input_rowsets, const RowIdConversion& rowid_conversion,
1644
        uint64_t start_version, uint64_t end_version, std::set<RowLocation>* missed_rows,
1645
        std::map<RowsetSharedPtr, std::list<std::pair<RowLocation, RowLocation>>>* location_map,
1646
0
        const DeleteBitmap& input_delete_bitmap, DeleteBitmap* output_rowset_delete_bitmap) {
1647
0
    RowLocation src;
1648
0
    RowLocation dst;
1649
0
    for (auto& rowset : input_rowsets) {
1650
0
        src.rowset_id = rowset->rowset_id();
1651
0
        for (uint32_t seg_id = 0; seg_id < rowset->num_segments(); ++seg_id) {
1652
0
            src.segment_id = seg_id;
1653
0
            DeleteBitmap subset_map(tablet_id());
1654
0
            input_delete_bitmap.subset({rowset->rowset_id(), seg_id, start_version},
1655
0
                                       {rowset->rowset_id(), seg_id, end_version}, &subset_map);
1656
            // traverse all versions and convert rowid
1657
0
            for (auto iter = subset_map.delete_bitmap.begin();
1658
0
                 iter != subset_map.delete_bitmap.end(); ++iter) {
1659
0
                auto cur_version = std::get<2>(iter->first);
1660
0
                for (auto index = iter->second.begin(); index != iter->second.end(); ++index) {
1661
0
                    src.row_id = *index;
1662
0
                    if (rowid_conversion.get(src, &dst) != 0) {
1663
0
                        VLOG_CRITICAL << "Can't find rowid, may be deleted by the delete_handler, "
1664
0
                                      << " src loaction: |" << src.rowset_id << "|"
1665
0
                                      << src.segment_id << "|" << src.row_id
1666
0
                                      << " version: " << cur_version;
1667
0
                        if (missed_rows) {
1668
0
                            missed_rows->insert(src);
1669
0
                        }
1670
0
                        continue;
1671
0
                    }
1672
0
                    VLOG_DEBUG << "calc_compaction_output_rowset_delete_bitmap dst location: |"
1673
0
                               << dst.rowset_id << "|" << dst.segment_id << "|" << dst.row_id
1674
0
                               << " src location: |" << src.rowset_id << "|" << src.segment_id
1675
0
                               << "|" << src.row_id << " start version: " << start_version
1676
0
                               << "end version" << end_version;
1677
0
                    if (location_map) {
1678
0
                        (*location_map)[rowset].emplace_back(src, dst);
1679
0
                    }
1680
0
                    output_rowset_delete_bitmap->add({dst.rowset_id, dst.segment_id, cur_version},
1681
0
                                                     dst.row_id);
1682
0
                }
1683
0
            }
1684
0
        }
1685
0
    }
1686
0
}
1687
1688
Status BaseTablet::check_rowid_conversion(
1689
        RowsetSharedPtr dst_rowset,
1690
        const std::map<RowsetSharedPtr, std::list<std::pair<RowLocation, RowLocation>>>&
1691
0
                location_map) {
1692
0
    if (location_map.empty()) {
1693
0
        VLOG_DEBUG << "check_rowid_conversion, location_map is empty";
1694
0
        return Status::OK();
1695
0
    }
1696
0
    std::vector<segment_v2::SegmentSharedPtr> dst_segments;
1697
1698
0
    RETURN_IF_ERROR(
1699
0
            std::dynamic_pointer_cast<BetaRowset>(dst_rowset)->load_segments(&dst_segments));
1700
0
    std::unordered_map<RowsetId, std::vector<segment_v2::SegmentSharedPtr>> input_rowsets_segment;
1701
1702
0
    VLOG_DEBUG << "check_rowid_conversion, dst_segments size: " << dst_segments.size();
1703
0
    for (auto [src_rowset, locations] : location_map) {
1704
0
        std::vector<segment_v2::SegmentSharedPtr>& segments =
1705
0
                input_rowsets_segment[src_rowset->rowset_id()];
1706
0
        if (segments.empty()) {
1707
0
            RETURN_IF_ERROR(
1708
0
                    std::dynamic_pointer_cast<BetaRowset>(src_rowset)->load_segments(&segments));
1709
0
        }
1710
0
        for (auto& [src, dst] : locations) {
1711
0
            std::string src_key;
1712
0
            std::string dst_key;
1713
0
            Status s = segments[src.segment_id]->read_key_by_rowid(src.row_id, &src_key);
1714
0
            if (UNLIKELY(s.is<NOT_IMPLEMENTED_ERROR>())) {
1715
0
                LOG(INFO) << "primary key index of old version does not "
1716
0
                             "support reading key by rowid";
1717
0
                break;
1718
0
            }
1719
0
            if (UNLIKELY(!s)) {
1720
0
                LOG(WARNING) << "failed to get src key: |" << src.rowset_id << "|" << src.segment_id
1721
0
                             << "|" << src.row_id << " status: " << s;
1722
0
                DCHECK(false);
1723
0
                return s;
1724
0
            }
1725
1726
0
            s = dst_segments[dst.segment_id]->read_key_by_rowid(dst.row_id, &dst_key);
1727
0
            if (UNLIKELY(!s)) {
1728
0
                LOG(WARNING) << "failed to get dst key: |" << dst.rowset_id << "|" << dst.segment_id
1729
0
                             << "|" << dst.row_id << " status: " << s;
1730
0
                DCHECK(false);
1731
0
                return s;
1732
0
            }
1733
1734
0
            VLOG_DEBUG << "check_rowid_conversion, src: |" << src.rowset_id << "|" << src.segment_id
1735
0
                       << "|" << src.row_id << "|" << src_key << " dst: |" << dst.rowset_id << "|"
1736
0
                       << dst.segment_id << "|" << dst.row_id << "|" << dst_key;
1737
0
            if (UNLIKELY(src_key.compare(dst_key) != 0)) {
1738
0
                LOG(WARNING) << "failed to check key, src key: |" << src.rowset_id << "|"
1739
0
                             << src.segment_id << "|" << src.row_id << "|" << src_key
1740
0
                             << " dst key: |" << dst.rowset_id << "|" << dst.segment_id << "|"
1741
0
                             << dst.row_id << "|" << dst_key;
1742
0
                DCHECK(false);
1743
0
                return Status::InternalError("failed to check rowid conversion");
1744
0
            }
1745
0
        }
1746
0
    }
1747
0
    return Status::OK();
1748
0
}
1749
1750
// The caller should hold _rowset_update_lock and _meta_lock lock.
1751
Status BaseTablet::update_delete_bitmap_without_lock(
1752
        const BaseTabletSPtr& self, const RowsetSharedPtr& rowset,
1753
0
        const std::vector<RowsetSharedPtr>* specified_base_rowsets) {
1754
0
    DBUG_EXECUTE_IF("BaseTablet.update_delete_bitmap_without_lock.random_failed", {
1755
0
        auto rnd = rand() % 100;
1756
0
        auto percent = dp->param("percent", 0.1);
1757
0
        if (rnd < (100 * percent)) {
1758
0
            LOG(WARNING) << "BaseTablet.update_delete_bitmap_without_lock.random_failed";
1759
0
            return Status::InternalError(
1760
0
                    "debug tablet update delete bitmap without lock random failed");
1761
0
        } else {
1762
0
            LOG(INFO) << "BaseTablet.update_delete_bitmap_without_lock.random_failed not "
1763
0
                         "triggered"
1764
0
                      << ", rnd:" << rnd << ", percent: " << percent;
1765
0
        }
1766
0
    });
1767
0
    int64_t cur_version = rowset->start_version();
1768
0
    std::vector<segment_v2::SegmentSharedPtr> segments;
1769
0
    RETURN_IF_ERROR(std::dynamic_pointer_cast<BetaRowset>(rowset)->load_segments(&segments));
1770
1771
    // If this rowset does not have a segment, there is no need for an update.
1772
0
    if (segments.empty()) {
1773
0
        LOG(INFO) << "[Schema Change or Clone] skip to construct delete bitmap tablet: "
1774
0
                  << self->tablet_id() << " cur max_version: " << cur_version;
1775
0
        return Status::OK();
1776
0
    }
1777
1778
    // calculate delete bitmap between segments if necessary.
1779
0
    DeleteBitmapPtr delete_bitmap = std::make_shared<DeleteBitmap>(self->tablet_id());
1780
0
    RETURN_IF_ERROR(self->calc_delete_bitmap_between_segments(
1781
0
            rowset->tablet_schema(), rowset->rowset_id(), segments, delete_bitmap));
1782
1783
    // get all base rowsets to calculate on
1784
0
    std::vector<RowsetSharedPtr> specified_rowsets;
1785
0
    RowsetIdUnorderedSet cur_rowset_ids;
1786
0
    if (specified_base_rowsets == nullptr) {
1787
0
        RETURN_IF_ERROR(self->get_all_rs_id_unlocked(cur_version - 1, &cur_rowset_ids));
1788
0
        specified_rowsets = self->get_rowset_by_ids(&cur_rowset_ids);
1789
0
    } else {
1790
0
        specified_rowsets = *specified_base_rowsets;
1791
0
    }
1792
1793
0
    OlapStopWatch watch;
1794
0
    auto token = self->calc_delete_bitmap_executor()->create_token();
1795
0
    RETURN_IF_ERROR(calc_delete_bitmap(self, rowset, segments, specified_rowsets, delete_bitmap,
1796
0
                                       cur_version - 1, token.get()));
1797
0
    RETURN_IF_ERROR(token->wait());
1798
0
    size_t total_rows = std::accumulate(
1799
0
            segments.begin(), segments.end(), 0,
1800
0
            [](size_t sum, const segment_v2::SegmentSharedPtr& s) { return sum += s->num_rows(); });
1801
0
    LOG(INFO) << "[Schema Change or Clone] construct delete bitmap tablet: " << self->tablet_id()
1802
0
              << ", rowset_ids: " << cur_rowset_ids.size() << ", cur max_version: " << cur_version
1803
0
              << ", transaction_id: " << -1 << ", cost: " << watch.get_elapse_time_us()
1804
0
              << "(us), total rows: " << total_rows;
1805
0
    if (config::enable_merge_on_write_correctness_check) {
1806
        // check if all the rowset has ROWSET_SENTINEL_MARK
1807
0
        auto st = self->check_delete_bitmap_correctness(delete_bitmap, cur_version - 1, -1,
1808
0
                                                        cur_rowset_ids, &specified_rowsets);
1809
0
        if (!st.ok()) {
1810
0
            LOG(WARNING) << fmt::format("delete bitmap correctness check failed in publish phase!");
1811
0
        }
1812
0
        delete_bitmap->remove_sentinel_marks();
1813
0
    }
1814
0
    for (auto& iter : delete_bitmap->delete_bitmap) {
1815
0
        self->_tablet_meta->delete_bitmap().merge(
1816
0
                {std::get<0>(iter.first), std::get<1>(iter.first), cur_version}, iter.second);
1817
0
    }
1818
1819
0
    return Status::OK();
1820
0
}
1821
1822
void BaseTablet::agg_delete_bitmap_for_stale_rowsets(
1823
0
        Version version, DeleteBitmapKeyRanges& remove_delete_bitmap_key_ranges) {
1824
0
    if (!config::enable_agg_and_remove_pre_rowsets_delete_bitmap) {
1825
0
        return;
1826
0
    }
1827
0
    if (!(keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write())) {
1828
0
        return;
1829
0
    }
1830
0
    int64_t start_version = version.first;
1831
0
    int64_t end_version = version.second;
1832
0
    if (start_version == end_version) {
1833
0
        return;
1834
0
    }
1835
0
    DCHECK(start_version < end_version)
1836
0
            << ". start_version: " << start_version << ", end_version: " << end_version;
1837
    // get pre rowsets
1838
0
    std::vector<RowsetSharedPtr> pre_rowsets {};
1839
0
    {
1840
0
        std::shared_lock rdlock(_meta_lock);
1841
0
        for (const auto& it2 : _rs_version_map) {
1842
0
            if (it2.first.second < start_version) {
1843
0
                pre_rowsets.emplace_back(it2.second);
1844
0
            }
1845
0
        }
1846
0
    }
1847
0
    std::sort(pre_rowsets.begin(), pre_rowsets.end(), Rowset::comparator);
1848
    // do agg for pre rowsets
1849
0
    DeleteBitmapPtr new_delete_bitmap = std::make_shared<DeleteBitmap>(tablet_id());
1850
0
    for (auto& rowset : pre_rowsets) {
1851
0
        for (uint32_t seg_id = 0; seg_id < rowset->num_segments(); ++seg_id) {
1852
0
            auto d = tablet_meta()->delete_bitmap().get_agg_without_cache(
1853
0
                    {rowset->rowset_id(), seg_id, end_version}, start_version);
1854
0
            if (d->isEmpty()) {
1855
0
                continue;
1856
0
            }
1857
0
            VLOG_DEBUG << "agg delete bitmap for tablet_id=" << tablet_id()
1858
0
                       << ", rowset_id=" << rowset->rowset_id() << ", seg_id=" << seg_id
1859
0
                       << ", rowset_version=" << rowset->version().to_string()
1860
0
                       << ". compaction start_version=" << start_version
1861
0
                       << ", end_version=" << end_version << ", delete_bitmap=" << d->cardinality();
1862
0
            DeleteBitmap::BitmapKey start_key {rowset->rowset_id(), seg_id, start_version};
1863
0
            DeleteBitmap::BitmapKey end_key {rowset->rowset_id(), seg_id, end_version};
1864
0
            new_delete_bitmap->set(end_key, *d);
1865
0
            remove_delete_bitmap_key_ranges.emplace_back(start_key, end_key);
1866
0
        }
1867
0
    }
1868
0
    DBUG_EXECUTE_IF("BaseTablet.agg_delete_bitmap_for_stale_rowsets.merge_delete_bitmap.block",
1869
0
                    DBUG_BLOCK);
1870
0
    tablet_meta()->delete_bitmap().merge(*new_delete_bitmap);
1871
0
}
1872
1873
void BaseTablet::check_agg_delete_bitmap_for_stale_rowsets(int64_t& useless_rowset_count,
1874
0
                                                           int64_t& useless_rowset_version_count) {
1875
0
    std::map<RowsetId, Version> rowset_ids;
1876
0
    std::set<int64_t> end_versions;
1877
0
    traverse_rowsets(
1878
0
            [&rowset_ids, &end_versions](const RowsetSharedPtr& rs) {
1879
0
                rowset_ids[rs->rowset_id()] = rs->version();
1880
0
                end_versions.emplace(rs->end_version());
1881
0
            },
1882
0
            true);
1883
1884
0
    std::set<RowsetId> useless_rowsets;
1885
0
    std::map<RowsetId, std::vector<int64_t>> useless_rowset_versions;
1886
0
    {
1887
0
        _tablet_meta->delete_bitmap().traverse_rowset_and_version(
1888
                // 0: rowset and rowset with version exists
1889
                // -1: rowset does not exist
1890
                // -2: find next <rowset, version>
1891
                //     rowset exist, rowset with version does not exist
1892
                //     sequence table
1893
0
                [&](const RowsetId& rowset_id, int64_t version) {
1894
0
                    auto rowset_it = rowset_ids.find(rowset_id);
1895
0
                    if (rowset_it == rowset_ids.end()) {
1896
0
                        useless_rowsets.emplace(rowset_id);
1897
0
                        return -1;
1898
0
                    }
1899
0
                    if (end_versions.find(version) == end_versions.end()) {
1900
0
                        if (tablet_schema()->has_sequence_col()) {
1901
0
                            auto rowset_version = rowset_it->second;
1902
0
                            if (version >= rowset_version.first &&
1903
0
                                version <= rowset_version.second) {
1904
0
                                return -2;
1905
0
                            }
1906
0
                        }
1907
0
                        if (useless_rowset_versions.find(rowset_id) ==
1908
0
                            useless_rowset_versions.end()) {
1909
0
                            useless_rowset_versions[rowset_id] = {};
1910
0
                        }
1911
0
                        useless_rowset_versions[rowset_id].emplace_back(version);
1912
0
                        return -2;
1913
0
                    }
1914
0
                    return 0;
1915
0
                });
1916
0
    }
1917
0
    useless_rowset_count = useless_rowsets.size();
1918
0
    useless_rowset_version_count = useless_rowset_versions.size();
1919
0
    if (!useless_rowsets.empty() || !useless_rowset_versions.empty()) {
1920
0
        std::stringstream ss;
1921
0
        if (!useless_rowsets.empty()) {
1922
0
            ss << "useless rowsets: {";
1923
0
            for (auto it = useless_rowsets.begin(); it != useless_rowsets.end(); ++it) {
1924
0
                if (it != useless_rowsets.begin()) {
1925
0
                    ss << ", ";
1926
0
                }
1927
0
                ss << it->to_string();
1928
0
            }
1929
0
            ss << "}. ";
1930
0
        }
1931
0
        if (!useless_rowset_versions.empty()) {
1932
0
            ss << "useless rowset versions: {";
1933
0
            for (auto iter = useless_rowset_versions.begin(); iter != useless_rowset_versions.end();
1934
0
                 ++iter) {
1935
0
                if (iter != useless_rowset_versions.begin()) {
1936
0
                    ss << ", ";
1937
0
                }
1938
0
                ss << iter->first.to_string() << ": [";
1939
                // some versions are continuous, such as [8, 9, 10, 11, 13, 17, 18]
1940
                // print as [8-11, 13, 17-18]
1941
0
                int64_t last_start_version = -1;
1942
0
                int64_t last_end_version = -1;
1943
0
                for (int64_t version : iter->second) {
1944
0
                    if (last_start_version == -1) {
1945
0
                        last_start_version = version;
1946
0
                        last_end_version = version;
1947
0
                        continue;
1948
0
                    }
1949
0
                    if (last_end_version + 1 == version) {
1950
0
                        last_end_version = version;
1951
0
                    } else {
1952
0
                        if (last_start_version == last_end_version) {
1953
0
                            ss << last_start_version << ", ";
1954
0
                        } else {
1955
0
                            ss << last_start_version << "-" << last_end_version << ", ";
1956
0
                        }
1957
0
                        last_start_version = version;
1958
0
                        last_end_version = version;
1959
0
                    }
1960
0
                }
1961
0
                if (last_start_version == last_end_version) {
1962
0
                    ss << last_start_version;
1963
0
                } else {
1964
0
                    ss << last_start_version << "-" << last_end_version;
1965
0
                }
1966
1967
0
                ss << "]";
1968
0
            }
1969
0
            ss << "}.";
1970
0
        }
1971
0
        LOG(WARNING) << "failed check_agg_delete_bitmap_for_stale_rowsets for tablet_id="
1972
0
                     << tablet_id() << ". " << ss.str();
1973
0
    } else {
1974
0
        LOG(INFO) << "succeed check_agg_delete_bitmap_for_stale_rowsets for tablet_id="
1975
0
                  << tablet_id();
1976
0
    }
1977
0
}
1978
1979
0
RowsetSharedPtr BaseTablet::get_rowset(const RowsetId& rowset_id) {
1980
0
    std::shared_lock rdlock(_meta_lock);
1981
0
    for (auto& version_rowset : _rs_version_map) {
1982
0
        if (version_rowset.second->rowset_id() == rowset_id) {
1983
0
            return version_rowset.second;
1984
0
        }
1985
0
    }
1986
0
    for (auto& stale_version_rowset : _stale_rs_version_map) {
1987
0
        if (stale_version_rowset.second->rowset_id() == rowset_id) {
1988
0
            return stale_version_rowset.second;
1989
0
        }
1990
0
    }
1991
0
    return nullptr;
1992
0
}
1993
1994
1
std::vector<RowsetSharedPtr> BaseTablet::get_snapshot_rowset(bool include_stale_rowset) const {
1995
1
    std::shared_lock rdlock(_meta_lock);
1996
1
    std::vector<RowsetSharedPtr> rowsets;
1997
1
    std::transform(_rs_version_map.cbegin(), _rs_version_map.cend(), std::back_inserter(rowsets),
1998
28
                   [](auto& kv) { return kv.second; });
1999
1
    if (include_stale_rowset) {
2000
0
        std::transform(_stale_rs_version_map.cbegin(), _stale_rs_version_map.cend(),
2001
0
                       std::back_inserter(rowsets), [](auto& kv) { return kv.second; });
2002
0
    }
2003
1
    return rowsets;
2004
1
}
2005
2006
void BaseTablet::calc_consecutive_empty_rowsets(
2007
        std::vector<RowsetSharedPtr>* empty_rowsets,
2008
4
        const std::vector<RowsetSharedPtr>& candidate_rowsets, int64_t limit) {
2009
4
    int len = cast_set<int>(candidate_rowsets.size());
2010
12
    for (int i = 0; i < len - 1; ++i) {
2011
9
        auto rowset = candidate_rowsets[i];
2012
9
        auto next_rowset = candidate_rowsets[i + 1];
2013
2014
        // identify two consecutive rowsets that are empty
2015
9
        if (rowset->num_segments() == 0 && next_rowset->num_segments() == 0 &&
2016
9
            !rowset->rowset_meta()->has_delete_predicate() &&
2017
9
            !next_rowset->rowset_meta()->has_delete_predicate() &&
2018
9
            rowset->end_version() == next_rowset->start_version() - 1) {
2019
1
            empty_rowsets->emplace_back(rowset);
2020
1
            empty_rowsets->emplace_back(next_rowset);
2021
1
            rowset = next_rowset;
2022
1
            int next_index = i + 2;
2023
2024
            // keep searching for consecutive empty rowsets
2025
6
            while (next_index < len && candidate_rowsets[next_index]->num_segments() == 0 &&
2026
6
                   !candidate_rowsets[next_index]->rowset_meta()->has_delete_predicate() &&
2027
6
                   rowset->end_version() == candidate_rowsets[next_index]->start_version() - 1) {
2028
5
                empty_rowsets->emplace_back(candidate_rowsets[next_index]);
2029
5
                rowset = candidate_rowsets[next_index++];
2030
5
            }
2031
            // if the number of consecutive empty rowset reach the limit,
2032
            // and there are still rowsets following them
2033
1
            if (empty_rowsets->size() >= limit && next_index < len) {
2034
1
                return;
2035
1
            } else {
2036
                // current rowset is not empty, start searching from that rowset in the next
2037
0
                i = next_index - 1;
2038
0
                empty_rowsets->clear();
2039
0
            }
2040
1
        }
2041
9
    }
2042
4
}
2043
2044
Status BaseTablet::calc_file_crc(uint32_t* crc_value, int64_t start_version, int64_t end_version,
2045
0
                                 uint32_t* rowset_count, int64_t* file_count) {
2046
0
    Version v(start_version, end_version);
2047
0
    std::vector<RowsetSharedPtr> rowsets;
2048
0
    traverse_rowsets([&rowsets, &v](const auto& rs) {
2049
        // get all rowsets
2050
0
        if (v.contains(rs->version())) {
2051
0
            rowsets.emplace_back(rs);
2052
0
        }
2053
0
    });
2054
0
    std::sort(rowsets.begin(), rowsets.end(), Rowset::comparator);
2055
0
    *rowset_count = cast_set<uint32_t>(rowsets.size());
2056
2057
0
    *crc_value = 0;
2058
0
    *file_count = 0;
2059
0
    for (const auto& rs : rowsets) {
2060
0
        uint32_t rs_crc_value = 0;
2061
0
        int64_t rs_file_count = 0;
2062
0
        auto rowset = std::static_pointer_cast<BetaRowset>(rs);
2063
0
        auto st = rowset->calc_file_crc(&rs_crc_value, &rs_file_count);
2064
0
        if (!st.ok()) {
2065
0
            return st;
2066
0
        }
2067
        // crc_value is calculated based on the crc_value of each rowset.
2068
0
        *crc_value = crc32c::Extend(*crc_value, reinterpret_cast<const char*>(&rs_crc_value),
2069
0
                                    sizeof(rs_crc_value));
2070
0
        *file_count += rs_file_count;
2071
0
    }
2072
0
    return Status::OK();
2073
0
}
2074
2075
0
Status BaseTablet::show_nested_index_file(std::string* json_meta) {
2076
0
    Version v(0, max_version_unlocked());
2077
0
    std::vector<RowsetSharedPtr> rowsets;
2078
0
    traverse_rowsets([&rowsets, &v](const auto& rs) {
2079
        // get all rowsets
2080
0
        if (v.contains(rs->version())) {
2081
0
            rowsets.emplace_back(rs);
2082
0
        }
2083
0
    });
2084
0
    std::sort(rowsets.begin(), rowsets.end(), Rowset::comparator);
2085
2086
0
    rapidjson::Document doc;
2087
0
    doc.SetObject();
2088
0
    rapidjson::Document::AllocatorType& allocator = doc.GetAllocator();
2089
0
    rapidjson::Value tabletIdValue(tablet_id());
2090
0
    doc.AddMember("tablet_id", tabletIdValue, allocator);
2091
2092
0
    rapidjson::Value rowsets_value(rapidjson::kArrayType);
2093
2094
0
    for (const auto& rs : rowsets) {
2095
0
        rapidjson::Value rowset_value(rapidjson::kObjectType);
2096
2097
0
        auto rowset = std::static_pointer_cast<BetaRowset>(rs);
2098
0
        RETURN_IF_ERROR(rowset->show_nested_index_file(&rowset_value, allocator));
2099
0
        rowsets_value.PushBack(rowset_value, allocator);
2100
0
    }
2101
0
    doc.AddMember("rowsets", rowsets_value, allocator);
2102
2103
0
    rapidjson::StringBuffer buffer;
2104
0
    rapidjson::PrettyWriter<rapidjson::StringBuffer> writer(buffer);
2105
0
    doc.Accept(writer);
2106
0
    *json_meta = std::string(buffer.GetString());
2107
2108
0
    return Status::OK();
2109
0
}
2110
2111
void BaseTablet::get_base_rowset_delete_bitmap_count(
2112
        uint64_t* max_base_rowset_delete_bitmap_score,
2113
0
        int64_t* max_base_rowset_delete_bitmap_score_tablet_id) {
2114
0
    std::vector<RowsetSharedPtr> rowsets_;
2115
0
    std::string base_rowset_id_str;
2116
0
    {
2117
0
        std::shared_lock rowset_ldlock(this->get_header_lock());
2118
0
        for (const auto& it : _rs_version_map) {
2119
0
            rowsets_.emplace_back(it.second);
2120
0
        }
2121
0
    }
2122
0
    std::sort(rowsets_.begin(), rowsets_.end(), Rowset::comparator);
2123
0
    if (!rowsets_.empty()) {
2124
0
        bool base_found = false;
2125
0
        for (auto& rowset : rowsets_) {
2126
0
            if (rowset->start_version() > 2) {
2127
0
                break;
2128
0
            }
2129
0
            base_found = true;
2130
0
            uint64_t base_rowset_delete_bitmap_count =
2131
0
                    this->tablet_meta()->delete_bitmap().get_count_with_range(
2132
0
                            {rowset->rowset_id(), 0, 0},
2133
0
                            {rowset->rowset_id(), UINT32_MAX, UINT64_MAX});
2134
0
            if (base_rowset_delete_bitmap_count > *max_base_rowset_delete_bitmap_score) {
2135
0
                *max_base_rowset_delete_bitmap_score = base_rowset_delete_bitmap_count;
2136
0
                *max_base_rowset_delete_bitmap_score_tablet_id = this->tablet_id();
2137
0
            }
2138
0
        }
2139
0
        if (!base_found) {
2140
0
            LOG(WARNING) << "can not found base rowset for tablet " << tablet_id();
2141
0
        }
2142
0
    }
2143
0
}
2144
2145
32
int32_t BaseTablet::max_version_config() {
2146
32
    int32_t max_version = tablet_meta()->compaction_policy() == CUMULATIVE_TIME_SERIES_POLICY
2147
32
                                  ? std::max(config::time_series_max_tablet_version_num,
2148
0
                                             config::max_tablet_version_num)
2149
32
                                  : config::max_tablet_version_num;
2150
32
    return max_version;
2151
32
}
2152
2153
0
void BaseTablet::prefill_dbm_agg_cache(const RowsetSharedPtr& rowset, int64_t version) {
2154
0
    for (std::size_t i = 0; i < rowset->num_segments(); i++) {
2155
0
        tablet_meta()->delete_bitmap().get_agg({rowset->rowset_id(), i, version});
2156
0
    }
2157
0
}
2158
2159
0
void BaseTablet::prefill_dbm_agg_cache_after_compaction(const RowsetSharedPtr& output_rowset) {
2160
0
    if (keys_type() == KeysType::UNIQUE_KEYS && enable_unique_key_merge_on_write() &&
2161
0
        (config::enable_prefill_output_dbm_agg_cache_after_compaction ||
2162
0
         config::enable_prefill_all_dbm_agg_cache_after_compaction)) {
2163
0
        int64_t cur_max_version {-1};
2164
0
        {
2165
0
            std::shared_lock rlock(get_header_lock());
2166
0
            cur_max_version = max_version_unlocked();
2167
0
        }
2168
0
        if (config::enable_prefill_all_dbm_agg_cache_after_compaction) {
2169
0
            traverse_rowsets(
2170
0
                    [&](const RowsetSharedPtr& rs) { prefill_dbm_agg_cache(rs, cur_max_version); },
2171
0
                    false);
2172
0
        } else if (config::enable_prefill_output_dbm_agg_cache_after_compaction) {
2173
0
            prefill_dbm_agg_cache(output_rowset, cur_max_version);
2174
0
        }
2175
0
    }
2176
0
}
2177
2178
} // namespace doris