Coverage Report

Created: 2025-12-31 17:27

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
/root/doris/be/src/olap/tablet.cpp
Line
Count
Source
1
// Licensed to the Apache Software Foundation (ASF) under one
2
// or more contributor license agreements.  See the NOTICE file
3
// distributed with this work for additional information
4
// regarding copyright ownership.  The ASF licenses this file
5
// to you under the Apache License, Version 2.0 (the
6
// "License"); you may not use this file except in compliance
7
// with the License.  You may obtain a copy of the License at
8
//
9
//   http://www.apache.org/licenses/LICENSE-2.0
10
//
11
// Unless required by applicable law or agreed to in writing,
12
// software distributed under the License is distributed on an
13
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
14
// KIND, either express or implied.  See the License for the
15
// specific language governing permissions and limitations
16
// under the License.
17
18
#include "olap/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/stopwatch.hpp"
117
#include "util/threadpool.h"
118
#include "util/time.h"
119
#include "util/trace.h"
120
#include "util/uid_util.h"
121
#include "util/work_thread_pool.hpp"
122
#include "vec/columns/column.h"
123
#include "vec/columns/column_string.h"
124
#include "vec/common/custom_allocator.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
#include "common/compile_check_begin.h"
134
class TupleDescriptor;
135
136
namespace vectorized {
137
class Block;
138
} // namespace vectorized
139
140
using namespace ErrorCode;
141
using namespace std::chrono_literals;
142
143
using std::pair;
144
using std::string;
145
using std::vector;
146
using io::FileSystemSPtr;
147
148
namespace {
149
150
bvar::Adder<uint64_t> exceed_version_limit_counter;
151
bvar::Window<bvar::Adder<uint64_t>> exceed_version_limit_counter_minute(
152
        &exceed_version_limit_counter, 60);
153
bvar::Adder<uint64_t> cooldown_pending_task("cooldown_pending_task");
154
bvar::Adder<uint64_t> cooldown_processing_task("cooldown_processing_task");
155
156
70
void set_last_failure_time(Tablet* tablet, const Compaction& compaction, int64_t ms) {
157
70
    switch (compaction.compaction_type()) {
158
64
    case ReaderType::READER_CUMULATIVE_COMPACTION:
159
64
        tablet->set_last_cumu_compaction_failure_time(ms);
160
64
        return;
161
6
    case ReaderType::READER_BASE_COMPACTION:
162
6
        tablet->set_last_base_compaction_failure_time(ms);
163
6
        return;
164
0
    case ReaderType::READER_FULL_COMPACTION:
165
0
        tablet->set_last_full_compaction_failure_time(ms);
166
0
        return;
167
0
    default:
168
0
        LOG(FATAL) << "invalid compaction type " << compaction.compaction_name()
169
0
                   << " tablet_id: " << tablet->tablet_id();
170
70
    }
171
70
};
172
173
} // namespace
174
175
bvar::Adder<uint64_t> unused_remote_rowset_num("unused_remote_rowset_num");
176
177
WriteCooldownMetaExecutors::WriteCooldownMetaExecutors(size_t executor_nums)
178
8
        : _executor_nums(executor_nums) {
179
48
    for (size_t i = 0; i < _executor_nums; i++) {
180
40
        std::unique_ptr<PriorityThreadPool> pool;
181
40
        static_cast<void>(ThreadPoolBuilder("WriteCooldownMetaExecutor")
182
40
                                  .set_min_threads(1)
183
40
                                  .set_max_threads(1)
184
40
                                  .set_max_queue_size(std::numeric_limits<int>::max())
185
40
                                  .build(&pool));
186
40
        _executors.emplace_back(std::move(pool));
187
40
    }
188
8
}
189
190
3
void WriteCooldownMetaExecutors::stop() {
191
15
    for (auto& pool_ptr : _executors) {
192
15
        if (pool_ptr) {
193
15
            pool_ptr->shutdown();
194
15
        }
195
15
    }
196
3
}
197
198
291k
void WriteCooldownMetaExecutors::WriteCooldownMetaExecutors::submit(TabletSharedPtr tablet) {
199
291k
    auto tablet_id = tablet->tablet_id();
200
201
291k
    {
202
291k
        std::shared_lock rdlock(tablet->get_header_lock());
203
291k
        if (!tablet->tablet_meta()->cooldown_meta_id().initialized()) {
204
0
            VLOG_NOTICE << "tablet " << tablet_id << " is not cooldown replica";
205
0
            return;
206
0
        }
207
291k
        if (tablet->tablet_state() == TABLET_SHUTDOWN) [[unlikely]] {
208
0
            LOG_INFO("tablet {} has been dropped, don't do cooldown", tablet_id);
209
0
            return;
210
0
        }
211
291k
    }
212
291k
    {
213
        // one tablet could at most have one cooldown task to be done
214
291k
        std::unique_lock<std::mutex> lck {_latch};
215
291k
        if (_pending_tablets.count(tablet_id) > 0) {
216
0
            return;
217
0
        }
218
291k
        _pending_tablets.insert(tablet_id);
219
291k
    }
220
221
291k
    auto async_write_task = [this, t = std::move(tablet)]() {
222
291k
        {
223
291k
            std::unique_lock<std::mutex> lck {_latch};
224
291k
            _pending_tablets.erase(t->tablet_id());
225
291k
        }
226
291k
        SCOPED_ATTACH_TASK(ExecEnv::GetInstance()->orphan_mem_tracker());
227
291k
        auto s = t->write_cooldown_meta();
228
291k
        if (s.ok()) {
229
9
            return;
230
9
        }
231
291k
        if (!s.is<ABORTED>()) {
232
291k
            LOG_EVERY_SECOND(WARNING)
233
28
                    << "write tablet " << t->tablet_id() << " cooldown meta failed because: " << s;
234
291k
            submit(t);
235
291k
            return;
236
291k
        }
237
0
        VLOG_DEBUG << "tablet " << t->tablet_id() << " is not cooldown replica";
238
0
    };
239
240
291k
    cooldown_pending_task << 1;
241
291k
    _executors[_get_executor_pos(tablet_id)]->offer([task = std::move(async_write_task)]() {
242
291k
        cooldown_pending_task << -1;
243
291k
        cooldown_processing_task << 1;
244
291k
        task();
245
291k
        cooldown_processing_task << -1;
246
291k
    });
247
291k
}
248
249
Tablet::Tablet(StorageEngine& engine, TabletMetaSharedPtr tablet_meta, DataDir* data_dir,
250
               const std::string_view& cumulative_compaction_type)
251
339k
        : BaseTablet(std::move(tablet_meta)),
252
339k
          _engine(engine),
253
339k
          _data_dir(data_dir),
254
339k
          _is_bad(false),
255
339k
          _last_cumu_compaction_failure_millis(0),
256
339k
          _last_base_compaction_failure_millis(0),
257
339k
          _last_full_compaction_failure_millis(0),
258
339k
          _last_cumu_compaction_success_millis(0),
259
339k
          _last_base_compaction_success_millis(0),
260
339k
          _last_full_compaction_success_millis(0),
261
339k
          _cumulative_point(K_INVALID_CUMULATIVE_POINT),
262
339k
          _newly_created_rowset_num(0),
263
339k
          _last_checkpoint_time(0),
264
339k
          _cumulative_compaction_type(cumulative_compaction_type),
265
339k
          _is_tablet_path_exists(true),
266
339k
          _last_missed_version(-1),
267
339k
          _last_missed_time_s(0) {
268
339k
    if (_data_dir != nullptr) {
269
339k
        _tablet_path = fmt::format("{}/{}/{}/{}/{}", _data_dir->path(), DATA_PREFIX,
270
339k
                                   _tablet_meta->shard_id(), tablet_id(), schema_hash());
271
339k
    }
272
339k
}
273
274
339k
bool Tablet::set_tablet_schema_into_rowset_meta() {
275
339k
    bool flag = false;
276
593k
    for (auto& [_, rowset_meta] : _tablet_meta->all_mutable_rs_metas()) {
277
593k
        if (!rowset_meta->tablet_schema()) {
278
0
            rowset_meta->set_tablet_schema(_tablet_meta->tablet_schema());
279
0
            flag = true;
280
0
        }
281
593k
    }
282
339k
    return flag;
283
339k
}
284
285
339k
Status Tablet::_init_once_action() {
286
339k
    Status res = Status::OK();
287
339k
    VLOG_NOTICE << "begin to load tablet. tablet=" << tablet_id()
288
416
                << ", version_size=" << _tablet_meta->version_count();
289
290
#ifdef BE_TEST
291
    // init cumulative compaction policy by type
292
    _cumulative_compaction_policy =
293
            CumulativeCompactionPolicyFactory::create_cumulative_compaction_policy(
294
                    _tablet_meta->compaction_policy());
295
#endif
296
297
596k
    for (const auto& [_, rs_meta] : _tablet_meta->all_rs_metas()) {
298
596k
        Version version = rs_meta->version();
299
596k
        RowsetSharedPtr rowset;
300
596k
        res = create_rowset(rs_meta, &rowset);
301
596k
        if (!res.ok()) {
302
0
            LOG(WARNING) << "fail to init rowset. tablet_id=" << tablet_id()
303
0
                         << ", schema_hash=" << schema_hash() << ", version=" << version
304
0
                         << ", res=" << res;
305
0
            return res;
306
0
        }
307
596k
        _rs_version_map[version] = std::move(rowset);
308
596k
    }
309
310
    // init stale rowset
311
339k
    int64_t now = ::time(nullptr);
312
339k
    for (const auto& [_, stale_rs_meta] : _tablet_meta->all_stale_rs_metas()) {
313
20.5k
        Version version = stale_rs_meta->version();
314
315
20.5k
        if (!stale_rs_meta->has_stale_at()) {
316
20.5k
            stale_rs_meta->set_stale_at(now);
317
20.5k
        }
318
319
20.5k
        RowsetSharedPtr rowset;
320
20.5k
        res = create_rowset(stale_rs_meta, &rowset);
321
20.5k
        if (!res.ok()) {
322
0
            LOG(WARNING) << "fail to init stale rowset. tablet_id:" << tablet_id()
323
0
                         << ", schema_hash:" << schema_hash() << ", version=" << version
324
0
                         << ", res:" << res;
325
0
            return res;
326
0
        }
327
20.5k
        _stale_rs_version_map[version] = std::move(rowset);
328
20.5k
    }
329
330
339k
    return res;
331
339k
}
332
333
339k
Status Tablet::init() {
334
339k
    return _init_once.call([this] { return _init_once_action(); });
335
339k
}
336
337
// should save tablet meta to remote meta store
338
// if it's a primary replica
339
35.2k
void Tablet::save_meta() {
340
35.2k
    check_table_size_correctness();
341
35.2k
    auto res = _tablet_meta->save_meta(_data_dir);
342
35.2k
    CHECK_EQ(res, Status::OK()) << "fail to save tablet_meta. res=" << res
343
0
                                << ", root=" << _data_dir->path();
344
35.2k
}
345
346
// Caller should hold _meta_lock.
347
Status Tablet::revise_tablet_meta(const std::vector<RowsetSharedPtr>& to_add,
348
                                  const std::vector<RowsetSharedPtr>& to_delete,
349
0
                                  bool is_incremental_clone) {
350
0
    LOG(INFO) << "begin to revise tablet. tablet_id=" << tablet_id();
351
    // 1. for incremental clone, we have to add the rowsets first to make it easy to compute
352
    //    all the delete bitmaps, and it's easy to delete them if we end up with a failure
353
    // 2. for full clone, we can calculate delete bitmaps on the cloned rowsets directly.
354
0
    if (is_incremental_clone) {
355
0
        CHECK(to_delete.empty()); // don't need to delete rowsets
356
0
        add_rowsets(to_add);
357
        // reconstruct from tablet meta
358
0
        _timestamped_version_tracker.construct_versioned_tracker(_tablet_meta->all_rs_metas());
359
0
    }
360
361
0
    Status calc_bm_status;
362
0
    std::vector<RowsetSharedPtr> base_rowsets_for_full_clone = to_add; // copy vector
363
0
    while (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write()) {
364
0
        std::vector<RowsetSharedPtr> calc_delete_bitmap_rowsets;
365
0
        int64_t to_add_min_version = INT64_MAX;
366
0
        int64_t to_add_max_version = INT64_MIN;
367
0
        for (auto& rs : to_add) {
368
0
            if (to_add_min_version > rs->start_version()) {
369
0
                to_add_min_version = rs->start_version();
370
0
            }
371
0
            if (to_add_max_version < rs->end_version()) {
372
0
                to_add_max_version = rs->end_version();
373
0
            }
374
0
        }
375
0
        Version calc_delete_bitmap_ver;
376
0
        if (is_incremental_clone) {
377
            // From the rowset of to_add with smallest version, all other rowsets
378
            // need to recalculate the delete bitmap
379
            // For example:
380
            // local tablet: [0-1] [2-5] [6-6] [9-10]
381
            // clone tablet: [7-7] [8-8]
382
            // new tablet:   [0-1] [2-5] [6-6] [7-7] [8-8] [9-10]
383
            // [7-7] [8-8] [9-10] need to recalculate delete bitmap
384
0
            calc_delete_bitmap_ver = Version(to_add_min_version, max_version_unlocked());
385
0
        } else {
386
            // the delete bitmap of to_add's rowsets has clone from remote when full clone.
387
            // only other rowsets in local need to recalculate the delete bitmap.
388
            // For example:
389
            // local tablet: [0-1]x [2-5]x [6-6]x [7-7]x [9-10]
390
            // clone tablet: [0-1]  [2-4]  [5-6]  [7-8]
391
            // new tablet:   [0-1]  [2-4]  [5-6]  [7-8] [9-10]
392
            // only [9-10] need to recalculate delete bitmap
393
0
            CHECK_EQ(to_add_min_version, 0) << "to_add_min_version is: " << to_add_min_version;
394
0
            calc_delete_bitmap_ver = Version(to_add_max_version + 1, max_version_unlocked());
395
0
        }
396
397
0
        if (calc_delete_bitmap_ver.first <= calc_delete_bitmap_ver.second) {
398
0
            auto ret = capture_consistent_rowsets_unlocked(calc_delete_bitmap_ver,
399
0
                                                           CaptureRowsetOps {});
400
0
            if (!ret) {
401
0
                LOG(WARNING) << "fail to capture_consistent_rowsets, res: " << ret.error();
402
0
                calc_bm_status = std::move(ret.error());
403
0
                break;
404
0
            }
405
0
            calc_delete_bitmap_rowsets = std::move(ret->rowsets);
406
            // FIXME(plat1ko): Use `const TabletSharedPtr&` as parameter
407
0
            auto self = _engine.tablet_manager()->get_tablet(tablet_id());
408
0
            CHECK(self);
409
0
            for (auto rs : calc_delete_bitmap_rowsets) {
410
0
                if (is_incremental_clone) {
411
0
                    calc_bm_status = update_delete_bitmap_without_lock(self, rs);
412
0
                } else {
413
0
                    calc_bm_status = update_delete_bitmap_without_lock(
414
0
                            self, rs, &base_rowsets_for_full_clone);
415
0
                    base_rowsets_for_full_clone.push_back(rs);
416
0
                }
417
0
                if (!calc_bm_status.ok()) {
418
0
                    LOG(WARNING) << "fail to update_delete_bitmap_without_lock, res: "
419
0
                                 << calc_bm_status;
420
0
                    break;
421
0
                }
422
0
            }
423
0
        }
424
0
        break; // while (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write())
425
0
    }
426
427
0
    DBUG_EXECUTE_IF("Tablet.revise_tablet_meta_fail", {
428
0
        auto ptablet_id = dp->param("tablet_id", 0);
429
0
        if (tablet_id() == ptablet_id) {
430
0
            LOG(INFO) << "injected revies_tablet_meta failure for tabelt: " << ptablet_id;
431
0
            calc_bm_status = Status::InternalError("fault injection error");
432
0
        }
433
0
    });
434
435
    // error handling
436
0
    if (!calc_bm_status.ok()) {
437
0
        if (is_incremental_clone) {
438
0
            RETURN_IF_ERROR(delete_rowsets(to_add, false));
439
0
            LOG(WARNING) << "incremental clone on tablet: " << tablet_id() << " failed due to "
440
0
                         << calc_bm_status.msg() << ", revert " << to_add.size()
441
0
                         << " rowsets added before.";
442
0
        } else {
443
0
            LOG(WARNING) << "full clone on tablet: " << tablet_id() << " failed due to "
444
0
                         << calc_bm_status.msg() << ", will not update tablet meta.";
445
0
        }
446
0
        return calc_bm_status;
447
0
    }
448
449
    // full clone, calculate delete bitmap succeeded, update rowset
450
0
    if (!is_incremental_clone) {
451
0
        RETURN_IF_ERROR(delete_rowsets(to_delete, false));
452
0
        add_rowsets(to_add);
453
        // reconstruct from tablet meta
454
0
        _timestamped_version_tracker.construct_versioned_tracker(_tablet_meta->all_rs_metas());
455
456
        // check the rowsets used for delete bitmap calculation is equal to the rowsets
457
        // that we can capture by version
458
0
        if (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write()) {
459
0
            Version full_version = Version(0, max_version_unlocked());
460
0
            auto ret = capture_consistent_rowsets_unlocked(full_version, CaptureRowsetOps {});
461
0
            DCHECK(ret) << ret.error();
462
0
            DCHECK_EQ(base_rowsets_for_full_clone.size(), ret->rowsets.size());
463
464
0
            if (ret && base_rowsets_for_full_clone.size() != ret->rowsets.size()) [[unlikely]] {
465
0
                LOG(WARNING) << "full clone succeeded, but the count("
466
0
                             << base_rowsets_for_full_clone.size()
467
0
                             << ") of base rowsets used for delete bitmap calculation is not match "
468
0
                                "expect count("
469
0
                             << ret->rowsets.size() << ") we capture from tablet meta";
470
0
            }
471
0
        }
472
0
    }
473
474
    // clear stale rowset
475
0
    for (auto& [v, rs] : _stale_rs_version_map) {
476
0
        _engine.add_unused_rowset(rs);
477
0
    }
478
0
    _stale_rs_version_map.clear();
479
0
    _tablet_meta->clear_stale_rowset();
480
0
    save_meta();
481
482
0
    LOG(INFO) << "finish to revise tablet. tablet_id=" << tablet_id();
483
0
    return Status::OK();
484
0
}
485
486
21.0k
Status Tablet::add_rowset(RowsetSharedPtr rowset) {
487
21.0k
    DCHECK(rowset != nullptr);
488
21.0k
    std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
489
21.0k
    SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
490
    // If the rowset already exist, just return directly.  The rowset_id is an unique-id,
491
    // we can use it to check this situation.
492
21.0k
    if (_contains_rowset(rowset->rowset_id())) {
493
705
        return Status::OK();
494
705
    }
495
    // Otherwise, the version should be not contained in any existing rowset.
496
20.3k
    RETURN_IF_ERROR(_contains_version(rowset->version()));
497
498
19.6k
    RETURN_IF_ERROR(_tablet_meta->add_rs_meta(rowset->rowset_meta()));
499
19.6k
    _rs_version_map[rowset->version()] = rowset;
500
19.6k
    _timestamped_version_tracker.add_version(rowset->version());
501
19.6k
    add_compaction_score(rowset->rowset_meta()->get_compaction_score());
502
503
19.6k
    std::vector<RowsetSharedPtr> rowsets_to_delete;
504
    // yiguolei: temp code, should remove the rowset contains by this rowset
505
    // but it should be removed in multi path version
506
40.8k
    for (auto& it : _rs_version_map) {
507
40.8k
        if (rowset->version().contains(it.first) && rowset->version() != it.first) {
508
0
            CHECK(it.second != nullptr)
509
0
                    << "there exist a version=" << it.first
510
0
                    << " contains the input rs with version=" << rowset->version()
511
0
                    << ", but the related rs is null";
512
0
            rowsets_to_delete.push_back(it.second);
513
0
        }
514
40.8k
    }
515
19.6k
    std::vector<RowsetSharedPtr> empty_vec;
516
19.6k
    RETURN_IF_ERROR(modify_rowsets(empty_vec, rowsets_to_delete));
517
19.6k
    ++_newly_created_rowset_num;
518
19.6k
    return Status::OK();
519
19.6k
}
520
521
460
bool Tablet::rowset_exists_unlocked(const RowsetSharedPtr& rowset) {
522
460
    if (auto it = _rs_version_map.find(rowset->version()); it == _rs_version_map.end()) {
523
0
        return false;
524
460
    } else if (rowset->rowset_id() != it->second->rowset_id()) {
525
0
        return false;
526
0
    }
527
460
    return true;
528
460
}
529
530
Status Tablet::modify_rowsets(std::vector<RowsetSharedPtr>& to_add,
531
19.7k
                              std::vector<RowsetSharedPtr>& to_delete, bool check_delete) {
532
    // the compaction process allow to compact the single version, eg: version[4-4].
533
    // this kind of "single version compaction" has same "input version" and "output version".
534
    // which means "to_add->version()" equals to "to_delete->version()".
535
    // So we should delete the "to_delete" before adding the "to_add",
536
    // otherwise, the "to_add" will be deleted from _rs_version_map, eventually.
537
    //
538
    // And if the version of "to_add" and "to_delete" are exactly same. eg:
539
    // to_add:      [7-7]
540
    // to_delete:   [7-7]
541
    // In this case, we no longer need to add the rowset in "to_delete" to
542
    // _stale_rs_version_map, but can delete it directly.
543
544
19.7k
    if (to_add.empty() && to_delete.empty()) {
545
19.6k
        return Status::OK();
546
19.6k
    }
547
548
87
    if (check_delete) {
549
566
        for (auto&& rs : to_delete) {
550
566
            if (auto it = _rs_version_map.find(rs->version()); it == _rs_version_map.end()) {
551
0
                return Status::Error<DELETE_VERSION_ERROR>(
552
0
                        "try to delete not exist version {} from {}", rs->version().to_string(),
553
0
                        tablet_id());
554
566
            } else if (rs->rowset_id() != it->second->rowset_id()) {
555
0
                return Status::Error<DELETE_VERSION_ERROR>(
556
0
                        "try to delete version {} from {}, but rowset id changed, delete "
557
0
                        "rowset id "
558
0
                        "is {}, exists rowsetid is {}",
559
0
                        rs->version().to_string(), tablet_id(), rs->rowset_id().to_string(),
560
0
                        it->second->rowset_id().to_string());
561
0
            }
562
566
        }
563
86
    }
564
565
87
    bool same_version = true;
566
87
    std::sort(to_add.begin(), to_add.end(), Rowset::comparator);
567
87
    std::sort(to_delete.begin(), to_delete.end(), Rowset::comparator);
568
87
    if (to_add.size() == to_delete.size()) {
569
32
        for (int i = 0; i < to_add.size(); ++i) {
570
16
            if (to_add[i]->version() != to_delete[i]->version()) {
571
0
                same_version = false;
572
0
                break;
573
0
            }
574
16
        }
575
71
    } else {
576
71
        same_version = false;
577
71
    }
578
579
87
    std::vector<RowsetMetaSharedPtr> rs_metas_to_delete;
580
87
    int64_t now = ::time(nullptr);
581
566
    for (auto& rs : to_delete) {
582
566
        rs_metas_to_delete.push_back(rs->rowset_meta());
583
566
        _rs_version_map.erase(rs->version());
584
585
566
        if (!same_version) {
586
550
            rs->rowset_meta()->set_stale_at(now);
587
            // put compaction rowsets in _stale_rs_version_map.
588
550
            _stale_rs_version_map[rs->version()] = rs;
589
550
        }
590
566
    }
591
592
87
    std::vector<RowsetMetaSharedPtr> rs_metas_to_add;
593
87
    for (auto& rs : to_add) {
594
87
        rs_metas_to_add.push_back(rs->rowset_meta());
595
87
        _rs_version_map[rs->version()] = rs;
596
597
87
        if (!same_version) {
598
            // If version are same, then _timestamped_version_tracker
599
            // already has this version, no need to add again.
600
71
            _timestamped_version_tracker.add_version(rs->version());
601
71
        }
602
87
        ++_newly_created_rowset_num;
603
87
    }
604
605
87
    _tablet_meta->modify_rs_metas(rs_metas_to_add, rs_metas_to_delete, same_version);
606
607
87
    if (!same_version) {
608
        // add rs_metas_to_delete to tracker
609
71
        _timestamped_version_tracker.add_stale_path_version(rs_metas_to_delete);
610
71
    } else {
611
        // delete rowset in "to_delete" directly
612
16
        for (auto& rs : to_delete) {
613
16
            LOG(INFO) << "add unused rowset " << rs->rowset_id() << " because of same version";
614
16
            if (rs->is_local()) {
615
16
                _engine.add_unused_rowset(rs);
616
16
            }
617
16
        }
618
16
    }
619
620
87
    int32_t add_score = 0;
621
87
    for (auto rs : to_add) {
622
87
        add_score += rs->rowset_meta()->get_compaction_score();
623
87
    }
624
87
    int32_t sub_score = 0;
625
566
    for (auto rs : to_delete) {
626
566
        sub_score += rs->rowset_meta()->get_compaction_score();
627
566
    }
628
87
    add_compaction_score(add_score - sub_score);
629
630
87
    return Status::OK();
631
87
}
632
633
5
void Tablet::add_rowsets(const std::vector<RowsetSharedPtr>& to_add) {
634
5
    if (to_add.empty()) {
635
0
        return;
636
0
    }
637
5
    std::vector<RowsetMetaSharedPtr> rs_metas;
638
5
    rs_metas.reserve(to_add.size());
639
5
    for (auto& rs : to_add) {
640
5
        _rs_version_map.emplace(rs->version(), rs);
641
5
        _timestamped_version_tracker.add_version(rs->version());
642
5
        rs_metas.push_back(rs->rowset_meta());
643
5
    }
644
5
    _tablet_meta->modify_rs_metas(rs_metas, {});
645
5
}
646
647
5
Status Tablet::delete_rowsets(const std::vector<RowsetSharedPtr>& to_delete, bool move_to_stale) {
648
5
    if (to_delete.empty()) {
649
0
        return Status::OK();
650
0
    }
651
5
    std::vector<RowsetMetaSharedPtr> rs_metas;
652
5
    rs_metas.reserve(to_delete.size());
653
5
    int64_t now = ::time(nullptr);
654
5
    for (const auto& rs : to_delete) {
655
5
        if (move_to_stale) {
656
0
            rs->rowset_meta()->set_stale_at(now);
657
0
        }
658
5
        rs_metas.push_back(rs->rowset_meta());
659
5
        _rs_version_map.erase(rs->version());
660
5
    }
661
5
    _tablet_meta->modify_rs_metas({}, rs_metas, !move_to_stale);
662
5
    if (move_to_stale) {
663
0
        for (const auto& rs : to_delete) {
664
0
            _stale_rs_version_map[rs->version()] = rs;
665
0
        }
666
0
        _timestamped_version_tracker.add_stale_path_version(rs_metas);
667
5
    } else {
668
5
        for (const auto& rs : to_delete) {
669
5
            _timestamped_version_tracker.delete_version(rs->version());
670
5
            if (rs->is_local()) {
671
5
                _engine.add_unused_rowset(rs);
672
5
                RETURN_IF_ERROR(RowsetMetaManager::remove(_data_dir->get_meta(), tablet_uid(),
673
5
                                                          rs->rowset_meta()->rowset_id()));
674
5
            }
675
5
        }
676
5
    }
677
5
    return Status::OK();
678
5
}
679
680
0
RowsetSharedPtr Tablet::_rowset_with_largest_size() {
681
0
    RowsetSharedPtr largest_rowset = nullptr;
682
0
    for (auto& it : _rs_version_map) {
683
0
        if (it.second->empty() || it.second->zero_num_rows()) {
684
0
            continue;
685
0
        }
686
0
        if (largest_rowset == nullptr || it.second->rowset_meta()->index_disk_size() >
687
0
                                                 largest_rowset->rowset_meta()->index_disk_size()) {
688
0
            largest_rowset = it.second;
689
0
        }
690
0
    }
691
692
0
    return largest_rowset;
693
0
}
694
695
// add inc rowset should not persist tablet meta, because it will be persisted when publish txn.
696
10.9k
Status Tablet::add_inc_rowset(const RowsetSharedPtr& rowset) {
697
10.9k
    DCHECK(rowset != nullptr);
698
10.9k
    std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
699
10.9k
    SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
700
10.9k
    if (_contains_rowset(rowset->rowset_id())) {
701
0
        return Status::OK();
702
0
    }
703
10.9k
    RETURN_IF_ERROR(_contains_version(rowset->version()));
704
705
10.9k
    RETURN_IF_ERROR(_tablet_meta->add_rs_meta(rowset->rowset_meta()));
706
10.9k
    _rs_version_map[rowset->version()] = rowset;
707
708
10.9k
    _timestamped_version_tracker.add_version(rowset->version());
709
710
10.9k
    ++_newly_created_rowset_num;
711
712
10.9k
    add_compaction_score(rowset->rowset_meta()->get_compaction_score());
713
714
10.9k
    return Status::OK();
715
10.9k
}
716
717
16.6k
void Tablet::_delete_stale_rowset_by_version(const Version& version) {
718
16.6k
    RowsetMetaSharedPtr rowset_meta = _tablet_meta->acquire_stale_rs_meta_by_version(version);
719
16.6k
    if (rowset_meta == nullptr) {
720
8
        return;
721
8
    }
722
16.6k
    _tablet_meta->delete_stale_rs_meta_by_version(version);
723
16.6k
    VLOG_NOTICE << "delete stale rowset. tablet=" << tablet_id() << ", version=" << version;
724
16.6k
}
725
726
664k
void Tablet::delete_expired_stale_rowset() {
727
664k
    if (config::enable_mow_verbose_log) {
728
0
        LOG_INFO("begin delete_expired_stale_rowset for tablet={}", tablet_id());
729
0
    }
730
664k
    int64_t now = UnixSeconds();
731
664k
    std::vector<std::pair<Version, std::vector<RowsetId>>> deleted_stale_rowsets;
732
    // hold write lock while processing stable rowset
733
664k
    {
734
664k
        std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
735
664k
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
736
        // Compute the end time to delete rowsets, when a expired rowset createtime less then this time, it will be deleted.
737
664k
        int64_t expired_stale_sweep_endtime =
738
664k
                static_cast<int64_t>(::difftime(now, config::tablet_rowset_stale_sweep_time_sec));
739
664k
        if (config::tablet_rowset_stale_sweep_by_size) {
740
0
            expired_stale_sweep_endtime = now;
741
0
        }
742
743
664k
        std::vector<int64_t> path_id_vec;
744
        // capture the path version to delete
745
664k
        _timestamped_version_tracker.capture_expired_paths(
746
664k
                static_cast<int64_t>(expired_stale_sweep_endtime), &path_id_vec);
747
748
664k
        if (path_id_vec.empty()) {
749
662k
            return;
750
662k
        }
751
752
1.92k
        const RowsetSharedPtr lastest_delta = get_rowset_with_max_version();
753
1.92k
        if (lastest_delta == nullptr) {
754
0
            LOG(WARNING) << "lastest_delta is null " << tablet_id();
755
0
            return;
756
0
        }
757
758
        // fetch missing version before delete
759
1.92k
        Versions missed_versions = get_missed_versions_unlocked(lastest_delta->end_version());
760
1.92k
        if (!missed_versions.empty()) {
761
0
            LOG(WARNING) << "tablet:" << tablet_id()
762
0
                         << ", missed version for version:" << lastest_delta->end_version();
763
0
            _print_missed_versions(missed_versions);
764
0
            return;
765
0
        }
766
767
        // do check consistent operation
768
1.92k
        auto path_id_iter = path_id_vec.begin();
769
770
1.92k
        std::map<int64_t, PathVersionListSharedPtr> stale_version_path_map;
771
4.92k
        while (path_id_iter != path_id_vec.end()) {
772
3.00k
            PathVersionListSharedPtr version_path =
773
3.00k
                    _timestamped_version_tracker.fetch_and_delete_path_by_id(*path_id_iter);
774
775
3.00k
            Version test_version = Version(0, lastest_delta->end_version());
776
3.00k
            stale_version_path_map[*path_id_iter] = version_path;
777
778
3.00k
            auto ret = capture_consistent_versions_unlocked(test_version, {});
779
            // 1. When there is no consistent versions, we must reconstruct the tracker.
780
3.00k
            if (!ret) {
781
                // 2. fetch missing version after delete
782
0
                Versions after_missed_versions =
783
0
                        get_missed_versions_unlocked(lastest_delta->end_version());
784
785
                // 2.1 check whether missed_versions and after_missed_versions are the same.
786
                // when they are the same, it means we can delete the path securely.
787
0
                bool is_missing = missed_versions.size() != after_missed_versions.size();
788
789
0
                if (!is_missing) {
790
0
                    for (int ver_index = 0; ver_index < missed_versions.size(); ver_index++) {
791
0
                        if (missed_versions[ver_index] != after_missed_versions[ver_index]) {
792
0
                            is_missing = true;
793
0
                            break;
794
0
                        }
795
0
                    }
796
0
                }
797
798
0
                if (is_missing) {
799
0
                    LOG(WARNING) << "The consistent version check fails, there are bugs. "
800
0
                                 << "Reconstruct the tracker to recover versions in tablet="
801
0
                                 << tablet_id();
802
803
                    // 3. try to recover
804
0
                    _timestamped_version_tracker.recover_versioned_tracker(stale_version_path_map);
805
806
                    // 4. double check the consistent versions
807
                    // fetch missing version after recover
808
0
                    Versions recover_missed_versions =
809
0
                            get_missed_versions_unlocked(lastest_delta->end_version());
810
811
                    // 4.1 check whether missed_versions and recover_missed_versions are the same.
812
                    // when they are the same, it means we recover successfully.
813
0
                    bool is_recover_missing =
814
0
                            missed_versions.size() != recover_missed_versions.size();
815
816
0
                    if (!is_recover_missing) {
817
0
                        for (int ver_index = 0; ver_index < missed_versions.size(); ver_index++) {
818
0
                            if (missed_versions[ver_index] != recover_missed_versions[ver_index]) {
819
0
                                is_recover_missing = true;
820
0
                                break;
821
0
                            }
822
0
                        }
823
0
                    }
824
825
                    // 5. check recover fail, version is mission
826
0
                    if (is_recover_missing) {
827
0
                        if (!config::ignore_rowset_stale_unconsistent_delete) {
828
0
                            LOG(FATAL)
829
0
                                    << "rowset stale unconsistent delete. tablet= " << tablet_id();
830
0
                        } else {
831
0
                            LOG(WARNING)
832
0
                                    << "rowset stale unconsistent delete. tablet= " << tablet_id();
833
0
                        }
834
0
                    }
835
0
                }
836
0
                return;
837
0
            }
838
3.00k
            path_id_iter++;
839
3.00k
        }
840
841
1.92k
        auto old_size = _stale_rs_version_map.size();
842
1.92k
        auto old_meta_size = _tablet_meta->all_stale_rs_metas().size();
843
844
        // do delete operation
845
1.92k
        auto to_delete_iter = stale_version_path_map.begin();
846
4.92k
        while (to_delete_iter != stale_version_path_map.end()) {
847
3.00k
            std::vector<TimestampedVersionSharedPtr>& to_delete_version =
848
3.00k
                    to_delete_iter->second->timestamped_versions();
849
3.00k
            int64_t start_version = -1;
850
3.00k
            int64_t end_version = -1;
851
3.00k
            std::vector<RowsetId> remove_rowset_ids;
852
16.6k
            for (auto& timestampedVersion : to_delete_version) {
853
16.6k
                auto it = _stale_rs_version_map.find(timestampedVersion->version());
854
16.6k
                if (it != _stale_rs_version_map.end()) {
855
16.6k
                    it->second->clear_cache();
856
                    // delete rowset
857
16.6k
                    if (it->second->is_local()) {
858
16.6k
                        _engine.add_unused_rowset(it->second);
859
16.6k
                        if (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write()) {
860
                            // mow does not support cold data in object storage
861
0
                            remove_rowset_ids.emplace_back(it->second->rowset_id());
862
0
                        }
863
16.6k
                    }
864
16.6k
                    _stale_rs_version_map.erase(it);
865
16.6k
                    VLOG_NOTICE << "delete stale rowset tablet=" << tablet_id() << " version["
866
0
                                << timestampedVersion->version().first << ","
867
0
                                << timestampedVersion->version().second
868
0
                                << "] move to unused_rowset success " << std::fixed
869
0
                                << expired_stale_sweep_endtime;
870
16.6k
                } else {
871
8
                    LOG(WARNING) << "delete stale rowset tablet=" << tablet_id() << " version["
872
8
                                 << timestampedVersion->version().first << ","
873
8
                                 << timestampedVersion->version().second
874
8
                                 << "] not find in stale rs version map";
875
8
                }
876
16.6k
                if (start_version < 0) {
877
3.00k
                    start_version = timestampedVersion->version().first;
878
3.00k
                }
879
16.6k
                end_version = timestampedVersion->version().second;
880
16.6k
                _delete_stale_rowset_by_version(timestampedVersion->version());
881
16.6k
            }
882
3.00k
            Version version(start_version, end_version);
883
3.00k
            to_delete_iter++;
884
3.00k
            if (!remove_rowset_ids.empty()) {
885
0
                deleted_stale_rowsets.emplace_back(version, remove_rowset_ids);
886
0
            }
887
3.00k
        }
888
889
1.92k
        bool reconstructed = _reconstruct_version_tracker_if_necessary();
890
891
1.92k
        VLOG_NOTICE << "delete stale rowset _stale_rs_version_map tablet=" << tablet_id()
892
1
                    << " current_size=" << _stale_rs_version_map.size() << " old_size=" << old_size
893
1
                    << " current_meta_size=" << _tablet_meta->all_stale_rs_metas().size()
894
1
                    << " old_meta_size=" << old_meta_size << " sweep endtime " << std::fixed
895
1
                    << expired_stale_sweep_endtime << ", reconstructed=" << reconstructed;
896
1.92k
    }
897
1.92k
    if (config::enable_agg_and_remove_pre_rowsets_delete_bitmap && !deleted_stale_rowsets.empty()) {
898
        // agg delete bitmap for pre rowsets; record unused delete bitmap key ranges
899
0
        OlapStopWatch watch;
900
0
        for (const auto& [version, remove_rowset_ids] : deleted_stale_rowsets) {
901
            // agg delete bitmap for pre rowset
902
0
            DeleteBitmapKeyRanges remove_delete_bitmap_key_ranges;
903
0
            agg_delete_bitmap_for_stale_rowsets(version, remove_delete_bitmap_key_ranges);
904
            // add remove delete bitmap
905
0
            if (!remove_delete_bitmap_key_ranges.empty()) {
906
0
                _engine.add_unused_delete_bitmap_key_ranges(tablet_id(), remove_rowset_ids,
907
0
                                                            remove_delete_bitmap_key_ranges);
908
0
            }
909
0
        }
910
0
        LOG(INFO) << "agg pre rowsets delete bitmap. tablet_id=" << tablet_id()
911
0
                  << ", size=" << deleted_stale_rowsets.size()
912
0
                  << ", cost(us)=" << watch.get_elapse_time_us();
913
0
    }
914
1.92k
#ifndef BE_TEST
915
1.92k
    {
916
1.92k
        std::shared_lock<std::shared_mutex> rlock(_meta_lock);
917
1.92k
        save_meta();
918
1.92k
    }
919
1.92k
#endif
920
1.92k
    if (config::enable_mow_verbose_log) {
921
0
        LOG_INFO("finish delete_expired_stale_rowset for tablet={}", tablet_id());
922
0
    }
923
1.92k
    DBUG_EXECUTE_IF("Tablet.delete_expired_stale_rowset.start_delete_unused_rowset", {
924
1.92k
        _engine.start_delete_unused_rowset();
925
1.92k
        [[maybe_unused]] auto st = _engine.start_trash_sweep(nullptr);
926
1.92k
    });
927
1.92k
}
928
929
0
Status Tablet::check_version_integrity(const Version& version, bool quiet) {
930
0
    std::shared_lock rdlock(_meta_lock);
931
0
    [[maybe_unused]] auto _versions = DORIS_TRY(
932
0
            capture_consistent_versions_unlocked(version, CaptureRowsetOps {.quiet = quiet}));
933
0
    return Status::OK();
934
0
}
935
936
453
bool Tablet::exceed_version_limit(int32_t limit) {
937
453
    if (_tablet_meta->version_count() > limit) {
938
0
        exceed_version_limit_counter << 1;
939
0
        return true;
940
0
    }
941
453
    return false;
942
453
}
943
944
// If any rowset contains the specific version, it means the version already exist
945
454
bool Tablet::check_version_exist(const Version& version) const {
946
454
    std::shared_lock rdlock(_meta_lock);
947
845
    for (auto& it : _rs_version_map) {
948
845
        if (it.first.contains(version)) {
949
453
            return true;
950
453
        }
951
845
    }
952
1
    return false;
953
454
}
954
955
// The meta read lock should be held before calling
956
void Tablet::acquire_version_and_rowsets(
957
0
        std::vector<std::pair<Version, RowsetSharedPtr>>* version_rowsets) const {
958
0
    for (const auto& it : _rs_version_map) {
959
0
        version_rowsets->emplace_back(it.first, it.second);
960
0
    }
961
0
}
962
963
Status Tablet::capture_rs_readers(const Version& spec_version, std::vector<RowSetSplits>* rs_splits,
964
2
                                  const CaptureRowsetOps& opts) {
965
2
    std::shared_lock rlock(_meta_lock);
966
2
    std::vector<Version> version_path;
967
2
    *rs_splits = DORIS_TRY(capture_rs_readers_unlocked(
968
1
            spec_version, CaptureRowsetOps {.skip_missing_versions = opts.skip_missing_versions}));
969
1
    return Status::OK();
970
2
}
971
972
1.92k
Versions Tablet::calc_missed_versions(int64_t spec_version, Versions existing_versions) const {
973
1.92k
    DCHECK(spec_version > 0) << "invalid spec_version: " << spec_version;
974
975
    // sort the existing versions in ascending order
976
1.92k
    std::sort(existing_versions.begin(), existing_versions.end(),
977
9.38k
              [](const Version& a, const Version& b) {
978
                  // simple because 2 versions are certainly not overlapping
979
9.38k
                  return a.first < b.first;
980
9.38k
              });
981
982
    // From the first version(=0),  find the missing version until spec_version
983
1.92k
    int64_t last_version = -1;
984
1.92k
    Versions missed_versions;
985
5.85k
    for (const Version& version : existing_versions) {
986
5.85k
        if (version.first > last_version + 1) {
987
4
            for (int64_t i = last_version + 1; i < version.first && i <= spec_version; ++i) {
988
                // Don't merge missed_versions because clone & snapshot use single version.
989
                // For example, if miss 4 ~ 6, clone need [4, 4], [5, 5], [6, 6], but not [4, 6].
990
2
                missed_versions.emplace_back(i, i);
991
2
            }
992
2
        }
993
5.85k
        last_version = version.second;
994
5.85k
        if (last_version >= spec_version) {
995
1.92k
            break;
996
1.92k
        }
997
5.85k
    }
998
1.92k
    for (int64_t i = last_version + 1; i <= spec_version; ++i) {
999
0
        missed_versions.emplace_back(i, i);
1000
0
    }
1001
1002
1.92k
    return missed_versions;
1003
1.92k
}
1004
1005
22.5M
bool Tablet::can_do_compaction(size_t path_hash, CompactionType compaction_type) {
1006
22.5M
    if (compaction_type == CompactionType::BASE_COMPACTION && tablet_state() != TABLET_RUNNING) {
1007
        // base compaction can only be done for tablet in TABLET_RUNNING state.
1008
        // but cumulative compaction can be done for TABLET_NOTREADY, such as tablet under alter process.
1009
0
        return false;
1010
0
    }
1011
1012
22.5M
    if (data_dir()->path_hash() != path_hash || !is_used() || !init_succeeded()) {
1013
7.58M
        return false;
1014
7.58M
    }
1015
1016
    // In TABLET_NOTREADY, we keep last 10 versions in new tablet so base tablet max_version
1017
    // not merged in new tablet and then we can do compaction
1018
14.9M
    return tablet_state() == TABLET_RUNNING || tablet_state() == TABLET_NOTREADY;
1019
22.5M
}
1020
1021
14.4M
uint32_t Tablet::calc_compaction_score() {
1022
14.4M
    if (_score_check_cnt++ % config::check_score_rounds_num != 0) {
1023
14.0M
        std::shared_lock rdlock(_meta_lock);
1024
14.0M
        if (_compaction_score > 0) {
1025
14.0M
            return _compaction_score;
1026
14.0M
        }
1027
14.0M
    }
1028
1029
337k
    {
1030
        // Need meta lock, because it will iterator "all_rs_metas" of tablet meta.
1031
337k
        std::shared_lock rdlock(_meta_lock);
1032
337k
        int32_t score = get_real_compaction_score();
1033
337k
        if (_compaction_score > 0 && _compaction_score != score) {
1034
0
            LOG(WARNING) << "cumu cache score not equal real score, cache score; "
1035
0
                         << _compaction_score << ", real score: " << score
1036
0
                         << ", tablet: " << tablet_id();
1037
0
        }
1038
337k
        _compaction_score = score;
1039
337k
        return score;
1040
14.4M
    }
1041
14.4M
}
1042
1043
bool Tablet::suitable_for_compaction(
1044
        CompactionType compaction_type,
1045
670k
        std::shared_ptr<CumulativeCompactionPolicy> cumulative_compaction_policy) {
1046
    // Need meta lock, because it will iterator "all_rs_metas" of tablet meta.
1047
670k
    std::shared_lock rdlock(_meta_lock);
1048
670k
    int32_t score = -1;
1049
670k
    if (compaction_type == CompactionType::CUMULATIVE_COMPACTION) {
1050
104k
        score = _calc_cumulative_compaction_score(cumulative_compaction_policy);
1051
566k
    } else {
1052
566k
        DCHECK_EQ(compaction_type, CompactionType::BASE_COMPACTION);
1053
566k
        score = _calc_base_compaction_score();
1054
566k
    }
1055
670k
    return score > 0;
1056
670k
}
1057
1058
16
uint32_t Tablet::calc_cold_data_compaction_score() const {
1059
16
    uint32_t score = 0;
1060
16
    std::vector<RowsetMetaSharedPtr> cooldowned_rowsets;
1061
16
    int64_t max_delete_version = 0;
1062
16
    {
1063
16
        std::shared_lock rlock(_meta_lock);
1064
68
        for (const auto& [_, rs_meta] : _tablet_meta->all_rs_metas()) {
1065
68
            if (!rs_meta->is_local()) {
1066
68
                cooldowned_rowsets.push_back(rs_meta);
1067
68
                if (rs_meta->has_delete_predicate() &&
1068
68
                    rs_meta->end_version() > max_delete_version) {
1069
0
                    max_delete_version = rs_meta->end_version();
1070
0
                }
1071
68
            }
1072
68
        }
1073
16
    }
1074
68
    for (auto& rs_meta : cooldowned_rowsets) {
1075
68
        if (rs_meta->end_version() < max_delete_version) {
1076
0
            score += rs_meta->num_segments();
1077
68
        } else {
1078
68
            score += rs_meta->get_compaction_score();
1079
68
        }
1080
68
    }
1081
16
    return (keys_type() != KeysType::DUP_KEYS) ? score * 2 : score;
1082
16
}
1083
1084
uint32_t Tablet::_calc_cumulative_compaction_score(
1085
103k
        std::shared_ptr<CumulativeCompactionPolicy> cumulative_compaction_policy) {
1086
103k
    if (cumulative_compaction_policy == nullptr) [[unlikely]] {
1087
0
        return 0;
1088
0
    }
1089
103k
#ifndef BE_TEST
1090
103k
    if (_cumulative_compaction_policy == nullptr ||
1091
103k
        _cumulative_compaction_policy->name() != cumulative_compaction_policy->name()) {
1092
4.53k
        _cumulative_compaction_policy = cumulative_compaction_policy;
1093
4.53k
    }
1094
103k
#endif
1095
103k
    DBUG_EXECUTE_IF("Tablet._calc_cumulative_compaction_score.return", {
1096
103k
        LOG_WARNING("Tablet._calc_cumulative_compaction_score.return")
1097
103k
                .tag("tablet id", tablet_id());
1098
103k
        return 0;
1099
103k
    });
1100
103k
    return _cumulative_compaction_policy->calc_cumulative_compaction_score(this);
1101
103k
}
1102
1103
566k
uint32_t Tablet::_calc_base_compaction_score() const {
1104
566k
    uint32_t score = 0;
1105
566k
    const int64_t point = cumulative_layer_point();
1106
566k
    bool base_rowset_exist = false;
1107
566k
    bool has_delete = false;
1108
999k
    for (const auto& [_, rs_meta] : _tablet_meta->all_rs_metas()) {
1109
999k
        if (rs_meta->start_version() == 0) {
1110
566k
            base_rowset_exist = true;
1111
566k
        }
1112
999k
        if (rs_meta->start_version() >= point || !rs_meta->is_local()) {
1113
            // all_rs_metas() is not sorted, so we use _continue_ other than _break_ here.
1114
984k
            continue;
1115
984k
        }
1116
14.5k
        if (rs_meta->has_delete_predicate()) {
1117
481
            has_delete = true;
1118
481
        }
1119
14.5k
        score += rs_meta->get_compaction_score();
1120
14.5k
    }
1121
1122
    // In the time series compaction policy, we want the base compaction to be triggered
1123
    // when there are delete versions present.
1124
566k
    if (_tablet_meta->compaction_policy() == CUMULATIVE_TIME_SERIES_POLICY) {
1125
70
        return (base_rowset_exist && has_delete) ? score : 0;
1126
70
    }
1127
1128
    // base不存在可能是tablet正在做alter table,先不选它,设score=0
1129
566k
    return base_rowset_exist ? score : 0;
1130
566k
}
1131
1132
0
void Tablet::max_continuous_version_from_beginning(Version* version, Version* max_version) {
1133
0
    bool has_version_cross;
1134
0
    std::shared_lock rdlock(_meta_lock);
1135
0
    _max_continuous_version_from_beginning_unlocked(version, max_version, &has_version_cross);
1136
0
}
1137
1138
void Tablet::_max_continuous_version_from_beginning_unlocked(Version* version, Version* max_version,
1139
1.31M
                                                             bool* has_version_cross) const {
1140
1.31M
    std::vector<Version> existing_versions;
1141
1.31M
    *has_version_cross = false;
1142
2.35M
    for (const auto& [ver, _] : _tablet_meta->all_rs_metas()) {
1143
2.35M
        existing_versions.emplace_back(ver);
1144
2.35M
    }
1145
1146
    // sort the existing versions in ascending order
1147
1.31M
    std::sort(existing_versions.begin(), existing_versions.end(),
1148
2.23M
              [](const Version& left, const Version& right) {
1149
                  // simple because 2 versions are certainly not overlapping
1150
2.23M
                  return left.first < right.first;
1151
2.23M
              });
1152
1153
1.31M
    Version max_continuous_version = {-1, -1};
1154
3.66M
    for (int i = 0; i < existing_versions.size(); ++i) {
1155
2.35M
        if (existing_versions[i].first > max_continuous_version.second + 1) {
1156
0
            break;
1157
2.35M
        } else if (existing_versions[i].first <= max_continuous_version.second) {
1158
0
            *has_version_cross = true;
1159
0
        }
1160
2.35M
        max_continuous_version = existing_versions[i];
1161
2.35M
    }
1162
1.31M
    *version = max_continuous_version;
1163
    // tablet may not has rowset, eg, tablet has just been clear for restore.
1164
1.31M
    if (max_version != nullptr && !existing_versions.empty()) {
1165
1.31M
        *max_version = existing_versions.back();
1166
1.31M
    }
1167
1.31M
}
1168
1169
65.7k
void Tablet::calculate_cumulative_point() {
1170
65.7k
    std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
1171
65.7k
    SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
1172
65.7k
    int64_t ret_cumulative_point;
1173
65.7k
    _cumulative_compaction_policy->calculate_cumulative_point(
1174
65.7k
            this, _tablet_meta->all_rs_metas(), _cumulative_point, &ret_cumulative_point);
1175
1176
65.7k
    if (ret_cumulative_point == K_INVALID_CUMULATIVE_POINT) {
1177
63.1k
        return;
1178
63.1k
    }
1179
2.63k
    set_cumulative_layer_point(ret_cumulative_point);
1180
2.63k
}
1181
1182
// NOTE: only used when create_table, so it is sure that there is no concurrent reader and writer.
1183
0
void Tablet::delete_all_files() {
1184
    // Release resources like memory and disk space.
1185
0
    std::shared_lock rdlock(_meta_lock);
1186
0
    for (auto it : _rs_version_map) {
1187
0
        static_cast<void>(it.second->remove());
1188
0
    }
1189
0
    _rs_version_map.clear();
1190
1191
0
    for (auto it : _stale_rs_version_map) {
1192
0
        static_cast<void>(it.second->remove());
1193
0
    }
1194
0
    _stale_rs_version_map.clear();
1195
0
}
1196
1197
0
void Tablet::check_tablet_path_exists() {
1198
0
    if (!tablet_path().empty()) {
1199
0
        std::error_code ec;
1200
0
        if (std::filesystem::is_directory(tablet_path(), ec)) {
1201
0
            _is_tablet_path_exists.store(true, std::memory_order_relaxed);
1202
0
        } else if (ec.value() == ENOENT || ec.value() == 0) {
1203
0
            _is_tablet_path_exists.store(false, std::memory_order_relaxed);
1204
0
        }
1205
0
    }
1206
0
}
1207
1208
31.3k
Status Tablet::_contains_version(const Version& version) {
1209
    // check if there exist a rowset contains the added rowset
1210
381k
    for (auto& it : _rs_version_map) {
1211
381k
        if (it.first.contains(version)) {
1212
            // TODO(lingbin): Is this check unnecessary?
1213
            // because the value type is std::shared_ptr, when will it be nullptr?
1214
            // In addition, in this class, there are many places that do not make this judgment
1215
            // when access _rs_version_map's value.
1216
712
            CHECK(it.second != nullptr) << "there exist a version=" << it.first
1217
0
                                        << " contains the input rs with version=" << version
1218
0
                                        << ", but the related rs is null";
1219
712
            return Status::Error<PUSH_VERSION_ALREADY_EXIST>("Tablet push duplicate version {}",
1220
712
                                                             version.to_string());
1221
712
        }
1222
381k
    }
1223
1224
30.6k
    return Status::OK();
1225
31.3k
}
1226
1227
65.7k
std::vector<RowsetSharedPtr> Tablet::pick_candidate_rowsets_to_cumulative_compaction() {
1228
65.7k
    std::vector<RowsetSharedPtr> candidate_rowsets;
1229
65.7k
    if (_cumulative_point == K_INVALID_CUMULATIVE_POINT) {
1230
0
        return candidate_rowsets;
1231
0
    }
1232
65.7k
    return _pick_visible_rowsets_to_compaction(_cumulative_point,
1233
65.7k
                                               std::numeric_limits<int64_t>::max());
1234
65.7k
}
1235
1236
11.7k
std::vector<RowsetSharedPtr> Tablet::pick_candidate_rowsets_to_base_compaction() {
1237
11.7k
    return _pick_visible_rowsets_to_compaction(std::numeric_limits<int64_t>::min(),
1238
11.7k
                                               _cumulative_point - 1);
1239
11.7k
}
1240
1241
std::vector<RowsetSharedPtr> Tablet::_pick_visible_rowsets_to_compaction(
1242
77.4k
        int64_t min_start_version, int64_t max_start_version) {
1243
77.4k
    auto [visible_version, update_ts] = get_visible_version_and_time();
1244
77.4k
    bool update_time_long = MonotonicMillis() - update_ts >
1245
77.4k
                            config::compaction_keep_invisible_version_timeout_sec * 1000L;
1246
77.4k
    int32_t keep_invisible_version_limit =
1247
77.4k
            update_time_long ? config::compaction_keep_invisible_version_min_count
1248
77.4k
                             : config::compaction_keep_invisible_version_max_count;
1249
1250
77.4k
    std::vector<RowsetSharedPtr> candidate_rowsets;
1251
77.4k
    {
1252
77.4k
        std::shared_lock rlock(_meta_lock);
1253
377k
        for (const auto& [version, rs] : _rs_version_map) {
1254
377k
            int64_t version_start = version.first;
1255
            // rowset is remote or rowset is not in given range
1256
377k
            if (!rs->is_local() || version_start < min_start_version ||
1257
377k
                version_start > max_start_version) {
1258
108k
                continue;
1259
108k
            }
1260
1261
            // can compact, met one of the conditions:
1262
            // 1. had been visible;
1263
            // 2. exceeds the limit of keep invisible versions.
1264
269k
            int64_t version_end = version.second;
1265
269k
            if (version_end <= visible_version ||
1266
269k
                version_end > visible_version + keep_invisible_version_limit) {
1267
233k
                candidate_rowsets.push_back(rs);
1268
233k
            }
1269
269k
        }
1270
77.4k
    }
1271
77.4k
    std::sort(candidate_rowsets.begin(), candidate_rowsets.end(), Rowset::comparator);
1272
77.4k
    return candidate_rowsets;
1273
77.4k
}
1274
1275
0
std::vector<RowsetSharedPtr> Tablet::pick_candidate_rowsets_to_full_compaction() {
1276
0
    std::vector<RowsetSharedPtr> candidate_rowsets;
1277
0
    traverse_rowsets([&candidate_rowsets](const auto& rs) {
1278
        // Do full compaction on all local rowsets.
1279
0
        if (rs->is_local()) {
1280
0
            candidate_rowsets.emplace_back(rs);
1281
0
        }
1282
0
    });
1283
0
    std::sort(candidate_rowsets.begin(), candidate_rowsets.end(), Rowset::comparator);
1284
0
    return candidate_rowsets;
1285
0
}
1286
1287
std::vector<RowsetSharedPtr> Tablet::pick_candidate_rowsets_to_build_inverted_index(
1288
21
        const std::set<int64_t>& alter_index_uids, bool is_drop_op) {
1289
21
    std::vector<RowsetSharedPtr> candidate_rowsets;
1290
21
    {
1291
21
        std::shared_lock rlock(_meta_lock);
1292
41
        auto has_alter_inverted_index = [&](RowsetSharedPtr rowset) -> bool {
1293
45
            for (const auto& index_id : alter_index_uids) {
1294
45
                if (rowset->tablet_schema()->has_inverted_index_with_index_id(index_id)) {
1295
8
                    return true;
1296
8
                }
1297
45
            }
1298
33
            return false;
1299
41
        };
1300
1301
21
        for (const auto& [version, rs] : _rs_version_map) {
1302
21
            if (!has_alter_inverted_index(rs) && is_drop_op) {
1303
1
                continue;
1304
1
            }
1305
20
            if (has_alter_inverted_index(rs) && !is_drop_op) {
1306
0
                continue;
1307
0
            }
1308
1309
20
            if (rs->is_local()) {
1310
20
                candidate_rowsets.push_back(rs);
1311
20
            }
1312
20
        }
1313
21
    }
1314
21
    std::sort(candidate_rowsets.begin(), candidate_rowsets.end(), Rowset::comparator);
1315
21
    return candidate_rowsets;
1316
21
}
1317
1318
1.39M
std::tuple<int64_t, int64_t> Tablet::get_visible_version_and_time() const {
1319
    // some old tablet has bug, its partition_id is 0, fe couldn't update its visible version.
1320
    // so let this tablet's visible version become int64 max.
1321
1.39M
    auto version_info = _visible_version.load();
1322
1.39M
    if (version_info != nullptr && partition_id() != 0) {
1323
1.39M
        return std::make_tuple(version_info->version.load(std::memory_order_relaxed),
1324
1.39M
                               version_info->update_ts);
1325
1.39M
    } else {
1326
90
        return std::make_tuple(std::numeric_limits<int64_t>::max(),
1327
90
                               std::numeric_limits<int64_t>::max());
1328
90
    }
1329
1.39M
}
1330
1331
// For http compaction action
1332
0
void Tablet::get_compaction_status(std::string* json_result) {
1333
0
    rapidjson::Document root;
1334
0
    root.SetObject();
1335
1336
0
    rapidjson::Document path_arr;
1337
0
    path_arr.SetArray();
1338
1339
0
    std::vector<RowsetSharedPtr> rowsets;
1340
0
    std::vector<RowsetSharedPtr> stale_rowsets;
1341
0
    std::vector<bool> delete_flags;
1342
0
    {
1343
0
        std::shared_lock rdlock(_meta_lock);
1344
0
        rowsets.reserve(_rs_version_map.size());
1345
0
        for (auto& it : _rs_version_map) {
1346
0
            rowsets.push_back(it.second);
1347
0
        }
1348
0
        std::sort(rowsets.begin(), rowsets.end(), Rowset::comparator);
1349
1350
0
        stale_rowsets.reserve(_stale_rs_version_map.size());
1351
0
        for (auto& it : _stale_rs_version_map) {
1352
0
            stale_rowsets.push_back(it.second);
1353
0
        }
1354
0
        std::sort(stale_rowsets.begin(), stale_rowsets.end(), Rowset::comparator);
1355
1356
0
        delete_flags.reserve(rowsets.size());
1357
0
        for (auto& rs : rowsets) {
1358
0
            delete_flags.push_back(rs->rowset_meta()->has_delete_predicate());
1359
0
        }
1360
        // get snapshot version path json_doc
1361
0
        _timestamped_version_tracker.get_stale_version_path_json_doc(path_arr);
1362
0
    }
1363
0
    rapidjson::Value cumulative_policy_type;
1364
0
    std::string policy_type_str = "cumulative compaction policy not initializied";
1365
0
    if (_cumulative_compaction_policy != nullptr) {
1366
0
        policy_type_str = _cumulative_compaction_policy->name();
1367
0
    }
1368
0
    cumulative_policy_type.SetString(policy_type_str.c_str(),
1369
0
                                     cast_set<uint32_t>(policy_type_str.length()),
1370
0
                                     root.GetAllocator());
1371
0
    root.AddMember("cumulative policy type", cumulative_policy_type, root.GetAllocator());
1372
0
    root.AddMember("cumulative point", _cumulative_point.load(), root.GetAllocator());
1373
1374
0
#define FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, key, unixmillis_value)                 \
1375
0
    {                                                                                \
1376
0
        rapidjson::Value value;                                                      \
1377
0
        std::string format_str = ToStringFromUnixMillis(unixmillis_value.load());    \
1378
0
        value.SetString(format_str.c_str(), cast_set<uint32_t>(format_str.length()), \
1379
0
                        root.GetAllocator());                                        \
1380
0
        root.AddMember(key, value, root.GetAllocator());                             \
1381
0
    }
1382
0
#define FORMAT_STRING_ADD_JSON_NODE(root, key, str_value)                          \
1383
0
    {                                                                              \
1384
0
        rapidjson::Value value;                                                    \
1385
0
        value.SetString(str_value.c_str(), cast_set<uint32_t>(str_value.length()), \
1386
0
                        root.GetAllocator());                                      \
1387
0
        root.AddMember(key, value, root.GetAllocator());                           \
1388
0
    }
1389
1390
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last cumulative failure time",
1391
0
                                    _last_cumu_compaction_failure_millis)
1392
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last base failure time",
1393
0
                                    _last_base_compaction_failure_millis)
1394
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last full failure time",
1395
0
                                    _last_full_compaction_failure_millis)
1396
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last cumulative success time",
1397
0
                                    _last_cumu_compaction_success_millis)
1398
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last base success time",
1399
0
                                    _last_base_compaction_success_millis)
1400
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last full success time",
1401
0
                                    _last_full_compaction_success_millis)
1402
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last cumulative schedule time",
1403
0
                                    _last_cumu_compaction_schedule_millis)
1404
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last base schedule time",
1405
0
                                    _last_base_compaction_schedule_millis)
1406
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last full schedule time",
1407
0
                                    _last_full_compaction_schedule_millis)
1408
0
    FORMAT_STRING_ADD_JSON_NODE(root, "last cumulative status", _last_cumu_compaction_status)
1409
0
    FORMAT_STRING_ADD_JSON_NODE(root, "last base status", _last_base_compaction_status)
1410
0
    FORMAT_STRING_ADD_JSON_NODE(root, "last full status", _last_full_compaction_status)
1411
1412
    // last single replica compaction status
1413
    // "single replica compaction status": {
1414
    //     "remote peer": "172.100.1.0:10875",
1415
    //     "last failure status": "",
1416
    //     "last fetched rowset": "[8-10]"
1417
    // }
1418
0
    rapidjson::Document status;
1419
0
    status.SetObject();
1420
0
    TReplicaInfo replica_info;
1421
0
    std::string dummp_token;
1422
0
    if (tablet_meta()->tablet_schema()->enable_single_replica_compaction() &&
1423
0
        _engine.get_peer_replica_info(tablet_id(), &replica_info, &dummp_token)) {
1424
        // remote peer
1425
0
        rapidjson::Value peer_addr;
1426
0
        std::string addr = replica_info.host + ":" + std::to_string(replica_info.brpc_port);
1427
0
        peer_addr.SetString(addr.c_str(), cast_set<uint32_t>(addr.length()), status.GetAllocator());
1428
0
        status.AddMember("remote peer", peer_addr, status.GetAllocator());
1429
        // last failure status
1430
0
        rapidjson::Value compaction_status;
1431
0
        compaction_status.SetString(
1432
0
                _last_single_compaction_failure_status.c_str(),
1433
0
                cast_set<uint32_t>(_last_single_compaction_failure_status.length()),
1434
0
                status.GetAllocator());
1435
0
        status.AddMember("last failure status", compaction_status, status.GetAllocator());
1436
        // last fetched rowset
1437
0
        rapidjson::Value version;
1438
0
        std::string fetched_version = _last_fetched_version.to_string();
1439
0
        version.SetString(fetched_version.c_str(), cast_set<uint32_t>(fetched_version.length()),
1440
0
                          status.GetAllocator());
1441
0
        status.AddMember("last fetched rowset", version, status.GetAllocator());
1442
0
        root.AddMember("single replica compaction status", status, root.GetAllocator());
1443
0
    }
1444
1445
    // print all rowsets' version as an array
1446
0
    rapidjson::Document versions_arr;
1447
0
    rapidjson::Document missing_versions_arr;
1448
0
    versions_arr.SetArray();
1449
0
    missing_versions_arr.SetArray();
1450
0
    int64_t last_version = -1;
1451
0
    for (auto& rowset : rowsets) {
1452
0
        const Version& ver = rowset->version();
1453
0
        if (ver.first != last_version + 1) {
1454
0
            rapidjson::Value miss_value;
1455
0
            miss_value.SetString(fmt::format("[{}-{}]", last_version + 1, ver.first - 1).c_str(),
1456
0
                                 missing_versions_arr.GetAllocator());
1457
0
            missing_versions_arr.PushBack(miss_value, missing_versions_arr.GetAllocator());
1458
0
        }
1459
0
        rapidjson::Value value;
1460
0
        std::string version_str = rowset->get_rowset_info_str();
1461
0
        value.SetString(version_str.c_str(), cast_set<uint32_t>(version_str.length()),
1462
0
                        versions_arr.GetAllocator());
1463
0
        versions_arr.PushBack(value, versions_arr.GetAllocator());
1464
0
        last_version = ver.second;
1465
0
    }
1466
0
    root.AddMember("rowsets", versions_arr, root.GetAllocator());
1467
0
    root.AddMember("missing_rowsets", missing_versions_arr, root.GetAllocator());
1468
1469
    // print all stale rowsets' version as an array
1470
0
    rapidjson::Document stale_versions_arr;
1471
0
    stale_versions_arr.SetArray();
1472
0
    for (auto& rowset : stale_rowsets) {
1473
0
        rapidjson::Value value;
1474
0
        std::string version_str = rowset->get_rowset_info_str();
1475
0
        value.SetString(version_str.c_str(), cast_set<uint32_t>(version_str.length()),
1476
0
                        stale_versions_arr.GetAllocator());
1477
0
        stale_versions_arr.PushBack(value, stale_versions_arr.GetAllocator());
1478
0
    }
1479
0
    root.AddMember("stale_rowsets", stale_versions_arr, root.GetAllocator());
1480
1481
    // add stale version rowsets
1482
0
    root.AddMember("stale version path", path_arr, root.GetAllocator());
1483
1484
    // to json string
1485
0
    rapidjson::StringBuffer strbuf;
1486
0
    rapidjson::PrettyWriter<rapidjson::StringBuffer> writer(strbuf);
1487
0
    root.Accept(writer);
1488
0
    *json_result = std::string(strbuf.GetString());
1489
0
}
1490
1491
339k
bool Tablet::do_tablet_meta_checkpoint() {
1492
339k
    std::lock_guard<std::shared_mutex> store_lock(_meta_store_lock);
1493
339k
    if (_newly_created_rowset_num == 0) {
1494
320k
        return false;
1495
320k
    }
1496
18.4k
    if (UnixMillis() - _last_checkpoint_time <
1497
18.4k
                config::tablet_meta_checkpoint_min_interval_secs * 1000 &&
1498
18.4k
        _newly_created_rowset_num < config::tablet_meta_checkpoint_min_new_rowsets_num) {
1499
0
        return false;
1500
0
    }
1501
    // hold read-lock other than write-lock, because it will not modify meta structure
1502
18.4k
    std::shared_lock rdlock(_meta_lock);
1503
18.4k
    if (tablet_state() != TABLET_RUNNING) {
1504
1
        LOG(INFO) << "tablet is under state=" << tablet_state()
1505
1
                  << ", not running, skip do checkpoint"
1506
1
                  << ", tablet=" << tablet_id();
1507
1
        return false;
1508
1
    }
1509
18.4k
    VLOG_NOTICE << "start to do tablet meta checkpoint, tablet=" << tablet_id();
1510
18.4k
    save_meta();
1511
    // if save meta successfully, then should remove the rowset meta existing in tablet
1512
    // meta from rowset meta store
1513
38.2k
    for (const auto& [_, rs_meta] : _tablet_meta->all_rs_metas()) {
1514
        // If we delete it from rowset manager's meta explicitly in previous checkpoint, just skip.
1515
38.2k
        if (rs_meta->is_remove_from_rowset_meta()) {
1516
0
            continue;
1517
0
        }
1518
38.2k
        if (RowsetMetaManager::check_rowset_meta(_data_dir->get_meta(), tablet_uid(),
1519
38.2k
                                                 rs_meta->rowset_id())) {
1520
32.9k
            RETURN_FALSE_IF_ERROR(RowsetMetaManager::remove(_data_dir->get_meta(), tablet_uid(),
1521
32.9k
                                                            rs_meta->rowset_id()));
1522
32.9k
            VLOG_NOTICE << "remove rowset id from meta store because it is already persistent with "
1523
0
                        << "tablet meta, rowset_id=" << rs_meta->rowset_id();
1524
32.9k
        }
1525
38.2k
        rs_meta->set_remove_from_rowset_meta();
1526
38.2k
    }
1527
1528
    // check _stale_rs_version_map to remove meta from rowset meta store
1529
18.4k
    for (const auto& [_, rs_meta] : _tablet_meta->all_stale_rs_metas()) {
1530
        // If we delete it from rowset manager's meta explicitly in previous checkpoint, just skip.
1531
3.52k
        if (rs_meta->is_remove_from_rowset_meta()) {
1532
0
            continue;
1533
0
        }
1534
3.52k
        if (RowsetMetaManager::check_rowset_meta(_data_dir->get_meta(), tablet_uid(),
1535
3.52k
                                                 rs_meta->rowset_id())) {
1536
3.42k
            RETURN_FALSE_IF_ERROR(RowsetMetaManager::remove(_data_dir->get_meta(), tablet_uid(),
1537
3.42k
                                                            rs_meta->rowset_id()));
1538
3.42k
            VLOG_NOTICE << "remove rowset id from meta store because it is already persistent with "
1539
0
                        << "tablet meta, rowset_id=" << rs_meta->rowset_id();
1540
3.42k
        }
1541
3.52k
        rs_meta->set_remove_from_rowset_meta();
1542
3.52k
    }
1543
1544
18.4k
    if (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write()) {
1545
92
        RETURN_FALSE_IF_ERROR(TabletMetaManager::remove_old_version_delete_bitmap(
1546
92
                _data_dir, tablet_id(), max_version_unlocked()));
1547
92
    }
1548
1549
18.4k
    _newly_created_rowset_num = 0;
1550
18.4k
    _last_checkpoint_time = UnixMillis();
1551
18.4k
    return true;
1552
18.4k
}
1553
1554
8.88k
bool Tablet::rowset_meta_is_useful(RowsetMetaSharedPtr rowset_meta) {
1555
8.88k
    std::shared_lock rdlock(_meta_lock);
1556
8.88k
    bool find_version = false;
1557
11.9k
    for (auto& version_rowset : _rs_version_map) {
1558
11.9k
        if (version_rowset.second->rowset_id() == rowset_meta->rowset_id()) {
1559
7.33k
            return true;
1560
7.33k
        }
1561
4.65k
        if (version_rowset.second->contains_version(rowset_meta->version())) {
1562
1.54k
            find_version = true;
1563
1.54k
        }
1564
4.65k
    }
1565
3.46k
    for (auto& stale_version_rowset : _stale_rs_version_map) {
1566
3.46k
        if (stale_version_rowset.second->rowset_id() == rowset_meta->rowset_id()) {
1567
818
            return true;
1568
818
        }
1569
2.64k
        if (stale_version_rowset.second->contains_version(rowset_meta->version())) {
1570
302
            find_version = true;
1571
302
        }
1572
2.64k
    }
1573
728
    return !find_version;
1574
1.54k
}
1575
1576
32.0k
bool Tablet::_contains_rowset(const RowsetId rowset_id) {
1577
384k
    for (auto& version_rowset : _rs_version_map) {
1578
384k
        if (version_rowset.second->rowset_id() == rowset_id) {
1579
55
            return true;
1580
55
        }
1581
384k
    }
1582
31.9k
    for (auto& stale_version_rowset : _stale_rs_version_map) {
1583
20.3k
        if (stale_version_rowset.second->rowset_id() == rowset_id) {
1584
650
            return true;
1585
650
        }
1586
20.3k
    }
1587
31.3k
    return false;
1588
31.9k
}
1589
1590
// need check if consecutive version missing in full report
1591
// alter tablet will ignore this check
1592
void Tablet::build_tablet_report_info(TTabletInfo* tablet_info,
1593
                                      bool enable_consecutive_missing_check,
1594
1.31M
                                      bool enable_path_check) {
1595
1.31M
    std::shared_lock rdlock(_meta_lock);
1596
1.31M
    tablet_info->__set_tablet_id(_tablet_meta->tablet_id());
1597
1.31M
    tablet_info->__set_schema_hash(_tablet_meta->schema_hash());
1598
1.31M
    tablet_info->__set_row_count(_tablet_meta->num_rows());
1599
1.31M
    tablet_info->__set_data_size(_tablet_meta->tablet_local_size());
1600
1601
    // Here we need to report to FE if there are any missing versions of tablet.
1602
    // We start from the initial version and traverse backwards until we meet a discontinuous version.
1603
1.31M
    Version cversion;
1604
1.31M
    Version max_version;
1605
1.31M
    bool has_version_cross;
1606
1.31M
    _max_continuous_version_from_beginning_unlocked(&cversion, &max_version, &has_version_cross);
1607
    // cause publish version task runs concurrently, version may be flying
1608
    // so we add a consecutive miss check to solve this problem:
1609
    // if publish version 5 arrives but version 4 flying, we may judge replica miss version
1610
    // and set version miss in tablet_info, which makes fe treat this replica as unhealth
1611
    // and lead to other problems
1612
1.31M
    if (enable_consecutive_missing_check) {
1613
1.31M
        if (cversion.second < max_version.second) {
1614
0
            if (_last_missed_version == cversion.second + 1) {
1615
0
                if (MonotonicSeconds() - _last_missed_time_s >= 60) {
1616
                    // version missed for over 60 seconds
1617
0
                    tablet_info->__set_version_miss(true);
1618
0
                    _last_missed_version = -1;
1619
0
                    _last_missed_time_s = 0;
1620
0
                }
1621
0
            } else {
1622
0
                _last_missed_version = cversion.second + 1;
1623
0
                _last_missed_time_s = MonotonicSeconds();
1624
0
            }
1625
0
        }
1626
1.31M
    } else {
1627
0
        tablet_info->__set_version_miss(cversion.second < max_version.second);
1628
0
    }
1629
1630
1.31M
    DBUG_EXECUTE_IF("Tablet.build_tablet_report_info.version_miss", {
1631
1.31M
        auto tablet_id = dp->param<int64_t>("tablet_id", -1);
1632
1.31M
        if (tablet_id != -1 && tablet_id == _tablet_meta->tablet_id()) {
1633
1.31M
            auto miss = dp->param<bool>("version_miss", true);
1634
1.31M
            tablet_info->__set_version_miss(miss);
1635
1.31M
        }
1636
1.31M
    });
1637
1638
    // find rowset with max version
1639
1.31M
    auto iter = _rs_version_map.find(max_version);
1640
1.31M
    if (iter == _rs_version_map.end()) {
1641
        // If the tablet is in running state, it must not be doing schema-change. so if we can not
1642
        // access its rowsets, it means that the tablet is bad and needs to be reported to the FE
1643
        // for subsequent repairs (through the cloning task)
1644
0
        if (tablet_state() == TABLET_RUNNING) {
1645
0
            tablet_info->__set_used(false);
1646
0
        }
1647
        // For other states, FE knows that the tablet is in a certain change process, so here
1648
        // still sets the state to normal when reporting. Note that every task has an timeout,
1649
        // so if the task corresponding to this change hangs, when the task timeout, FE will know
1650
        // and perform state modification operations.
1651
0
    }
1652
1653
1.31M
    if (tablet_state() == TABLET_RUNNING) {
1654
1.31M
        if (has_version_cross || is_io_error_too_times() || !data_dir()->is_used()) {
1655
0
            LOG(INFO) << "report " << tablet_id() << " as bad, version_cross=" << has_version_cross
1656
0
                      << ", ioe times=" << get_io_error_times() << ", data_dir used "
1657
0
                      << data_dir()->is_used();
1658
0
            tablet_info->__set_used(false);
1659
0
        }
1660
1661
1.31M
        if (enable_path_check) {
1662
1.31M
            if (!_is_tablet_path_exists.exchange(true, std::memory_order_relaxed)) {
1663
0
                LOG(INFO) << "report " << tablet_id() << " as bad, tablet directory not found";
1664
0
                tablet_info->__set_used(false);
1665
0
            }
1666
1.31M
        }
1667
1.31M
    }
1668
1669
    // There are two cases when tablet state is TABLET_NOTREADY
1670
    // case 1: tablet is doing schema change. Fe knows it's state, doing nothing.
1671
    // case 2: tablet has finished schema change, but failed. Fe will perform recovery.
1672
1.31M
    if (tablet_state() == TABLET_NOTREADY && is_alter_failed()) {
1673
0
        tablet_info->__set_used(false);
1674
0
    }
1675
1676
1.31M
    if (tablet_state() == TABLET_SHUTDOWN) {
1677
0
        tablet_info->__set_used(false);
1678
0
    }
1679
1680
1.31M
    DBUG_EXECUTE_IF("Tablet.build_tablet_report_info.used", {
1681
1.31M
        auto tablet_id = dp->param<int64_t>("tablet_id", -1);
1682
1.31M
        if (tablet_id != -1 && tablet_id == _tablet_meta->tablet_id()) {
1683
1.31M
            auto used = dp->param<bool>("used", true);
1684
1.31M
            LOG_WARNING("Tablet.build_tablet_report_info.used")
1685
1.31M
                    .tag("tablet id", tablet_id)
1686
1.31M
                    .tag("used", used);
1687
1.31M
            tablet_info->__set_used(used);
1688
1.31M
        } else {
1689
1.31M
            LOG_WARNING("Tablet.build_tablet_report_info.used").tag("tablet id", tablet_id);
1690
1.31M
        }
1691
1.31M
    });
1692
1693
1.31M
    int64_t total_version_count = _tablet_meta->version_count();
1694
1695
    // For compatibility.
1696
    // For old fe, it wouldn't send visible version request to be, then be's visible version is always 0.
1697
    // Let visible_version_count set to total_version_count in be's report.
1698
1.31M
    int64_t visible_version_count = total_version_count;
1699
1.31M
    if (auto [visible_version, _] = get_visible_version_and_time(); visible_version > 0) {
1700
959k
        visible_version_count = _tablet_meta->version_count_cross_with_range({0, visible_version});
1701
959k
    }
1702
    // the report version is the largest continuous version, same logic as in FE side
1703
1.31M
    tablet_info->__set_version(cversion.second);
1704
    // Useless but it is a required filed in TTabletInfo
1705
1.31M
    tablet_info->__set_version_hash(0);
1706
1.31M
    tablet_info->__set_partition_id(_tablet_meta->partition_id());
1707
1.31M
    tablet_info->__set_storage_medium(_data_dir->storage_medium());
1708
1.31M
    tablet_info->__set_total_version_count(total_version_count);
1709
1.31M
    tablet_info->__set_visible_version_count(visible_version_count);
1710
1.31M
    tablet_info->__set_path_hash(_data_dir->path_hash());
1711
1.31M
    tablet_info->__set_is_in_memory(_tablet_meta->tablet_schema()->is_in_memory());
1712
1.31M
    tablet_info->__set_replica_id(replica_id());
1713
1.31M
    tablet_info->__set_remote_data_size(_tablet_meta->tablet_remote_size());
1714
1.31M
    if (_tablet_meta->cooldown_meta_id().initialized()) { // has cooldowned data
1715
19
        tablet_info->__set_cooldown_term(_cooldown_conf.term);
1716
19
        tablet_info->__set_cooldown_meta_id(_tablet_meta->cooldown_meta_id().to_thrift());
1717
19
    }
1718
1.31M
    if (tablet_state() == TABLET_RUNNING && _tablet_meta->storage_policy_id() > 0) {
1719
        // tablet may not have cooldowned data, but the storage policy is set
1720
1.31k
        tablet_info->__set_cooldown_term(_cooldown_conf.term);
1721
1.31k
    }
1722
1.31M
    tablet_info->__set_local_index_size(_tablet_meta->tablet_local_index_size());
1723
1.31M
    tablet_info->__set_local_segment_size(_tablet_meta->tablet_local_segment_size());
1724
1.31M
    tablet_info->__set_remote_index_size(_tablet_meta->tablet_remote_index_size());
1725
1.31M
    tablet_info->__set_remote_segment_size(_tablet_meta->tablet_remote_segment_size());
1726
1.31M
}
1727
1728
1
void Tablet::report_error(const Status& st) {
1729
1
    if (st.is<ErrorCode::IO_ERROR>()) {
1730
0
        ++_io_error_times;
1731
1
    } else if (st.is<ErrorCode::CORRUPTION>()) {
1732
1
        _io_error_times = config::max_tablet_io_errors + 1;
1733
1
    } else if (st.is<ErrorCode::NOT_FOUND>()) {
1734
0
        check_tablet_path_exists();
1735
0
        if (!_is_tablet_path_exists.load(std::memory_order_relaxed)) {
1736
0
            _io_error_times = config::max_tablet_io_errors + 1;
1737
0
        }
1738
0
    }
1739
1
}
1740
1741
Status Tablet::prepare_compaction_and_calculate_permits(
1742
        CompactionType compaction_type, const TabletSharedPtr& tablet,
1743
77.4k
        std::shared_ptr<CompactionMixin>& compaction, int64_t& permits) {
1744
77.4k
    if (compaction_type == CompactionType::CUMULATIVE_COMPACTION) {
1745
65.6k
        MonotonicStopWatch watch;
1746
65.6k
        watch.start();
1747
1748
65.6k
        compaction = std::make_shared<CumulativeCompaction>(tablet->_engine, tablet);
1749
65.6k
        DorisMetrics::instance()->cumulative_compaction_request_total->increment(1);
1750
65.6k
        Status res = compaction->prepare_compact();
1751
65.6k
        if (!config::disable_compaction_trace_log &&
1752
65.6k
            static_cast<double>(watch.elapsed_time()) / 1e9 >
1753
0
                    config::cumulative_compaction_trace_threshold) {
1754
0
            std::stringstream ss;
1755
0
            compaction->runtime_profile()->pretty_print(&ss);
1756
0
            LOG(WARNING) << "prepare cumulative compaction cost "
1757
0
                         << static_cast<double>(watch.elapsed_time()) / 1e9 << std::endl
1758
0
                         << ss.str();
1759
0
        }
1760
1761
65.6k
        if (!res.ok()) {
1762
65.6k
            permits = 0;
1763
            // if we meet a delete version, should increase the cumulative point to let base compaction handle the delete version.
1764
            // no need to wait 5s.
1765
65.6k
            if (!res.is<ErrorCode::CUMULATIVE_MEET_DELETE_VERSION>() ||
1766
65.6k
                config::enable_sleep_between_delete_cumu_compaction) {
1767
65.5k
                tablet->set_last_cumu_compaction_failure_time(UnixMillis());
1768
65.5k
            }
1769
65.6k
            if (!res.is<CUMULATIVE_NO_SUITABLE_VERSION>() &&
1770
65.6k
                !res.is<ErrorCode::CUMULATIVE_MEET_DELETE_VERSION>()) {
1771
0
                DorisMetrics::instance()->cumulative_compaction_request_failed->increment(1);
1772
0
                return Status::InternalError("prepare cumulative compaction with err: {}",
1773
0
                                             res.to_string());
1774
0
            }
1775
            // return OK if OLAP_ERR_CUMULATIVE_NO_SUITABLE_VERSION, so that we don't need to
1776
            // print too much useless logs.
1777
            // And because we set permits to 0, so even if we return OK here, nothing will be done.
1778
65.6k
            return Status::OK();
1779
65.6k
        }
1780
65.6k
    } else if (compaction_type == CompactionType::BASE_COMPACTION) {
1781
11.7k
        MonotonicStopWatch watch;
1782
11.7k
        watch.start();
1783
1784
11.7k
        compaction = std::make_shared<BaseCompaction>(tablet->_engine, tablet);
1785
11.7k
        DorisMetrics::instance()->base_compaction_request_total->increment(1);
1786
11.7k
        Status res = compaction->prepare_compact();
1787
11.7k
        if (!config::disable_compaction_trace_log &&
1788
11.7k
            static_cast<double>(watch.elapsed_time()) / 1e9 >
1789
0
                    config::base_compaction_trace_threshold) {
1790
0
            std::stringstream ss;
1791
0
            compaction->runtime_profile()->pretty_print(&ss);
1792
0
            LOG(WARNING) << "prepare base compaction cost "
1793
0
                         << static_cast<double>(watch.elapsed_time()) / 1e9 << std::endl
1794
0
                         << ss.str();
1795
0
        }
1796
1797
11.7k
        tablet->set_last_base_compaction_status(res.to_string());
1798
11.7k
        if (!res.ok()) {
1799
11.7k
            tablet->set_last_base_compaction_failure_time(UnixMillis());
1800
11.7k
            permits = 0;
1801
11.7k
            if (!res.is<BE_NO_SUITABLE_VERSION>()) {
1802
0
                DorisMetrics::instance()->base_compaction_request_failed->increment(1);
1803
0
                return Status::InternalError("prepare base compaction with err: {}",
1804
0
                                             res.to_string());
1805
0
            }
1806
            // return OK if OLAP_ERR_BE_NO_SUITABLE_VERSION, so that we don't need to
1807
            // print too much useless logs.
1808
            // And because we set permits to 0, so even if we return OK here, nothing will be done.
1809
11.7k
            return Status::OK();
1810
11.7k
        }
1811
11.7k
    } else {
1812
0
        DCHECK_EQ(compaction_type, CompactionType::FULL_COMPACTION);
1813
1814
0
        compaction = std::make_shared<FullCompaction>(tablet->_engine, tablet);
1815
0
        Status res = compaction->prepare_compact();
1816
0
        if (!res.ok()) {
1817
0
            tablet->set_last_full_compaction_failure_time(UnixMillis());
1818
0
            permits = 0;
1819
0
            if (!res.is<FULL_NO_SUITABLE_VERSION>()) {
1820
0
                return Status::InternalError("prepare full compaction with err: {}",
1821
0
                                             res.to_string());
1822
0
            }
1823
            // return OK if OLAP_ERR_BE_NO_SUITABLE_VERSION, so that we don't need to
1824
            // print too much useless logs.
1825
            // And because we set permits to 0, so even if we return OK here, nothing will be done.
1826
0
            return Status::OK();
1827
0
        }
1828
0
    }
1829
1830
    // Time series policy does not rely on permits, it uses goal size to control memory
1831
90
    if (tablet->tablet_meta()->compaction_policy() == CUMULATIVE_TIME_SERIES_POLICY) {
1832
        // permits = 0 means that prepare_compaction failed
1833
0
        permits = 1;
1834
90
    } else {
1835
90
        permits = compaction->get_compaction_permits();
1836
90
    }
1837
90
    return Status::OK();
1838
77.4k
}
1839
1840
0
void Tablet::execute_single_replica_compaction(SingleReplicaCompaction& compaction) {
1841
0
    Status res = compaction.execute_compact();
1842
0
    if (!res.ok()) {
1843
0
        set_last_failure_time(this, compaction, UnixMillis());
1844
0
        set_last_single_compaction_failure_status(res.to_string());
1845
0
        if (res.is<CANCELLED>()) {
1846
0
            DorisMetrics::instance()->single_compaction_request_cancelled->increment(1);
1847
            // "CANCELLED" indicates that the peer has not performed compaction,
1848
            // wait for the peer to perform compaction
1849
0
            set_skip_compaction(true, compaction.real_compact_type(), UnixSeconds());
1850
0
            VLOG_CRITICAL << "Cannel fetching from the remote peer. res=" << res
1851
0
                          << ", tablet=" << tablet_id();
1852
0
        } else {
1853
0
            DorisMetrics::instance()->single_compaction_request_failed->increment(1);
1854
0
            LOG(WARNING) << "failed to do single replica compaction. res=" << res
1855
0
                         << ", tablet=" << tablet_id();
1856
0
        }
1857
0
        return;
1858
0
    }
1859
0
    set_last_failure_time(this, compaction, 0);
1860
0
}
1861
1862
28.8M
bool Tablet::should_fetch_from_peer() {
1863
28.8M
    return tablet_meta()->tablet_schema()->enable_single_replica_compaction() &&
1864
28.8M
           _engine.should_fetch_from_peer(tablet_id());
1865
28.8M
}
1866
1867
3
std::vector<Version> Tablet::get_all_local_versions() {
1868
3
    std::vector<Version> local_versions;
1869
3
    {
1870
3
        std::shared_lock rlock(_meta_lock);
1871
63
        for (const auto& [version, rs] : _rs_version_map) {
1872
63
            if (rs->is_local()) {
1873
53
                local_versions.emplace_back(version);
1874
53
            }
1875
63
        }
1876
3
    }
1877
3
    std::sort(local_versions.begin(), local_versions.end(),
1878
324
              [](const Version& left, const Version& right) { return left.first < right.first; });
1879
3
    return local_versions;
1880
3
}
1881
1882
70
void Tablet::execute_compaction(CompactionMixin& compaction) {
1883
70
    signal::tablet_id = tablet_id();
1884
1885
70
    MonotonicStopWatch watch;
1886
70
    watch.start();
1887
1888
70
    Status res = [&]() { RETURN_IF_CATCH_EXCEPTION({ return compaction.execute_compact(); }); }();
1889
1890
70
    if (!res.ok()) [[unlikely]] {
1891
0
        set_last_failure_time(this, compaction, UnixMillis());
1892
0
        LOG(WARNING) << "failed to do " << compaction.compaction_name()
1893
0
                     << ", tablet=" << tablet_id() << " : " << res;
1894
70
    } else {
1895
70
        set_last_failure_time(this, compaction, 0);
1896
70
    }
1897
1898
70
    if (!config::disable_compaction_trace_log) {
1899
0
        auto need_trace = [&compaction, &watch] {
1900
0
            return compaction.compaction_type() == ReaderType::READER_CUMULATIVE_COMPACTION
1901
0
                           ? static_cast<double>(watch.elapsed_time()) / 1e9 >
1902
0
                                     config::cumulative_compaction_trace_threshold
1903
0
                   : compaction.compaction_type() == ReaderType::READER_BASE_COMPACTION
1904
0
                           ? static_cast<double>(watch.elapsed_time()) / 1e9 >
1905
0
                                     config::base_compaction_trace_threshold
1906
0
                           : false;
1907
0
        };
1908
0
        if (need_trace()) {
1909
0
            std::stringstream ss;
1910
0
            compaction.runtime_profile()->pretty_print(&ss);
1911
0
            LOG(WARNING) << "execute " << compaction.compaction_name() << " cost "
1912
0
                         << static_cast<double>(watch.elapsed_time()) / 1e9 << std::endl
1913
0
                         << ss.str();
1914
0
        }
1915
0
    }
1916
70
}
1917
1918
304
Status Tablet::create_initial_rowset(const int64_t req_version) {
1919
304
    if (req_version < 1) {
1920
0
        return Status::Error<CE_CMD_PARAMS_ERROR>(
1921
0
                "init version of tablet should at least 1. req.ver={}", req_version);
1922
0
    }
1923
304
    Version version(0, req_version);
1924
304
    RowsetSharedPtr new_rowset;
1925
    // there is no data in init rowset, so overlapping info is unknown.
1926
304
    RowsetWriterContext context;
1927
304
    context.version = version;
1928
304
    context.rowset_state = VISIBLE;
1929
304
    context.segments_overlap = OVERLAP_UNKNOWN;
1930
304
    context.tablet_schema = tablet_schema();
1931
304
    context.newest_write_timestamp = UnixSeconds();
1932
304
    auto rs_writer = DORIS_TRY(create_rowset_writer(context, false));
1933
304
    RETURN_IF_ERROR(rs_writer->flush());
1934
304
    RETURN_IF_ERROR(rs_writer->build(new_rowset));
1935
304
    RETURN_IF_ERROR(add_rowset(std::move(new_rowset)));
1936
304
    set_cumulative_layer_point(req_version + 1);
1937
304
    return Status::OK();
1938
304
}
1939
1940
Result<std::unique_ptr<RowsetWriter>> Tablet::create_rowset_writer(RowsetWriterContext& context,
1941
916
                                                                   bool vertical) {
1942
916
    context.rowset_id = _engine.next_rowset_id();
1943
916
    _init_context_common_fields(context);
1944
916
    return RowsetFactory::create_rowset_writer(_engine, context, vertical);
1945
916
}
1946
1947
// create a rowset writer with rowset_id and seg_id
1948
// after writer, merge this transient rowset with original rowset
1949
Result<std::unique_ptr<RowsetWriter>> Tablet::create_transient_rowset_writer(
1950
        const Rowset& rowset, std::shared_ptr<PartialUpdateInfo> partial_update_info,
1951
42
        int64_t txn_expiration) {
1952
42
    RowsetWriterContext context;
1953
42
    context.rowset_state = PREPARED;
1954
42
    context.segments_overlap = OVERLAPPING;
1955
42
    context.tablet_schema = std::make_shared<TabletSchema>();
1956
    // During a partial update, the extracted columns of a variant should not be included in the tablet schema.
1957
    // This is because the partial update for a variant needs to ignore the extracted columns.
1958
    // Otherwise, the schema types in different rowsets might be inconsistent. When performing a partial update,
1959
    // the complete variant is constructed by reading all the sub-columns of the variant.
1960
42
    context.tablet_schema = rowset.tablet_schema()->copy_without_variant_extracted_columns();
1961
42
    context.newest_write_timestamp = UnixSeconds();
1962
42
    context.tablet_id = table_id();
1963
42
    context.enable_segcompaction = false;
1964
    // ATTN: context.tablet is a shared_ptr, can't simply set it's value to `this`. We should
1965
    // get the shared_ptr from tablet_manager.
1966
42
    auto tablet = _engine.tablet_manager()->get_tablet(tablet_id());
1967
42
    if (!tablet) {
1968
0
        LOG(WARNING) << "cant find tablet by tablet_id=" << tablet_id();
1969
0
        return ResultError(Status::NotFound("cant find tablet by tablet_id={}", tablet_id()));
1970
0
    }
1971
42
    context.tablet = tablet;
1972
42
    context.write_type = DataWriteType::TYPE_DIRECT;
1973
42
    context.partial_update_info = std::move(partial_update_info);
1974
42
    context.is_transient_rowset_writer = true;
1975
42
    return create_transient_rowset_writer(context, rowset.rowset_id())
1976
42
            .transform([&](auto&& writer) {
1977
42
                writer->set_segment_start_id(cast_set<int32_t>(rowset.num_segments()));
1978
42
                return writer;
1979
42
            });
1980
42
}
1981
1982
Result<std::unique_ptr<RowsetWriter>> Tablet::create_transient_rowset_writer(
1983
42
        RowsetWriterContext& context, const RowsetId& rowset_id) {
1984
42
    context.rowset_id = rowset_id;
1985
42
    _init_context_common_fields(context);
1986
42
    return RowsetFactory::create_rowset_writer(_engine, context, false);
1987
42
}
1988
1989
958
void Tablet::_init_context_common_fields(RowsetWriterContext& context) {
1990
958
    context.tablet_uid = tablet_uid();
1991
958
    context.tablet_id = tablet_id();
1992
958
    context.partition_id = partition_id();
1993
958
    context.tablet_schema_hash = schema_hash();
1994
958
    context.rowset_type = tablet_meta()->preferred_rowset_type();
1995
    // Alpha Rowset will be removed in the future, so that if the tablet's default rowset type is
1996
    // alpha rowset, then set the newly created rowset to storage engine's default rowset.
1997
958
    if (context.rowset_type == ALPHA_ROWSET) {
1998
0
        context.rowset_type = _engine.default_rowset_type();
1999
0
    }
2000
2001
958
    if (context.is_local_rowset()) {
2002
958
        context.tablet_path = _tablet_path;
2003
958
    }
2004
2005
958
    context.data_dir = data_dir();
2006
958
    context.enable_unique_key_merge_on_write = enable_unique_key_merge_on_write();
2007
2008
958
    context.encrypt_algorithm = tablet_meta()->encryption_algorithm();
2009
958
}
2010
2011
637k
Status Tablet::create_rowset(const RowsetMetaSharedPtr& rowset_meta, RowsetSharedPtr* rowset) {
2012
637k
    return RowsetFactory::create_rowset(_tablet_meta->tablet_schema(),
2013
637k
                                        rowset_meta->is_local() ? _tablet_path : "", rowset_meta,
2014
637k
                                        rowset);
2015
637k
}
2016
2017
7
Status Tablet::cooldown(RowsetSharedPtr rowset) {
2018
7
    std::unique_lock schema_change_lock(_schema_change_lock, std::try_to_lock);
2019
7
    if (!schema_change_lock.owns_lock()) {
2020
0
        return Status::Error<TRY_LOCK_FAILED>(
2021
0
                "try schema_change_lock failed, schema change running or inverted index built "
2022
0
                "on "
2023
0
                "this tablet={}",
2024
0
                tablet_id());
2025
0
    }
2026
    // Check executing serially with compaction task.
2027
7
    std::unique_lock base_compaction_lock(_base_compaction_lock, std::try_to_lock);
2028
7
    if (!base_compaction_lock.owns_lock()) {
2029
0
        return Status::Error<TRY_LOCK_FAILED>("try base_compaction_lock failed");
2030
0
    }
2031
7
    std::unique_lock cumu_compaction_lock(_cumulative_compaction_lock, std::try_to_lock);
2032
7
    if (!cumu_compaction_lock.owns_lock()) {
2033
0
        return Status::Error<TRY_LOCK_FAILED>("try cumu_compaction_lock failed");
2034
0
    }
2035
7
    std::shared_lock cooldown_conf_rlock(_cooldown_conf_lock);
2036
7
    if (_cooldown_conf.cooldown_replica_id <= 0) { // wait for FE to push cooldown conf
2037
2
        return Status::InternalError("invalid cooldown_replica_id");
2038
2
    }
2039
2040
5
    auto mem_tracker = MemTrackerLimiter::create_shared(
2041
5
            MemTrackerLimiter::Type::OTHER,
2042
5
            fmt::format("Tablet::cooldown#tableId={}:replicaId={}", std::to_string(tablet_id()),
2043
5
                        std::to_string(replica_id())));
2044
5
    SCOPED_ATTACH_TASK(mem_tracker);
2045
2046
5
    if (_cooldown_conf.cooldown_replica_id == replica_id()) {
2047
        // this replica is cooldown replica
2048
5
        RETURN_IF_ERROR(_cooldown_data(std::move(rowset)));
2049
5
    } else {
2050
0
        Status st = _follow_cooldowned_data();
2051
0
        if (UNLIKELY(!st.ok())) {
2052
0
            _last_failed_follow_cooldown_time = time(nullptr);
2053
0
            return st;
2054
0
        }
2055
0
        _last_failed_follow_cooldown_time = 0;
2056
0
    }
2057
5
    return Status::OK();
2058
5
}
2059
2060
// hold SHARED `cooldown_conf_lock`
2061
5
Status Tablet::_cooldown_data(RowsetSharedPtr rowset) {
2062
5
    DCHECK(_cooldown_conf.cooldown_replica_id == replica_id());
2063
2064
5
    auto storage_resource = DORIS_TRY(get_resource_by_storage_policy_id(storage_policy_id()));
2065
5
    RowsetSharedPtr old_rowset = nullptr;
2066
2067
5
    if (rowset) {
2068
0
        const auto& rowset_id = rowset->rowset_id();
2069
0
        const auto& rowset_version = rowset->version();
2070
0
        std::shared_lock meta_rlock(_meta_lock);
2071
0
        auto iter = _rs_version_map.find(rowset_version);
2072
0
        if (iter != _rs_version_map.end() && iter->second->rowset_id() == rowset_id) {
2073
0
            old_rowset = rowset;
2074
0
        }
2075
0
    }
2076
2077
5
    if (!old_rowset) {
2078
5
        old_rowset = pick_cooldown_rowset();
2079
5
    }
2080
2081
5
    if (!old_rowset) {
2082
0
        LOG(INFO) << "cannot pick cooldown rowset in tablet " << tablet_id();
2083
0
        return Status::OK();
2084
0
    }
2085
2086
5
    RowsetId new_rowset_id = _engine.next_rowset_id();
2087
5
    auto pending_rs_guard = _engine.pending_remote_rowsets().add(new_rowset_id);
2088
5
    Status st;
2089
5
    Defer defer {[&] {
2090
5
        if (!st.ok()) {
2091
            // reclaim the incomplete rowset data in remote storage
2092
0
            record_unused_remote_rowset(new_rowset_id, storage_resource.fs->id(),
2093
0
                                        old_rowset->num_segments());
2094
0
        }
2095
5
    }};
2096
5
    auto start = std::chrono::steady_clock::now();
2097
5
    if (st = old_rowset->upload_to(storage_resource, new_rowset_id); !st.ok()) {
2098
0
        return st;
2099
0
    }
2100
2101
5
    auto duration = std::chrono::duration<double>(std::chrono::steady_clock::now() - start);
2102
5
    LOG(INFO) << "Upload rowset " << old_rowset->version() << " " << new_rowset_id.to_string()
2103
5
              << " to " << storage_resource.fs->root_path().native()
2104
5
              << ", tablet_id=" << tablet_id() << ", duration=" << duration.count()
2105
5
              << ", capacity=" << old_rowset->total_disk_size()
2106
5
              << ", tp=" << static_cast<double>(old_rowset->total_disk_size()) / duration.count()
2107
5
              << ", old rowset_id=" << old_rowset->rowset_id().to_string();
2108
2109
    // gen a new rowset
2110
5
    auto new_rowset_meta = std::make_shared<RowsetMeta>();
2111
5
    new_rowset_meta->init(old_rowset->rowset_meta().get());
2112
5
    new_rowset_meta->set_rowset_id(new_rowset_id);
2113
5
    new_rowset_meta->set_remote_storage_resource(std::move(storage_resource));
2114
5
    new_rowset_meta->set_creation_time(time(nullptr));
2115
5
    UniqueId cooldown_meta_id = UniqueId::gen_uid();
2116
5
    RowsetSharedPtr new_rowset;
2117
5
    RETURN_IF_ERROR(RowsetFactory::create_rowset(_tablet_meta->tablet_schema(), "", new_rowset_meta,
2118
5
                                                 &new_rowset));
2119
2120
5
    {
2121
5
        std::unique_lock meta_wlock(_meta_lock);
2122
5
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2123
5
        if (tablet_state() == TABLET_RUNNING) {
2124
5
            RETURN_IF_ERROR(delete_rowsets({std::move(old_rowset)}, false));
2125
5
            add_rowsets({std::move(new_rowset)});
2126
            // TODO(plat1ko): process primary key
2127
5
            _tablet_meta->set_cooldown_meta_id(cooldown_meta_id);
2128
5
        }
2129
5
    }
2130
5
    {
2131
5
        std::shared_lock meta_rlock(_meta_lock);
2132
5
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2133
5
        save_meta();
2134
5
    }
2135
    // Upload cooldowned rowset meta to remote fs
2136
    // ATTN: Even if it is an empty rowset, in order for the followers to synchronize, the coolown meta must be
2137
    // uploaded, otherwise followers may never completely cooldown.
2138
5
    if (auto t = _engine.tablet_manager()->get_tablet(tablet_id());
2139
5
        t != nullptr) { // `t` can be nullptr if it has been dropped
2140
5
        async_write_cooldown_meta(std::move(t));
2141
5
    }
2142
5
    return Status::OK();
2143
5
}
2144
2145
// hold SHARED `cooldown_conf_lock`
2146
Status Tablet::_read_cooldown_meta(const StorageResource& storage_resource,
2147
0
                                   TabletMetaPB* tablet_meta_pb) {
2148
0
    std::string remote_meta_path = storage_resource.cooldown_tablet_meta_path(
2149
0
            tablet_id(), _cooldown_conf.cooldown_replica_id, _cooldown_conf.term);
2150
0
    io::FileReaderSPtr tablet_meta_reader;
2151
0
    RETURN_IF_ERROR(storage_resource.fs->open_file(remote_meta_path, &tablet_meta_reader));
2152
0
    auto file_size = tablet_meta_reader->size();
2153
0
    size_t bytes_read;
2154
0
    auto buf = make_unique_buffer<uint8_t>(file_size);
2155
0
    RETURN_IF_ERROR(tablet_meta_reader->read_at(0, {buf.get(), file_size}, &bytes_read));
2156
0
    RETURN_IF_ERROR(tablet_meta_reader->close());
2157
0
    if (!tablet_meta_pb->ParseFromArray(buf.get(), cast_set<int>(file_size))) {
2158
0
        return Status::InternalError("malformed tablet meta, path={}/{}",
2159
0
                                     storage_resource.fs->root_path().native(), remote_meta_path);
2160
0
    }
2161
0
    return Status::OK();
2162
0
}
2163
2164
// `rs_metas` MUST already be sorted by `RowsetMeta::comparator`
2165
9
Status check_version_continuity(const std::vector<RowsetMetaSharedPtr>& rs_metas) {
2166
9
    if (rs_metas.size() < 2) {
2167
2
        return Status::OK();
2168
2
    }
2169
7
    auto prev = rs_metas.begin();
2170
23
    for (auto it = rs_metas.begin() + 1; it != rs_metas.end(); ++it) {
2171
16
        if ((*prev)->end_version() + 1 != (*it)->start_version()) {
2172
0
            return Status::InternalError("versions are not continuity: prev={} cur={}",
2173
0
                                         (*prev)->version().to_string(),
2174
0
                                         (*it)->version().to_string());
2175
0
        }
2176
16
        prev = it;
2177
16
    }
2178
7
    return Status::OK();
2179
7
}
2180
2181
// It's guaranteed the write cooldown meta task would be invoked at the end unless BE crashes
2182
// one tablet would at most have one async task to be done
2183
9
void Tablet::async_write_cooldown_meta(TabletSharedPtr tablet) {
2184
9
    ExecEnv::GetInstance()->write_cooldown_meta_executors()->submit(std::move(tablet));
2185
9
}
2186
2187
332
bool Tablet::update_cooldown_conf(int64_t cooldown_term, int64_t cooldown_replica_id) {
2188
332
    std::unique_lock wlock(_cooldown_conf_lock, std::try_to_lock);
2189
332
    if (!wlock.owns_lock()) {
2190
0
        LOG(INFO) << "try cooldown_conf_lock failed, tablet_id=" << tablet_id();
2191
0
        return false;
2192
0
    }
2193
332
    if (cooldown_term <= _cooldown_conf.term) {
2194
0
        return false;
2195
0
    }
2196
332
    LOG(INFO) << "update cooldown conf. tablet_id=" << tablet_id()
2197
332
              << " cooldown_replica_id: " << _cooldown_conf.cooldown_replica_id << " -> "
2198
332
              << cooldown_replica_id << ", cooldown_term: " << _cooldown_conf.term << " -> "
2199
332
              << cooldown_term;
2200
332
    _cooldown_conf.cooldown_replica_id = cooldown_replica_id;
2201
332
    _cooldown_conf.term = cooldown_term;
2202
332
    return true;
2203
332
}
2204
2205
291k
Status Tablet::write_cooldown_meta() {
2206
291k
    std::shared_lock rlock(_cooldown_conf_lock);
2207
291k
    if (_cooldown_conf.cooldown_replica_id != _tablet_meta->replica_id()) {
2208
0
        return Status::Aborted<false>("not cooldown replica({} vs {}) tablet_id={}",
2209
0
                                      _tablet_meta->replica_id(),
2210
0
                                      _cooldown_conf.cooldown_replica_id, tablet_id());
2211
0
    }
2212
2213
291k
    auto storage_resource = DORIS_TRY(get_resource_by_storage_policy_id(storage_policy_id()));
2214
2215
9
    std::vector<RowsetMetaSharedPtr> cooldowned_rs_metas;
2216
9
    UniqueId cooldown_meta_id;
2217
9
    {
2218
9
        std::shared_lock meta_rlock(_meta_lock);
2219
27
        for (const auto& [_, rs_meta] : _tablet_meta->all_rs_metas()) {
2220
27
            if (!rs_meta->is_local()) {
2221
25
                cooldowned_rs_metas.push_back(rs_meta);
2222
25
            }
2223
27
        }
2224
9
        cooldown_meta_id = _tablet_meta->cooldown_meta_id();
2225
9
    }
2226
9
    if (cooldowned_rs_metas.empty()) {
2227
0
        LOG(INFO) << "no cooldown meta to write, tablet_id=" << tablet_id();
2228
0
        return Status::OK();
2229
0
    }
2230
9
    std::sort(cooldowned_rs_metas.begin(), cooldowned_rs_metas.end(), RowsetMeta::comparator);
2231
9
    DCHECK(cooldowned_rs_metas.front()->start_version() == 0);
2232
    // If version not continuous, it must be a bug
2233
9
    if (auto st = check_version_continuity(cooldowned_rs_metas); !st.ok()) {
2234
0
        DCHECK(st.ok()) << st << " tablet_id=" << tablet_id();
2235
0
        st.set_code(ABORTED);
2236
0
        return st;
2237
0
    }
2238
2239
9
    TabletMetaPB tablet_meta_pb;
2240
9
    auto* rs_metas = tablet_meta_pb.mutable_rs_metas();
2241
9
    rs_metas->Reserve(cast_set<int>(cooldowned_rs_metas.size()));
2242
25
    for (auto& rs_meta : cooldowned_rs_metas) {
2243
25
        rs_metas->Add(rs_meta->get_rowset_pb());
2244
25
    }
2245
9
    tablet_meta_pb.mutable_cooldown_meta_id()->set_hi(cooldown_meta_id.hi);
2246
9
    tablet_meta_pb.mutable_cooldown_meta_id()->set_lo(cooldown_meta_id.lo);
2247
2248
9
    std::string remote_meta_path = storage_resource.cooldown_tablet_meta_path(
2249
9
            tablet_id(), _cooldown_conf.cooldown_replica_id, _cooldown_conf.term);
2250
9
    io::FileWriterPtr tablet_meta_writer;
2251
    // FIXME(plat1ko): What if object store permanently unavailable?
2252
9
    RETURN_IF_ERROR(storage_resource.fs->create_file(remote_meta_path, &tablet_meta_writer));
2253
9
    auto val = tablet_meta_pb.SerializeAsString();
2254
9
    RETURN_IF_ERROR(tablet_meta_writer->append({val.data(), val.size()}));
2255
9
    return tablet_meta_writer->close();
2256
9
}
2257
2258
// hold SHARED `cooldown_conf_lock`
2259
0
Status Tablet::_follow_cooldowned_data() {
2260
0
    DCHECK(_cooldown_conf.cooldown_replica_id != replica_id());
2261
0
    LOG(INFO) << "try to follow cooldowned data. tablet_id=" << tablet_id()
2262
0
              << " cooldown_replica_id=" << _cooldown_conf.cooldown_replica_id
2263
0
              << " local replica=" << replica_id();
2264
2265
0
    auto storage_resource = DORIS_TRY(get_resource_by_storage_policy_id(storage_policy_id()));
2266
    // MUST executing serially with cold data compaction, because compaction input rowsets may be deleted by this function
2267
0
    std::unique_lock cold_compaction_lock(_cold_compaction_lock, std::try_to_lock);
2268
0
    if (!cold_compaction_lock.owns_lock()) {
2269
0
        return Status::Error<TRY_LOCK_FAILED>("try cold_compaction_lock failed");
2270
0
    }
2271
2272
0
    TabletMetaPB cooldown_meta_pb;
2273
0
    auto st = _read_cooldown_meta(storage_resource, &cooldown_meta_pb);
2274
0
    if (!st.ok()) {
2275
0
        LOG(INFO) << "cannot read cooldown meta: " << st;
2276
0
        return Status::InternalError<false>("cannot read cooldown meta");
2277
0
    }
2278
2279
0
    if (cooldown_meta_pb.rs_metas_size() <= 0) {
2280
0
        LOG(WARNING)
2281
0
                << "Cooldown meta file exists but rs_metas is empty for tablet " << tablet_id()
2282
0
                << ". Cooldown meta id: " << cooldown_meta_pb.cooldown_meta_id()
2283
0
                << ". This may indicate a cooldown meta synchronization issue or an invalid file.";
2284
0
        return Status::InternalError<false>("Cooldown meta rs_metas is empty");
2285
0
    }
2286
2287
0
    if (_tablet_meta->cooldown_meta_id() == cooldown_meta_pb.cooldown_meta_id()) {
2288
        // cooldowned rowsets are same, no need to follow
2289
0
        return Status::OK();
2290
0
    }
2291
2292
0
    int64_t cooldowned_version = cooldown_meta_pb.rs_metas().rbegin()->end_version();
2293
2294
0
    std::vector<RowsetSharedPtr> overlap_rowsets;
2295
0
    bool version_aligned = false;
2296
2297
    // Holding these to delete rowsets' shared ptr until save meta can avoid trash sweeping thread
2298
    // deleting these rowsets' files before rowset meta has been removed from disk, which may cause
2299
    // data loss when BE reboot before save meta to disk.
2300
0
    std::vector<RowsetSharedPtr> to_delete;
2301
0
    std::vector<RowsetSharedPtr> to_add;
2302
2303
0
    {
2304
0
        std::lock_guard wlock(_meta_lock);
2305
0
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2306
0
        if (tablet_state() != TABLET_RUNNING) {
2307
0
            return Status::InternalError<false>("tablet not running");
2308
0
        }
2309
2310
0
        for (auto& [v, rs] : _rs_version_map) {
2311
0
            if (v.second <= cooldowned_version) {
2312
0
                overlap_rowsets.push_back(rs);
2313
0
                if (!version_aligned && v.second == cooldowned_version) {
2314
0
                    version_aligned = true;
2315
0
                }
2316
0
            } else if (!rs->is_local()) {
2317
0
                return Status::InternalError<false>(
2318
0
                        "cooldowned version larger than that to follow with cooldown version "
2319
0
                        "{}",
2320
0
                        cooldowned_version);
2321
0
            }
2322
0
        }
2323
2324
0
        if (!version_aligned) {
2325
0
            return Status::InternalError<false>("cooldowned version is not aligned with version {}",
2326
0
                                                cooldowned_version);
2327
0
        }
2328
2329
0
        std::sort(overlap_rowsets.begin(), overlap_rowsets.end(), Rowset::comparator);
2330
2331
        // Find different rowset in `overlap_rowsets` and `cooldown_meta_pb.rs_metas`
2332
0
        auto rs_pb_it = cooldown_meta_pb.rs_metas().begin();
2333
0
        auto rs_it = overlap_rowsets.begin();
2334
0
        for (; rs_pb_it != cooldown_meta_pb.rs_metas().end() && rs_it != overlap_rowsets.end();
2335
0
             ++rs_pb_it, ++rs_it) {
2336
0
            if (rs_pb_it->rowset_id_v2() != (*rs_it)->rowset_id().to_string()) {
2337
0
                break;
2338
0
            }
2339
0
        }
2340
2341
0
        to_delete.assign(rs_it, overlap_rowsets.end());
2342
0
        to_add.reserve(cooldown_meta_pb.rs_metas().end() - rs_pb_it);
2343
0
        for (; rs_pb_it != cooldown_meta_pb.rs_metas().end(); ++rs_pb_it) {
2344
0
            auto rs_meta = std::make_shared<RowsetMeta>();
2345
0
            rs_meta->init_from_pb(*rs_pb_it);
2346
0
            RowsetSharedPtr rs;
2347
0
            RETURN_IF_ERROR(
2348
0
                    RowsetFactory::create_rowset(_tablet_meta->tablet_schema(), "", rs_meta, &rs));
2349
0
            to_add.push_back(std::move(rs));
2350
0
        }
2351
        // Note: We CANNOT call `modify_rowsets` here because `modify_rowsets` cannot process version graph correctly.
2352
0
        RETURN_IF_ERROR(delete_rowsets(to_delete, false));
2353
0
        add_rowsets(to_add);
2354
        // TODO(plat1ko): process primary key
2355
0
        _tablet_meta->set_cooldown_meta_id(cooldown_meta_pb.cooldown_meta_id());
2356
0
    }
2357
2358
0
    {
2359
0
        std::lock_guard rlock(_meta_lock);
2360
0
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2361
0
        save_meta();
2362
0
    }
2363
2364
0
    if (!to_add.empty()) {
2365
0
        LOG(INFO) << "modify rowsets when follow cooldowned data, tablet_id=" << tablet_id()
2366
0
                  << [&]() {
2367
0
                         std::stringstream ss;
2368
0
                         ss << " delete rowsets:\n";
2369
0
                         for (auto&& rs : to_delete) {
2370
0
                             ss << rs->version() << ' ' << rs->rowset_id() << '\n';
2371
0
                         }
2372
0
                         ss << "add rowsets:\n";
2373
0
                         for (auto&& rs : to_add) {
2374
0
                             ss << rs->version() << ' ' << rs->rowset_id() << '\n';
2375
0
                         }
2376
0
                         return ss.str();
2377
0
                     }();
2378
0
    }
2379
2380
0
    return Status::OK();
2381
0
}
2382
2383
3.96M
bool Tablet::_has_data_to_cooldown() {
2384
3.96M
    int64_t min_local_version = std::numeric_limits<int64_t>::max();
2385
3.96M
    RowsetSharedPtr rowset;
2386
3.96M
    std::shared_lock meta_rlock(_meta_lock);
2387
    // Ususally once the tablet has done cooldown successfully then the first
2388
    // rowset would always be remote rowset
2389
3.96M
    bool has_cooldowned = false;
2390
7.10M
    for (const auto& [_, rs] : _rs_version_map) {
2391
7.10M
        if (!rs->is_local()) {
2392
59
            has_cooldowned = true;
2393
59
            break;
2394
59
        }
2395
7.10M
    }
2396
7.10M
    for (auto& [v, rs] : _rs_version_map) {
2397
7.10M
        auto predicate = rs->is_local() && v.first < min_local_version;
2398
7.10M
        if (!has_cooldowned) {
2399
7.10M
            predicate = predicate && (rs->data_disk_size() > 0);
2400
7.10M
        }
2401
7.10M
        if (predicate) {
2402
            // this is a local rowset and has data
2403
700k
            min_local_version = v.first;
2404
700k
            rowset = rs;
2405
700k
        }
2406
7.10M
    }
2407
2408
3.96M
    int64_t newest_cooldown_time = 0;
2409
3.96M
    if (rowset != nullptr) {
2410
636k
        newest_cooldown_time = _get_newest_cooldown_time(rowset);
2411
636k
    }
2412
2413
3.96M
    return (newest_cooldown_time != 0) && (newest_cooldown_time < UnixSeconds());
2414
3.96M
}
2415
2416
3.96M
RowsetSharedPtr Tablet::pick_cooldown_rowset() {
2417
3.96M
    RowsetSharedPtr rowset;
2418
2419
3.96M
    if (!_has_data_to_cooldown()) {
2420
3.96M
        return nullptr;
2421
3.96M
    }
2422
2423
    // TODO(plat1ko): should we maintain `cooldowned_version` in `Tablet`?
2424
8
    int64_t cooldowned_version = -1;
2425
    // We pick the rowset with smallest start version in local.
2426
8
    int64_t min_local_version = std::numeric_limits<int64_t>::max();
2427
8
    {
2428
8
        std::shared_lock meta_rlock(_meta_lock);
2429
26
        for (auto& [v, rs] : _rs_version_map) {
2430
26
            if (!rs->is_local()) {
2431
4
                cooldowned_version = std::max(cooldowned_version, v.second);
2432
22
            } else if (v.first < min_local_version) { // this is a local rowset
2433
10
                min_local_version = v.first;
2434
10
                rowset = rs;
2435
10
            }
2436
26
        }
2437
8
    }
2438
8
    if (!rowset) {
2439
0
        return nullptr;
2440
0
    }
2441
8
    if (tablet_footprint() == 0) {
2442
0
        VLOG_DEBUG << "skip cooldown due to empty tablet_id = " << tablet_id();
2443
0
        return nullptr;
2444
0
    }
2445
8
    if (min_local_version != cooldowned_version + 1) { // ensure version continuity
2446
0
        if (UNLIKELY(cooldowned_version != -1)) {
2447
0
            LOG(WARNING) << "version not continuous. tablet_id=" << tablet_id()
2448
0
                         << " cooldowned_version=" << cooldowned_version
2449
0
                         << " min_local_version=" << min_local_version;
2450
0
        }
2451
0
        return nullptr;
2452
0
    }
2453
8
    return rowset;
2454
8
}
2455
2456
636k
int64_t Tablet::_get_newest_cooldown_time(const RowsetSharedPtr& rowset) {
2457
636k
    int64_t id = storage_policy_id();
2458
636k
    if (id <= 0) {
2459
636k
        VLOG_DEBUG << "tablet does not need cooldown, tablet id: " << tablet_id();
2460
636k
        return 0;
2461
636k
    }
2462
12
    auto storage_policy = get_storage_policy(id);
2463
12
    if (!storage_policy) {
2464
0
        LOG(WARNING) << "Cannot get storage policy: " << id;
2465
0
        return 0;
2466
0
    }
2467
12
    auto cooldown_ttl_sec = storage_policy->cooldown_ttl;
2468
12
    auto cooldown_datetime = storage_policy->cooldown_datetime;
2469
12
    int64_t newest_cooldown_time = std::numeric_limits<int64_t>::max();
2470
2471
12
    if (cooldown_ttl_sec >= 0) {
2472
9
        newest_cooldown_time = rowset->newest_write_timestamp() + cooldown_ttl_sec;
2473
9
    }
2474
12
    if (cooldown_datetime > 0) {
2475
10
        newest_cooldown_time = std::min(newest_cooldown_time, cooldown_datetime);
2476
10
    }
2477
2478
12
    return newest_cooldown_time;
2479
12
}
2480
2481
3.96M
RowsetSharedPtr Tablet::need_cooldown(int64_t* cooldown_timestamp, size_t* file_size) {
2482
3.96M
    RowsetSharedPtr rowset = pick_cooldown_rowset();
2483
3.96M
    if (!rowset) {
2484
3.96M
        VLOG_DEBUG << "pick cooldown rowset, get null, tablet id: " << tablet_id();
2485
3.96M
        return nullptr;
2486
3.96M
    }
2487
2488
3
    auto newest_cooldown_time = _get_newest_cooldown_time(rowset);
2489
2490
    // the rowset should do cooldown job only if it's cooldown ttl plus newest write time is less than
2491
    // current time or it's datatime is less than current time
2492
3
    if (newest_cooldown_time != 0 && newest_cooldown_time < UnixSeconds()) {
2493
3
        *cooldown_timestamp = newest_cooldown_time;
2494
3
        *file_size = rowset->total_disk_size();
2495
3
        VLOG_DEBUG << "tablet need cooldown, tablet id: " << tablet_id()
2496
0
                   << " file_size: " << *file_size;
2497
3
        return rowset;
2498
3
    }
2499
2500
0
    VLOG_DEBUG << "tablet does not need cooldown, tablet id: " << tablet_id()
2501
0
               << " newest write time: " << rowset->newest_write_timestamp();
2502
0
    return nullptr;
2503
3
}
2504
2505
void Tablet::record_unused_remote_rowset(const RowsetId& rowset_id, const std::string& resource,
2506
0
                                         int64_t num_segments) {
2507
0
    auto gc_key = REMOTE_ROWSET_GC_PREFIX + rowset_id.to_string();
2508
0
    RemoteRowsetGcPB gc_pb;
2509
0
    gc_pb.set_resource_id(resource);
2510
0
    gc_pb.set_tablet_id(tablet_id());
2511
0
    gc_pb.set_num_segments(num_segments);
2512
0
    auto st =
2513
0
            _data_dir->get_meta()->put(META_COLUMN_FAMILY_INDEX, gc_key, gc_pb.SerializeAsString());
2514
0
    if (!st.ok()) {
2515
0
        LOG(WARNING) << "failed to record unused remote rowset. tablet_id=" << tablet_id()
2516
0
                     << " rowset_id=" << rowset_id << " resource_id=" << resource;
2517
0
    }
2518
0
    unused_remote_rowset_num << 1;
2519
0
}
2520
2521
14.1k
Status Tablet::remove_all_remote_rowsets() {
2522
14.1k
    DCHECK(tablet_state() == TABLET_SHUTDOWN);
2523
14.1k
    std::set<std::string> resource_ids;
2524
31.4k
    for (const auto& [_, rs_meta] : _tablet_meta->all_rs_metas()) {
2525
31.4k
        if (!rs_meta->is_local()) {
2526
0
            resource_ids.insert(rs_meta->resource_id());
2527
0
        }
2528
31.4k
    }
2529
14.1k
    if (resource_ids.empty()) {
2530
14.1k
        return Status::OK();
2531
14.1k
    }
2532
0
    auto tablet_gc_key = REMOTE_TABLET_GC_PREFIX + std::to_string(tablet_id());
2533
0
    RemoteTabletGcPB gc_pb;
2534
0
    for (auto& resource_id : resource_ids) {
2535
0
        gc_pb.add_resource_ids(resource_id);
2536
0
    }
2537
0
    return _data_dir->get_meta()->put(META_COLUMN_FAMILY_INDEX, tablet_gc_key,
2538
0
                                      gc_pb.SerializeAsString());
2539
14.1k
}
2540
2541
0
void Tablet::update_max_version_schema(const TabletSchemaSPtr& tablet_schema) {
2542
0
    std::lock_guard wrlock(_meta_lock);
2543
0
    SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2544
    // Double Check for concurrent update
2545
0
    if (!_max_version_schema ||
2546
0
        tablet_schema->schema_version() > _max_version_schema->schema_version()) {
2547
0
        _max_version_schema = tablet_schema;
2548
0
    }
2549
0
}
2550
2551
0
CalcDeleteBitmapExecutor* Tablet::calc_delete_bitmap_executor() {
2552
0
    return _engine.calc_delete_bitmap_executor();
2553
0
}
2554
2555
Status Tablet::save_delete_bitmap(const TabletTxnInfo* txn_info, int64_t txn_id,
2556
                                  DeleteBitmapPtr delete_bitmap, RowsetWriter* rowset_writer,
2557
                                  const RowsetIdUnorderedSet& cur_rowset_ids, int64_t lock_id,
2558
444
                                  int64_t next_visible_version) {
2559
444
    RowsetSharedPtr rowset = txn_info->rowset;
2560
444
    int64_t cur_version = rowset->start_version();
2561
2562
    // update version without write lock, compaction and publish_txn
2563
    // will update delete bitmap, handle compaction with _rowset_update_lock
2564
    // and publish_txn runs sequential so no need to lock here
2565
1.39k
    for (auto& [key, bitmap] : delete_bitmap->delete_bitmap) {
2566
        // skip sentinel mark, which is used for delete bitmap correctness check
2567
1.39k
        if (std::get<1>(key) != DeleteBitmap::INVALID_SEGMENT_ID) {
2568
46
            _tablet_meta->delete_bitmap().merge({std::get<0>(key), std::get<1>(key), cur_version},
2569
46
                                                bitmap);
2570
46
        }
2571
1.39k
    }
2572
2573
444
    return Status::OK();
2574
444
}
2575
2576
47
void Tablet::merge_delete_bitmap(const DeleteBitmap& delete_bitmap) {
2577
47
    _tablet_meta->delete_bitmap().merge(delete_bitmap);
2578
47
}
2579
2580
0
bool Tablet::check_all_rowset_segment() {
2581
0
    std::shared_lock rdlock(_meta_lock);
2582
0
    for (auto& version_rowset : _rs_version_map) {
2583
0
        RowsetSharedPtr rowset = version_rowset.second;
2584
0
        if (!rowset->check_rowset_segment()) {
2585
0
            LOG(WARNING) << "Tablet Segment Check. find a bad tablet, tablet_id=" << tablet_id();
2586
0
            return false;
2587
0
        }
2588
0
    }
2589
0
    return true;
2590
0
}
2591
2592
13.4M
void Tablet::set_skip_compaction(bool skip, CompactionType compaction_type, int64_t start) {
2593
13.4M
    if (!skip) {
2594
0
        _skip_cumu_compaction = false;
2595
0
        _skip_base_compaction = false;
2596
0
        return;
2597
0
    }
2598
13.4M
    if (compaction_type == CompactionType::CUMULATIVE_COMPACTION) {
2599
7.62M
        _skip_cumu_compaction = true;
2600
7.62M
        _skip_cumu_compaction_ts = start;
2601
7.62M
    } else {
2602
5.87M
        DCHECK(compaction_type == CompactionType::BASE_COMPACTION);
2603
5.87M
        _skip_base_compaction = true;
2604
5.87M
        _skip_base_compaction_ts = start;
2605
5.87M
    }
2606
13.4M
}
2607
2608
315M
bool Tablet::should_skip_compaction(CompactionType compaction_type, int64_t now) {
2609
315M
    if (compaction_type == CompactionType::CUMULATIVE_COMPACTION && _skip_cumu_compaction &&
2610
315M
        now < _skip_cumu_compaction_ts + config::skip_tablet_compaction_second) {
2611
270M
        return true;
2612
270M
    } else if (compaction_type == CompactionType::BASE_COMPACTION && _skip_base_compaction &&
2613
44.8M
               now < _skip_base_compaction_ts + config::skip_tablet_compaction_second) {
2614
22.3M
        return true;
2615
22.3M
    }
2616
22.5M
    return false;
2617
315M
}
2618
2619
0
std::pair<std::string, int64_t> Tablet::get_binlog_info(std::string_view binlog_version) const {
2620
0
    return RowsetMetaManager::get_binlog_info(_data_dir->get_meta(), tablet_uid(), binlog_version);
2621
0
}
2622
2623
std::string Tablet::get_rowset_binlog_meta(std::string_view binlog_version,
2624
0
                                           std::string_view rowset_id) const {
2625
0
    return RowsetMetaManager::get_rowset_binlog_meta(_data_dir->get_meta(), tablet_uid(),
2626
0
                                                     binlog_version, rowset_id);
2627
0
}
2628
2629
Status Tablet::get_rowset_binlog_metas(const std::vector<int64_t>& binlog_versions,
2630
0
                                       RowsetBinlogMetasPB* metas_pb) {
2631
0
    return RowsetMetaManager::get_rowset_binlog_metas(_data_dir->get_meta(), tablet_uid(),
2632
0
                                                      binlog_versions, metas_pb);
2633
0
}
2634
2635
4
Status Tablet::get_rowset_binlog_metas(Version binlog_versions, RowsetBinlogMetasPB* metas_pb) {
2636
4
    return RowsetMetaManager::get_rowset_binlog_metas(_data_dir->get_meta(), tablet_uid(),
2637
4
                                                      binlog_versions, metas_pb);
2638
4
}
2639
2640
std::string Tablet::get_segment_filepath(std::string_view rowset_id,
2641
0
                                         std::string_view segment_index) const {
2642
0
    return fmt::format("{}/_binlog/{}_{}.dat", _tablet_path, rowset_id, segment_index);
2643
0
}
2644
2645
1
std::string Tablet::get_segment_filepath(std::string_view rowset_id, int64_t segment_index) const {
2646
1
    return fmt::format("{}/_binlog/{}_{}.dat", _tablet_path, rowset_id, segment_index);
2647
1
}
2648
2649
0
std::vector<std::string> Tablet::get_binlog_filepath(std::string_view binlog_version) const {
2650
0
    const auto& [rowset_id, num_segments] = get_binlog_info(binlog_version);
2651
0
    std::vector<std::string> binlog_filepath;
2652
0
    for (int i = 0; i < num_segments; ++i) {
2653
        // TODO(Drogon): rewrite by filesystem path
2654
0
        auto segment_file = fmt::format("{}_{}.dat", rowset_id, i);
2655
0
        binlog_filepath.emplace_back(fmt::format("{}/_binlog/{}", _tablet_path, segment_file));
2656
0
    }
2657
0
    return binlog_filepath;
2658
0
}
2659
2660
0
bool Tablet::can_add_binlog(uint64_t total_binlog_size) const {
2661
0
    return !_data_dir->reach_capacity_limit(total_binlog_size);
2662
0
}
2663
2664
467
bool Tablet::is_enable_binlog() {
2665
467
    return config::enable_feature_binlog && tablet_meta()->binlog_config().is_enable();
2666
467
}
2667
2668
0
void Tablet::set_binlog_config(BinlogConfig binlog_config) {
2669
0
    tablet_meta()->set_binlog_config(binlog_config);
2670
0
}
2671
2672
2
void Tablet::gc_binlogs(int64_t version) {
2673
2
    auto meta = _data_dir->get_meta();
2674
2
    DCHECK(meta != nullptr);
2675
2676
2
    const auto& tablet_uid = this->tablet_uid();
2677
2
    const auto tablet_id = this->tablet_id();
2678
2
    std::string begin_key = make_binlog_meta_key_prefix(tablet_uid);
2679
2
    std::string end_key = make_binlog_meta_key_prefix(tablet_uid, version + 1);
2680
2
    LOG(INFO) << fmt::format("gc binlog meta, tablet_id:{}, begin_key:{}, end_key:{}", tablet_id,
2681
2
                             begin_key, end_key);
2682
2683
2
    std::vector<std::string> wait_for_deleted_binlog_keys;
2684
2
    std::vector<std::string> wait_for_deleted_binlog_files;
2685
2
    auto add_to_wait_for_deleted = [&](std::string_view key, std::string_view rowset_id,
2686
2
                                       int64_t num_segments) {
2687
        // add binlog meta key and binlog data key
2688
1
        wait_for_deleted_binlog_keys.emplace_back(key);
2689
1
        wait_for_deleted_binlog_keys.push_back(get_binlog_data_key_from_meta_key(key));
2690
2691
        // add binlog segment files and index files
2692
2
        for (int64_t i = 0; i < num_segments; ++i) {
2693
1
            auto segment_file_path = get_segment_filepath(rowset_id, i);
2694
1
            wait_for_deleted_binlog_files.emplace_back(segment_file_path);
2695
2696
            // index files
2697
1
            if (tablet_schema()->has_inverted_index()) {
2698
1
                if (tablet_schema()->get_inverted_index_storage_format() ==
2699
1
                    InvertedIndexStorageFormatPB::V1) {
2700
0
                    for (const auto& index : tablet_schema()->inverted_indexes()) {
2701
0
                        auto index_file = InvertedIndexDescriptor::get_index_file_path_v1(
2702
0
                                InvertedIndexDescriptor::get_index_file_path_prefix(
2703
0
                                        segment_file_path),
2704
0
                                index->index_id(), index->get_index_suffix());
2705
0
                        wait_for_deleted_binlog_files.emplace_back(index_file);
2706
0
                    }
2707
1
                } else {
2708
1
                    auto index_file = InvertedIndexDescriptor::get_index_file_path_v2(
2709
1
                            InvertedIndexDescriptor::get_index_file_path_prefix(segment_file_path));
2710
1
                    wait_for_deleted_binlog_files.emplace_back(index_file);
2711
1
                }
2712
1
            }
2713
1
        }
2714
1
    };
2715
2716
2
    auto check_binlog_ttl = [&](std::string_view key, std::string_view value) mutable -> bool {
2717
2
        if (key >= end_key) {
2718
1
            return false;
2719
1
        }
2720
2721
1
        BinlogMetaEntryPB binlog_meta_entry_pb;
2722
1
        if (!binlog_meta_entry_pb.ParseFromArray(value.data(), cast_set<int>(value.size()))) {
2723
0
            LOG(WARNING) << "failed to parse binlog meta entry, key:" << key;
2724
0
            return true;
2725
0
        }
2726
2727
1
        auto num_segments = binlog_meta_entry_pb.num_segments();
2728
1
        std::string rowset_id;
2729
1
        if (binlog_meta_entry_pb.has_rowset_id_v2()) {
2730
1
            rowset_id = binlog_meta_entry_pb.rowset_id_v2();
2731
1
        } else {
2732
            // key is 'binlog_meta_6943f1585fe834b5-e542c2b83a21d0b7_00000000000000000069_020000000000000135449d7cd7eadfe672aa0f928fa99593', extract last part '020000000000000135449d7cd7eadfe672aa0f928fa99593'
2733
0
            auto pos = key.rfind('_');
2734
0
            if (pos == std::string::npos) {
2735
0
                LOG(WARNING) << fmt::format("invalid binlog meta key:{}", key);
2736
0
                return false;
2737
0
            }
2738
0
            rowset_id = key.substr(pos + 1);
2739
0
        }
2740
1
        add_to_wait_for_deleted(key, rowset_id, num_segments);
2741
2742
1
        return true;
2743
1
    };
2744
2745
2
    auto status = meta->iterate(META_COLUMN_FAMILY_INDEX, begin_key, check_binlog_ttl);
2746
2
    if (!status.ok()) {
2747
0
        LOG(WARNING) << "failed to iterate binlog meta, status:" << status;
2748
0
        return;
2749
0
    }
2750
2751
    // first remove binlog files, if failed, just break, then retry next time
2752
    // this keep binlog meta in meta store, so that binlog can be removed next time
2753
2
    bool remove_binlog_files_failed = false;
2754
2
    for (auto& file : wait_for_deleted_binlog_files) {
2755
2
        if (unlink(file.c_str()) != 0) {
2756
            // file not exist, continue
2757
0
            if (errno == ENOENT) {
2758
0
                continue;
2759
0
            }
2760
2761
0
            remove_binlog_files_failed = true;
2762
0
            LOG(WARNING) << "failed to remove binlog file:" << file << ", errno:" << errno;
2763
0
            break;
2764
0
        }
2765
2
    }
2766
2
    if (!remove_binlog_files_failed) {
2767
2
        static_cast<void>(meta->remove(META_COLUMN_FAMILY_INDEX, wait_for_deleted_binlog_keys));
2768
2
    }
2769
2
}
2770
2771
0
Status Tablet::ingest_binlog_metas(RowsetBinlogMetasPB* metas_pb) {
2772
0
    return RowsetMetaManager::ingest_binlog_metas(_data_dir->get_meta(), tablet_uid(), metas_pb);
2773
0
}
2774
2775
26.2k
void Tablet::clear_cache() {
2776
26.2k
    std::vector<RowsetSharedPtr> rowsets;
2777
26.2k
    {
2778
26.2k
        std::shared_lock rlock(get_header_lock());
2779
26.2k
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2780
2781
76.9k
        for (auto& [_, rowset] : rowset_map()) {
2782
76.9k
            rowsets.push_back(rowset);
2783
76.9k
        }
2784
26.2k
        for (auto& [_, rowset] : stale_rowset_map()) {
2785
6.91k
            rowsets.push_back(rowset);
2786
6.91k
        }
2787
26.2k
    }
2788
83.8k
    for (auto& rowset : rowsets) {
2789
83.8k
        rowset->clear_cache();
2790
83.8k
    }
2791
26.2k
}
2792
2793
35.2k
void Tablet::check_table_size_correctness() {
2794
35.2k
    if (!config::enable_table_size_correctness_check) {
2795
35.2k
        return;
2796
35.2k
    }
2797
18.4E
    for (const auto& [_, rs_meta] : _tablet_meta->all_rs_metas()) {
2798
0
        int64_t total_segment_size = get_segment_file_size(rs_meta);
2799
0
        int64_t total_inverted_index_size = get_inverted_index_file_size(rs_meta);
2800
0
        if (rs_meta->data_disk_size() != total_segment_size ||
2801
0
            rs_meta->index_disk_size() != total_inverted_index_size ||
2802
0
            rs_meta->data_disk_size() + rs_meta->index_disk_size() != rs_meta->total_disk_size()) {
2803
0
            LOG(WARNING) << "[Local table table size check failed]:"
2804
0
                         << " tablet id: " << rs_meta->tablet_id()
2805
0
                         << ", rowset id:" << rs_meta->rowset_id()
2806
0
                         << ", rowset data disk size:" << rs_meta->data_disk_size()
2807
0
                         << ", rowset real data disk size:" << total_segment_size
2808
0
                         << ", rowset index disk size:" << rs_meta->index_disk_size()
2809
0
                         << ", rowset real index disk size:" << total_inverted_index_size
2810
0
                         << ", rowset total disk size:" << rs_meta->total_disk_size()
2811
0
                         << ", rowset segment path:"
2812
0
                         << StorageResource().remote_segment_path(
2813
0
                                    rs_meta->tablet_id(), rs_meta->rowset_id().to_string(), 0);
2814
0
            DCHECK(false);
2815
0
        }
2816
0
    }
2817
18.4E
}
2818
2819
0
std::string Tablet::get_segment_path(const RowsetMetaSharedPtr& rs_meta, int64_t seg_id) {
2820
0
    std::string segment_path;
2821
0
    if (rs_meta->is_local()) {
2822
0
        segment_path = local_segment_path(_tablet_path, rs_meta->rowset_id().to_string(), seg_id);
2823
0
    } else {
2824
0
        segment_path = rs_meta->remote_storage_resource().value()->remote_segment_path(
2825
0
                rs_meta->tablet_id(), rs_meta->rowset_id().to_string(), seg_id);
2826
0
    }
2827
0
    return segment_path;
2828
0
}
2829
2830
0
int64_t Tablet::get_segment_file_size(const RowsetMetaSharedPtr& rs_meta) {
2831
0
    const auto& fs = rs_meta->fs();
2832
0
    if (!fs) {
2833
0
        LOG(WARNING) << "get fs failed, resource_id={}" << rs_meta->resource_id();
2834
0
    }
2835
0
    int64_t total_segment_size = 0;
2836
0
    for (int64_t seg_id = 0; seg_id < rs_meta->num_segments(); seg_id++) {
2837
0
        std::string segment_path = get_segment_path(rs_meta, seg_id);
2838
0
        int64_t segment_file_size = 0;
2839
0
        auto st = fs->file_size(segment_path, &segment_file_size);
2840
0
        if (!st.ok()) {
2841
0
            segment_file_size = 0;
2842
0
            LOG(WARNING) << "table size correctness check get segment size failed! msg:"
2843
0
                         << st.to_string() << ", segment path:" << segment_path;
2844
0
        }
2845
0
        total_segment_size += segment_file_size;
2846
0
    }
2847
0
    return total_segment_size;
2848
0
}
2849
2850
0
int64_t Tablet::get_inverted_index_file_size(const RowsetMetaSharedPtr& rs_meta) {
2851
0
    const auto& fs = rs_meta->fs();
2852
0
    if (!fs) {
2853
0
        LOG(WARNING) << "get fs failed, resource_id={}" << rs_meta->resource_id();
2854
0
    }
2855
0
    int64_t total_inverted_index_size = 0;
2856
2857
0
    if (rs_meta->tablet_schema()->get_inverted_index_storage_format() ==
2858
0
        InvertedIndexStorageFormatPB::V1) {
2859
0
        const auto& indices = rs_meta->tablet_schema()->inverted_indexes();
2860
0
        for (auto& index : indices) {
2861
0
            for (int seg_id = 0; seg_id < rs_meta->num_segments(); ++seg_id) {
2862
0
                std::string segment_path = get_segment_path(rs_meta, seg_id);
2863
0
                int64_t file_size = 0;
2864
2865
0
                std::string inverted_index_file_path =
2866
0
                        InvertedIndexDescriptor::get_index_file_path_v1(
2867
0
                                InvertedIndexDescriptor::get_index_file_path_prefix(segment_path),
2868
0
                                index->index_id(), index->get_index_suffix());
2869
0
                auto st = fs->file_size(inverted_index_file_path, &file_size);
2870
0
                if (!st.ok()) {
2871
0
                    file_size = 0;
2872
0
                    LOG(WARNING) << " tablet id: " << get_tablet_info().tablet_id
2873
0
                                 << ", rowset id:" << rs_meta->rowset_id()
2874
0
                                 << ", table size correctness check get inverted index v1 "
2875
0
                                    "size failed! msg:"
2876
0
                                 << st.to_string()
2877
0
                                 << ", inverted index path:" << inverted_index_file_path;
2878
0
                }
2879
0
                total_inverted_index_size += file_size;
2880
0
            }
2881
0
        }
2882
0
    } else {
2883
0
        for (int seg_id = 0; seg_id < rs_meta->num_segments(); ++seg_id) {
2884
0
            int64_t file_size = 0;
2885
0
            std::string segment_path = get_segment_path(rs_meta, seg_id);
2886
0
            std::string inverted_index_file_path = InvertedIndexDescriptor::get_index_file_path_v2(
2887
0
                    InvertedIndexDescriptor::get_index_file_path_prefix(segment_path));
2888
0
            auto st = fs->file_size(inverted_index_file_path, &file_size);
2889
0
            if (!st.ok()) {
2890
0
                file_size = 0;
2891
0
                if (st.is<NOT_FOUND>()) {
2892
0
                    LOG(INFO) << " tablet id: " << get_tablet_info().tablet_id
2893
0
                              << ", rowset id:" << rs_meta->rowset_id()
2894
0
                              << ", table size correctness check get inverted index v2 failed "
2895
0
                                 "because file not exist:"
2896
0
                              << inverted_index_file_path;
2897
0
                } else {
2898
0
                    LOG(WARNING) << " tablet id: " << get_tablet_info().tablet_id
2899
0
                                 << ", rowset id:" << rs_meta->rowset_id()
2900
0
                                 << ", table size correctness check get inverted index v2 "
2901
0
                                    "size failed! msg:"
2902
0
                                 << st.to_string()
2903
0
                                 << ", inverted index path:" << inverted_index_file_path;
2904
0
                }
2905
0
            }
2906
0
            total_inverted_index_size += file_size;
2907
0
        }
2908
0
    }
2909
0
    return total_inverted_index_size;
2910
0
}
2911
2912
Status Tablet::prepare_txn(TPartitionId partition_id, TTransactionId transaction_id,
2913
481
                           const PUniqueId& load_id, bool ingest) {
2914
481
    std::shared_lock base_migration_lock(get_migration_lock(), std::defer_lock);
2915
2916
    // TODO: rename debugpoint.
2917
481
    DBUG_EXECUTE_IF("PushHandler::_do_streaming_ingestion.try_lock_fail", {
2918
481
        return Status::ObtainLockFailed("PushHandler::_do_streaming_ingestion get lock failed");
2919
481
    })
2920
2921
481
    if (!base_migration_lock.try_lock_for(
2922
481
                std::chrono::milliseconds(config::migration_lock_timeout_ms))) {
2923
0
        return Status::ObtainLockFailed("try_lock migration lock failed after {}ms",
2924
0
                                        config::migration_lock_timeout_ms);
2925
0
    }
2926
2927
481
    std::lock_guard<std::mutex> push_lock(get_push_lock());
2928
481
    return _engine.txn_manager()->prepare_txn(partition_id, transaction_id, tablet_id(),
2929
481
                                              tablet_uid(), load_id, ingest);
2930
481
}
2931
2932
#include "common/compile_check_end.h"
2933
} // namespace doris