Coverage Report

Created: 2025-05-12 16:11

/root/doris/be/src/olap/tablet.cpp
Line
Count
Source (jump to first uncovered line)
1
// Licensed to the Apache Software Foundation (ASF) under one
2
// or more contributor license agreements.  See the NOTICE file
3
// distributed with this work for additional information
4
// regarding copyright ownership.  The ASF licenses this file
5
// to you under the Apache License, Version 2.0 (the
6
// "License"); you may not use this file except in compliance
7
// with the License.  You may obtain a copy of the License at
8
//
9
//   http://www.apache.org/licenses/LICENSE-2.0
10
//
11
// Unless required by applicable law or agreed to in writing,
12
// software distributed under the License is distributed on an
13
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
14
// KIND, either express or implied.  See the License for the
15
// specific language governing permissions and limitations
16
// under the License.
17
18
#include "olap/tablet.h"
19
20
#include <butil/logging.h>
21
#include <bvar/reducer.h>
22
#include <bvar/window.h>
23
#include <fmt/format.h>
24
#include <gen_cpp/FrontendService_types.h>
25
#include <gen_cpp/MasterService_types.h>
26
#include <gen_cpp/Metrics_types.h>
27
#include <gen_cpp/olap_file.pb.h>
28
#include <gen_cpp/types.pb.h>
29
#include <glog/logging.h>
30
#include <rapidjson/document.h>
31
#include <rapidjson/encodings.h>
32
#include <rapidjson/prettywriter.h>
33
#include <rapidjson/rapidjson.h>
34
#include <rapidjson/stringbuffer.h>
35
36
#include <algorithm>
37
#include <atomic>
38
#include <boost/container/detail/std_fwd.hpp>
39
#include <cstdint>
40
#include <roaring/roaring.hh>
41
42
#include "common/compiler_util.h" // IWYU pragma: keep
43
// IWYU pragma: no_include <bits/chrono.h>
44
#include <chrono> // IWYU pragma: keep
45
#include <filesystem>
46
#include <iterator>
47
#include <limits>
48
#include <map>
49
#include <memory>
50
#include <mutex>
51
#include <set>
52
#include <shared_mutex>
53
#include <string>
54
#include <tuple>
55
#include <type_traits>
56
#include <unordered_map>
57
#include <unordered_set>
58
59
#include "agent/utils.h"
60
#include "common/config.h"
61
#include "common/consts.h"
62
#include "common/logging.h"
63
#include "common/signal_handler.h"
64
#include "common/status.h"
65
#include "io/fs/file_reader.h"
66
#include "io/fs/file_reader_writer_fwd.h"
67
#include "io/fs/file_system.h"
68
#include "io/fs/file_writer.h"
69
#include "io/fs/path.h"
70
#include "io/fs/remote_file_system.h"
71
#include "io/io_common.h"
72
#include "olap/base_compaction.h"
73
#include "olap/base_tablet.h"
74
#include "olap/binlog.h"
75
#include "olap/cumulative_compaction.h"
76
#include "olap/cumulative_compaction_policy.h"
77
#include "olap/cumulative_compaction_time_series_policy.h"
78
#include "olap/delete_bitmap_calculator.h"
79
#include "olap/full_compaction.h"
80
#include "olap/memtable.h"
81
#include "olap/olap_common.h"
82
#include "olap/olap_define.h"
83
#include "olap/olap_meta.h"
84
#include "olap/primary_key_index.h"
85
#include "olap/rowset/beta_rowset.h"
86
#include "olap/rowset/rowset.h"
87
#include "olap/rowset/rowset_factory.h"
88
#include "olap/rowset/rowset_fwd.h"
89
#include "olap/rowset/rowset_meta.h"
90
#include "olap/rowset/rowset_meta_manager.h"
91
#include "olap/rowset/rowset_writer.h"
92
#include "olap/rowset/rowset_writer_context.h"
93
#include "olap/rowset/segment_v2/column_reader.h"
94
#include "olap/rowset/segment_v2/common.h"
95
#include "olap/rowset/segment_v2/indexed_column_reader.h"
96
#include "olap/rowset/vertical_beta_rowset_writer.h"
97
#include "olap/schema_change.h"
98
#include "olap/single_replica_compaction.h"
99
#include "olap/storage_engine.h"
100
#include "olap/storage_policy.h"
101
#include "olap/tablet_manager.h"
102
#include "olap/tablet_meta.h"
103
#include "olap/tablet_meta_manager.h"
104
#include "olap/tablet_schema.h"
105
#include "olap/txn_manager.h"
106
#include "olap/types.h"
107
#include "olap/utils.h"
108
#include "segment_loader.h"
109
#include "service/point_query_executor.h"
110
#include "tablet.h"
111
#include "util/bvar_helper.h"
112
#include "util/debug_points.h"
113
#include "util/defer_op.h"
114
#include "util/doris_metrics.h"
115
#include "util/pretty_printer.h"
116
#include "util/scoped_cleanup.h"
117
#include "util/stopwatch.hpp"
118
#include "util/threadpool.h"
119
#include "util/time.h"
120
#include "util/trace.h"
121
#include "util/uid_util.h"
122
#include "util/work_thread_pool.hpp"
123
#include "vec/columns/column.h"
124
#include "vec/columns/column_string.h"
125
#include "vec/common/schema_util.h"
126
#include "vec/common/string_ref.h"
127
#include "vec/data_types/data_type.h"
128
#include "vec/data_types/data_type_factory.hpp"
129
#include "vec/data_types/serde/data_type_serde.h"
130
#include "vec/jsonb/serialize.h"
131
132
namespace doris {
133
class TupleDescriptor;
134
135
namespace vectorized {
136
class Block;
137
} // namespace vectorized
138
139
using namespace ErrorCode;
140
using namespace std::chrono_literals;
141
142
using std::pair;
143
using std::string;
144
using std::vector;
145
using io::FileSystemSPtr;
146
147
namespace {
148
149
bvar::Adder<uint64_t> exceed_version_limit_counter;
150
bvar::Window<bvar::Adder<uint64_t>> exceed_version_limit_counter_minute(
151
        &exceed_version_limit_counter, 60);
152
bvar::Adder<uint64_t> cooldown_pending_task("cooldown_pending_task");
153
bvar::Adder<uint64_t> cooldown_processing_task("cooldown_processing_task");
154
155
12.7k
void set_last_failure_time(Tablet* tablet, const Compaction& compaction, int64_t ms) {
156
12.7k
    switch (compaction.compaction_type()) {
157
12.6k
    case ReaderType::READER_CUMULATIVE_COMPACTION:
158
12.6k
        tablet->set_last_cumu_compaction_failure_time(ms);
159
12.6k
        return;
160
96
    case ReaderType::READER_BASE_COMPACTION:
161
96
        tablet->set_last_base_compaction_failure_time(ms);
162
96
        return;
163
10
    case ReaderType::READER_FULL_COMPACTION:
164
10
        tablet->set_last_full_compaction_failure_time(ms);
165
10
        return;
166
0
    default:
167
0
        LOG(FATAL) << "invalid compaction type " << compaction.compaction_name()
168
0
                   << " tablet_id: " << tablet->tablet_id();
169
12.7k
    }
170
12.7k
};
171
172
} // namespace
173
174
bvar::Adder<uint64_t> unused_remote_rowset_num("unused_remote_rowset_num");
175
176
WriteCooldownMetaExecutors::WriteCooldownMetaExecutors(size_t executor_nums)
177
4
        : _executor_nums(executor_nums) {
178
24
    for (size_t i = 0; i < _executor_nums; i++) {
179
20
        std::unique_ptr<PriorityThreadPool> pool;
180
20
        static_cast<void>(ThreadPoolBuilder("WriteCooldownMetaExecutor")
181
20
                                  .set_min_threads(1)
182
20
                                  .set_max_threads(1)
183
20
                                  .set_max_queue_size(std::numeric_limits<int>::max())
184
20
                                  .build(&pool));
185
20
        _executors.emplace_back(std::move(pool));
186
20
    }
187
4
}
188
189
1
void WriteCooldownMetaExecutors::stop() {
190
5
    for (auto& pool_ptr : _executors) {
191
5
        if (pool_ptr) {
192
5
            pool_ptr->shutdown();
193
5
        }
194
5
    }
195
1
}
196
197
17
void WriteCooldownMetaExecutors::WriteCooldownMetaExecutors::submit(TabletSharedPtr tablet) {
198
17
    auto tablet_id = tablet->tablet_id();
199
200
17
    {
201
17
        std::shared_lock rdlock(tablet->get_header_lock());
202
17
        if (!tablet->tablet_meta()->cooldown_meta_id().initialized()) {
203
0
            VLOG_NOTICE << "tablet " << tablet_id << " is not cooldown replica";
204
0
            return;
205
0
        }
206
17
        if (tablet->tablet_state() == TABLET_SHUTDOWN) [[unlikely]] {
207
0
            LOG_INFO("tablet {} has been dropped, don't do cooldown", tablet_id);
208
0
            return;
209
0
        }
210
17
    }
211
17
    {
212
        // one tablet could at most have one cooldown task to be done
213
17
        std::unique_lock<std::mutex> lck {_latch};
214
17
        if (_pending_tablets.count(tablet_id) > 0) {
215
0
            return;
216
0
        }
217
17
        _pending_tablets.insert(tablet_id);
218
17
    }
219
220
17
    auto async_write_task = [this, t = std::move(tablet)]() {
221
17
        {
222
17
            std::unique_lock<std::mutex> lck {_latch};
223
17
            _pending_tablets.erase(t->tablet_id());
224
17
        }
225
17
        SCOPED_ATTACH_TASK(ExecEnv::GetInstance()->orphan_mem_tracker());
226
17
        auto s = t->write_cooldown_meta();
227
17
        if (s.ok()) {
228
17
            return;
229
17
        }
230
0
        if (!s.is<ABORTED>()) {
231
0
            LOG_EVERY_SECOND(WARNING)
232
0
                    << "write tablet " << t->tablet_id() << " cooldown meta failed because: " << s;
233
0
            submit(t);
234
0
            return;
235
0
        }
236
0
        VLOG_DEBUG << "tablet " << t->tablet_id() << " is not cooldown replica";
237
0
    };
238
239
17
    cooldown_pending_task << 1;
240
17
    _executors[_get_executor_pos(tablet_id)]->offer([task = std::move(async_write_task)]() {
241
17
        cooldown_pending_task << -1;
242
17
        cooldown_processing_task << 1;
243
17
        task();
244
17
        cooldown_processing_task << -1;
245
17
    });
246
17
}
247
248
Tablet::Tablet(StorageEngine& engine, TabletMetaSharedPtr tablet_meta, DataDir* data_dir,
249
               const std::string_view& cumulative_compaction_type)
250
        : BaseTablet(std::move(tablet_meta)),
251
          _engine(engine),
252
          _data_dir(data_dir),
253
          _is_bad(false),
254
          _last_cumu_compaction_failure_millis(0),
255
          _last_base_compaction_failure_millis(0),
256
          _last_full_compaction_failure_millis(0),
257
          _last_cumu_compaction_success_millis(0),
258
          _last_base_compaction_success_millis(0),
259
          _last_full_compaction_success_millis(0),
260
          _cumulative_point(K_INVALID_CUMULATIVE_POINT),
261
          _newly_created_rowset_num(0),
262
          _last_checkpoint_time(0),
263
          _cumulative_compaction_type(cumulative_compaction_type),
264
          _is_tablet_path_exists(true),
265
          _last_missed_version(-1),
266
233k
          _last_missed_time_s(0) {
267
234k
    if (_data_dir != nullptr) {
268
234k
        _tablet_path = fmt::format("{}/{}/{}/{}/{}", _data_dir->path(), DATA_PREFIX,
269
234k
                                   _tablet_meta->shard_id(), tablet_id(), schema_hash());
270
234k
    }
271
233k
}
272
273
1.28k
bool Tablet::set_tablet_schema_into_rowset_meta() {
274
1.28k
    bool flag = false;
275
2.44k
    for (auto&& rowset_meta : _tablet_meta->all_mutable_rs_metas()) {
276
2.44k
        if (!rowset_meta->tablet_schema()) {
277
0
            rowset_meta->set_tablet_schema(_tablet_meta->tablet_schema());
278
0
            flag = true;
279
0
        }
280
2.44k
    }
281
1.28k
    return flag;
282
1.28k
}
283
284
233k
Status Tablet::_init_once_action() {
285
233k
    Status res = Status::OK();
286
18.4E
    VLOG_NOTICE << "begin to load tablet. tablet=" << tablet_id()
287
18.4E
                << ", version_size=" << _tablet_meta->version_count();
288
289
#ifdef BE_TEST
290
    // init cumulative compaction policy by type
291
    _cumulative_compaction_policy =
292
            CumulativeCompactionPolicyFactory::create_cumulative_compaction_policy(
293
                    _tablet_meta->compaction_policy());
294
#endif
295
296
233k
    for (const auto& rs_meta : _tablet_meta->all_rs_metas()) {
297
4.36k
        Version version = rs_meta->version();
298
4.36k
        RowsetSharedPtr rowset;
299
4.36k
        res = create_rowset(rs_meta, &rowset);
300
4.36k
        if (!res.ok()) {
301
0
            LOG(WARNING) << "fail to init rowset. tablet_id=" << tablet_id()
302
0
                         << ", schema_hash=" << schema_hash() << ", version=" << version
303
0
                         << ", res=" << res;
304
0
            return res;
305
0
        }
306
4.36k
        _rs_version_map[version] = std::move(rowset);
307
4.36k
    }
308
309
    // init stale rowset
310
233k
    for (const auto& stale_rs_meta : _tablet_meta->all_stale_rs_metas()) {
311
50
        Version version = stale_rs_meta->version();
312
50
        RowsetSharedPtr rowset;
313
50
        res = create_rowset(stale_rs_meta, &rowset);
314
50
        if (!res.ok()) {
315
0
            LOG(WARNING) << "fail to init stale rowset. tablet_id:" << tablet_id()
316
0
                         << ", schema_hash:" << schema_hash() << ", version=" << version
317
0
                         << ", res:" << res;
318
0
            return res;
319
0
        }
320
50
        _stale_rs_version_map[version] = std::move(rowset);
321
50
    }
322
323
233k
    return res;
324
233k
}
325
326
234k
Status Tablet::init() {
327
234k
    return _init_once.call([this] { return _init_once_action(); });
328
234k
}
329
330
// should save tablet meta to remote meta store
331
// if it's a primary replica
332
578k
void Tablet::save_meta() {
333
578k
    check_table_size_correctness();
334
578k
    auto res = _tablet_meta->save_meta(_data_dir);
335
578k
    CHECK_EQ(res, Status::OK()) << "fail to save tablet_meta. res=" << res
336
0
                                << ", root=" << _data_dir->path();
337
578k
}
338
339
// Caller should hold _meta_lock.
340
Status Tablet::revise_tablet_meta(const std::vector<RowsetSharedPtr>& to_add,
341
                                  const std::vector<RowsetSharedPtr>& to_delete,
342
0
                                  bool is_incremental_clone) {
343
0
    LOG(INFO) << "begin to revise tablet. tablet_id=" << tablet_id();
344
    // 1. for incremental clone, we have to add the rowsets first to make it easy to compute
345
    //    all the delete bitmaps, and it's easy to delete them if we end up with a failure
346
    // 2. for full clone, we can calculate delete bitmaps on the cloned rowsets directly.
347
0
    if (is_incremental_clone) {
348
0
        CHECK(to_delete.empty()); // don't need to delete rowsets
349
0
        add_rowsets(to_add);
350
        // reconstruct from tablet meta
351
0
        _timestamped_version_tracker.construct_versioned_tracker(_tablet_meta->all_rs_metas());
352
0
    }
353
354
0
    Status calc_bm_status;
355
0
    std::vector<RowsetSharedPtr> base_rowsets_for_full_clone = to_add; // copy vector
356
0
    while (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write()) {
357
0
        std::vector<RowsetSharedPtr> calc_delete_bitmap_rowsets;
358
0
        int64_t to_add_min_version = INT64_MAX;
359
0
        int64_t to_add_max_version = INT64_MIN;
360
0
        for (auto& rs : to_add) {
361
0
            if (to_add_min_version > rs->start_version()) {
362
0
                to_add_min_version = rs->start_version();
363
0
            }
364
0
            if (to_add_max_version < rs->end_version()) {
365
0
                to_add_max_version = rs->end_version();
366
0
            }
367
0
        }
368
0
        Version calc_delete_bitmap_ver;
369
0
        if (is_incremental_clone) {
370
            // From the rowset of to_add with smallest version, all other rowsets
371
            // need to recalculate the delete bitmap
372
            // For example:
373
            // local tablet: [0-1] [2-5] [6-6] [9-10]
374
            // clone tablet: [7-7] [8-8]
375
            // new tablet:   [0-1] [2-5] [6-6] [7-7] [8-8] [9-10]
376
            // [7-7] [8-8] [9-10] need to recalculate delete bitmap
377
0
            calc_delete_bitmap_ver = Version(to_add_min_version, max_version_unlocked());
378
0
        } else {
379
            // the delete bitmap of to_add's rowsets has clone from remote when full clone.
380
            // only other rowsets in local need to recalculate the delete bitmap.
381
            // For example:
382
            // local tablet: [0-1]x [2-5]x [6-6]x [7-7]x [9-10]
383
            // clone tablet: [0-1]  [2-4]  [5-6]  [7-8]
384
            // new tablet:   [0-1]  [2-4]  [5-6]  [7-8] [9-10]
385
            // only [9-10] need to recalculate delete bitmap
386
0
            CHECK_EQ(to_add_min_version, 0) << "to_add_min_version is: " << to_add_min_version;
387
0
            calc_delete_bitmap_ver = Version(to_add_max_version + 1, max_version_unlocked());
388
0
        }
389
390
0
        if (calc_delete_bitmap_ver.first <= calc_delete_bitmap_ver.second) {
391
0
            calc_bm_status = capture_consistent_rowsets_unlocked(calc_delete_bitmap_ver,
392
0
                                                                 &calc_delete_bitmap_rowsets);
393
0
            if (!calc_bm_status.ok()) {
394
0
                LOG(WARNING) << "fail to capture_consistent_rowsets, res: " << calc_bm_status;
395
0
                break;
396
0
            }
397
            // FIXME(plat1ko): Use `const TabletSharedPtr&` as parameter
398
0
            auto self = _engine.tablet_manager()->get_tablet(tablet_id());
399
0
            CHECK(self);
400
0
            for (auto rs : calc_delete_bitmap_rowsets) {
401
0
                if (is_incremental_clone) {
402
0
                    calc_bm_status = update_delete_bitmap_without_lock(self, rs);
403
0
                } else {
404
0
                    calc_bm_status = update_delete_bitmap_without_lock(
405
0
                            self, rs, &base_rowsets_for_full_clone);
406
0
                    base_rowsets_for_full_clone.push_back(rs);
407
0
                }
408
0
                if (!calc_bm_status.ok()) {
409
0
                    LOG(WARNING) << "fail to update_delete_bitmap_without_lock, res: "
410
0
                                 << calc_bm_status;
411
0
                    break;
412
0
                }
413
0
            }
414
0
        }
415
0
        break; // while (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write())
416
0
    }
417
418
0
    DBUG_EXECUTE_IF("Tablet.revise_tablet_meta_fail", {
419
0
        auto ptablet_id = dp->param("tablet_id", 0);
420
0
        if (tablet_id() == ptablet_id) {
421
0
            LOG(INFO) << "injected revies_tablet_meta failure for tabelt: " << ptablet_id;
422
0
            calc_bm_status = Status::InternalError("fault injection error");
423
0
        }
424
0
    });
425
426
    // error handling
427
0
    if (!calc_bm_status.ok()) {
428
0
        if (is_incremental_clone) {
429
0
            RETURN_IF_ERROR(delete_rowsets(to_add, false));
430
0
            LOG(WARNING) << "incremental clone on tablet: " << tablet_id() << " failed due to "
431
0
                         << calc_bm_status.msg() << ", revert " << to_add.size()
432
0
                         << " rowsets added before.";
433
0
        } else {
434
0
            LOG(WARNING) << "full clone on tablet: " << tablet_id() << " failed due to "
435
0
                         << calc_bm_status.msg() << ", will not update tablet meta.";
436
0
        }
437
0
        return calc_bm_status;
438
0
    }
439
440
    // full clone, calculate delete bitmap succeeded, update rowset
441
0
    if (!is_incremental_clone) {
442
0
        RETURN_IF_ERROR(delete_rowsets(to_delete, false));
443
0
        add_rowsets(to_add);
444
        // reconstruct from tablet meta
445
0
        _timestamped_version_tracker.construct_versioned_tracker(_tablet_meta->all_rs_metas());
446
447
        // check the rowsets used for delete bitmap calculation is equal to the rowsets
448
        // that we can capture by version
449
0
        if (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write()) {
450
0
            Version full_version = Version(0, max_version_unlocked());
451
0
            std::vector<RowsetSharedPtr> expected_rowsets;
452
0
            auto st = capture_consistent_rowsets_unlocked(full_version, &expected_rowsets);
453
0
            DCHECK(st.ok()) << st;
454
0
            DCHECK_EQ(base_rowsets_for_full_clone.size(), expected_rowsets.size());
455
0
            if (st.ok() && base_rowsets_for_full_clone.size() != expected_rowsets.size())
456
0
                    [[unlikely]] {
457
0
                LOG(WARNING) << "full clone succeeded, but the count("
458
0
                             << base_rowsets_for_full_clone.size()
459
0
                             << ") of base rowsets used for delete bitmap calculation is not match "
460
0
                                "expect count("
461
0
                             << expected_rowsets.size() << ") we capture from tablet meta";
462
0
            }
463
0
        }
464
0
    }
465
466
    // clear stale rowset
467
0
    for (auto& [v, rs] : _stale_rs_version_map) {
468
0
        _engine.add_unused_rowset(rs);
469
0
    }
470
0
    _stale_rs_version_map.clear();
471
0
    _tablet_meta->clear_stale_rowset();
472
0
    save_meta();
473
474
0
    LOG(INFO) << "finish to revise tablet. tablet_id=" << tablet_id();
475
0
    return Status::OK();
476
0
}
477
478
247k
Status Tablet::add_rowset(RowsetSharedPtr rowset) {
479
247k
    DCHECK(rowset != nullptr);
480
247k
    std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
481
247k
    SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
482
    // If the rowset already exist, just return directly.  The rowset_id is an unique-id,
483
    // we can use it to check this situation.
484
247k
    if (_contains_rowset(rowset->rowset_id())) {
485
24
        return Status::OK();
486
24
    }
487
    // Otherwise, the version should be not contained in any existing rowset.
488
247k
    RETURN_IF_ERROR(_contains_version(rowset->version()));
489
490
247k
    RETURN_IF_ERROR(_tablet_meta->add_rs_meta(rowset->rowset_meta()));
491
247k
    _rs_version_map[rowset->version()] = rowset;
492
247k
    _timestamped_version_tracker.add_version(rowset->version());
493
247k
    add_compaction_score(rowset->rowset_meta()->get_compaction_score());
494
495
247k
    std::vector<RowsetSharedPtr> rowsets_to_delete;
496
    // yiguolei: temp code, should remove the rowset contains by this rowset
497
    // but it should be removed in multi path version
498
268k
    for (auto& it : _rs_version_map) {
499
268k
        if (rowset->version().contains(it.first) && rowset->version() != it.first) {
500
0
            CHECK(it.second != nullptr)
501
0
                    << "there exist a version=" << it.first
502
0
                    << " contains the input rs with version=" << rowset->version()
503
0
                    << ", but the related rs is null";
504
0
            rowsets_to_delete.push_back(it.second);
505
0
        }
506
268k
    }
507
247k
    std::vector<RowsetSharedPtr> empty_vec;
508
247k
    RETURN_IF_ERROR(modify_rowsets(empty_vec, rowsets_to_delete));
509
247k
    ++_newly_created_rowset_num;
510
247k
    return Status::OK();
511
247k
}
512
513
48.6k
bool Tablet::rowset_exists_unlocked(const RowsetSharedPtr& rowset) {
514
48.6k
    if (auto it = _rs_version_map.find(rowset->version()); it == _rs_version_map.end()) {
515
0
        return false;
516
48.6k
    } else if (rowset->rowset_id() != it->second->rowset_id()) {
517
0
        return false;
518
0
    }
519
48.6k
    return true;
520
48.6k
}
521
522
Status Tablet::modify_rowsets(std::vector<RowsetSharedPtr>& to_add,
523
261k
                              std::vector<RowsetSharedPtr>& to_delete, bool check_delete) {
524
    // the compaction process allow to compact the single version, eg: version[4-4].
525
    // this kind of "single version compaction" has same "input version" and "output version".
526
    // which means "to_add->version()" equals to "to_delete->version()".
527
    // So we should delete the "to_delete" before adding the "to_add",
528
    // otherwise, the "to_add" will be deleted from _rs_version_map, eventually.
529
    //
530
    // And if the version of "to_add" and "to_delete" are exactly same. eg:
531
    // to_add:      [7-7]
532
    // to_delete:   [7-7]
533
    // In this case, we no longer need to add the rowset in "to_delete" to
534
    // _stale_rs_version_map, but can delete it directly.
535
536
261k
    if (to_add.empty() && to_delete.empty()) {
537
247k
        return Status::OK();
538
247k
    }
539
540
14.4k
    if (check_delete) {
541
128k
        for (auto&& rs : to_delete) {
542
128k
            if (auto it = _rs_version_map.find(rs->version()); it == _rs_version_map.end()) {
543
0
                return Status::Error<DELETE_VERSION_ERROR>(
544
0
                        "try to delete not exist version {} from {}", rs->version().to_string(),
545
0
                        tablet_id());
546
128k
            } else if (rs->rowset_id() != it->second->rowset_id()) {
547
0
                return Status::Error<DELETE_VERSION_ERROR>(
548
0
                        "try to delete version {} from {}, but rowset id changed, delete rowset id "
549
0
                        "is {}, exists rowsetid is {}",
550
0
                        rs->version().to_string(), tablet_id(), rs->rowset_id().to_string(),
551
0
                        it->second->rowset_id().to_string());
552
0
            }
553
128k
        }
554
14.2k
    }
555
556
14.4k
    bool same_version = true;
557
14.4k
    std::sort(to_add.begin(), to_add.end(), Rowset::comparator);
558
14.4k
    std::sort(to_delete.begin(), to_delete.end(), Rowset::comparator);
559
14.4k
    if (to_add.size() == to_delete.size()) {
560
4.43k
        for (int i = 0; i < to_add.size(); ++i) {
561
3.22k
            if (to_add[i]->version() != to_delete[i]->version()) {
562
0
                same_version = false;
563
0
                break;
564
0
            }
565
3.22k
        }
566
13.2k
    } else {
567
13.2k
        same_version = false;
568
13.2k
    }
569
570
14.4k
    std::vector<RowsetMetaSharedPtr> rs_metas_to_delete;
571
128k
    for (auto& rs : to_delete) {
572
128k
        rs_metas_to_delete.push_back(rs->rowset_meta());
573
128k
        _rs_version_map.erase(rs->version());
574
575
128k
        if (!same_version) {
576
            // put compaction rowsets in _stale_rs_version_map.
577
124k
            _stale_rs_version_map[rs->version()] = rs;
578
124k
        }
579
128k
    }
580
581
14.4k
    std::vector<RowsetMetaSharedPtr> rs_metas_to_add;
582
16.3k
    for (auto& rs : to_add) {
583
16.3k
        rs_metas_to_add.push_back(rs->rowset_meta());
584
16.3k
        _rs_version_map[rs->version()] = rs;
585
586
16.3k
        if (!same_version) {
587
            // If version are same, then _timestamped_version_tracker
588
            // already has this version, no need to add again.
589
13.0k
            _timestamped_version_tracker.add_version(rs->version());
590
13.0k
        }
591
16.3k
        ++_newly_created_rowset_num;
592
16.3k
    }
593
594
14.4k
    _tablet_meta->modify_rs_metas(rs_metas_to_add, rs_metas_to_delete, same_version);
595
596
14.4k
    if (!same_version) {
597
        // add rs_metas_to_delete to tracker
598
13.0k
        _timestamped_version_tracker.add_stale_path_version(rs_metas_to_delete);
599
13.0k
    } else {
600
        // delete rowset in "to_delete" directly
601
3.22k
        for (auto& rs : to_delete) {
602
3.22k
            LOG(INFO) << "add unused rowset " << rs->rowset_id() << " because of same version";
603
3.22k
            if (rs->is_local()) {
604
3.22k
                _engine.add_unused_rowset(rs);
605
3.22k
            }
606
3.22k
        }
607
1.34k
    }
608
609
14.4k
    int32_t add_score = 0;
610
16.3k
    for (auto rs : to_add) {
611
16.3k
        add_score += rs->rowset_meta()->get_compaction_score();
612
16.3k
    }
613
14.4k
    int32_t sub_score = 0;
614
128k
    for (auto rs : to_delete) {
615
128k
        sub_score += rs->rowset_meta()->get_compaction_score();
616
128k
    }
617
14.4k
    add_compaction_score(add_score - sub_score);
618
619
14.4k
    return Status::OK();
620
14.4k
}
621
622
17
void Tablet::add_rowsets(const std::vector<RowsetSharedPtr>& to_add) {
623
17
    if (to_add.empty()) {
624
0
        return;
625
0
    }
626
17
    std::vector<RowsetMetaSharedPtr> rs_metas;
627
17
    rs_metas.reserve(to_add.size());
628
17
    for (auto& rs : to_add) {
629
17
        _rs_version_map.emplace(rs->version(), rs);
630
17
        _timestamped_version_tracker.add_version(rs->version());
631
17
        rs_metas.push_back(rs->rowset_meta());
632
17
    }
633
17
    _tablet_meta->modify_rs_metas(rs_metas, {});
634
17
}
635
636
8.17k
Status Tablet::delete_rowsets(const std::vector<RowsetSharedPtr>& to_delete, bool move_to_stale) {
637
8.17k
    if (to_delete.empty()) {
638
0
        return Status::OK();
639
0
    }
640
8.17k
    std::vector<RowsetMetaSharedPtr> rs_metas;
641
8.17k
    rs_metas.reserve(to_delete.size());
642
8.20k
    for (const auto& rs : to_delete) {
643
8.20k
        rs_metas.push_back(rs->rowset_meta());
644
8.20k
        _rs_version_map.erase(rs->version());
645
8.20k
    }
646
8.17k
    _tablet_meta->modify_rs_metas({}, rs_metas, !move_to_stale);
647
8.17k
    if (move_to_stale) {
648
0
        for (const auto& rs : to_delete) {
649
0
            _stale_rs_version_map[rs->version()] = rs;
650
0
        }
651
0
        _timestamped_version_tracker.add_stale_path_version(rs_metas);
652
8.17k
    } else {
653
8.19k
        for (const auto& rs : to_delete) {
654
8.19k
            _timestamped_version_tracker.delete_version(rs->version());
655
8.19k
            if (rs->is_local()) {
656
8.19k
                _engine.add_unused_rowset(rs);
657
8.19k
                RETURN_IF_ERROR(RowsetMetaManager::remove(_data_dir->get_meta(), tablet_uid(),
658
8.19k
                                                          rs->rowset_meta()->rowset_id()));
659
8.19k
            }
660
8.19k
        }
661
8.17k
    }
662
8.17k
    return Status::OK();
663
8.17k
}
664
665
0
RowsetSharedPtr Tablet::_rowset_with_largest_size() {
666
0
    RowsetSharedPtr largest_rowset = nullptr;
667
0
    for (auto& it : _rs_version_map) {
668
0
        if (it.second->empty() || it.second->zero_num_rows()) {
669
0
            continue;
670
0
        }
671
0
        if (largest_rowset == nullptr || it.second->rowset_meta()->index_disk_size() >
672
0
                                                 largest_rowset->rowset_meta()->index_disk_size()) {
673
0
            largest_rowset = it.second;
674
0
        }
675
0
    }
676
677
0
    return largest_rowset;
678
0
}
679
680
// add inc rowset should not persist tablet meta, because it will be persisted when publish txn.
681
271k
Status Tablet::add_inc_rowset(const RowsetSharedPtr& rowset) {
682
271k
    DCHECK(rowset != nullptr);
683
271k
    std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
684
271k
    SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
685
271k
    if (_contains_rowset(rowset->rowset_id())) {
686
0
        return Status::OK();
687
0
    }
688
271k
    RETURN_IF_ERROR(_contains_version(rowset->version()));
689
690
271k
    RETURN_IF_ERROR(_tablet_meta->add_rs_meta(rowset->rowset_meta()));
691
271k
    _rs_version_map[rowset->version()] = rowset;
692
693
271k
    _timestamped_version_tracker.add_version(rowset->version());
694
695
271k
    ++_newly_created_rowset_num;
696
697
271k
    add_compaction_score(rowset->rowset_meta()->get_compaction_score());
698
699
271k
    return Status::OK();
700
271k
}
701
702
96.3k
void Tablet::_delete_stale_rowset_by_version(const Version& version) {
703
96.3k
    RowsetMetaSharedPtr rowset_meta = _tablet_meta->acquire_stale_rs_meta_by_version(version);
704
96.3k
    if (rowset_meta == nullptr) {
705
8
        return;
706
8
    }
707
96.3k
    _tablet_meta->delete_stale_rs_meta_by_version(version);
708
96.3k
    VLOG_NOTICE << "delete stale rowset. tablet=" << tablet_id() << ", version=" << version;
709
96.3k
}
710
711
1.45M
void Tablet::delete_expired_stale_rowset() {
712
1.45M
    if (config::enable_mow_verbose_log) {
713
0
        LOG_INFO("begin delete_expired_stale_rowset for tablet={}", tablet_id());
714
0
    }
715
1.45M
    int64_t now = UnixSeconds();
716
1.45M
    std::vector<std::pair<Version, std::vector<RowsetId>>> deleted_stale_rowsets;
717
    // hold write lock while processing stable rowset
718
1.45M
    {
719
1.45M
        std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
720
1.45M
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
721
        // Compute the end time to delete rowsets, when a expired rowset createtime less then this time, it will be deleted.
722
1.45M
        double expired_stale_sweep_endtime =
723
1.45M
                ::difftime(now, config::tablet_rowset_stale_sweep_time_sec);
724
1.45M
        if (config::tablet_rowset_stale_sweep_by_size) {
725
0
            expired_stale_sweep_endtime = now;
726
0
        }
727
728
1.45M
        std::vector<int64_t> path_id_vec;
729
        // capture the path version to delete
730
1.45M
        _timestamped_version_tracker.capture_expired_paths(
731
1.45M
                static_cast<int64_t>(expired_stale_sweep_endtime), &path_id_vec);
732
733
1.45M
        if (path_id_vec.empty()) {
734
1.45M
            return;
735
1.45M
        }
736
737
6.40k
        const RowsetSharedPtr lastest_delta = get_rowset_with_max_version();
738
6.40k
        if (lastest_delta == nullptr) {
739
0
            LOG(WARNING) << "lastest_delta is null " << tablet_id();
740
0
            return;
741
0
        }
742
743
        // fetch missing version before delete
744
6.40k
        Versions missed_versions = get_missed_versions_unlocked(lastest_delta->end_version());
745
6.40k
        if (!missed_versions.empty()) {
746
0
            LOG(WARNING) << "tablet:" << tablet_id()
747
0
                         << ", missed version for version:" << lastest_delta->end_version();
748
0
            _print_missed_versions(missed_versions);
749
0
            return;
750
0
        }
751
752
        // do check consistent operation
753
6.40k
        auto path_id_iter = path_id_vec.begin();
754
755
6.40k
        std::map<int64_t, PathVersionListSharedPtr> stale_version_path_map;
756
15.4k
        while (path_id_iter != path_id_vec.end()) {
757
9.09k
            PathVersionListSharedPtr version_path =
758
9.09k
                    _timestamped_version_tracker.fetch_and_delete_path_by_id(*path_id_iter);
759
760
9.09k
            Version test_version = Version(0, lastest_delta->end_version());
761
9.09k
            stale_version_path_map[*path_id_iter] = version_path;
762
763
9.09k
            Status status =
764
9.09k
                    capture_consistent_versions_unlocked(test_version, nullptr, false, false);
765
            // 1. When there is no consistent versions, we must reconstruct the tracker.
766
9.09k
            if (!status.ok()) {
767
                // 2. fetch missing version after delete
768
0
                Versions after_missed_versions =
769
0
                        get_missed_versions_unlocked(lastest_delta->end_version());
770
771
                // 2.1 check whether missed_versions and after_missed_versions are the same.
772
                // when they are the same, it means we can delete the path securely.
773
0
                bool is_missing = missed_versions.size() != after_missed_versions.size();
774
775
0
                if (!is_missing) {
776
0
                    for (int ver_index = 0; ver_index < missed_versions.size(); ver_index++) {
777
0
                        if (missed_versions[ver_index] != after_missed_versions[ver_index]) {
778
0
                            is_missing = true;
779
0
                            break;
780
0
                        }
781
0
                    }
782
0
                }
783
784
0
                if (is_missing) {
785
0
                    LOG(WARNING) << "The consistent version check fails, there are bugs. "
786
0
                                 << "Reconstruct the tracker to recover versions in tablet="
787
0
                                 << tablet_id();
788
789
                    // 3. try to recover
790
0
                    _timestamped_version_tracker.recover_versioned_tracker(stale_version_path_map);
791
792
                    // 4. double check the consistent versions
793
                    // fetch missing version after recover
794
0
                    Versions recover_missed_versions =
795
0
                            get_missed_versions_unlocked(lastest_delta->end_version());
796
797
                    // 4.1 check whether missed_versions and recover_missed_versions are the same.
798
                    // when they are the same, it means we recover successfully.
799
0
                    bool is_recover_missing =
800
0
                            missed_versions.size() != recover_missed_versions.size();
801
802
0
                    if (!is_recover_missing) {
803
0
                        for (int ver_index = 0; ver_index < missed_versions.size(); ver_index++) {
804
0
                            if (missed_versions[ver_index] != recover_missed_versions[ver_index]) {
805
0
                                is_recover_missing = true;
806
0
                                break;
807
0
                            }
808
0
                        }
809
0
                    }
810
811
                    // 5. check recover fail, version is mission
812
0
                    if (is_recover_missing) {
813
0
                        if (!config::ignore_rowset_stale_unconsistent_delete) {
814
0
                            LOG(FATAL)
815
0
                                    << "rowset stale unconsistent delete. tablet= " << tablet_id();
816
0
                        } else {
817
0
                            LOG(WARNING)
818
0
                                    << "rowset stale unconsistent delete. tablet= " << tablet_id();
819
0
                        }
820
0
                    }
821
0
                }
822
0
                return;
823
0
            }
824
9.09k
            path_id_iter++;
825
9.09k
        }
826
827
6.40k
        auto old_size = _stale_rs_version_map.size();
828
6.40k
        auto old_meta_size = _tablet_meta->all_stale_rs_metas().size();
829
830
        // do delete operation
831
6.40k
        std::vector<std::string> version_to_delete;
832
6.40k
        auto to_delete_iter = stale_version_path_map.begin();
833
15.4k
        while (to_delete_iter != stale_version_path_map.end()) {
834
9.09k
            std::vector<TimestampedVersionSharedPtr>& to_delete_version =
835
9.09k
                    to_delete_iter->second->timestamped_versions();
836
9.09k
            int64_t start_version = -1;
837
9.09k
            int64_t end_version = -1;
838
9.09k
            std::vector<RowsetId> remove_rowset_ids;
839
96.3k
            for (auto& timestampedVersion : to_delete_version) {
840
96.3k
                auto it = _stale_rs_version_map.find(timestampedVersion->version());
841
96.3k
                if (it != _stale_rs_version_map.end()) {
842
96.3k
                    it->second->clear_cache();
843
                    // delete rowset
844
96.3k
                    if (it->second->is_local()) {
845
96.3k
                        _engine.add_unused_rowset(it->second);
846
96.3k
                        if (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write()) {
847
                            // mow does not support cold data in object storage
848
44.0k
                            remove_rowset_ids.emplace_back(it->second->rowset_id());
849
44.0k
                        }
850
96.3k
                    }
851
96.3k
                    _stale_rs_version_map.erase(it);
852
96.3k
                    VLOG_NOTICE << "delete stale rowset tablet=" << tablet_id() << " version["
853
0
                                << timestampedVersion->version().first << ","
854
0
                                << timestampedVersion->version().second
855
0
                                << "] move to unused_rowset success " << std::fixed
856
0
                                << expired_stale_sweep_endtime;
857
96.3k
                } else {
858
8
                    LOG(WARNING) << "delete stale rowset tablet=" << tablet_id() << " version["
859
8
                                 << timestampedVersion->version().first << ","
860
8
                                 << timestampedVersion->version().second
861
8
                                 << "] not find in stale rs version map";
862
8
                }
863
96.3k
                if (start_version < 0) {
864
9.09k
                    start_version = timestampedVersion->version().first;
865
9.09k
                }
866
96.3k
                end_version = timestampedVersion->version().second;
867
96.3k
                _delete_stale_rowset_by_version(timestampedVersion->version());
868
96.3k
            }
869
9.09k
            Version version(start_version, end_version);
870
9.09k
            version_to_delete.emplace_back(version.to_string());
871
9.09k
            to_delete_iter++;
872
9.09k
            if (!remove_rowset_ids.empty()) {
873
2.64k
                deleted_stale_rowsets.emplace_back(version, remove_rowset_ids);
874
2.64k
            }
875
9.09k
        }
876
6.40k
        _tablet_meta->delete_bitmap().remove_stale_delete_bitmap_from_queue(version_to_delete);
877
878
6.40k
        bool reconstructed = _reconstruct_version_tracker_if_necessary();
879
880
6.40k
        VLOG_NOTICE << "delete stale rowset _stale_rs_version_map tablet=" << tablet_id()
881
0
                    << " current_size=" << _stale_rs_version_map.size() << " old_size=" << old_size
882
0
                    << " current_meta_size=" << _tablet_meta->all_stale_rs_metas().size()
883
0
                    << " old_meta_size=" << old_meta_size << " sweep endtime " << std::fixed
884
0
                    << expired_stale_sweep_endtime << ", reconstructed=" << reconstructed;
885
6.40k
    }
886
6.40k
    if (!deleted_stale_rowsets.empty()) {
887
        // agg delete bitmap for pre rowsets; record unused delete bitmap key ranges
888
429
        OlapStopWatch watch;
889
2.64k
        for (const auto& [version, remove_rowset_ids] : deleted_stale_rowsets) {
890
            // agg delete bitmap for pre rowset
891
2.64k
            DeleteBitmapKeyRanges remove_delete_bitmap_key_ranges;
892
2.64k
            agg_delete_bitmap_for_stale_rowsets(version, remove_delete_bitmap_key_ranges);
893
            // add remove delete bitmap
894
2.64k
            if (!remove_delete_bitmap_key_ranges.empty()) {
895
147
                _engine.add_unused_delete_bitmap_key_ranges(tablet_id(), remove_rowset_ids,
896
147
                                                            remove_delete_bitmap_key_ranges);
897
147
            }
898
2.64k
        }
899
429
        LOG(INFO) << "agg pre rowsets delete bitmap. tablet_id=" << tablet_id()
900
429
                  << ", size=" << deleted_stale_rowsets.size()
901
429
                  << ", cost(us)=" << watch.get_elapse_time_us();
902
429
    }
903
6.40k
#ifndef BE_TEST
904
6.40k
    {
905
6.40k
        std::shared_lock<std::shared_mutex> rlock(_meta_lock);
906
6.40k
        save_meta();
907
6.40k
    }
908
6.40k
#endif
909
6.40k
    if (config::enable_mow_verbose_log) {
910
0
        LOG_INFO("finish delete_expired_stale_rowset for tablet={}", tablet_id());
911
0
    }
912
6.40k
    DBUG_EXECUTE_IF("Tablet.delete_expired_stale_rowset.start_delete_unused_rowset",
913
6.40k
                    { _engine.start_delete_unused_rowset(); });
914
6.40k
}
915
916
Status Tablet::capture_consistent_versions_unlocked(const Version& spec_version,
917
                                                    Versions* version_path,
918
3.16M
                                                    bool skip_missing_version, bool quiet) const {
919
3.16M
    Status status =
920
3.16M
            _timestamped_version_tracker.capture_consistent_versions(spec_version, version_path);
921
3.16M
    if (!status.ok() && !quiet) {
922
1
        Versions missed_versions = get_missed_versions_unlocked(spec_version.second);
923
1
        if (missed_versions.empty()) {
924
            // if version_path is null, it may be a compaction check logic.
925
            // so to avoid print too many logs.
926
0
            if (version_path != nullptr) {
927
0
                LOG(WARNING) << "tablet:" << tablet_id()
928
0
                             << ", version already has been merged. spec_version: " << spec_version
929
0
                             << ", max_version: " << max_version_unlocked();
930
0
            }
931
0
            status = Status::Error<VERSION_ALREADY_MERGED, false>(
932
0
                    "versions are already compacted, spec_version "
933
0
                    "{}, max_version {}, tablet_id {}",
934
0
                    spec_version.second, max_version_unlocked(), tablet_id());
935
1
        } else {
936
1
            if (version_path != nullptr) {
937
1
                LOG(WARNING) << "status:" << status << ", tablet:" << tablet_id()
938
1
                             << ", missed version for version:" << spec_version;
939
1
                _print_missed_versions(missed_versions);
940
1
                if (skip_missing_version) {
941
0
                    LOG(WARNING) << "force skipping missing version for tablet:" << tablet_id();
942
0
                    return Status::OK();
943
0
                }
944
1
            }
945
1
        }
946
1
    }
947
948
3.16M
    DBUG_EXECUTE_IF("TTablet::capture_consistent_versions.inject_failure", {
949
3.16M
        auto tablet_id = dp->param<int64>("tablet_id", -1);
950
3.16M
        if (tablet_id != -1 && tablet_id == _tablet_meta->tablet_id()) {
951
3.16M
            status = Status::Error<VERSION_ALREADY_MERGED>("version already merged");
952
3.16M
        }
953
3.16M
    });
954
955
3.16M
    return status;
956
3.16M
}
957
958
8.08k
Status Tablet::check_version_integrity(const Version& version, bool quiet) {
959
8.08k
    std::shared_lock rdlock(_meta_lock);
960
8.08k
    return capture_consistent_versions_unlocked(version, nullptr, false, quiet);
961
8.08k
}
962
963
517k
bool Tablet::exceed_version_limit(int32_t limit) {
964
517k
    if (_tablet_meta->version_count() > limit) {
965
0
        exceed_version_limit_counter << 1;
966
0
        return true;
967
0
    }
968
517k
    return false;
969
517k
}
970
971
// If any rowset contains the specific version, it means the version already exist
972
1.15M
bool Tablet::check_version_exist(const Version& version) const {
973
1.15M
    std::shared_lock rdlock(_meta_lock);
974
5.01M
    for (auto& it : _rs_version_map) {
975
5.01M
        if (it.first.contains(version)) {
976
878k
            return true;
977
878k
        }
978
5.01M
    }
979
277k
    return false;
980
1.15M
}
981
982
// The meta read lock should be held before calling
983
void Tablet::acquire_version_and_rowsets(
984
279k
        std::vector<std::pair<Version, RowsetSharedPtr>>* version_rowsets) const {
985
440k
    for (const auto& it : _rs_version_map) {
986
440k
        version_rowsets->emplace_back(it.first, it.second);
987
440k
    }
988
279k
}
989
990
Status Tablet::capture_consistent_rowsets_unlocked(const Version& spec_version,
991
1.95k
                                                   std::vector<RowsetSharedPtr>* rowsets) const {
992
1.95k
    std::vector<Version> version_path;
993
1.95k
    RETURN_IF_ERROR(
994
1.95k
            capture_consistent_versions_unlocked(spec_version, &version_path, false, false));
995
1.95k
    RETURN_IF_ERROR(_capture_consistent_rowsets_unlocked(version_path, rowsets));
996
1.95k
    return Status::OK();
997
1.95k
}
998
999
Status Tablet::capture_rs_readers(const Version& spec_version, std::vector<RowSetSplits>* rs_splits,
1000
3.14M
                                  bool skip_missing_version) {
1001
3.14M
    std::shared_lock rlock(_meta_lock);
1002
3.14M
    std::vector<Version> version_path;
1003
3.14M
    RETURN_IF_ERROR(capture_consistent_versions_unlocked(spec_version, &version_path,
1004
3.14M
                                                         skip_missing_version, false));
1005
3.14M
    RETURN_IF_ERROR(capture_rs_readers_unlocked(version_path, rs_splits));
1006
3.14M
    return Status::OK();
1007
3.14M
}
1008
1009
6.40k
Versions Tablet::calc_missed_versions(int64_t spec_version, Versions existing_versions) const {
1010
6.40k
    DCHECK(spec_version > 0) << "invalid spec_version: " << spec_version;
1011
1012
    // sort the existing versions in ascending order
1013
6.40k
    std::sort(existing_versions.begin(), existing_versions.end(),
1014
44.8k
              [](const Version& a, const Version& b) {
1015
                  // simple because 2 versions are certainly not overlapping
1016
44.8k
                  return a.first < b.first;
1017
44.8k
              });
1018
1019
    // From the first version(=0),  find the missing version until spec_version
1020
6.40k
    int64_t last_version = -1;
1021
6.40k
    Versions missed_versions;
1022
19.3k
    for (const Version& version : existing_versions) {
1023
19.3k
        if (version.first > last_version + 1) {
1024
4
            for (int64_t i = last_version + 1; i < version.first && i <= spec_version; ++i) {
1025
                // Don't merge missed_versions because clone & snapshot use single version.
1026
                // For example, if miss 4 ~ 6, clone need [4, 4], [5, 5], [6, 6], but not [4, 6].
1027
2
                missed_versions.emplace_back(i, i);
1028
2
            }
1029
2
        }
1030
19.3k
        last_version = version.second;
1031
19.3k
        if (last_version >= spec_version) {
1032
6.40k
            break;
1033
6.40k
        }
1034
19.3k
    }
1035
6.40k
    for (int64_t i = last_version + 1; i <= spec_version; ++i) {
1036
0
        missed_versions.emplace_back(i, i);
1037
0
    }
1038
1039
6.40k
    return missed_versions;
1040
6.40k
}
1041
1042
126M
bool Tablet::can_do_compaction(size_t path_hash, CompactionType compaction_type) {
1043
126M
    if (compaction_type == CompactionType::BASE_COMPACTION && tablet_state() != TABLET_RUNNING) {
1044
        // base compaction can only be done for tablet in TABLET_RUNNING state.
1045
        // but cumulative compaction can be done for TABLET_NOTREADY, such as tablet under alter process.
1046
32.1k
        return false;
1047
32.1k
    }
1048
1049
126M
    if (data_dir()->path_hash() != path_hash || !is_used() || !init_succeeded()) {
1050
47.1M
        return false;
1051
47.1M
    }
1052
1053
    // In TABLET_NOTREADY, we keep last 10 versions in new tablet so base tablet max_version
1054
    // not merged in new tablet and then we can do compaction
1055
79.5M
    return tablet_state() == TABLET_RUNNING || tablet_state() == TABLET_NOTREADY;
1056
126M
}
1057
1058
78.9M
uint32_t Tablet::calc_compaction_score() {
1059
78.9M
    if (_score_check_cnt++ % config::check_score_rounds_num != 0) {
1060
78.6M
        std::shared_lock rdlock(_meta_lock);
1061
78.6M
        if (_compaction_score > 0) {
1062
78.6M
            return _compaction_score;
1063
78.6M
        }
1064
78.6M
    }
1065
1066
242k
    {
1067
        // Need meta lock, because it will iterator "all_rs_metas" of tablet meta.
1068
242k
        std::shared_lock rdlock(_meta_lock);
1069
242k
        int32_t score = get_real_compaction_score();
1070
242k
        if (_compaction_score > 0 && _compaction_score != score) {
1071
236
            LOG(WARNING) << "cumu cache score not equal real score, cache score; "
1072
236
                         << _compaction_score << ", real score: " << score
1073
236
                         << ", tablet: " << tablet_id();
1074
236
        }
1075
242k
        _compaction_score = score;
1076
242k
        return score;
1077
78.9M
    }
1078
78.9M
}
1079
1080
bool Tablet::suitable_for_compaction(
1081
        CompactionType compaction_type,
1082
185k
        std::shared_ptr<CumulativeCompactionPolicy> cumulative_compaction_policy) {
1083
    // Need meta lock, because it will iterator "all_rs_metas" of tablet meta.
1084
185k
    std::shared_lock rdlock(_meta_lock);
1085
185k
    int32_t score = -1;
1086
185k
    if (compaction_type == CompactionType::CUMULATIVE_COMPACTION) {
1087
156k
        score = _calc_cumulative_compaction_score(cumulative_compaction_policy);
1088
156k
    } else {
1089
28.8k
        DCHECK_EQ(compaction_type, CompactionType::BASE_COMPACTION);
1090
28.8k
        score = _calc_base_compaction_score();
1091
28.8k
    }
1092
185k
    return score > 0;
1093
185k
}
1094
1095
81
uint32_t Tablet::calc_cold_data_compaction_score() const {
1096
81
    uint32_t score = 0;
1097
81
    std::vector<RowsetMetaSharedPtr> cooldowned_rowsets;
1098
81
    int64_t max_delete_version = 0;
1099
81
    {
1100
81
        std::shared_lock rlock(_meta_lock);
1101
92
        for (auto& rs_meta : _tablet_meta->all_rs_metas()) {
1102
92
            if (!rs_meta->is_local()) {
1103
91
                cooldowned_rowsets.push_back(rs_meta);
1104
91
                if (rs_meta->has_delete_predicate() &&
1105
91
                    rs_meta->end_version() > max_delete_version) {
1106
0
                    max_delete_version = rs_meta->end_version();
1107
0
                }
1108
91
            }
1109
92
        }
1110
81
    }
1111
91
    for (auto& rs_meta : cooldowned_rowsets) {
1112
91
        if (rs_meta->end_version() < max_delete_version) {
1113
0
            score += rs_meta->num_segments();
1114
91
        } else {
1115
91
            score += rs_meta->get_compaction_score();
1116
91
        }
1117
91
    }
1118
81
    return (keys_type() != KeysType::DUP_KEYS) ? score * 2 : score;
1119
81
}
1120
1121
uint32_t Tablet::_calc_cumulative_compaction_score(
1122
156k
        std::shared_ptr<CumulativeCompactionPolicy> cumulative_compaction_policy) {
1123
156k
    if (cumulative_compaction_policy == nullptr) [[unlikely]] {
1124
0
        return 0;
1125
0
    }
1126
156k
#ifndef BE_TEST
1127
156k
    if (_cumulative_compaction_policy == nullptr ||
1128
156k
        _cumulative_compaction_policy->name() != cumulative_compaction_policy->name()) {
1129
18.1k
        _cumulative_compaction_policy = cumulative_compaction_policy;
1130
18.1k
    }
1131
156k
#endif
1132
156k
    DBUG_EXECUTE_IF("Tablet._calc_cumulative_compaction_score.return", {
1133
156k
        LOG_WARNING("Tablet._calc_cumulative_compaction_score.return")
1134
156k
                .tag("tablet id", tablet_id());
1135
156k
        return 0;
1136
156k
    });
1137
156k
    return _cumulative_compaction_policy->calc_cumulative_compaction_score(this);
1138
156k
}
1139
1140
28.8k
uint32_t Tablet::_calc_base_compaction_score() const {
1141
28.8k
    uint32_t score = 0;
1142
28.8k
    const int64_t point = cumulative_layer_point();
1143
28.8k
    bool base_rowset_exist = false;
1144
28.8k
    bool has_delete = false;
1145
197k
    for (auto& rs_meta : _tablet_meta->all_rs_metas()) {
1146
197k
        if (rs_meta->start_version() == 0) {
1147
28.8k
            base_rowset_exist = true;
1148
28.8k
        }
1149
197k
        if (rs_meta->start_version() >= point || !rs_meta->is_local()) {
1150
            // all_rs_metas() is not sorted, so we use _continue_ other than _break_ here.
1151
177k
            continue;
1152
177k
        }
1153
20.1k
        if (rs_meta->has_delete_predicate()) {
1154
557
            has_delete = true;
1155
557
        }
1156
20.1k
        score += rs_meta->get_compaction_score();
1157
20.1k
    }
1158
1159
    // In the time series compaction policy, we want the base compaction to be triggered
1160
    // when there are delete versions present.
1161
28.8k
    if (_tablet_meta->compaction_policy() == CUMULATIVE_TIME_SERIES_POLICY) {
1162
0
        return (base_rowset_exist && has_delete) ? score : 0;
1163
0
    }
1164
1165
    // base不存在可能是tablet正在做alter table,先不选它,设score=0
1166
28.8k
    return base_rowset_exist ? score : 0;
1167
28.8k
}
1168
1169
8.51k
void Tablet::max_continuous_version_from_beginning(Version* version, Version* max_version) {
1170
8.51k
    bool has_version_cross;
1171
8.51k
    std::shared_lock rdlock(_meta_lock);
1172
8.51k
    _max_continuous_version_from_beginning_unlocked(version, max_version, &has_version_cross);
1173
8.51k
}
1174
1175
void Tablet::_max_continuous_version_from_beginning_unlocked(Version* version, Version* max_version,
1176
20.5M
                                                             bool* has_version_cross) const {
1177
20.5M
    std::vector<Version> existing_versions;
1178
20.5M
    *has_version_cross = false;
1179
37.3M
    for (auto& rs : _tablet_meta->all_rs_metas()) {
1180
37.3M
        existing_versions.emplace_back(rs->version());
1181
37.3M
    }
1182
1183
    // sort the existing versions in ascending order
1184
20.5M
    std::sort(existing_versions.begin(), existing_versions.end(),
1185
38.1M
              [](const Version& left, const Version& right) {
1186
                  // simple because 2 versions are certainly not overlapping
1187
38.1M
                  return left.first < right.first;
1188
38.1M
              });
1189
1190
20.5M
    Version max_continuous_version = {-1, -1};
1191
57.9M
    for (int i = 0; i < existing_versions.size(); ++i) {
1192
37.3M
        if (existing_versions[i].first > max_continuous_version.second + 1) {
1193
8
            break;
1194
37.3M
        } else if (existing_versions[i].first <= max_continuous_version.second) {
1195
0
            *has_version_cross = true;
1196
0
        }
1197
37.3M
        max_continuous_version = existing_versions[i];
1198
37.3M
    }
1199
20.5M
    *version = max_continuous_version;
1200
    // tablet may not has rowset, eg, tablet has just been clear for restore.
1201
20.5M
    if (max_version != nullptr && !existing_versions.empty()) {
1202
20.5M
        *max_version = existing_versions.back();
1203
20.5M
    }
1204
20.5M
}
1205
1206
36.9k
void Tablet::calculate_cumulative_point() {
1207
36.9k
    std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
1208
36.9k
    SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
1209
36.9k
    int64_t ret_cumulative_point;
1210
36.9k
    _cumulative_compaction_policy->calculate_cumulative_point(
1211
36.9k
            this, _tablet_meta->all_rs_metas(), _cumulative_point, &ret_cumulative_point);
1212
1213
36.9k
    if (ret_cumulative_point == K_INVALID_CUMULATIVE_POINT) {
1214
35.5k
        return;
1215
35.5k
    }
1216
1.40k
    set_cumulative_layer_point(ret_cumulative_point);
1217
1.40k
}
1218
1219
// NOTE: only used when create_table, so it is sure that there is no concurrent reader and writer.
1220
0
void Tablet::delete_all_files() {
1221
    // Release resources like memory and disk space.
1222
0
    std::shared_lock rdlock(_meta_lock);
1223
0
    for (auto it : _rs_version_map) {
1224
0
        static_cast<void>(it.second->remove());
1225
0
    }
1226
0
    _rs_version_map.clear();
1227
1228
0
    for (auto it : _stale_rs_version_map) {
1229
0
        static_cast<void>(it.second->remove());
1230
0
    }
1231
0
    _stale_rs_version_map.clear();
1232
0
}
1233
1234
0
void Tablet::check_tablet_path_exists() {
1235
0
    if (!tablet_path().empty()) {
1236
0
        std::error_code ec;
1237
0
        if (std::filesystem::is_directory(tablet_path(), ec)) {
1238
0
            _is_tablet_path_exists.store(true, std::memory_order_relaxed);
1239
0
        } else if (ec.value() == ENOENT || ec.value() == 0) {
1240
0
            _is_tablet_path_exists.store(false, std::memory_order_relaxed);
1241
0
        }
1242
0
    }
1243
0
}
1244
1245
518k
Status Tablet::_contains_version(const Version& version) {
1246
    // check if there exist a rowset contains the added rowset
1247
2.50M
    for (auto& it : _rs_version_map) {
1248
2.50M
        if (it.first.contains(version)) {
1249
            // TODO(lingbin): Is this check unnecessary?
1250
            // because the value type is std::shared_ptr, when will it be nullptr?
1251
            // In addition, in this class, there are many places that do not make this judgment
1252
            // when access _rs_version_map's value.
1253
260
            CHECK(it.second != nullptr) << "there exist a version=" << it.first
1254
0
                                        << " contains the input rs with version=" << version
1255
0
                                        << ", but the related rs is null";
1256
260
            return Status::Error<PUSH_VERSION_ALREADY_EXIST>("Tablet push duplicate version {}",
1257
260
                                                             version.to_string());
1258
260
        }
1259
2.50M
    }
1260
1261
518k
    return Status::OK();
1262
518k
}
1263
1264
36.9k
std::vector<RowsetSharedPtr> Tablet::pick_candidate_rowsets_to_cumulative_compaction() {
1265
36.9k
    std::vector<RowsetSharedPtr> candidate_rowsets;
1266
36.9k
    if (_cumulative_point == K_INVALID_CUMULATIVE_POINT) {
1267
0
        return candidate_rowsets;
1268
0
    }
1269
36.9k
    return _pick_visible_rowsets_to_compaction(_cumulative_point,
1270
36.9k
                                               std::numeric_limits<int64_t>::max());
1271
36.9k
}
1272
1273
4.52k
std::vector<RowsetSharedPtr> Tablet::pick_candidate_rowsets_to_base_compaction() {
1274
4.52k
    return _pick_visible_rowsets_to_compaction(std::numeric_limits<int64_t>::min(),
1275
4.52k
                                               _cumulative_point - 1);
1276
4.52k
}
1277
1278
std::vector<RowsetSharedPtr> Tablet::_pick_visible_rowsets_to_compaction(
1279
41.4k
        int64_t min_start_version, int64_t max_start_version) {
1280
41.4k
    auto [visible_version, update_ts] = get_visible_version_and_time();
1281
41.4k
    bool update_time_long = MonotonicMillis() - update_ts >
1282
41.4k
                            config::compaction_keep_invisible_version_timeout_sec * 1000L;
1283
41.4k
    int32_t keep_invisible_version_limit =
1284
41.4k
            update_time_long ? config::compaction_keep_invisible_version_min_count
1285
41.4k
                             : config::compaction_keep_invisible_version_max_count;
1286
1287
41.4k
    std::vector<RowsetSharedPtr> candidate_rowsets;
1288
41.4k
    {
1289
41.4k
        std::shared_lock rlock(_meta_lock);
1290
328k
        for (const auto& [version, rs] : _rs_version_map) {
1291
328k
            int64_t version_start = version.first;
1292
            // rowset is remote or rowset is not in given range
1293
328k
            if (!rs->is_local() || version_start < min_start_version ||
1294
328k
                version_start > max_start_version) {
1295
94.1k
                continue;
1296
94.1k
            }
1297
1298
            // can compact, met one of the conditions:
1299
            // 1. had been visible;
1300
            // 2. exceeds the limit of keep invisible versions.
1301
234k
            int64_t version_end = version.second;
1302
234k
            if (version_end <= visible_version ||
1303
234k
                version_end > visible_version + keep_invisible_version_limit) {
1304
223k
                candidate_rowsets.push_back(rs);
1305
223k
            }
1306
234k
        }
1307
41.4k
    }
1308
41.4k
    std::sort(candidate_rowsets.begin(), candidate_rowsets.end(), Rowset::comparator);
1309
41.4k
    return candidate_rowsets;
1310
41.4k
}
1311
1312
154
std::vector<RowsetSharedPtr> Tablet::pick_candidate_rowsets_to_full_compaction() {
1313
154
    std::vector<RowsetSharedPtr> candidate_rowsets;
1314
624
    traverse_rowsets([&candidate_rowsets](const auto& rs) {
1315
        // Do full compaction on all local rowsets.
1316
624
        if (rs->is_local()) {
1317
624
            candidate_rowsets.emplace_back(rs);
1318
624
        }
1319
624
    });
1320
154
    std::sort(candidate_rowsets.begin(), candidate_rowsets.end(), Rowset::comparator);
1321
154
    return candidate_rowsets;
1322
154
}
1323
1324
std::vector<RowsetSharedPtr> Tablet::pick_candidate_rowsets_to_build_inverted_index(
1325
1.35k
        const std::set<int64_t>& alter_index_uids, bool is_drop_op) {
1326
1.35k
    std::vector<RowsetSharedPtr> candidate_rowsets;
1327
1.35k
    {
1328
1.35k
        std::shared_lock rlock(_meta_lock);
1329
6.80k
        auto has_alter_inverted_index = [&](RowsetSharedPtr rowset) -> bool {
1330
6.80k
            for (const auto& index_id : alter_index_uids) {
1331
6.80k
                if (rowset->tablet_schema()->has_inverted_index_with_index_id(index_id)) {
1332
3.53k
                    return true;
1333
3.53k
                }
1334
6.80k
            }
1335
3.27k
            return false;
1336
6.80k
        };
1337
1338
3.50k
        for (const auto& [version, rs] : _rs_version_map) {
1339
3.50k
            if (!has_alter_inverted_index(rs) && is_drop_op) {
1340
198
                continue;
1341
198
            }
1342
3.30k
            if (has_alter_inverted_index(rs) && !is_drop_op) {
1343
72
                continue;
1344
72
            }
1345
1346
3.23k
            if (rs->is_local()) {
1347
3.23k
                candidate_rowsets.push_back(rs);
1348
3.23k
            }
1349
3.23k
        }
1350
1.35k
    }
1351
1.35k
    std::sort(candidate_rowsets.begin(), candidate_rowsets.end(), Rowset::comparator);
1352
1.35k
    return candidate_rowsets;
1353
1.35k
}
1354
1355
20.6M
std::tuple<int64_t, int64_t> Tablet::get_visible_version_and_time() const {
1356
    // some old tablet has bug, its partition_id is 0, fe couldn't update its visible version.
1357
    // so let this tablet's visible version become int64 max.
1358
20.6M
    auto version_info = std::atomic_load_explicit(&_visible_version, std::memory_order_relaxed);
1359
20.6M
    if (version_info != nullptr && partition_id() != 0) {
1360
20.6M
        return std::make_tuple(version_info->version.load(std::memory_order_relaxed),
1361
20.6M
                               version_info->update_ts);
1362
20.6M
    } else {
1363
33
        return std::make_tuple(std::numeric_limits<int64_t>::max(),
1364
33
                               std::numeric_limits<int64_t>::max());
1365
33
    }
1366
20.6M
}
1367
1368
// For http compaction action
1369
2.13k
void Tablet::get_compaction_status(std::string* json_result) {
1370
2.13k
    rapidjson::Document root;
1371
2.13k
    root.SetObject();
1372
1373
2.13k
    rapidjson::Document path_arr;
1374
2.13k
    path_arr.SetArray();
1375
1376
2.13k
    std::vector<RowsetSharedPtr> rowsets;
1377
2.13k
    std::vector<RowsetSharedPtr> stale_rowsets;
1378
2.13k
    std::vector<bool> delete_flags;
1379
2.13k
    {
1380
2.13k
        std::shared_lock rdlock(_meta_lock);
1381
2.13k
        rowsets.reserve(_rs_version_map.size());
1382
10.6k
        for (auto& it : _rs_version_map) {
1383
10.6k
            rowsets.push_back(it.second);
1384
10.6k
        }
1385
2.13k
        std::sort(rowsets.begin(), rowsets.end(), Rowset::comparator);
1386
1387
2.13k
        stale_rowsets.reserve(_stale_rs_version_map.size());
1388
7.71k
        for (auto& it : _stale_rs_version_map) {
1389
7.71k
            stale_rowsets.push_back(it.second);
1390
7.71k
        }
1391
2.13k
        std::sort(stale_rowsets.begin(), stale_rowsets.end(), Rowset::comparator);
1392
1393
2.13k
        delete_flags.reserve(rowsets.size());
1394
10.6k
        for (auto& rs : rowsets) {
1395
10.6k
            delete_flags.push_back(rs->rowset_meta()->has_delete_predicate());
1396
10.6k
        }
1397
        // get snapshot version path json_doc
1398
2.13k
        _timestamped_version_tracker.get_stale_version_path_json_doc(path_arr);
1399
2.13k
    }
1400
2.13k
    rapidjson::Value cumulative_policy_type;
1401
2.13k
    std::string policy_type_str = "cumulative compaction policy not initializied";
1402
2.13k
    if (_cumulative_compaction_policy != nullptr) {
1403
758
        policy_type_str = _cumulative_compaction_policy->name();
1404
758
    }
1405
2.13k
    cumulative_policy_type.SetString(policy_type_str.c_str(), policy_type_str.length(),
1406
2.13k
                                     root.GetAllocator());
1407
2.13k
    root.AddMember("cumulative policy type", cumulative_policy_type, root.GetAllocator());
1408
2.13k
    root.AddMember("cumulative point", _cumulative_point.load(), root.GetAllocator());
1409
2.13k
    rapidjson::Value cumu_value;
1410
2.13k
    std::string format_str = ToStringFromUnixMillis(_last_cumu_compaction_failure_millis.load());
1411
2.13k
    cumu_value.SetString(format_str.c_str(), format_str.length(), root.GetAllocator());
1412
2.13k
    root.AddMember("last cumulative failure time", cumu_value, root.GetAllocator());
1413
2.13k
    rapidjson::Value base_value;
1414
2.13k
    format_str = ToStringFromUnixMillis(_last_base_compaction_failure_millis.load());
1415
2.13k
    base_value.SetString(format_str.c_str(), format_str.length(), root.GetAllocator());
1416
2.13k
    root.AddMember("last base failure time", base_value, root.GetAllocator());
1417
2.13k
    rapidjson::Value full_value;
1418
2.13k
    format_str = ToStringFromUnixMillis(_last_full_compaction_failure_millis.load());
1419
2.13k
    full_value.SetString(format_str.c_str(), format_str.length(), root.GetAllocator());
1420
2.13k
    root.AddMember("last full failure time", full_value, root.GetAllocator());
1421
2.13k
    rapidjson::Value cumu_success_value;
1422
2.13k
    format_str = ToStringFromUnixMillis(_last_cumu_compaction_success_millis.load());
1423
2.13k
    cumu_success_value.SetString(format_str.c_str(), format_str.length(), root.GetAllocator());
1424
2.13k
    root.AddMember("last cumulative success time", cumu_success_value, root.GetAllocator());
1425
2.13k
    rapidjson::Value base_success_value;
1426
2.13k
    format_str = ToStringFromUnixMillis(_last_base_compaction_success_millis.load());
1427
2.13k
    base_success_value.SetString(format_str.c_str(), format_str.length(), root.GetAllocator());
1428
2.13k
    root.AddMember("last base success time", base_success_value, root.GetAllocator());
1429
2.13k
    rapidjson::Value full_success_value;
1430
2.13k
    format_str = ToStringFromUnixMillis(_last_full_compaction_success_millis.load());
1431
2.13k
    full_success_value.SetString(format_str.c_str(), format_str.length(), root.GetAllocator());
1432
2.13k
    root.AddMember("last full success time", full_success_value, root.GetAllocator());
1433
2.13k
    rapidjson::Value base_schedule_value;
1434
2.13k
    format_str = ToStringFromUnixMillis(_last_base_compaction_schedule_millis.load());
1435
2.13k
    base_schedule_value.SetString(format_str.c_str(), format_str.length(), root.GetAllocator());
1436
2.13k
    root.AddMember("last base schedule time", base_schedule_value, root.GetAllocator());
1437
2.13k
    rapidjson::Value base_compaction_status_value;
1438
2.13k
    base_compaction_status_value.SetString(_last_base_compaction_status.c_str(),
1439
2.13k
                                           _last_base_compaction_status.length(),
1440
2.13k
                                           root.GetAllocator());
1441
2.13k
    root.AddMember("last base status", base_compaction_status_value, root.GetAllocator());
1442
1443
    // last single replica compaction status
1444
    // "single replica compaction status": {
1445
    //     "remote peer": "172.100.1.0:10875",
1446
    //     "last failure status": "",
1447
    //     "last fetched rowset": "[8-10]"
1448
    // }
1449
2.13k
    rapidjson::Document status;
1450
2.13k
    status.SetObject();
1451
2.13k
    TReplicaInfo replica_info;
1452
2.13k
    std::string dummp_token;
1453
2.13k
    if (tablet_meta()->tablet_schema()->enable_single_replica_compaction() &&
1454
2.13k
        _engine.get_peer_replica_info(tablet_id(), &replica_info, &dummp_token)) {
1455
        // remote peer
1456
0
        rapidjson::Value peer_addr;
1457
0
        std::string addr = replica_info.host + ":" + std::to_string(replica_info.brpc_port);
1458
0
        peer_addr.SetString(addr.c_str(), addr.length(), status.GetAllocator());
1459
0
        status.AddMember("remote peer", peer_addr, status.GetAllocator());
1460
        // last failure status
1461
0
        rapidjson::Value compaction_status;
1462
0
        compaction_status.SetString(_last_single_compaction_failure_status.c_str(),
1463
0
                                    _last_single_compaction_failure_status.length(),
1464
0
                                    status.GetAllocator());
1465
0
        status.AddMember("last failure status", compaction_status, status.GetAllocator());
1466
        // last fetched rowset
1467
0
        rapidjson::Value version;
1468
0
        std::string fetched_version = _last_fetched_version.to_string();
1469
0
        version.SetString(fetched_version.c_str(), fetched_version.length(), status.GetAllocator());
1470
0
        status.AddMember("last fetched rowset", version, status.GetAllocator());
1471
0
        root.AddMember("single replica compaction status", status, root.GetAllocator());
1472
0
    }
1473
1474
    // print all rowsets' version as an array
1475
2.13k
    rapidjson::Document versions_arr;
1476
2.13k
    rapidjson::Document missing_versions_arr;
1477
2.13k
    versions_arr.SetArray();
1478
2.13k
    missing_versions_arr.SetArray();
1479
2.13k
    int64_t last_version = -1;
1480
10.6k
    for (auto& rowset : rowsets) {
1481
10.6k
        const Version& ver = rowset->version();
1482
10.6k
        if (ver.first != last_version + 1) {
1483
0
            rapidjson::Value miss_value;
1484
0
            miss_value.SetString(fmt::format("[{}-{}]", last_version + 1, ver.first - 1).c_str(),
1485
0
                                 missing_versions_arr.GetAllocator());
1486
0
            missing_versions_arr.PushBack(miss_value, missing_versions_arr.GetAllocator());
1487
0
        }
1488
10.6k
        rapidjson::Value value;
1489
10.6k
        std::string version_str = rowset->get_rowset_info_str();
1490
10.6k
        value.SetString(version_str.c_str(), version_str.length(), versions_arr.GetAllocator());
1491
10.6k
        versions_arr.PushBack(value, versions_arr.GetAllocator());
1492
10.6k
        last_version = ver.second;
1493
10.6k
    }
1494
2.13k
    root.AddMember("rowsets", versions_arr, root.GetAllocator());
1495
2.13k
    root.AddMember("missing_rowsets", missing_versions_arr, root.GetAllocator());
1496
1497
    // print all stale rowsets' version as an array
1498
2.13k
    rapidjson::Document stale_versions_arr;
1499
2.13k
    stale_versions_arr.SetArray();
1500
7.71k
    for (auto& rowset : stale_rowsets) {
1501
7.71k
        rapidjson::Value value;
1502
7.71k
        std::string version_str = rowset->get_rowset_info_str();
1503
7.71k
        value.SetString(version_str.c_str(), version_str.length(),
1504
7.71k
                        stale_versions_arr.GetAllocator());
1505
7.71k
        stale_versions_arr.PushBack(value, stale_versions_arr.GetAllocator());
1506
7.71k
    }
1507
2.13k
    root.AddMember("stale_rowsets", stale_versions_arr, root.GetAllocator());
1508
1509
    // add stale version rowsets
1510
2.13k
    root.AddMember("stale version path", path_arr, root.GetAllocator());
1511
1512
    // to json string
1513
2.13k
    rapidjson::StringBuffer strbuf;
1514
2.13k
    rapidjson::PrettyWriter<rapidjson::StringBuffer> writer(strbuf);
1515
2.13k
    root.Accept(writer);
1516
2.13k
    *json_result = std::string(strbuf.GetString());
1517
2.13k
}
1518
1519
635k
bool Tablet::do_tablet_meta_checkpoint() {
1520
635k
    std::lock_guard<std::shared_mutex> store_lock(_meta_store_lock);
1521
635k
    if (_newly_created_rowset_num == 0) {
1522
427k
        return false;
1523
427k
    }
1524
207k
    if (UnixMillis() - _last_checkpoint_time <
1525
207k
                config::tablet_meta_checkpoint_min_interval_secs * 1000 &&
1526
207k
        _newly_created_rowset_num < config::tablet_meta_checkpoint_min_new_rowsets_num) {
1527
1.76k
        return false;
1528
1.76k
    }
1529
    // hold read-lock other than write-lock, because it will not modify meta structure
1530
206k
    std::shared_lock rdlock(_meta_lock);
1531
206k
    if (tablet_state() != TABLET_RUNNING) {
1532
951
        LOG(INFO) << "tablet is under state=" << tablet_state()
1533
951
                  << ", not running, skip do checkpoint"
1534
951
                  << ", tablet=" << tablet_id();
1535
951
        return false;
1536
951
    }
1537
205k
    VLOG_NOTICE << "start to do tablet meta checkpoint, tablet=" << tablet_id();
1538
205k
    save_meta();
1539
    // if save meta successfully, then should remove the rowset meta existing in tablet
1540
    // meta from rowset meta store
1541
373k
    for (auto& rs_meta : _tablet_meta->all_rs_metas()) {
1542
        // If we delete it from rowset manager's meta explicitly in previous checkpoint, just skip.
1543
373k
        if (rs_meta->is_remove_from_rowset_meta()) {
1544
7.50k
            continue;
1545
7.50k
        }
1546
365k
        if (RowsetMetaManager::check_rowset_meta(_data_dir->get_meta(), tablet_uid(),
1547
365k
                                                 rs_meta->rowset_id())) {
1548
363k
            RETURN_FALSE_IF_ERROR(RowsetMetaManager::remove(_data_dir->get_meta(), tablet_uid(),
1549
363k
                                                            rs_meta->rowset_id()));
1550
363k
            VLOG_NOTICE << "remove rowset id from meta store because it is already persistent with "
1551
0
                        << "tablet meta, rowset_id=" << rs_meta->rowset_id();
1552
363k
        }
1553
365k
        rs_meta->set_remove_from_rowset_meta();
1554
365k
    }
1555
1556
    // check _stale_rs_version_map to remove meta from rowset meta store
1557
205k
    for (auto& rs_meta : _tablet_meta->all_stale_rs_metas()) {
1558
        // If we delete it from rowset manager's meta explicitly in previous checkpoint, just skip.
1559
98.1k
        if (rs_meta->is_remove_from_rowset_meta()) {
1560
12.8k
            continue;
1561
12.8k
        }
1562
85.3k
        if (RowsetMetaManager::check_rowset_meta(_data_dir->get_meta(), tablet_uid(),
1563
85.3k
                                                 rs_meta->rowset_id())) {
1564
85.2k
            RETURN_FALSE_IF_ERROR(RowsetMetaManager::remove(_data_dir->get_meta(), tablet_uid(),
1565
85.2k
                                                            rs_meta->rowset_id()));
1566
85.2k
            VLOG_NOTICE << "remove rowset id from meta store because it is already persistent with "
1567
0
                        << "tablet meta, rowset_id=" << rs_meta->rowset_id();
1568
85.2k
        }
1569
85.3k
        rs_meta->set_remove_from_rowset_meta();
1570
85.3k
    }
1571
1572
205k
    if (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write()) {
1573
37.0k
        RETURN_FALSE_IF_ERROR(TabletMetaManager::remove_old_version_delete_bitmap(
1574
37.0k
                _data_dir, tablet_id(), max_version_unlocked()));
1575
37.0k
    }
1576
1577
205k
    _newly_created_rowset_num = 0;
1578
205k
    _last_checkpoint_time = UnixMillis();
1579
205k
    return true;
1580
205k
}
1581
1582
333k
bool Tablet::rowset_meta_is_useful(RowsetMetaSharedPtr rowset_meta) {
1583
333k
    std::shared_lock rdlock(_meta_lock);
1584
333k
    bool find_version = false;
1585
1.83M
    for (auto& version_rowset : _rs_version_map) {
1586
1.83M
        if (version_rowset.second->rowset_id() == rowset_meta->rowset_id()) {
1587
217k
            return true;
1588
217k
        }
1589
1.61M
        if (version_rowset.second->contains_version(rowset_meta->version())) {
1590
115k
            find_version = true;
1591
115k
        }
1592
1.61M
    }
1593
31.6M
    for (auto& stale_version_rowset : _stale_rs_version_map) {
1594
31.6M
        if (stale_version_rowset.second->rowset_id() == rowset_meta->rowset_id()) {
1595
111k
            return true;
1596
111k
        }
1597
31.5M
        if (stale_version_rowset.second->contains_version(rowset_meta->version())) {
1598
414k
            find_version = true;
1599
414k
        }
1600
31.5M
    }
1601
4.19k
    return !find_version;
1602
115k
}
1603
1604
518k
bool Tablet::_contains_rowset(const RowsetId rowset_id) {
1605
2.49M
    for (auto& version_rowset : _rs_version_map) {
1606
2.49M
        if (version_rowset.second->rowset_id() == rowset_id) {
1607
4
            return true;
1608
4
        }
1609
2.49M
    }
1610
32.1M
    for (auto& stale_version_rowset : _stale_rs_version_map) {
1611
32.1M
        if (stale_version_rowset.second->rowset_id() == rowset_id) {
1612
20
            return true;
1613
20
        }
1614
32.1M
    }
1615
518k
    return false;
1616
518k
}
1617
1618
// need check if consecutive version missing in full report
1619
// alter tablet will ignore this check
1620
void Tablet::build_tablet_report_info(TTabletInfo* tablet_info,
1621
                                      bool enable_consecutive_missing_check,
1622
20.5M
                                      bool enable_path_check) {
1623
20.5M
    std::shared_lock rdlock(_meta_lock);
1624
20.5M
    tablet_info->__set_tablet_id(_tablet_meta->tablet_id());
1625
20.5M
    tablet_info->__set_schema_hash(_tablet_meta->schema_hash());
1626
20.5M
    tablet_info->__set_row_count(_tablet_meta->num_rows());
1627
20.5M
    tablet_info->__set_data_size(_tablet_meta->tablet_local_size());
1628
1629
    // Here we need to report to FE if there are any missing versions of tablet.
1630
    // We start from the initial version and traverse backwards until we meet a discontinuous version.
1631
20.5M
    Version cversion;
1632
20.5M
    Version max_version;
1633
20.5M
    bool has_version_cross;
1634
20.5M
    _max_continuous_version_from_beginning_unlocked(&cversion, &max_version, &has_version_cross);
1635
    // cause publish version task runs concurrently, version may be flying
1636
    // so we add a consecutive miss check to solve this problem:
1637
    // if publish version 5 arrives but version 4 flying, we may judge replica miss version
1638
    // and set version miss in tablet_info, which makes fe treat this replica as unhealth
1639
    // and lead to other problems
1640
20.5M
    if (enable_consecutive_missing_check) {
1641
20.5M
        if (cversion.second < max_version.second) {
1642
17
            if (_last_missed_version == cversion.second + 1) {
1643
4
                if (MonotonicSeconds() - _last_missed_time_s >= 60) {
1644
                    // version missed for over 60 seconds
1645
0
                    tablet_info->__set_version_miss(true);
1646
0
                    _last_missed_version = -1;
1647
0
                    _last_missed_time_s = 0;
1648
0
                }
1649
13
            } else {
1650
13
                _last_missed_version = cversion.second + 1;
1651
13
                _last_missed_time_s = MonotonicSeconds();
1652
13
            }
1653
17
        }
1654
20.5M
    } else {
1655
11.3k
        tablet_info->__set_version_miss(cversion.second < max_version.second);
1656
11.3k
    }
1657
1658
20.5M
    DBUG_EXECUTE_IF("Tablet.build_tablet_report_info.version_miss", {
1659
20.5M
        auto tablet_id = dp->param<int64>("tablet_id", -1);
1660
20.5M
        if (tablet_id != -1 && tablet_id == _tablet_meta->tablet_id()) {
1661
20.5M
            auto miss = dp->param<bool>("version_miss", true);
1662
20.5M
            tablet_info->__set_version_miss(miss);
1663
20.5M
        }
1664
20.5M
    });
1665
1666
    // find rowset with max version
1667
20.5M
    auto iter = _rs_version_map.find(max_version);
1668
20.5M
    if (iter == _rs_version_map.end()) {
1669
        // If the tablet is in running state, it must not be doing schema-change. so if we can not
1670
        // access its rowsets, it means that the tablet is bad and needs to be reported to the FE
1671
        // for subsequent repairs (through the cloning task)
1672
9
        if (tablet_state() == TABLET_RUNNING) {
1673
0
            tablet_info->__set_used(false);
1674
0
        }
1675
        // For other states, FE knows that the tablet is in a certain change process, so here
1676
        // still sets the state to normal when reporting. Note that every task has an timeout,
1677
        // so if the task corresponding to this change hangs, when the task timeout, FE will know
1678
        // and perform state modification operations.
1679
9
    }
1680
1681
20.5M
    if (tablet_state() == TABLET_RUNNING) {
1682
20.5M
        if (has_version_cross || is_io_error_too_times() || !data_dir()->is_used()) {
1683
0
            LOG(INFO) << "report " << tablet_id() << " as bad, version_cross=" << has_version_cross
1684
0
                      << ", ioe times=" << get_io_error_times() << ", data_dir used "
1685
0
                      << data_dir()->is_used();
1686
0
            tablet_info->__set_used(false);
1687
0
        }
1688
1689
20.5M
        if (enable_path_check) {
1690
20.5M
            if (!_is_tablet_path_exists.exchange(true, std::memory_order_relaxed)) {
1691
0
                LOG(INFO) << "report " << tablet_id() << " as bad, tablet directory not found";
1692
0
                tablet_info->__set_used(false);
1693
0
            }
1694
20.5M
        }
1695
20.5M
    }
1696
1697
    // There are two cases when tablet state is TABLET_NOTREADY
1698
    // case 1: tablet is doing schema change. Fe knows it's state, doing nothing.
1699
    // case 2: tablet has finished schema change, but failed. Fe will perform recovery.
1700
20.5M
    if (tablet_state() == TABLET_NOTREADY && is_alter_failed()) {
1701
1.11k
        tablet_info->__set_used(false);
1702
1.11k
    }
1703
1704
20.5M
    if (tablet_state() == TABLET_SHUTDOWN) {
1705
27
        tablet_info->__set_used(false);
1706
27
    }
1707
1708
20.5M
    DBUG_EXECUTE_IF("Tablet.build_tablet_report_info.used", {
1709
20.5M
        auto tablet_id = dp->param<int64>("tablet_id", -1);
1710
20.5M
        if (tablet_id != -1 && tablet_id == _tablet_meta->tablet_id()) {
1711
20.5M
            auto used = dp->param<bool>("used", true);
1712
20.5M
            LOG_WARNING("Tablet.build_tablet_report_info.used")
1713
20.5M
                    .tag("tablet id", tablet_id)
1714
20.5M
                    .tag("used", used);
1715
20.5M
            tablet_info->__set_used(used);
1716
20.5M
        } else {
1717
20.5M
            LOG_WARNING("Tablet.build_tablet_report_info.used").tag("tablet id", tablet_id);
1718
20.5M
        }
1719
20.5M
    });
1720
1721
20.5M
    int64_t total_version_count = _tablet_meta->version_count();
1722
1723
    // For compatibility.
1724
    // For old fe, it wouldn't send visible version request to be, then be's visible version is always 0.
1725
    // Let visible_version_count set to total_version_count in be's report.
1726
20.5M
    int64_t visible_version_count = total_version_count;
1727
20.5M
    if (auto [visible_version, _] = get_visible_version_and_time(); visible_version > 0) {
1728
17.9M
        visible_version_count = _tablet_meta->version_count_cross_with_range({0, visible_version});
1729
17.9M
    }
1730
    // the report version is the largest continuous version, same logic as in FE side
1731
20.5M
    tablet_info->__set_version(cversion.second);
1732
    // Useless but it is a required filed in TTabletInfo
1733
20.5M
    tablet_info->__set_version_hash(0);
1734
20.5M
    tablet_info->__set_partition_id(_tablet_meta->partition_id());
1735
20.5M
    tablet_info->__set_storage_medium(_data_dir->storage_medium());
1736
20.5M
    tablet_info->__set_total_version_count(total_version_count);
1737
20.5M
    tablet_info->__set_visible_version_count(visible_version_count);
1738
20.5M
    tablet_info->__set_path_hash(_data_dir->path_hash());
1739
20.5M
    tablet_info->__set_is_in_memory(_tablet_meta->tablet_schema()->is_in_memory());
1740
20.5M
    tablet_info->__set_replica_id(replica_id());
1741
20.5M
    tablet_info->__set_remote_data_size(_tablet_meta->tablet_remote_size());
1742
20.5M
    if (_tablet_meta->cooldown_meta_id().initialized()) { // has cooldowned data
1743
265
        tablet_info->__set_cooldown_term(_cooldown_conf.term);
1744
265
        tablet_info->__set_cooldown_meta_id(_tablet_meta->cooldown_meta_id().to_thrift());
1745
265
    }
1746
20.5M
    if (tablet_state() == TABLET_RUNNING && _tablet_meta->storage_policy_id() > 0) {
1747
        // tablet may not have cooldowned data, but the storage policy is set
1748
33.4k
        tablet_info->__set_cooldown_term(_cooldown_conf.term);
1749
33.4k
    }
1750
20.5M
    tablet_info->__set_local_index_size(_tablet_meta->tablet_local_index_size());
1751
20.5M
    tablet_info->__set_local_segment_size(_tablet_meta->tablet_local_segment_size());
1752
20.5M
    tablet_info->__set_remote_index_size(_tablet_meta->tablet_remote_index_size());
1753
20.5M
    tablet_info->__set_remote_segment_size(_tablet_meta->tablet_remote_segment_size());
1754
20.5M
}
1755
1756
595
void Tablet::report_error(const Status& st) {
1757
595
    if (st.is<ErrorCode::IO_ERROR>()) {
1758
0
        ++_io_error_times;
1759
595
    } else if (st.is<ErrorCode::CORRUPTION>()) {
1760
1
        _io_error_times = config::max_tablet_io_errors + 1;
1761
594
    } else if (st.is<ErrorCode::NOT_FOUND>()) {
1762
0
        check_tablet_path_exists();
1763
0
        if (!_is_tablet_path_exists.load(std::memory_order_relaxed)) {
1764
0
            _io_error_times = config::max_tablet_io_errors + 1;
1765
0
        }
1766
0
    }
1767
595
}
1768
1769
Status Tablet::prepare_compaction_and_calculate_permits(
1770
        CompactionType compaction_type, const TabletSharedPtr& tablet,
1771
41.2k
        std::shared_ptr<CompactionMixin>& compaction, int64_t& permits) {
1772
41.2k
    if (compaction_type == CompactionType::CUMULATIVE_COMPACTION) {
1773
36.5k
        MonotonicStopWatch watch;
1774
36.5k
        watch.start();
1775
1776
36.5k
        compaction = std::make_shared<CumulativeCompaction>(tablet->_engine, tablet);
1777
36.5k
        DorisMetrics::instance()->cumulative_compaction_request_total->increment(1);
1778
36.5k
        Status res = compaction->prepare_compact();
1779
36.5k
        if (!config::disable_compaction_trace_log &&
1780
36.5k
            watch.elapsed_time() / 1e9 > config::cumulative_compaction_trace_threshold) {
1781
0
            std::stringstream ss;
1782
0
            compaction->runtime_profile()->pretty_print(&ss);
1783
0
            LOG(WARNING) << "prepare cumulative compaction cost " << watch.elapsed_time() / 1e9
1784
0
                         << std::endl
1785
0
                         << ss.str();
1786
0
        }
1787
1788
36.5k
        if (!res.ok()) {
1789
23.9k
            permits = 0;
1790
            // if we meet a delete version, should increase the cumulative point to let base compaction handle the delete version.
1791
            // no need to wait 5s.
1792
23.9k
            if (!(res.msg() == "_last_delete_version.first not equal to -1") ||
1793
23.9k
                config::enable_sleep_between_delete_cumu_compaction) {
1794
23.8k
                tablet->set_last_cumu_compaction_failure_time(UnixMillis());
1795
23.8k
            }
1796
23.9k
            if (!res.is<CUMULATIVE_NO_SUITABLE_VERSION>()) {
1797
1
                DorisMetrics::instance()->cumulative_compaction_request_failed->increment(1);
1798
1
                return Status::InternalError("prepare cumulative compaction with err: {}",
1799
1
                                             res.to_string());
1800
1
            }
1801
            // return OK if OLAP_ERR_CUMULATIVE_NO_SUITABLE_VERSION, so that we don't need to
1802
            // print too much useless logs.
1803
            // And because we set permits to 0, so even if we return OK here, nothing will be done.
1804
23.9k
            return Status::OK();
1805
23.9k
        }
1806
36.5k
    } else if (compaction_type == CompactionType::BASE_COMPACTION) {
1807
4.52k
        MonotonicStopWatch watch;
1808
4.52k
        watch.start();
1809
1810
4.52k
        compaction = std::make_shared<BaseCompaction>(tablet->_engine, tablet);
1811
4.52k
        DorisMetrics::instance()->base_compaction_request_total->increment(1);
1812
4.52k
        Status res = compaction->prepare_compact();
1813
4.52k
        if (!config::disable_compaction_trace_log &&
1814
4.52k
            watch.elapsed_time() / 1e9 > config::base_compaction_trace_threshold) {
1815
0
            std::stringstream ss;
1816
0
            compaction->runtime_profile()->pretty_print(&ss);
1817
0
            LOG(WARNING) << "prepare base compaction cost " << watch.elapsed_time() / 1e9
1818
0
                         << std::endl
1819
0
                         << ss.str();
1820
0
        }
1821
1822
4.52k
        tablet->set_last_base_compaction_status(res.to_string());
1823
4.52k
        if (!res.ok()) {
1824
4.43k
            tablet->set_last_base_compaction_failure_time(UnixMillis());
1825
4.43k
            permits = 0;
1826
4.43k
            if (!res.is<BE_NO_SUITABLE_VERSION>()) {
1827
1
                DorisMetrics::instance()->base_compaction_request_failed->increment(1);
1828
1
                return Status::InternalError("prepare base compaction with err: {}",
1829
1
                                             res.to_string());
1830
1
            }
1831
            // return OK if OLAP_ERR_BE_NO_SUITABLE_VERSION, so that we don't need to
1832
            // print too much useless logs.
1833
            // And because we set permits to 0, so even if we return OK here, nothing will be done.
1834
4.42k
            return Status::OK();
1835
4.43k
        }
1836
4.52k
    } else {
1837
112
        DCHECK_EQ(compaction_type, CompactionType::FULL_COMPACTION);
1838
1839
112
        compaction = std::make_shared<FullCompaction>(tablet->_engine, tablet);
1840
112
        Status res = compaction->prepare_compact();
1841
112
        if (!res.ok()) {
1842
102
            tablet->set_last_full_compaction_failure_time(UnixMillis());
1843
102
            permits = 0;
1844
102
            if (!res.is<FULL_NO_SUITABLE_VERSION>()) {
1845
0
                return Status::InternalError("prepare full compaction with err: {}",
1846
0
                                             res.to_string());
1847
0
            }
1848
            // return OK if OLAP_ERR_BE_NO_SUITABLE_VERSION, so that we don't need to
1849
            // print too much useless logs.
1850
            // And because we set permits to 0, so even if we return OK here, nothing will be done.
1851
102
            return Status::OK();
1852
102
        }
1853
112
    }
1854
1855
    // Time series policy does not rely on permits, it uses goal size to control memory
1856
12.7k
    if (tablet->tablet_meta()->compaction_policy() == CUMULATIVE_TIME_SERIES_POLICY) {
1857
        // permits = 0 means that prepare_compaction failed
1858
0
        permits = 1;
1859
12.7k
    } else {
1860
12.7k
        permits = compaction->get_compaction_permits();
1861
12.7k
    }
1862
12.7k
    return Status::OK();
1863
41.2k
}
1864
1865
0
void Tablet::execute_single_replica_compaction(SingleReplicaCompaction& compaction) {
1866
0
    Status res = compaction.execute_compact();
1867
0
    if (!res.ok()) {
1868
0
        set_last_failure_time(this, compaction, UnixMillis());
1869
0
        set_last_single_compaction_failure_status(res.to_string());
1870
0
        if (res.is<CANCELLED>()) {
1871
0
            DorisMetrics::instance()->single_compaction_request_cancelled->increment(1);
1872
            // "CANCELLED" indicates that the peer has not performed compaction,
1873
            // wait for the peer to perform compaction
1874
0
            set_skip_compaction(true, compaction.real_compact_type(), UnixSeconds());
1875
0
            VLOG_CRITICAL << "Cannel fetching from the remote peer. res=" << res
1876
0
                          << ", tablet=" << tablet_id();
1877
0
        } else {
1878
0
            DorisMetrics::instance()->single_compaction_request_failed->increment(1);
1879
0
            LOG(WARNING) << "failed to do single replica compaction. res=" << res
1880
0
                         << ", tablet=" << tablet_id();
1881
0
        }
1882
0
        return;
1883
0
    }
1884
0
    set_last_failure_time(this, compaction, 0);
1885
0
}
1886
1887
157M
bool Tablet::should_fetch_from_peer() {
1888
157M
    return tablet_meta()->tablet_schema()->enable_single_replica_compaction() &&
1889
157M
           _engine.should_fetch_from_peer(tablet_id());
1890
157M
}
1891
1892
3
std::vector<Version> Tablet::get_all_local_versions() {
1893
3
    std::vector<Version> local_versions;
1894
3
    {
1895
3
        std::shared_lock rlock(_meta_lock);
1896
63
        for (const auto& [version, rs] : _rs_version_map) {
1897
63
            if (rs->is_local()) {
1898
53
                local_versions.emplace_back(version);
1899
53
            }
1900
63
        }
1901
3
    }
1902
3
    std::sort(local_versions.begin(), local_versions.end(),
1903
318
              [](const Version& left, const Version& right) { return left.first < right.first; });
1904
3
    return local_versions;
1905
3
}
1906
1907
12.7k
void Tablet::execute_compaction(CompactionMixin& compaction) {
1908
12.7k
    signal::tablet_id = tablet_id();
1909
1910
12.7k
    MonotonicStopWatch watch;
1911
12.7k
    watch.start();
1912
1913
12.7k
    Status res = [&]() { RETURN_IF_CATCH_EXCEPTION({ return compaction.execute_compact(); }); }();
1914
1915
12.7k
    if (!res.ok()) [[unlikely]] {
1916
0
        set_last_failure_time(this, compaction, UnixMillis());
1917
0
        LOG(WARNING) << "failed to do " << compaction.compaction_name()
1918
0
                     << ", tablet=" << tablet_id() << " : " << res;
1919
12.7k
    } else {
1920
12.7k
        set_last_failure_time(this, compaction, 0);
1921
12.7k
    }
1922
1923
12.7k
    if (!config::disable_compaction_trace_log) {
1924
0
        auto need_trace = [&compaction, &watch] {
1925
0
            return compaction.compaction_type() == ReaderType::READER_CUMULATIVE_COMPACTION
1926
0
                           ? watch.elapsed_time() / 1e9 >
1927
0
                                     config::cumulative_compaction_trace_threshold
1928
0
                   : compaction.compaction_type() == ReaderType::READER_BASE_COMPACTION
1929
0
                           ? watch.elapsed_time() / 1e9 > config::base_compaction_trace_threshold
1930
0
                           : false;
1931
0
        };
1932
0
        if (need_trace()) {
1933
0
            std::stringstream ss;
1934
0
            compaction.runtime_profile()->pretty_print(&ss);
1935
0
            LOG(WARNING) << "execute " << compaction.compaction_name() << " cost "
1936
0
                         << watch.elapsed_time() / 1e9 << std::endl
1937
0
                         << ss.str();
1938
0
        }
1939
0
    }
1940
12.7k
}
1941
1942
231k
Status Tablet::create_initial_rowset(const int64_t req_version) {
1943
231k
    if (req_version < 1) {
1944
0
        return Status::Error<CE_CMD_PARAMS_ERROR>(
1945
0
                "init version of tablet should at least 1. req.ver={}", req_version);
1946
0
    }
1947
231k
    Version version(0, req_version);
1948
231k
    RowsetSharedPtr new_rowset;
1949
    // there is no data in init rowset, so overlapping info is unknown.
1950
231k
    RowsetWriterContext context;
1951
231k
    context.version = version;
1952
231k
    context.rowset_state = VISIBLE;
1953
231k
    context.segments_overlap = OVERLAP_UNKNOWN;
1954
231k
    context.tablet_schema = tablet_schema();
1955
231k
    context.newest_write_timestamp = UnixSeconds();
1956
231k
    auto rs_writer = DORIS_TRY(create_rowset_writer(context, false));
1957
231k
    RETURN_IF_ERROR(rs_writer->flush());
1958
231k
    RETURN_IF_ERROR(rs_writer->build(new_rowset));
1959
231k
    RETURN_IF_ERROR(add_rowset(std::move(new_rowset)));
1960
231k
    set_cumulative_layer_point(req_version + 1);
1961
231k
    return Status::OK();
1962
231k
}
1963
1964
Result<std::unique_ptr<RowsetWriter>> Tablet::create_rowset_writer(RowsetWriterContext& context,
1965
528k
                                                                   bool vertical) {
1966
528k
    context.rowset_id = _engine.next_rowset_id();
1967
528k
    _init_context_common_fields(context);
1968
528k
    return RowsetFactory::create_rowset_writer(_engine, context, vertical);
1969
528k
}
1970
1971
// create a rowset writer with rowset_id and seg_id
1972
// after writer, merge this transient rowset with original rowset
1973
Result<std::unique_ptr<RowsetWriter>> Tablet::create_transient_rowset_writer(
1974
        const Rowset& rowset, std::shared_ptr<PartialUpdateInfo> partial_update_info,
1975
2.53k
        int64_t txn_expiration) {
1976
2.53k
    RowsetWriterContext context;
1977
2.53k
    context.rowset_state = PREPARED;
1978
2.53k
    context.segments_overlap = OVERLAPPING;
1979
2.53k
    context.tablet_schema = std::make_shared<TabletSchema>();
1980
    // During a partial update, the extracted columns of a variant should not be included in the tablet schema.
1981
    // This is because the partial update for a variant needs to ignore the extracted columns.
1982
    // Otherwise, the schema types in different rowsets might be inconsistent. When performing a partial update,
1983
    // the complete variant is constructed by reading all the sub-columns of the variant.
1984
2.53k
    context.tablet_schema = rowset.tablet_schema()->copy_without_variant_extracted_columns();
1985
2.53k
    context.newest_write_timestamp = UnixSeconds();
1986
2.53k
    context.tablet_id = table_id();
1987
2.53k
    context.enable_segcompaction = false;
1988
    // ATTN: context.tablet is a shared_ptr, can't simply set it's value to `this`. We should
1989
    // get the shared_ptr from tablet_manager.
1990
2.53k
    auto tablet = _engine.tablet_manager()->get_tablet(tablet_id());
1991
2.53k
    if (!tablet) {
1992
0
        LOG(WARNING) << "cant find tablet by tablet_id=" << tablet_id();
1993
0
        return ResultError(Status::NotFound("cant find tablet by tablet_id={}", tablet_id()));
1994
0
    }
1995
2.53k
    context.tablet = tablet;
1996
2.53k
    context.write_type = DataWriteType::TYPE_DIRECT;
1997
2.53k
    context.partial_update_info = std::move(partial_update_info);
1998
2.53k
    context.is_transient_rowset_writer = true;
1999
2.53k
    return create_transient_rowset_writer(context, rowset.rowset_id())
2000
2.53k
            .transform([&](auto&& writer) {
2001
2.53k
                writer->set_segment_start_id(rowset.num_segments());
2002
2.53k
                return writer;
2003
2.53k
            });
2004
2.53k
}
2005
2006
Result<std::unique_ptr<RowsetWriter>> Tablet::create_transient_rowset_writer(
2007
2.53k
        RowsetWriterContext& context, const RowsetId& rowset_id) {
2008
2.53k
    context.rowset_id = rowset_id;
2009
2.53k
    _init_context_common_fields(context);
2010
2.53k
    return RowsetFactory::create_rowset_writer(_engine, context, false);
2011
2.53k
}
2012
2013
530k
void Tablet::_init_context_common_fields(RowsetWriterContext& context) {
2014
530k
    context.tablet_uid = tablet_uid();
2015
530k
    context.tablet_id = tablet_id();
2016
530k
    context.partition_id = partition_id();
2017
530k
    context.tablet_schema_hash = schema_hash();
2018
530k
    context.rowset_type = tablet_meta()->preferred_rowset_type();
2019
    // Alpha Rowset will be removed in the future, so that if the tablet's default rowset type is
2020
    // alpha rowset, then set the newly created rowset to storage engine's default rowset.
2021
530k
    if (context.rowset_type == ALPHA_ROWSET) {
2022
0
        context.rowset_type = _engine.default_rowset_type();
2023
0
    }
2024
2025
530k
    if (context.is_local_rowset()) {
2026
530k
        context.tablet_path = _tablet_path;
2027
530k
    }
2028
2029
530k
    context.data_dir = data_dir();
2030
530k
    context.enable_unique_key_merge_on_write = enable_unique_key_merge_on_write();
2031
530k
}
2032
2033
4.71k
Status Tablet::create_rowset(const RowsetMetaSharedPtr& rowset_meta, RowsetSharedPtr* rowset) {
2034
4.71k
    return RowsetFactory::create_rowset(_tablet_meta->tablet_schema(),
2035
4.71k
                                        rowset_meta->is_local() ? _tablet_path : "", rowset_meta,
2036
4.71k
                                        rowset);
2037
4.71k
}
2038
2039
29
Status Tablet::cooldown(RowsetSharedPtr rowset) {
2040
29
    std::unique_lock schema_change_lock(_schema_change_lock, std::try_to_lock);
2041
29
    if (!schema_change_lock.owns_lock()) {
2042
0
        return Status::Error<TRY_LOCK_FAILED>(
2043
0
                "try schema_change_lock failed, schema change running or inverted index built on "
2044
0
                "this tablet={}",
2045
0
                tablet_id());
2046
0
    }
2047
    // Check executing serially with compaction task.
2048
29
    std::unique_lock base_compaction_lock(_base_compaction_lock, std::try_to_lock);
2049
29
    if (!base_compaction_lock.owns_lock()) {
2050
0
        return Status::Error<TRY_LOCK_FAILED>("try base_compaction_lock failed");
2051
0
    }
2052
29
    std::unique_lock cumu_compaction_lock(_cumulative_compaction_lock, std::try_to_lock);
2053
29
    if (!cumu_compaction_lock.owns_lock()) {
2054
0
        return Status::Error<TRY_LOCK_FAILED>("try cumu_compaction_lock failed");
2055
0
    }
2056
29
    std::shared_lock cooldown_conf_rlock(_cooldown_conf_lock);
2057
29
    if (_cooldown_conf.cooldown_replica_id <= 0) { // wait for FE to push cooldown conf
2058
14
        return Status::InternalError("invalid cooldown_replica_id");
2059
14
    }
2060
2061
15
    auto mem_tracker = MemTrackerLimiter::create_shared(
2062
15
            MemTrackerLimiter::Type::OTHER,
2063
15
            fmt::format("Tablet::cooldown#tableId={}:replicaId={}", std::to_string(tablet_id()),
2064
15
                        std::to_string(replica_id())));
2065
15
    SCOPED_ATTACH_TASK(mem_tracker);
2066
2067
15
    if (_cooldown_conf.cooldown_replica_id == replica_id()) {
2068
        // this replica is cooldown replica
2069
15
        RETURN_IF_ERROR(_cooldown_data(std::move(rowset)));
2070
15
    } else {
2071
0
        Status st = _follow_cooldowned_data();
2072
0
        if (UNLIKELY(!st.ok())) {
2073
0
            _last_failed_follow_cooldown_time = time(nullptr);
2074
0
            return st;
2075
0
        }
2076
0
        _last_failed_follow_cooldown_time = 0;
2077
0
    }
2078
15
    return Status::OK();
2079
15
}
2080
2081
// hold SHARED `cooldown_conf_lock`
2082
15
Status Tablet::_cooldown_data(RowsetSharedPtr rowset) {
2083
15
    DCHECK(_cooldown_conf.cooldown_replica_id == replica_id());
2084
2085
15
    auto storage_resource = DORIS_TRY(get_resource_by_storage_policy_id(storage_policy_id()));
2086
15
    RowsetSharedPtr old_rowset = nullptr;
2087
2088
15
    if (rowset) {
2089
0
        const auto& rowset_id = rowset->rowset_id();
2090
0
        const auto& rowset_version = rowset->version();
2091
0
        std::shared_lock meta_rlock(_meta_lock);
2092
0
        auto iter = _rs_version_map.find(rowset_version);
2093
0
        if (iter != _rs_version_map.end() && iter->second->rowset_id() == rowset_id) {
2094
0
            old_rowset = rowset;
2095
0
        }
2096
0
    }
2097
2098
15
    if (!old_rowset) {
2099
15
        old_rowset = pick_cooldown_rowset();
2100
15
    }
2101
2102
15
    if (!old_rowset) {
2103
0
        LOG(INFO) << "cannot pick cooldown rowset in tablet " << tablet_id();
2104
0
        return Status::OK();
2105
0
    }
2106
2107
15
    RowsetId new_rowset_id = _engine.next_rowset_id();
2108
15
    auto pending_rs_guard = _engine.pending_remote_rowsets().add(new_rowset_id);
2109
15
    Status st;
2110
15
    Defer defer {[&] {
2111
15
        if (!st.ok()) {
2112
            // reclaim the incomplete rowset data in remote storage
2113
0
            record_unused_remote_rowset(new_rowset_id, storage_resource.fs->id(),
2114
0
                                        old_rowset->num_segments());
2115
0
        }
2116
15
    }};
2117
15
    auto start = std::chrono::steady_clock::now();
2118
15
    if (st = old_rowset->upload_to(storage_resource, new_rowset_id); !st.ok()) {
2119
0
        return st;
2120
0
    }
2121
2122
15
    auto duration = std::chrono::duration<float>(std::chrono::steady_clock::now() - start);
2123
15
    LOG(INFO) << "Upload rowset " << old_rowset->version() << " " << new_rowset_id.to_string()
2124
15
              << " to " << storage_resource.fs->root_path().native()
2125
15
              << ", tablet_id=" << tablet_id() << ", duration=" << duration.count()
2126
15
              << ", capacity=" << old_rowset->total_disk_size()
2127
15
              << ", tp=" << old_rowset->total_disk_size() / duration.count()
2128
15
              << ", old rowset_id=" << old_rowset->rowset_id().to_string();
2129
2130
    // gen a new rowset
2131
15
    auto new_rowset_meta = std::make_shared<RowsetMeta>();
2132
15
    new_rowset_meta->init(old_rowset->rowset_meta().get());
2133
15
    new_rowset_meta->set_rowset_id(new_rowset_id);
2134
15
    new_rowset_meta->set_remote_storage_resource(std::move(storage_resource));
2135
15
    new_rowset_meta->set_creation_time(time(nullptr));
2136
15
    UniqueId cooldown_meta_id = UniqueId::gen_uid();
2137
15
    RowsetSharedPtr new_rowset;
2138
15
    RETURN_IF_ERROR(RowsetFactory::create_rowset(_tablet_meta->tablet_schema(), "", new_rowset_meta,
2139
15
                                                 &new_rowset));
2140
2141
15
    {
2142
15
        std::unique_lock meta_wlock(_meta_lock);
2143
15
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2144
15
        if (tablet_state() == TABLET_RUNNING) {
2145
15
            RETURN_IF_ERROR(delete_rowsets({std::move(old_rowset)}, false));
2146
15
            add_rowsets({std::move(new_rowset)});
2147
            // TODO(plat1ko): process primary key
2148
15
            _tablet_meta->set_cooldown_meta_id(cooldown_meta_id);
2149
15
        }
2150
15
    }
2151
15
    {
2152
15
        std::shared_lock meta_rlock(_meta_lock);
2153
15
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2154
15
        save_meta();
2155
15
    }
2156
    // Upload cooldowned rowset meta to remote fs
2157
    // ATTN: Even if it is an empty rowset, in order for the followers to synchronize, the coolown meta must be
2158
    // uploaded, otherwise followers may never completely cooldown.
2159
15
    if (auto t = _engine.tablet_manager()->get_tablet(tablet_id());
2160
15
        t != nullptr) { // `t` can be nullptr if it has been dropped
2161
15
        async_write_cooldown_meta(std::move(t));
2162
15
    }
2163
15
    return Status::OK();
2164
15
}
2165
2166
// hold SHARED `cooldown_conf_lock`
2167
Status Tablet::_read_cooldown_meta(const StorageResource& storage_resource,
2168
0
                                   TabletMetaPB* tablet_meta_pb) {
2169
0
    std::string remote_meta_path = storage_resource.cooldown_tablet_meta_path(
2170
0
            tablet_id(), _cooldown_conf.cooldown_replica_id, _cooldown_conf.term);
2171
0
    io::FileReaderSPtr tablet_meta_reader;
2172
0
    RETURN_IF_ERROR(storage_resource.fs->open_file(remote_meta_path, &tablet_meta_reader));
2173
0
    auto file_size = tablet_meta_reader->size();
2174
0
    size_t bytes_read;
2175
0
    auto buf = std::unique_ptr<uint8_t[]>(new uint8_t[file_size]);
2176
0
    RETURN_IF_ERROR(tablet_meta_reader->read_at(0, {buf.get(), file_size}, &bytes_read));
2177
0
    RETURN_IF_ERROR(tablet_meta_reader->close());
2178
0
    if (!tablet_meta_pb->ParseFromArray(buf.get(), file_size)) {
2179
0
        return Status::InternalError("malformed tablet meta, path={}/{}",
2180
0
                                     storage_resource.fs->root_path().native(), remote_meta_path);
2181
0
    }
2182
0
    return Status::OK();
2183
0
}
2184
2185
// `rs_metas` MUST already be sorted by `RowsetMeta::comparator`
2186
17
Status check_version_continuity(const std::vector<RowsetMetaSharedPtr>& rs_metas) {
2187
17
    if (rs_metas.size() < 2) {
2188
6
        return Status::OK();
2189
6
    }
2190
11
    auto prev = rs_metas.begin();
2191
36
    for (auto it = rs_metas.begin() + 1; it != rs_metas.end(); ++it) {
2192
25
        if ((*prev)->end_version() + 1 != (*it)->start_version()) {
2193
0
            return Status::InternalError("versions are not continuity: prev={} cur={}",
2194
0
                                         (*prev)->version().to_string(),
2195
0
                                         (*it)->version().to_string());
2196
0
        }
2197
25
        prev = it;
2198
25
    }
2199
11
    return Status::OK();
2200
11
}
2201
2202
// It's guaranteed the write cooldown meta task would be invoked at the end unless BE crashes
2203
// one tablet would at most have one async task to be done
2204
17
void Tablet::async_write_cooldown_meta(TabletSharedPtr tablet) {
2205
17
    ExecEnv::GetInstance()->write_cooldown_meta_executors()->submit(std::move(tablet));
2206
17
}
2207
2208
166
bool Tablet::update_cooldown_conf(int64_t cooldown_term, int64_t cooldown_replica_id) {
2209
166
    std::unique_lock wlock(_cooldown_conf_lock, std::try_to_lock);
2210
166
    if (!wlock.owns_lock()) {
2211
0
        LOG(INFO) << "try cooldown_conf_lock failed, tablet_id=" << tablet_id();
2212
0
        return false;
2213
0
    }
2214
166
    if (cooldown_term <= _cooldown_conf.term) {
2215
0
        return false;
2216
0
    }
2217
166
    LOG(INFO) << "update cooldown conf. tablet_id=" << tablet_id()
2218
166
              << " cooldown_replica_id: " << _cooldown_conf.cooldown_replica_id << " -> "
2219
166
              << cooldown_replica_id << ", cooldown_term: " << _cooldown_conf.term << " -> "
2220
166
              << cooldown_term;
2221
166
    _cooldown_conf.cooldown_replica_id = cooldown_replica_id;
2222
166
    _cooldown_conf.term = cooldown_term;
2223
166
    return true;
2224
166
}
2225
2226
17
Status Tablet::write_cooldown_meta() {
2227
17
    std::shared_lock rlock(_cooldown_conf_lock);
2228
17
    if (_cooldown_conf.cooldown_replica_id != _tablet_meta->replica_id()) {
2229
0
        return Status::Aborted<false>("not cooldown replica({} vs {}) tablet_id={}",
2230
0
                                      _tablet_meta->replica_id(),
2231
0
                                      _cooldown_conf.cooldown_replica_id, tablet_id());
2232
0
    }
2233
2234
17
    auto storage_resource = DORIS_TRY(get_resource_by_storage_policy_id(storage_policy_id()));
2235
2236
17
    std::vector<RowsetMetaSharedPtr> cooldowned_rs_metas;
2237
17
    UniqueId cooldown_meta_id;
2238
17
    {
2239
17
        std::shared_lock meta_rlock(_meta_lock);
2240
65
        for (auto& rs_meta : _tablet_meta->all_rs_metas()) {
2241
65
            if (!rs_meta->is_local()) {
2242
42
                cooldowned_rs_metas.push_back(rs_meta);
2243
42
            }
2244
65
        }
2245
17
        cooldown_meta_id = _tablet_meta->cooldown_meta_id();
2246
17
    }
2247
17
    if (cooldowned_rs_metas.empty()) {
2248
0
        LOG(INFO) << "no cooldown meta to write, tablet_id=" << tablet_id();
2249
0
        return Status::OK();
2250
0
    }
2251
17
    std::sort(cooldowned_rs_metas.begin(), cooldowned_rs_metas.end(), RowsetMeta::comparator);
2252
17
    DCHECK(cooldowned_rs_metas.front()->start_version() == 0);
2253
    // If version not continuous, it must be a bug
2254
17
    if (auto st = check_version_continuity(cooldowned_rs_metas); !st.ok()) {
2255
0
        DCHECK(st.ok()) << st << " tablet_id=" << tablet_id();
2256
0
        st.set_code(ABORTED);
2257
0
        return st;
2258
0
    }
2259
2260
17
    TabletMetaPB tablet_meta_pb;
2261
17
    auto* rs_metas = tablet_meta_pb.mutable_rs_metas();
2262
17
    rs_metas->Reserve(cooldowned_rs_metas.size());
2263
42
    for (auto& rs_meta : cooldowned_rs_metas) {
2264
42
        rs_metas->Add(rs_meta->get_rowset_pb());
2265
42
    }
2266
17
    tablet_meta_pb.mutable_cooldown_meta_id()->set_hi(cooldown_meta_id.hi);
2267
17
    tablet_meta_pb.mutable_cooldown_meta_id()->set_lo(cooldown_meta_id.lo);
2268
2269
17
    std::string remote_meta_path = storage_resource.cooldown_tablet_meta_path(
2270
17
            tablet_id(), _cooldown_conf.cooldown_replica_id, _cooldown_conf.term);
2271
17
    io::FileWriterPtr tablet_meta_writer;
2272
    // FIXME(plat1ko): What if object store permanently unavailable?
2273
17
    RETURN_IF_ERROR(storage_resource.fs->create_file(remote_meta_path, &tablet_meta_writer));
2274
17
    auto val = tablet_meta_pb.SerializeAsString();
2275
17
    RETURN_IF_ERROR(tablet_meta_writer->append({val.data(), val.size()}));
2276
17
    return tablet_meta_writer->close();
2277
17
}
2278
2279
// hold SHARED `cooldown_conf_lock`
2280
0
Status Tablet::_follow_cooldowned_data() {
2281
0
    DCHECK(_cooldown_conf.cooldown_replica_id != replica_id());
2282
0
    LOG(INFO) << "try to follow cooldowned data. tablet_id=" << tablet_id()
2283
0
              << " cooldown_replica_id=" << _cooldown_conf.cooldown_replica_id
2284
0
              << " local replica=" << replica_id();
2285
2286
0
    auto storage_resource = DORIS_TRY(get_resource_by_storage_policy_id(storage_policy_id()));
2287
    // MUST executing serially with cold data compaction, because compaction input rowsets may be deleted by this function
2288
0
    std::unique_lock cold_compaction_lock(_cold_compaction_lock, std::try_to_lock);
2289
0
    if (!cold_compaction_lock.owns_lock()) {
2290
0
        return Status::Error<TRY_LOCK_FAILED>("try cold_compaction_lock failed");
2291
0
    }
2292
2293
0
    TabletMetaPB cooldown_meta_pb;
2294
0
    auto st = _read_cooldown_meta(storage_resource, &cooldown_meta_pb);
2295
0
    if (!st.ok()) {
2296
0
        LOG(INFO) << "cannot read cooldown meta: " << st;
2297
0
        return Status::InternalError<false>("cannot read cooldown meta");
2298
0
    }
2299
0
    DCHECK(cooldown_meta_pb.rs_metas_size() > 0);
2300
0
    if (_tablet_meta->cooldown_meta_id() == cooldown_meta_pb.cooldown_meta_id()) {
2301
        // cooldowned rowsets are same, no need to follow
2302
0
        return Status::OK();
2303
0
    }
2304
2305
0
    int64_t cooldowned_version = cooldown_meta_pb.rs_metas().rbegin()->end_version();
2306
2307
0
    std::vector<RowsetSharedPtr> overlap_rowsets;
2308
0
    bool version_aligned = false;
2309
2310
    // Holding these to delete rowsets' shared ptr until save meta can avoid trash sweeping thread
2311
    // deleting these rowsets' files before rowset meta has been removed from disk, which may cause
2312
    // data loss when BE reboot before save meta to disk.
2313
0
    std::vector<RowsetSharedPtr> to_delete;
2314
0
    std::vector<RowsetSharedPtr> to_add;
2315
2316
0
    {
2317
0
        std::lock_guard wlock(_meta_lock);
2318
0
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2319
0
        if (tablet_state() != TABLET_RUNNING) {
2320
0
            return Status::InternalError<false>("tablet not running");
2321
0
        }
2322
2323
0
        for (auto& [v, rs] : _rs_version_map) {
2324
0
            if (v.second <= cooldowned_version) {
2325
0
                overlap_rowsets.push_back(rs);
2326
0
                if (!version_aligned && v.second == cooldowned_version) {
2327
0
                    version_aligned = true;
2328
0
                }
2329
0
            } else if (!rs->is_local()) {
2330
0
                return Status::InternalError<false>(
2331
0
                        "cooldowned version larger than that to follow with cooldown version {}",
2332
0
                        cooldowned_version);
2333
0
            }
2334
0
        }
2335
2336
0
        if (!version_aligned) {
2337
0
            return Status::InternalError<false>("cooldowned version is not aligned with version {}",
2338
0
                                                cooldowned_version);
2339
0
        }
2340
2341
0
        std::sort(overlap_rowsets.begin(), overlap_rowsets.end(), Rowset::comparator);
2342
2343
        // Find different rowset in `overlap_rowsets` and `cooldown_meta_pb.rs_metas`
2344
0
        auto rs_pb_it = cooldown_meta_pb.rs_metas().begin();
2345
0
        auto rs_it = overlap_rowsets.begin();
2346
0
        for (; rs_pb_it != cooldown_meta_pb.rs_metas().end() && rs_it != overlap_rowsets.end();
2347
0
             ++rs_pb_it, ++rs_it) {
2348
0
            if (rs_pb_it->rowset_id_v2() != (*rs_it)->rowset_id().to_string()) {
2349
0
                break;
2350
0
            }
2351
0
        }
2352
2353
0
        to_delete.assign(rs_it, overlap_rowsets.end());
2354
0
        to_add.reserve(cooldown_meta_pb.rs_metas().end() - rs_pb_it);
2355
0
        for (; rs_pb_it != cooldown_meta_pb.rs_metas().end(); ++rs_pb_it) {
2356
0
            auto rs_meta = std::make_shared<RowsetMeta>();
2357
0
            rs_meta->init_from_pb(*rs_pb_it);
2358
0
            RowsetSharedPtr rs;
2359
0
            RETURN_IF_ERROR(
2360
0
                    RowsetFactory::create_rowset(_tablet_meta->tablet_schema(), "", rs_meta, &rs));
2361
0
            to_add.push_back(std::move(rs));
2362
0
        }
2363
        // Note: We CANNOT call `modify_rowsets` here because `modify_rowsets` cannot process version graph correctly.
2364
0
        RETURN_IF_ERROR(delete_rowsets(to_delete, false));
2365
0
        add_rowsets(to_add);
2366
        // TODO(plat1ko): process primary key
2367
0
        _tablet_meta->set_cooldown_meta_id(cooldown_meta_pb.cooldown_meta_id());
2368
0
    }
2369
2370
0
    {
2371
0
        std::lock_guard rlock(_meta_lock);
2372
0
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2373
0
        save_meta();
2374
0
    }
2375
2376
0
    if (!to_add.empty()) {
2377
0
        LOG(INFO) << "modify rowsets when follow cooldowned data, tablet_id=" << tablet_id()
2378
0
                  << [&]() {
2379
0
                         std::stringstream ss;
2380
0
                         ss << " delete rowsets:\n";
2381
0
                         for (auto&& rs : to_delete) {
2382
0
                             ss << rs->version() << ' ' << rs->rowset_id() << '\n';
2383
0
                         }
2384
0
                         ss << "add rowsets:\n";
2385
0
                         for (auto&& rs : to_add) {
2386
0
                             ss << rs->version() << ' ' << rs->rowset_id() << '\n';
2387
0
                         }
2388
0
                         return ss.str();
2389
0
                     }();
2390
0
    }
2391
2392
0
    return Status::OK();
2393
0
}
2394
2395
20.1M
bool Tablet::_has_data_to_cooldown() {
2396
20.1M
    int64_t min_local_version = std::numeric_limits<int64_t>::max();
2397
20.1M
    RowsetSharedPtr rowset;
2398
20.1M
    std::shared_lock meta_rlock(_meta_lock);
2399
    // Ususally once the tablet has done cooldown successfully then the first
2400
    // rowset would always be remote rowset
2401
20.1M
    bool has_cooldowned = false;
2402
36.4M
    for (const auto& [_, rs] : _rs_version_map) {
2403
36.4M
        if (!rs->is_local()) {
2404
251
            has_cooldowned = true;
2405
251
            break;
2406
251
        }
2407
36.4M
    }
2408
36.4M
    for (auto& [v, rs] : _rs_version_map) {
2409
36.4M
        auto predicate = rs->is_local() && v.first < min_local_version;
2410
36.4M
        if (!has_cooldowned) {
2411
36.4M
            predicate = predicate && (rs->data_disk_size() > 0);
2412
36.4M
        }
2413
36.4M
        if (predicate) {
2414
            // this is a local rowset and has data
2415
3.36M
            min_local_version = v.first;
2416
3.36M
            rowset = rs;
2417
3.36M
        }
2418
36.4M
    }
2419
2420
20.1M
    int64_t newest_cooldown_time = 0;
2421
20.1M
    if (rowset != nullptr) {
2422
3.03M
        newest_cooldown_time = _get_newest_cooldown_time(rowset);
2423
3.03M
    }
2424
2425
20.1M
    return (newest_cooldown_time != 0) && (newest_cooldown_time < UnixSeconds());
2426
20.1M
}
2427
2428
20.1M
RowsetSharedPtr Tablet::pick_cooldown_rowset() {
2429
20.1M
    RowsetSharedPtr rowset;
2430
2431
20.1M
    if (!_has_data_to_cooldown()) {
2432
20.1M
        return nullptr;
2433
20.1M
    }
2434
2435
    // TODO(plat1ko): should we maintain `cooldowned_version` in `Tablet`?
2436
40
    int64_t cooldowned_version = -1;
2437
    // We pick the rowset with smallest start version in local.
2438
40
    int64_t min_local_version = std::numeric_limits<int64_t>::max();
2439
40
    {
2440
40
        std::shared_lock meta_rlock(_meta_lock);
2441
192
        for (auto& [v, rs] : _rs_version_map) {
2442
192
            if (!rs->is_local()) {
2443
46
                cooldowned_version = std::max(cooldowned_version, v.second);
2444
146
            } else if (v.first < min_local_version) { // this is a local rowset
2445
113
                min_local_version = v.first;
2446
113
                rowset = rs;
2447
113
            }
2448
192
        }
2449
40
    }
2450
40
    if (!rowset) {
2451
0
        return nullptr;
2452
0
    }
2453
40
    if (tablet_footprint() == 0) {
2454
0
        VLOG_DEBUG << "skip cooldown due to empty tablet_id = " << tablet_id();
2455
0
        return nullptr;
2456
0
    }
2457
40
    if (min_local_version != cooldowned_version + 1) { // ensure version continuity
2458
0
        if (UNLIKELY(cooldowned_version != -1)) {
2459
0
            LOG(WARNING) << "version not continuous. tablet_id=" << tablet_id()
2460
0
                         << " cooldowned_version=" << cooldowned_version
2461
0
                         << " min_local_version=" << min_local_version;
2462
0
        }
2463
0
        return nullptr;
2464
0
    }
2465
40
    return rowset;
2466
40
}
2467
2468
3.03M
int64_t Tablet::_get_newest_cooldown_time(const RowsetSharedPtr& rowset) {
2469
3.03M
    int64_t id = storage_policy_id();
2470
3.03M
    if (id <= 0) {
2471
3.03M
        VLOG_DEBUG << "tablet does not need cooldown, tablet id: " << tablet_id();
2472
3.03M
        return 0;
2473
3.03M
    }
2474
109
    auto storage_policy = get_storage_policy(id);
2475
109
    if (!storage_policy) {
2476
43
        LOG(WARNING) << "Cannot get storage policy: " << id;
2477
43
        return 0;
2478
43
    }
2479
66
    auto cooldown_ttl_sec = storage_policy->cooldown_ttl;
2480
66
    auto cooldown_datetime = storage_policy->cooldown_datetime;
2481
66
    int64_t newest_cooldown_time = std::numeric_limits<int64_t>::max();
2482
2483
66
    if (cooldown_ttl_sec >= 0) {
2484
63
        newest_cooldown_time = rowset->newest_write_timestamp() + cooldown_ttl_sec;
2485
63
    }
2486
66
    if (cooldown_datetime > 0) {
2487
10
        newest_cooldown_time = std::min(newest_cooldown_time, cooldown_datetime);
2488
10
    }
2489
2490
66
    return newest_cooldown_time;
2491
109
}
2492
2493
20.1M
RowsetSharedPtr Tablet::need_cooldown(int64_t* cooldown_timestamp, size_t* file_size) {
2494
20.1M
    RowsetSharedPtr rowset = pick_cooldown_rowset();
2495
20.1M
    if (!rowset) {
2496
20.1M
        VLOG_DEBUG << "pick cooldown rowset, get null, tablet id: " << tablet_id();
2497
20.1M
        return nullptr;
2498
20.1M
    }
2499
2500
25
    auto newest_cooldown_time = _get_newest_cooldown_time(rowset);
2501
2502
    // the rowset should do cooldown job only if it's cooldown ttl plus newest write time is less than
2503
    // current time or it's datatime is less than current time
2504
25
    if (newest_cooldown_time != 0 && newest_cooldown_time < UnixSeconds()) {
2505
25
        *cooldown_timestamp = newest_cooldown_time;
2506
25
        *file_size = rowset->total_disk_size();
2507
25
        VLOG_DEBUG << "tablet need cooldown, tablet id: " << tablet_id()
2508
0
                   << " file_size: " << *file_size;
2509
25
        return rowset;
2510
25
    }
2511
2512
0
    VLOG_DEBUG << "tablet does not need cooldown, tablet id: " << tablet_id()
2513
0
               << " newest write time: " << rowset->newest_write_timestamp();
2514
0
    return nullptr;
2515
25
}
2516
2517
void Tablet::record_unused_remote_rowset(const RowsetId& rowset_id, const std::string& resource,
2518
0
                                         int64_t num_segments) {
2519
0
    auto gc_key = REMOTE_ROWSET_GC_PREFIX + rowset_id.to_string();
2520
0
    RemoteRowsetGcPB gc_pb;
2521
0
    gc_pb.set_resource_id(resource);
2522
0
    gc_pb.set_tablet_id(tablet_id());
2523
0
    gc_pb.set_num_segments(num_segments);
2524
0
    auto st =
2525
0
            _data_dir->get_meta()->put(META_COLUMN_FAMILY_INDEX, gc_key, gc_pb.SerializeAsString());
2526
0
    if (!st.ok()) {
2527
0
        LOG(WARNING) << "failed to record unused remote rowset. tablet_id=" << tablet_id()
2528
0
                     << " rowset_id=" << rowset_id << " resource_id=" << resource;
2529
0
    }
2530
0
    unused_remote_rowset_num << 1;
2531
0
}
2532
2533
95.3k
Status Tablet::remove_all_remote_rowsets() {
2534
95.3k
    DCHECK(tablet_state() == TABLET_SHUTDOWN);
2535
95.3k
    std::set<std::string> resource_ids;
2536
146k
    for (auto& rs_meta : _tablet_meta->all_rs_metas()) {
2537
146k
        if (!rs_meta->is_local()) {
2538
1
            resource_ids.insert(rs_meta->resource_id());
2539
1
        }
2540
146k
    }
2541
95.3k
    if (resource_ids.empty()) {
2542
95.3k
        return Status::OK();
2543
95.3k
    }
2544
1
    auto tablet_gc_key = REMOTE_TABLET_GC_PREFIX + std::to_string(tablet_id());
2545
1
    RemoteTabletGcPB gc_pb;
2546
1
    for (auto& resource_id : resource_ids) {
2547
1
        gc_pb.add_resource_ids(resource_id);
2548
1
    }
2549
1
    return _data_dir->get_meta()->put(META_COLUMN_FAMILY_INDEX, tablet_gc_key,
2550
1
                                      gc_pb.SerializeAsString());
2551
95.3k
}
2552
2553
0
void Tablet::update_max_version_schema(const TabletSchemaSPtr& tablet_schema) {
2554
0
    std::lock_guard wrlock(_meta_lock);
2555
0
    SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2556
    // Double Check for concurrent update
2557
0
    if (!_max_version_schema ||
2558
0
        tablet_schema->schema_version() > _max_version_schema->schema_version()) {
2559
0
        _max_version_schema = tablet_schema;
2560
0
    }
2561
0
}
2562
2563
94
CalcDeleteBitmapExecutor* Tablet::calc_delete_bitmap_executor() {
2564
94
    return _engine.calc_delete_bitmap_executor();
2565
94
}
2566
2567
Status Tablet::save_delete_bitmap(const TabletTxnInfo* txn_info, int64_t txn_id,
2568
                                  DeleteBitmapPtr delete_bitmap, RowsetWriter* rowset_writer,
2569
                                  const RowsetIdUnorderedSet& cur_rowset_ids, int64_t lock_id,
2570
66.2k
                                  int64_t next_visible_version) {
2571
66.2k
    RowsetSharedPtr rowset = txn_info->rowset;
2572
66.2k
    int64_t cur_version = rowset->start_version();
2573
2574
    // update version without write lock, compaction and publish_txn
2575
    // will update delete bitmap, handle compaction with _rowset_update_lock
2576
    // and publish_txn runs sequential so no need to lock here
2577
503k
    for (auto& [key, bitmap] : delete_bitmap->delete_bitmap) {
2578
        // skip sentinel mark, which is used for delete bitmap correctness check
2579
503k
        if (std::get<1>(key) != DeleteBitmap::INVALID_SEGMENT_ID) {
2580
8.02k
            _tablet_meta->delete_bitmap().merge({std::get<0>(key), std::get<1>(key), cur_version},
2581
8.02k
                                                bitmap);
2582
8.02k
        }
2583
503k
    }
2584
2585
66.2k
    return Status::OK();
2586
66.2k
}
2587
2588
3.18k
void Tablet::merge_delete_bitmap(const DeleteBitmap& delete_bitmap) {
2589
3.18k
    _tablet_meta->delete_bitmap().merge(delete_bitmap);
2590
3.18k
}
2591
2592
0
bool Tablet::check_all_rowset_segment() {
2593
0
    std::shared_lock rdlock(_meta_lock);
2594
0
    for (auto& version_rowset : _rs_version_map) {
2595
0
        RowsetSharedPtr rowset = version_rowset.second;
2596
0
        if (!rowset->check_rowset_segment()) {
2597
0
            LOG(WARNING) << "Tablet Segment Check. find a bad tablet, tablet_id=" << tablet_id();
2598
0
            return false;
2599
0
        }
2600
0
    }
2601
0
    return true;
2602
0
}
2603
2604
67.3M
void Tablet::set_skip_compaction(bool skip, CompactionType compaction_type, int64_t start) {
2605
67.3M
    if (!skip) {
2606
115
        _skip_cumu_compaction = false;
2607
115
        _skip_base_compaction = false;
2608
115
        return;
2609
115
    }
2610
67.3M
    if (compaction_type == CompactionType::CUMULATIVE_COMPACTION) {
2611
38.7M
        _skip_cumu_compaction = true;
2612
38.7M
        _skip_cumu_compaction_ts = start;
2613
38.7M
    } else {
2614
28.6M
        DCHECK(compaction_type == CompactionType::BASE_COMPACTION);
2615
28.6M
        _skip_base_compaction = true;
2616
28.6M
        _skip_base_compaction_ts = start;
2617
28.6M
    }
2618
67.3M
}
2619
2620
1.08G
bool Tablet::should_skip_compaction(CompactionType compaction_type, int64_t now) {
2621
1.08G
    if (compaction_type == CompactionType::CUMULATIVE_COMPACTION && _skip_cumu_compaction &&
2622
1.08G
        now < _skip_cumu_compaction_ts + config::skip_tablet_compaction_second) {
2623
891M
        return true;
2624
891M
    } else if (compaction_type == CompactionType::BASE_COMPACTION && _skip_base_compaction &&
2625
189M
               now < _skip_base_compaction_ts + config::skip_tablet_compaction_second) {
2626
62.4M
        return true;
2627
62.4M
    }
2628
126M
    return false;
2629
1.08G
}
2630
2631
240
std::pair<std::string, int64_t> Tablet::get_binlog_info(std::string_view binlog_version) const {
2632
240
    return RowsetMetaManager::get_binlog_info(_data_dir->get_meta(), tablet_uid(), binlog_version);
2633
240
}
2634
2635
std::string Tablet::get_rowset_binlog_meta(std::string_view binlog_version,
2636
222
                                           std::string_view rowset_id) const {
2637
222
    return RowsetMetaManager::get_rowset_binlog_meta(_data_dir->get_meta(), tablet_uid(),
2638
222
                                                     binlog_version, rowset_id);
2639
222
}
2640
2641
Status Tablet::get_rowset_binlog_metas(const std::vector<int64_t>& binlog_versions,
2642
0
                                       RowsetBinlogMetasPB* metas_pb) {
2643
0
    return RowsetMetaManager::get_rowset_binlog_metas(_data_dir->get_meta(), tablet_uid(),
2644
0
                                                      binlog_versions, metas_pb);
2645
0
}
2646
2647
4
Status Tablet::get_rowset_binlog_metas(Version binlog_versions, RowsetBinlogMetasPB* metas_pb) {
2648
4
    return RowsetMetaManager::get_rowset_binlog_metas(_data_dir->get_meta(), tablet_uid(),
2649
4
                                                      binlog_versions, metas_pb);
2650
4
}
2651
2652
std::string Tablet::get_segment_filepath(std::string_view rowset_id,
2653
436
                                         std::string_view segment_index) const {
2654
436
    return fmt::format("{}/_binlog/{}_{}.dat", _tablet_path, rowset_id, segment_index);
2655
436
}
2656
2657
1
std::string Tablet::get_segment_filepath(std::string_view rowset_id, int64_t segment_index) const {
2658
1
    return fmt::format("{}/_binlog/{}_{}.dat", _tablet_path, rowset_id, segment_index);
2659
1
}
2660
2661
0
std::vector<std::string> Tablet::get_binlog_filepath(std::string_view binlog_version) const {
2662
0
    const auto& [rowset_id, num_segments] = get_binlog_info(binlog_version);
2663
0
    std::vector<std::string> binlog_filepath;
2664
0
    for (int i = 0; i < num_segments; ++i) {
2665
        // TODO(Drogon): rewrite by filesystem path
2666
0
        auto segment_file = fmt::format("{}_{}.dat", rowset_id, i);
2667
0
        binlog_filepath.emplace_back(fmt::format("{}/_binlog/{}", _tablet_path, segment_file));
2668
0
    }
2669
0
    return binlog_filepath;
2670
0
}
2671
2672
444
bool Tablet::can_add_binlog(uint64_t total_binlog_size) const {
2673
444
    return !_data_dir->reach_capacity_limit(total_binlog_size);
2674
444
}
2675
2676
260k
bool Tablet::is_enable_binlog() {
2677
260k
    return config::enable_feature_binlog && tablet_meta()->binlog_config().is_enable();
2678
260k
}
2679
2680
44
void Tablet::set_binlog_config(BinlogConfig binlog_config) {
2681
44
    tablet_meta()->set_binlog_config(binlog_config);
2682
44
}
2683
2684
2
void Tablet::gc_binlogs(int64_t version) {
2685
2
    auto meta = _data_dir->get_meta();
2686
2
    DCHECK(meta != nullptr);
2687
2688
2
    const auto& tablet_uid = this->tablet_uid();
2689
2
    const auto tablet_id = this->tablet_id();
2690
2
    std::string begin_key = make_binlog_meta_key_prefix(tablet_uid);
2691
2
    std::string end_key = make_binlog_meta_key_prefix(tablet_uid, version + 1);
2692
2
    LOG(INFO) << fmt::format("gc binlog meta, tablet_id:{}, begin_key:{}, end_key:{}", tablet_id,
2693
2
                             begin_key, end_key);
2694
2695
2
    std::vector<std::string> wait_for_deleted_binlog_keys;
2696
2
    std::vector<std::string> wait_for_deleted_binlog_files;
2697
2
    auto add_to_wait_for_deleted = [&](std::string_view key, std::string_view rowset_id,
2698
2
                                       int64_t num_segments) {
2699
        // add binlog meta key and binlog data key
2700
0
        wait_for_deleted_binlog_keys.emplace_back(key);
2701
0
        wait_for_deleted_binlog_keys.push_back(get_binlog_data_key_from_meta_key(key));
2702
2703
        // add binlog segment files and index files
2704
0
        for (int64_t i = 0; i < num_segments; ++i) {
2705
0
            auto segment_file_path = get_segment_filepath(rowset_id, i);
2706
0
            wait_for_deleted_binlog_files.emplace_back(segment_file_path);
2707
2708
            // index files
2709
0
            if (tablet_schema()->has_inverted_index()) {
2710
0
                if (tablet_schema()->get_inverted_index_storage_format() ==
2711
0
                    InvertedIndexStorageFormatPB::V1) {
2712
0
                    for (const auto& index : tablet_schema()->inverted_indexes()) {
2713
0
                        auto index_file = InvertedIndexDescriptor::get_index_file_path_v1(
2714
0
                                InvertedIndexDescriptor::get_index_file_path_prefix(
2715
0
                                        segment_file_path),
2716
0
                                index->index_id(), index->get_index_suffix());
2717
0
                        wait_for_deleted_binlog_files.emplace_back(index_file);
2718
0
                    }
2719
0
                } else {
2720
0
                    auto index_file = InvertedIndexDescriptor::get_index_file_path_v2(
2721
0
                            InvertedIndexDescriptor::get_index_file_path_prefix(segment_file_path));
2722
0
                    wait_for_deleted_binlog_files.emplace_back(index_file);
2723
0
                }
2724
0
            }
2725
0
        }
2726
0
    };
2727
2728
2
    auto check_binlog_ttl = [&](std::string_view key, std::string_view value) mutable -> bool {
2729
0
        if (key >= end_key) {
2730
0
            return false;
2731
0
        }
2732
2733
0
        BinlogMetaEntryPB binlog_meta_entry_pb;
2734
0
        if (!binlog_meta_entry_pb.ParseFromArray(value.data(), value.size())) {
2735
0
            LOG(WARNING) << "failed to parse binlog meta entry, key:" << key;
2736
0
            return true;
2737
0
        }
2738
2739
0
        auto num_segments = binlog_meta_entry_pb.num_segments();
2740
0
        std::string rowset_id;
2741
0
        if (binlog_meta_entry_pb.has_rowset_id_v2()) {
2742
0
            rowset_id = binlog_meta_entry_pb.rowset_id_v2();
2743
0
        } else {
2744
            // key is 'binlog_meta_6943f1585fe834b5-e542c2b83a21d0b7_00000000000000000069_020000000000000135449d7cd7eadfe672aa0f928fa99593', extract last part '020000000000000135449d7cd7eadfe672aa0f928fa99593'
2745
0
            auto pos = key.rfind('_');
2746
0
            if (pos == std::string::npos) {
2747
0
                LOG(WARNING) << fmt::format("invalid binlog meta key:{}", key);
2748
0
                return false;
2749
0
            }
2750
0
            rowset_id = key.substr(pos + 1);
2751
0
        }
2752
0
        add_to_wait_for_deleted(key, rowset_id, num_segments);
2753
2754
0
        return true;
2755
0
    };
2756
2757
2
    auto status = meta->iterate(META_COLUMN_FAMILY_INDEX, begin_key, check_binlog_ttl);
2758
2
    if (!status.ok()) {
2759
0
        LOG(WARNING) << "failed to iterate binlog meta, status:" << status;
2760
0
        return;
2761
0
    }
2762
2763
    // first remove binlog files, if failed, just break, then retry next time
2764
    // this keep binlog meta in meta store, so that binlog can be removed next time
2765
2
    bool remove_binlog_files_failed = false;
2766
2
    for (auto& file : wait_for_deleted_binlog_files) {
2767
2
        if (unlink(file.c_str()) != 0) {
2768
            // file not exist, continue
2769
0
            if (errno == ENOENT) {
2770
0
                continue;
2771
0
            }
2772
2773
0
            remove_binlog_files_failed = true;
2774
0
            LOG(WARNING) << "failed to remove binlog file:" << file << ", errno:" << errno;
2775
0
            break;
2776
0
        }
2777
2
    }
2778
2
    if (!remove_binlog_files_failed) {
2779
2
        static_cast<void>(meta->remove(META_COLUMN_FAMILY_INDEX, wait_for_deleted_binlog_keys));
2780
2
    }
2781
2
}
2782
2783
0
Status Tablet::ingest_binlog_metas(RowsetBinlogMetasPB* metas_pb) {
2784
0
    return RowsetMetaManager::ingest_binlog_metas(_data_dir->get_meta(), tablet_uid(), metas_pb);
2785
0
}
2786
2787
190k
void Tablet::clear_cache() {
2788
190k
    std::vector<RowsetSharedPtr> rowsets;
2789
190k
    {
2790
190k
        std::shared_lock rlock(get_header_lock());
2791
190k
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2792
2793
311k
        for (auto& [_, rowset] : rowset_map()) {
2794
311k
            rowsets.push_back(rowset);
2795
311k
        }
2796
190k
        for (auto& [_, rowset] : stale_rowset_map()) {
2797
25.9k
            rowsets.push_back(rowset);
2798
25.9k
        }
2799
190k
    }
2800
337k
    for (auto& rowset : rowsets) {
2801
337k
        rowset->clear_cache();
2802
337k
    }
2803
190k
}
2804
2805
578k
void Tablet::check_table_size_correctness() {
2806
578k
    if (!config::enable_table_size_correctness_check) {
2807
578k
        return;
2808
578k
    }
2809
18.4E
    const std::vector<RowsetMetaSharedPtr>& all_rs_metas = _tablet_meta->all_rs_metas();
2810
18.4E
    for (const auto& rs_meta : all_rs_metas) {
2811
0
        int64_t total_segment_size = get_segment_file_size(rs_meta);
2812
0
        int64_t total_inverted_index_size = get_inverted_index_file_size(rs_meta);
2813
0
        if (rs_meta->data_disk_size() != total_segment_size ||
2814
0
            rs_meta->index_disk_size() != total_inverted_index_size ||
2815
0
            rs_meta->data_disk_size() + rs_meta->index_disk_size() != rs_meta->total_disk_size()) {
2816
0
            LOG(WARNING) << "[Local table table size check failed]:"
2817
0
                         << " tablet id: " << rs_meta->tablet_id()
2818
0
                         << ", rowset id:" << rs_meta->rowset_id()
2819
0
                         << ", rowset data disk size:" << rs_meta->data_disk_size()
2820
0
                         << ", rowset real data disk size:" << total_segment_size
2821
0
                         << ", rowset index disk size:" << rs_meta->index_disk_size()
2822
0
                         << ", rowset real index disk size:" << total_inverted_index_size
2823
0
                         << ", rowset total disk size:" << rs_meta->total_disk_size()
2824
0
                         << ", rowset segment path:"
2825
0
                         << StorageResource().remote_segment_path(
2826
0
                                    rs_meta->tablet_id(), rs_meta->rowset_id().to_string(), 0);
2827
0
            DCHECK(false);
2828
0
        }
2829
0
    }
2830
18.4E
}
2831
2832
0
std::string Tablet::get_segment_path(const RowsetMetaSharedPtr& rs_meta, int64_t seg_id) {
2833
0
    std::string segment_path;
2834
0
    if (rs_meta->is_local()) {
2835
0
        segment_path = local_segment_path(_tablet_path, rs_meta->rowset_id().to_string(), seg_id);
2836
0
    } else {
2837
0
        segment_path = rs_meta->remote_storage_resource().value()->remote_segment_path(
2838
0
                rs_meta->tablet_id(), rs_meta->rowset_id().to_string(), seg_id);
2839
0
    }
2840
0
    return segment_path;
2841
0
}
2842
2843
0
int64_t Tablet::get_segment_file_size(const RowsetMetaSharedPtr& rs_meta) {
2844
0
    const auto& fs = rs_meta->fs();
2845
0
    if (!fs) {
2846
0
        LOG(WARNING) << "get fs failed, resource_id={}" << rs_meta->resource_id();
2847
0
    }
2848
0
    int64_t total_segment_size = 0;
2849
0
    for (int64_t seg_id = 0; seg_id < rs_meta->num_segments(); seg_id++) {
2850
0
        std::string segment_path = get_segment_path(rs_meta, seg_id);
2851
0
        int64_t segment_file_size = 0;
2852
0
        auto st = fs->file_size(segment_path, &segment_file_size);
2853
0
        if (!st.ok()) {
2854
0
            segment_file_size = 0;
2855
0
            LOG(WARNING) << "table size correctness check get segment size failed! msg:"
2856
0
                         << st.to_string() << ", segment path:" << segment_path;
2857
0
        }
2858
0
        total_segment_size += segment_file_size;
2859
0
    }
2860
0
    return total_segment_size;
2861
0
}
2862
2863
0
int64_t Tablet::get_inverted_index_file_size(const RowsetMetaSharedPtr& rs_meta) {
2864
0
    const auto& fs = rs_meta->fs();
2865
0
    if (!fs) {
2866
0
        LOG(WARNING) << "get fs failed, resource_id={}" << rs_meta->resource_id();
2867
0
    }
2868
0
    int64_t total_inverted_index_size = 0;
2869
2870
0
    if (rs_meta->tablet_schema()->get_inverted_index_storage_format() ==
2871
0
        InvertedIndexStorageFormatPB::V1) {
2872
0
        const auto& indices = rs_meta->tablet_schema()->inverted_indexes();
2873
0
        for (auto& index : indices) {
2874
0
            for (int seg_id = 0; seg_id < rs_meta->num_segments(); ++seg_id) {
2875
0
                std::string segment_path = get_segment_path(rs_meta, seg_id);
2876
0
                int64_t file_size = 0;
2877
2878
0
                std::string inverted_index_file_path =
2879
0
                        InvertedIndexDescriptor::get_index_file_path_v1(
2880
0
                                InvertedIndexDescriptor::get_index_file_path_prefix(segment_path),
2881
0
                                index->index_id(), index->get_index_suffix());
2882
0
                auto st = fs->file_size(inverted_index_file_path, &file_size);
2883
0
                if (!st.ok()) {
2884
0
                    file_size = 0;
2885
0
                    LOG(WARNING) << " tablet id: " << get_tablet_info().tablet_id
2886
0
                                 << ", rowset id:" << rs_meta->rowset_id()
2887
0
                                 << ", table size correctness check get inverted index v1 "
2888
0
                                    "size failed! msg:"
2889
0
                                 << st.to_string()
2890
0
                                 << ", inverted index path:" << inverted_index_file_path;
2891
0
                }
2892
0
                total_inverted_index_size += file_size;
2893
0
            }
2894
0
        }
2895
0
    } else {
2896
0
        for (int seg_id = 0; seg_id < rs_meta->num_segments(); ++seg_id) {
2897
0
            int64_t file_size = 0;
2898
0
            std::string segment_path = get_segment_path(rs_meta, seg_id);
2899
0
            std::string inverted_index_file_path = InvertedIndexDescriptor::get_index_file_path_v2(
2900
0
                    InvertedIndexDescriptor::get_index_file_path_prefix(segment_path));
2901
0
            auto st = fs->file_size(inverted_index_file_path, &file_size);
2902
0
            if (!st.ok()) {
2903
0
                file_size = 0;
2904
0
                if (st.is<NOT_FOUND>()) {
2905
0
                    LOG(INFO) << " tablet id: " << get_tablet_info().tablet_id
2906
0
                              << ", rowset id:" << rs_meta->rowset_id()
2907
0
                              << ", table size correctness check get inverted index v2 failed "
2908
0
                                 "because file not exist:"
2909
0
                              << inverted_index_file_path;
2910
0
                } else {
2911
0
                    LOG(WARNING) << " tablet id: " << get_tablet_info().tablet_id
2912
0
                                 << ", rowset id:" << rs_meta->rowset_id()
2913
0
                                 << ", table size correctness check get inverted index v2 "
2914
0
                                    "size failed! msg:"
2915
0
                                 << st.to_string()
2916
0
                                 << ", inverted index path:" << inverted_index_file_path;
2917
0
                }
2918
0
            }
2919
0
            total_inverted_index_size += file_size;
2920
0
        }
2921
0
    }
2922
0
    return total_inverted_index_size;
2923
0
}
2924
2925
} // namespace doris