Coverage Report

Created: 2026-01-22 12:20

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/string_ref.h"
126
#include "vec/data_types/data_type.h"
127
#include "vec/data_types/data_type_factory.hpp"
128
#include "vec/data_types/serde/data_type_serde.h"
129
#include "vec/jsonb/serialize.h"
130
131
namespace doris {
132
#include "common/compile_check_begin.h"
133
class TupleDescriptor;
134
135
namespace vectorized {
136
class Block;
137
} // namespace vectorized
138
139
using namespace ErrorCode;
140
using namespace std::chrono_literals;
141
142
using std::pair;
143
using std::string;
144
using std::vector;
145
using io::FileSystemSPtr;
146
147
namespace {
148
149
bvar::Adder<uint64_t> exceed_version_limit_counter;
150
bvar::Window<bvar::Adder<uint64_t>> exceed_version_limit_counter_minute(
151
        &exceed_version_limit_counter, 60);
152
bvar::Adder<uint64_t> cooldown_pending_task("cooldown_pending_task");
153
bvar::Adder<uint64_t> cooldown_processing_task("cooldown_processing_task");
154
155
51
void set_last_failure_time(Tablet* tablet, const Compaction& compaction, int64_t ms) {
156
51
    switch (compaction.compaction_type()) {
157
51
    case ReaderType::READER_CUMULATIVE_COMPACTION:
158
51
        tablet->set_last_cumu_compaction_failure_time(ms);
159
51
        return;
160
0
    case ReaderType::READER_BASE_COMPACTION:
161
0
        tablet->set_last_base_compaction_failure_time(ms);
162
0
        return;
163
0
    case ReaderType::READER_FULL_COMPACTION:
164
0
        tablet->set_last_full_compaction_failure_time(ms);
165
0
        return;
166
0
    default:
167
0
        LOG(FATAL) << "invalid compaction type " << compaction.compaction_name()
168
0
                   << " tablet_id: " << tablet->tablet_id();
169
51
    }
170
51
};
171
172
} // namespace
173
174
bvar::Adder<uint64_t> unused_remote_rowset_num("unused_remote_rowset_num");
175
176
WriteCooldownMetaExecutors::WriteCooldownMetaExecutors(size_t executor_nums)
177
8
        : _executor_nums(executor_nums) {
178
48
    for (size_t i = 0; i < _executor_nums; i++) {
179
40
        std::unique_ptr<PriorityThreadPool> pool;
180
40
        static_cast<void>(ThreadPoolBuilder("WriteCooldownMetaExecutor")
181
40
                                  .set_min_threads(1)
182
40
                                  .set_max_threads(1)
183
40
                                  .set_max_queue_size(std::numeric_limits<int>::max())
184
40
                                  .build(&pool));
185
40
        _executors.emplace_back(std::move(pool));
186
40
    }
187
8
}
188
189
3
void WriteCooldownMetaExecutors::stop() {
190
15
    for (auto& pool_ptr : _executors) {
191
15
        if (pool_ptr) {
192
15
            pool_ptr->shutdown();
193
15
        }
194
15
    }
195
3
}
196
197
116k
void WriteCooldownMetaExecutors::WriteCooldownMetaExecutors::submit(TabletSharedPtr tablet) {
198
116k
    auto tablet_id = tablet->tablet_id();
199
200
116k
    {
201
116k
        std::shared_lock rdlock(tablet->get_header_lock());
202
116k
        if (!tablet->tablet_meta()->cooldown_meta_id().initialized()) {
203
0
            VLOG_NOTICE << "tablet " << tablet_id << " is not cooldown replica";
204
0
            return;
205
0
        }
206
116k
        if (tablet->tablet_state() == TABLET_SHUTDOWN) [[unlikely]] {
207
0
            LOG_INFO("tablet {} has been dropped, don't do cooldown", tablet_id);
208
0
            return;
209
0
        }
210
116k
    }
211
116k
    {
212
        // one tablet could at most have one cooldown task to be done
213
116k
        std::unique_lock<std::mutex> lck {_latch};
214
116k
        if (_pending_tablets.count(tablet_id) > 0) {
215
0
            return;
216
0
        }
217
116k
        _pending_tablets.insert(tablet_id);
218
116k
    }
219
220
116k
    auto async_write_task = [this, t = std::move(tablet)]() {
221
116k
        {
222
116k
            std::unique_lock<std::mutex> lck {_latch};
223
116k
            _pending_tablets.erase(t->tablet_id());
224
116k
        }
225
116k
        SCOPED_ATTACH_TASK(ExecEnv::GetInstance()->orphan_mem_tracker());
226
116k
        auto s = t->write_cooldown_meta();
227
116k
        if (s.ok()) {
228
7
            return;
229
7
        }
230
116k
        if (!s.is<ABORTED>()) {
231
116k
            LOG_EVERY_SECOND(WARNING)
232
14
                    << "write tablet " << t->tablet_id() << " cooldown meta failed because: " << s;
233
116k
            submit(t);
234
116k
            return;
235
116k
        }
236
0
        VLOG_DEBUG << "tablet " << t->tablet_id() << " is not cooldown replica";
237
0
    };
238
239
116k
    cooldown_pending_task << 1;
240
116k
    _executors[_get_executor_pos(tablet_id)]->offer([task = std::move(async_write_task)]() {
241
116k
        cooldown_pending_task << -1;
242
116k
        cooldown_processing_task << 1;
243
116k
        task();
244
116k
        cooldown_processing_task << -1;
245
116k
    });
246
116k
}
247
248
Tablet::Tablet(StorageEngine& engine, TabletMetaSharedPtr tablet_meta, DataDir* data_dir,
249
               const std::string_view& cumulative_compaction_type)
250
335k
        : BaseTablet(std::move(tablet_meta)),
251
335k
          _engine(engine),
252
335k
          _data_dir(data_dir),
253
335k
          _is_bad(false),
254
335k
          _last_cumu_compaction_failure_millis(0),
255
335k
          _last_base_compaction_failure_millis(0),
256
335k
          _last_full_compaction_failure_millis(0),
257
335k
          _last_cumu_compaction_success_millis(0),
258
335k
          _last_base_compaction_success_millis(0),
259
335k
          _last_full_compaction_success_millis(0),
260
335k
          _cumulative_point(K_INVALID_CUMULATIVE_POINT),
261
335k
          _newly_created_rowset_num(0),
262
335k
          _last_checkpoint_time(0),
263
335k
          _cumulative_compaction_type(cumulative_compaction_type),
264
335k
          _is_tablet_path_exists(true),
265
335k
          _last_missed_version(-1),
266
335k
          _last_missed_time_s(0) {
267
335k
    if (_data_dir != nullptr) {
268
335k
        _tablet_path = fmt::format("{}/{}/{}/{}/{}", _data_dir->path(), DATA_PREFIX,
269
335k
                                   _tablet_meta->shard_id(), tablet_id(), schema_hash());
270
335k
    }
271
335k
}
272
273
334k
bool Tablet::set_tablet_schema_into_rowset_meta() {
274
334k
    bool flag = false;
275
602k
    for (auto& [_, rowset_meta] : _tablet_meta->all_mutable_rs_metas()) {
276
602k
        if (!rowset_meta->tablet_schema()) {
277
0
            rowset_meta->set_tablet_schema(_tablet_meta->tablet_schema());
278
0
            flag = true;
279
0
        }
280
602k
    }
281
334k
    return flag;
282
334k
}
283
284
335k
Status Tablet::_init_once_action() {
285
335k
    Status res = Status::OK();
286
335k
    VLOG_NOTICE << "begin to load tablet. tablet=" << tablet_id()
287
419
                << ", version_size=" << _tablet_meta->version_count();
288
289
#ifdef BE_TEST
290
    // init cumulative compaction policy by type
291
    _cumulative_compaction_policy =
292
            CumulativeCompactionPolicyFactory::create_cumulative_compaction_policy(
293
                    _tablet_meta->compaction_policy());
294
#endif
295
296
605k
    for (const auto& [_, rs_meta] : _tablet_meta->all_rs_metas()) {
297
605k
        Version version = rs_meta->version();
298
605k
        RowsetSharedPtr rowset;
299
605k
        res = create_rowset(rs_meta, &rowset);
300
605k
        if (!res.ok()) {
301
0
            LOG(WARNING) << "fail to init rowset. tablet_id=" << tablet_id()
302
0
                         << ", schema_hash=" << schema_hash() << ", version=" << version
303
0
                         << ", res=" << res;
304
0
            return res;
305
0
        }
306
605k
        _rs_version_map[version] = std::move(rowset);
307
605k
    }
308
309
    // init stale rowset
310
335k
    int64_t now = ::time(nullptr);
311
335k
    for (const auto& [_, stale_rs_meta] : _tablet_meta->all_stale_rs_metas()) {
312
13.8k
        Version version = stale_rs_meta->version();
313
314
13.8k
        if (!stale_rs_meta->has_stale_at()) {
315
13.8k
            stale_rs_meta->set_stale_at(now);
316
13.8k
        }
317
318
13.8k
        RowsetSharedPtr rowset;
319
13.8k
        res = create_rowset(stale_rs_meta, &rowset);
320
13.8k
        if (!res.ok()) {
321
0
            LOG(WARNING) << "fail to init stale rowset. tablet_id:" << tablet_id()
322
0
                         << ", schema_hash:" << schema_hash() << ", version=" << version
323
0
                         << ", res:" << res;
324
0
            return res;
325
0
        }
326
13.8k
        _stale_rs_version_map[version] = std::move(rowset);
327
13.8k
    }
328
329
335k
    return res;
330
335k
}
331
332
335k
Status Tablet::init() {
333
335k
    return _init_once.call([this] { return _init_once_action(); });
334
335k
}
335
336
// should save tablet meta to remote meta store
337
// if it's a primary replica
338
11.4k
void Tablet::save_meta() {
339
11.4k
    check_table_size_correctness();
340
11.4k
    auto res = _tablet_meta->save_meta(_data_dir);
341
11.4k
    CHECK_EQ(res, Status::OK()) << "fail to save tablet_meta. res=" << res
342
0
                                << ", root=" << _data_dir->path();
343
11.4k
}
344
345
// Caller should hold _meta_lock.
346
Status Tablet::revise_tablet_meta(const std::vector<RowsetSharedPtr>& to_add,
347
                                  const std::vector<RowsetSharedPtr>& to_delete,
348
0
                                  bool is_incremental_clone) {
349
0
    LOG(INFO) << "begin to revise tablet. tablet_id=" << tablet_id();
350
    // 1. for incremental clone, we have to add the rowsets first to make it easy to compute
351
    //    all the delete bitmaps, and it's easy to delete them if we end up with a failure
352
    // 2. for full clone, we can calculate delete bitmaps on the cloned rowsets directly.
353
0
    if (is_incremental_clone) {
354
0
        CHECK(to_delete.empty()); // don't need to delete rowsets
355
0
        add_rowsets(to_add);
356
        // reconstruct from tablet meta
357
0
        _timestamped_version_tracker.construct_versioned_tracker(_tablet_meta->all_rs_metas());
358
0
    }
359
360
0
    Status calc_bm_status;
361
0
    std::vector<RowsetSharedPtr> base_rowsets_for_full_clone = to_add; // copy vector
362
0
    while (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write()) {
363
0
        std::vector<RowsetSharedPtr> calc_delete_bitmap_rowsets;
364
0
        int64_t to_add_min_version = INT64_MAX;
365
0
        int64_t to_add_max_version = INT64_MIN;
366
0
        for (auto& rs : to_add) {
367
0
            if (to_add_min_version > rs->start_version()) {
368
0
                to_add_min_version = rs->start_version();
369
0
            }
370
0
            if (to_add_max_version < rs->end_version()) {
371
0
                to_add_max_version = rs->end_version();
372
0
            }
373
0
        }
374
0
        Version calc_delete_bitmap_ver;
375
0
        if (is_incremental_clone) {
376
            // From the rowset of to_add with smallest version, all other rowsets
377
            // need to recalculate the delete bitmap
378
            // For example:
379
            // local tablet: [0-1] [2-5] [6-6] [9-10]
380
            // clone tablet: [7-7] [8-8]
381
            // new tablet:   [0-1] [2-5] [6-6] [7-7] [8-8] [9-10]
382
            // [7-7] [8-8] [9-10] need to recalculate delete bitmap
383
0
            calc_delete_bitmap_ver = Version(to_add_min_version, max_version_unlocked());
384
0
        } else {
385
            // the delete bitmap of to_add's rowsets has clone from remote when full clone.
386
            // only other rowsets in local need to recalculate the delete bitmap.
387
            // For example:
388
            // local tablet: [0-1]x [2-5]x [6-6]x [7-7]x [9-10]
389
            // clone tablet: [0-1]  [2-4]  [5-6]  [7-8]
390
            // new tablet:   [0-1]  [2-4]  [5-6]  [7-8] [9-10]
391
            // only [9-10] need to recalculate delete bitmap
392
0
            CHECK_EQ(to_add_min_version, 0) << "to_add_min_version is: " << to_add_min_version;
393
0
            calc_delete_bitmap_ver = Version(to_add_max_version + 1, max_version_unlocked());
394
0
        }
395
396
0
        if (calc_delete_bitmap_ver.first <= calc_delete_bitmap_ver.second) {
397
0
            auto ret = capture_consistent_rowsets_unlocked(calc_delete_bitmap_ver,
398
0
                                                           CaptureRowsetOps {});
399
0
            if (!ret) {
400
0
                LOG(WARNING) << "fail to capture_consistent_rowsets, res: " << ret.error();
401
0
                calc_bm_status = std::move(ret.error());
402
0
                break;
403
0
            }
404
0
            calc_delete_bitmap_rowsets = std::move(ret->rowsets);
405
            // FIXME(plat1ko): Use `const TabletSharedPtr&` as parameter
406
0
            auto self = _engine.tablet_manager()->get_tablet(tablet_id());
407
0
            CHECK(self);
408
0
            for (auto rs : calc_delete_bitmap_rowsets) {
409
0
                if (is_incremental_clone) {
410
0
                    calc_bm_status = update_delete_bitmap_without_lock(self, rs);
411
0
                } else {
412
0
                    calc_bm_status = update_delete_bitmap_without_lock(
413
0
                            self, rs, &base_rowsets_for_full_clone);
414
0
                    base_rowsets_for_full_clone.push_back(rs);
415
0
                }
416
0
                if (!calc_bm_status.ok()) {
417
0
                    LOG(WARNING) << "fail to update_delete_bitmap_without_lock, res: "
418
0
                                 << calc_bm_status;
419
0
                    break;
420
0
                }
421
0
            }
422
0
        }
423
0
        break; // while (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write())
424
0
    }
425
426
0
    DBUG_EXECUTE_IF("Tablet.revise_tablet_meta_fail", {
427
0
        auto ptablet_id = dp->param("tablet_id", 0);
428
0
        if (tablet_id() == ptablet_id) {
429
0
            LOG(INFO) << "injected revies_tablet_meta failure for tabelt: " << ptablet_id;
430
0
            calc_bm_status = Status::InternalError("fault injection error");
431
0
        }
432
0
    });
433
434
    // error handling
435
0
    if (!calc_bm_status.ok()) {
436
0
        if (is_incremental_clone) {
437
0
            RETURN_IF_ERROR(delete_rowsets(to_add, false));
438
0
            LOG(WARNING) << "incremental clone on tablet: " << tablet_id() << " failed due to "
439
0
                         << calc_bm_status.msg() << ", revert " << to_add.size()
440
0
                         << " rowsets added before.";
441
0
        } else {
442
0
            LOG(WARNING) << "full clone on tablet: " << tablet_id() << " failed due to "
443
0
                         << calc_bm_status.msg() << ", will not update tablet meta.";
444
0
        }
445
0
        return calc_bm_status;
446
0
    }
447
448
    // full clone, calculate delete bitmap succeeded, update rowset
449
0
    if (!is_incremental_clone) {
450
0
        RETURN_IF_ERROR(delete_rowsets(to_delete, false));
451
0
        add_rowsets(to_add);
452
        // reconstruct from tablet meta
453
0
        _timestamped_version_tracker.construct_versioned_tracker(_tablet_meta->all_rs_metas());
454
455
        // check the rowsets used for delete bitmap calculation is equal to the rowsets
456
        // that we can capture by version
457
0
        if (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write()) {
458
0
            Version full_version = Version(0, max_version_unlocked());
459
0
            auto ret = capture_consistent_rowsets_unlocked(full_version, CaptureRowsetOps {});
460
0
            DCHECK(ret) << ret.error();
461
0
            DCHECK_EQ(base_rowsets_for_full_clone.size(), ret->rowsets.size());
462
463
0
            if (ret && base_rowsets_for_full_clone.size() != ret->rowsets.size()) [[unlikely]] {
464
0
                LOG(WARNING) << "full clone succeeded, but the count("
465
0
                             << base_rowsets_for_full_clone.size()
466
0
                             << ") of base rowsets used for delete bitmap calculation is not match "
467
0
                                "expect count("
468
0
                             << ret->rowsets.size() << ") we capture from tablet meta";
469
0
            }
470
0
        }
471
0
    }
472
473
    // clear stale rowset
474
0
    for (auto& [v, rs] : _stale_rs_version_map) {
475
0
        _engine.add_unused_rowset(rs);
476
0
    }
477
0
    _stale_rs_version_map.clear();
478
0
    _tablet_meta->clear_stale_rowset();
479
0
    save_meta();
480
481
0
    LOG(INFO) << "finish to revise tablet. tablet_id=" << tablet_id();
482
0
    return Status::OK();
483
0
}
484
485
2.65k
Status Tablet::add_rowset(RowsetSharedPtr rowset) {
486
2.65k
    DCHECK(rowset != nullptr);
487
2.65k
    std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
488
2.65k
    SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
489
    // If the rowset already exist, just return directly.  The rowset_id is an unique-id,
490
    // we can use it to check this situation.
491
2.65k
    if (_contains_rowset(rowset->rowset_id())) {
492
309
        return Status::OK();
493
309
    }
494
    // Otherwise, the version should be not contained in any existing rowset.
495
2.34k
    RETURN_IF_ERROR(_contains_version(rowset->version()));
496
497
1.54k
    RETURN_IF_ERROR(_tablet_meta->add_rs_meta(rowset->rowset_meta()));
498
1.54k
    _rs_version_map[rowset->version()] = rowset;
499
1.54k
    _timestamped_version_tracker.add_version(rowset->version());
500
1.54k
    add_compaction_score(rowset->rowset_meta()->get_compaction_score());
501
502
1.54k
    std::vector<RowsetSharedPtr> rowsets_to_delete;
503
    // yiguolei: temp code, should remove the rowset contains by this rowset
504
    // but it should be removed in multi path version
505
3.05k
    for (auto& it : _rs_version_map) {
506
3.05k
        if (rowset->version().contains(it.first) && rowset->version() != it.first) {
507
0
            CHECK(it.second != nullptr)
508
0
                    << "there exist a version=" << it.first
509
0
                    << " contains the input rs with version=" << rowset->version()
510
0
                    << ", but the related rs is null";
511
0
            rowsets_to_delete.push_back(it.second);
512
0
        }
513
3.05k
    }
514
1.54k
    std::vector<RowsetSharedPtr> empty_vec;
515
1.54k
    RETURN_IF_ERROR(modify_rowsets(empty_vec, rowsets_to_delete));
516
1.54k
    ++_newly_created_rowset_num;
517
1.54k
    return Status::OK();
518
1.54k
}
519
520
114
bool Tablet::rowset_exists_unlocked(const RowsetSharedPtr& rowset) {
521
114
    if (auto it = _rs_version_map.find(rowset->version()); it == _rs_version_map.end()) {
522
0
        return false;
523
114
    } else if (rowset->rowset_id() != it->second->rowset_id()) {
524
0
        return false;
525
0
    }
526
114
    return true;
527
114
}
528
529
Status Tablet::modify_rowsets(std::vector<RowsetSharedPtr>& to_add,
530
1.60k
                              std::vector<RowsetSharedPtr>& to_delete, bool check_delete) {
531
    // the compaction process allow to compact the single version, eg: version[4-4].
532
    // this kind of "single version compaction" has same "input version" and "output version".
533
    // which means "to_add->version()" equals to "to_delete->version()".
534
    // So we should delete the "to_delete" before adding the "to_add",
535
    // otherwise, the "to_add" will be deleted from _rs_version_map, eventually.
536
    //
537
    // And if the version of "to_add" and "to_delete" are exactly same. eg:
538
    // to_add:      [7-7]
539
    // to_delete:   [7-7]
540
    // In this case, we no longer need to add the rowset in "to_delete" to
541
    // _stale_rs_version_map, but can delete it directly.
542
543
1.60k
    if (to_add.empty() && to_delete.empty()) {
544
1.54k
        return Status::OK();
545
1.54k
    }
546
547
68
    if (check_delete) {
548
232
        for (auto&& rs : to_delete) {
549
232
            if (auto it = _rs_version_map.find(rs->version()); it == _rs_version_map.end()) {
550
0
                return Status::Error<DELETE_VERSION_ERROR>(
551
0
                        "try to delete not exist version {} from {}", rs->version().to_string(),
552
0
                        tablet_id());
553
232
            } else if (rs->rowset_id() != it->second->rowset_id()) {
554
0
                return Status::Error<DELETE_VERSION_ERROR>(
555
0
                        "try to delete version {} from {}, but rowset id changed, delete "
556
0
                        "rowset id "
557
0
                        "is {}, exists rowsetid is {}",
558
0
                        rs->version().to_string(), tablet_id(), rs->rowset_id().to_string(),
559
0
                        it->second->rowset_id().to_string());
560
0
            }
561
232
        }
562
67
    }
563
564
68
    bool same_version = true;
565
68
    std::sort(to_add.begin(), to_add.end(), Rowset::comparator);
566
68
    std::sort(to_delete.begin(), to_delete.end(), Rowset::comparator);
567
68
    if (to_add.size() == to_delete.size()) {
568
32
        for (int i = 0; i < to_add.size(); ++i) {
569
16
            if (to_add[i]->version() != to_delete[i]->version()) {
570
0
                same_version = false;
571
0
                break;
572
0
            }
573
16
        }
574
52
    } else {
575
52
        same_version = false;
576
52
    }
577
578
68
    std::vector<RowsetMetaSharedPtr> rs_metas_to_delete;
579
68
    int64_t now = ::time(nullptr);
580
232
    for (auto& rs : to_delete) {
581
232
        rs_metas_to_delete.push_back(rs->rowset_meta());
582
232
        _rs_version_map.erase(rs->version());
583
584
232
        if (!same_version) {
585
216
            rs->rowset_meta()->set_stale_at(now);
586
            // put compaction rowsets in _stale_rs_version_map.
587
216
            _stale_rs_version_map[rs->version()] = rs;
588
216
        }
589
232
    }
590
591
68
    std::vector<RowsetMetaSharedPtr> rs_metas_to_add;
592
68
    for (auto& rs : to_add) {
593
68
        rs_metas_to_add.push_back(rs->rowset_meta());
594
68
        _rs_version_map[rs->version()] = rs;
595
596
68
        if (!same_version) {
597
            // If version are same, then _timestamped_version_tracker
598
            // already has this version, no need to add again.
599
52
            _timestamped_version_tracker.add_version(rs->version());
600
52
        }
601
68
        ++_newly_created_rowset_num;
602
68
    }
603
604
68
    _tablet_meta->modify_rs_metas(rs_metas_to_add, rs_metas_to_delete, same_version);
605
606
68
    if (!same_version) {
607
        // add rs_metas_to_delete to tracker
608
52
        _timestamped_version_tracker.add_stale_path_version(rs_metas_to_delete);
609
52
    } else {
610
        // delete rowset in "to_delete" directly
611
16
        for (auto& rs : to_delete) {
612
16
            LOG(INFO) << "add unused rowset " << rs->rowset_id() << " because of same version";
613
16
            if (rs->is_local()) {
614
16
                _engine.add_unused_rowset(rs);
615
16
            }
616
16
        }
617
16
    }
618
619
68
    int32_t add_score = 0;
620
68
    for (auto rs : to_add) {
621
68
        add_score += rs->rowset_meta()->get_compaction_score();
622
68
    }
623
68
    int32_t sub_score = 0;
624
232
    for (auto rs : to_delete) {
625
232
        sub_score += rs->rowset_meta()->get_compaction_score();
626
232
    }
627
68
    add_compaction_score(add_score - sub_score);
628
629
68
    return Status::OK();
630
68
}
631
632
5
void Tablet::add_rowsets(const std::vector<RowsetSharedPtr>& to_add) {
633
5
    if (to_add.empty()) {
634
0
        return;
635
0
    }
636
5
    std::vector<RowsetMetaSharedPtr> rs_metas;
637
5
    rs_metas.reserve(to_add.size());
638
5
    for (auto& rs : to_add) {
639
5
        _rs_version_map.emplace(rs->version(), rs);
640
5
        _timestamped_version_tracker.add_version(rs->version());
641
5
        rs_metas.push_back(rs->rowset_meta());
642
5
    }
643
5
    _tablet_meta->modify_rs_metas(rs_metas, {});
644
5
}
645
646
5
Status Tablet::delete_rowsets(const std::vector<RowsetSharedPtr>& to_delete, bool move_to_stale) {
647
5
    if (to_delete.empty()) {
648
0
        return Status::OK();
649
0
    }
650
5
    std::vector<RowsetMetaSharedPtr> rs_metas;
651
5
    rs_metas.reserve(to_delete.size());
652
5
    int64_t now = ::time(nullptr);
653
5
    for (const auto& rs : to_delete) {
654
5
        if (move_to_stale) {
655
0
            rs->rowset_meta()->set_stale_at(now);
656
0
        }
657
5
        rs_metas.push_back(rs->rowset_meta());
658
5
        _rs_version_map.erase(rs->version());
659
5
    }
660
5
    _tablet_meta->modify_rs_metas({}, rs_metas, !move_to_stale);
661
5
    if (move_to_stale) {
662
0
        for (const auto& rs : to_delete) {
663
0
            _stale_rs_version_map[rs->version()] = rs;
664
0
        }
665
0
        _timestamped_version_tracker.add_stale_path_version(rs_metas);
666
5
    } else {
667
5
        for (const auto& rs : to_delete) {
668
5
            _timestamped_version_tracker.delete_version(rs->version());
669
5
            if (rs->is_local()) {
670
5
                _engine.add_unused_rowset(rs);
671
5
                RETURN_IF_ERROR(RowsetMetaManager::remove(_data_dir->get_meta(), tablet_uid(),
672
5
                                                          rs->rowset_meta()->rowset_id()));
673
5
            }
674
5
        }
675
5
    }
676
5
    return Status::OK();
677
5
}
678
679
0
RowsetSharedPtr Tablet::_rowset_with_largest_size() {
680
0
    RowsetSharedPtr largest_rowset = nullptr;
681
0
    for (auto& it : _rs_version_map) {
682
0
        if (it.second->empty() || it.second->zero_num_rows()) {
683
0
            continue;
684
0
        }
685
0
        if (largest_rowset == nullptr || it.second->rowset_meta()->index_disk_size() >
686
0
                                                 largest_rowset->rowset_meta()->index_disk_size()) {
687
0
            largest_rowset = it.second;
688
0
        }
689
0
    }
690
691
0
    return largest_rowset;
692
0
}
693
694
// add inc rowset should not persist tablet meta, because it will be persisted when publish txn.
695
10.6k
Status Tablet::add_inc_rowset(const RowsetSharedPtr& rowset) {
696
10.6k
    DCHECK(rowset != nullptr);
697
10.6k
    std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
698
10.6k
    SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
699
10.6k
    if (_contains_rowset(rowset->rowset_id())) {
700
0
        return Status::OK();
701
0
    }
702
10.6k
    RETURN_IF_ERROR(_contains_version(rowset->version()));
703
704
10.6k
    RETURN_IF_ERROR(_tablet_meta->add_rs_meta(rowset->rowset_meta()));
705
10.6k
    _rs_version_map[rowset->version()] = rowset;
706
707
10.6k
    _timestamped_version_tracker.add_version(rowset->version());
708
709
10.6k
    ++_newly_created_rowset_num;
710
711
10.6k
    add_compaction_score(rowset->rowset_meta()->get_compaction_score());
712
713
10.6k
    return Status::OK();
714
10.6k
}
715
716
13.2k
void Tablet::_delete_stale_rowset_by_version(const Version& version) {
717
13.2k
    RowsetMetaSharedPtr rowset_meta = _tablet_meta->acquire_stale_rs_meta_by_version(version);
718
13.2k
    if (rowset_meta == nullptr) {
719
8
        return;
720
8
    }
721
13.2k
    _tablet_meta->delete_stale_rs_meta_by_version(version);
722
13.2k
    VLOG_NOTICE << "delete stale rowset. tablet=" << tablet_id() << ", version=" << version;
723
13.2k
}
724
725
660k
void Tablet::delete_expired_stale_rowset() {
726
660k
    if (config::enable_mow_verbose_log) {
727
0
        LOG_INFO("begin delete_expired_stale_rowset for tablet={}", tablet_id());
728
0
    }
729
660k
    int64_t now = UnixSeconds();
730
660k
    std::vector<std::pair<Version, std::vector<RowsetId>>> deleted_stale_rowsets;
731
    // hold write lock while processing stable rowset
732
660k
    {
733
660k
        std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
734
660k
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
735
        // Compute the end time to delete rowsets, when a expired rowset createtime less then this time, it will be deleted.
736
660k
        int64_t expired_stale_sweep_endtime =
737
660k
                static_cast<int64_t>(::difftime(now, config::tablet_rowset_stale_sweep_time_sec));
738
660k
        if (config::tablet_rowset_stale_sweep_by_size) {
739
0
            expired_stale_sweep_endtime = now;
740
0
        }
741
742
660k
        std::vector<int64_t> path_id_vec;
743
        // capture the path version to delete
744
660k
        _timestamped_version_tracker.capture_expired_paths(
745
660k
                static_cast<int64_t>(expired_stale_sweep_endtime), &path_id_vec);
746
747
660k
        if (path_id_vec.empty()) {
748
659k
            return;
749
659k
        }
750
751
1.49k
        const RowsetSharedPtr lastest_delta = get_rowset_with_max_version();
752
1.49k
        if (lastest_delta == nullptr) {
753
0
            LOG(WARNING) << "lastest_delta is null " << tablet_id();
754
0
            return;
755
0
        }
756
757
        // fetch missing version before delete
758
1.49k
        Versions missed_versions = get_missed_versions_unlocked(lastest_delta->end_version());
759
1.49k
        if (!missed_versions.empty()) {
760
0
            LOG(WARNING) << "tablet:" << tablet_id()
761
0
                         << ", missed version for version:" << lastest_delta->end_version();
762
0
            _print_missed_versions(missed_versions);
763
0
            return;
764
0
        }
765
766
        // do check consistent operation
767
1.49k
        auto path_id_iter = path_id_vec.begin();
768
769
1.49k
        std::map<int64_t, PathVersionListSharedPtr> stale_version_path_map;
770
3.88k
        while (path_id_iter != path_id_vec.end()) {
771
2.38k
            PathVersionListSharedPtr version_path =
772
2.38k
                    _timestamped_version_tracker.fetch_and_delete_path_by_id(*path_id_iter);
773
774
2.38k
            Version test_version = Version(0, lastest_delta->end_version());
775
2.38k
            stale_version_path_map[*path_id_iter] = version_path;
776
777
2.38k
            auto ret = capture_consistent_versions_unlocked(test_version, {});
778
            // 1. When there is no consistent versions, we must reconstruct the tracker.
779
2.38k
            if (!ret) {
780
                // 2. fetch missing version after delete
781
0
                Versions after_missed_versions =
782
0
                        get_missed_versions_unlocked(lastest_delta->end_version());
783
784
                // 2.1 check whether missed_versions and after_missed_versions are the same.
785
                // when they are the same, it means we can delete the path securely.
786
0
                bool is_missing = missed_versions.size() != after_missed_versions.size();
787
788
0
                if (!is_missing) {
789
0
                    for (int ver_index = 0; ver_index < missed_versions.size(); ver_index++) {
790
0
                        if (missed_versions[ver_index] != after_missed_versions[ver_index]) {
791
0
                            is_missing = true;
792
0
                            break;
793
0
                        }
794
0
                    }
795
0
                }
796
797
0
                if (is_missing) {
798
0
                    LOG(WARNING) << "The consistent version check fails, there are bugs. "
799
0
                                 << "Reconstruct the tracker to recover versions in tablet="
800
0
                                 << tablet_id();
801
802
                    // 3. try to recover
803
0
                    _timestamped_version_tracker.recover_versioned_tracker(stale_version_path_map);
804
805
                    // 4. double check the consistent versions
806
                    // fetch missing version after recover
807
0
                    Versions recover_missed_versions =
808
0
                            get_missed_versions_unlocked(lastest_delta->end_version());
809
810
                    // 4.1 check whether missed_versions and recover_missed_versions are the same.
811
                    // when they are the same, it means we recover successfully.
812
0
                    bool is_recover_missing =
813
0
                            missed_versions.size() != recover_missed_versions.size();
814
815
0
                    if (!is_recover_missing) {
816
0
                        for (int ver_index = 0; ver_index < missed_versions.size(); ver_index++) {
817
0
                            if (missed_versions[ver_index] != recover_missed_versions[ver_index]) {
818
0
                                is_recover_missing = true;
819
0
                                break;
820
0
                            }
821
0
                        }
822
0
                    }
823
824
                    // 5. check recover fail, version is mission
825
0
                    if (is_recover_missing) {
826
0
                        if (!config::ignore_rowset_stale_unconsistent_delete) {
827
0
                            LOG(FATAL)
828
0
                                    << "rowset stale unconsistent delete. tablet= " << tablet_id();
829
0
                        } else {
830
0
                            LOG(WARNING)
831
0
                                    << "rowset stale unconsistent delete. tablet= " << tablet_id();
832
0
                        }
833
0
                    }
834
0
                }
835
0
                return;
836
0
            }
837
2.38k
            path_id_iter++;
838
2.38k
        }
839
840
1.49k
        auto old_size = _stale_rs_version_map.size();
841
1.49k
        auto old_meta_size = _tablet_meta->all_stale_rs_metas().size();
842
843
        // do delete operation
844
1.49k
        auto to_delete_iter = stale_version_path_map.begin();
845
3.88k
        while (to_delete_iter != stale_version_path_map.end()) {
846
2.38k
            std::vector<TimestampedVersionSharedPtr>& to_delete_version =
847
2.38k
                    to_delete_iter->second->timestamped_versions();
848
2.38k
            int64_t start_version = -1;
849
2.38k
            int64_t end_version = -1;
850
2.38k
            std::vector<RowsetId> remove_rowset_ids;
851
13.2k
            for (auto& timestampedVersion : to_delete_version) {
852
13.2k
                auto it = _stale_rs_version_map.find(timestampedVersion->version());
853
13.2k
                if (it != _stale_rs_version_map.end()) {
854
13.2k
                    it->second->clear_cache();
855
                    // delete rowset
856
13.2k
                    if (it->second->is_local()) {
857
13.2k
                        _engine.add_unused_rowset(it->second);
858
13.2k
                        if (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write()) {
859
                            // mow does not support cold data in object storage
860
0
                            remove_rowset_ids.emplace_back(it->second->rowset_id());
861
0
                        }
862
13.2k
                    }
863
13.2k
                    _stale_rs_version_map.erase(it);
864
13.2k
                    VLOG_NOTICE << "delete stale rowset tablet=" << tablet_id() << " version["
865
0
                                << timestampedVersion->version().first << ","
866
0
                                << timestampedVersion->version().second
867
0
                                << "] move to unused_rowset success " << std::fixed
868
0
                                << expired_stale_sweep_endtime;
869
13.2k
                } else {
870
8
                    LOG(WARNING) << "delete stale rowset tablet=" << tablet_id() << " version["
871
8
                                 << timestampedVersion->version().first << ","
872
8
                                 << timestampedVersion->version().second
873
8
                                 << "] not find in stale rs version map";
874
8
                }
875
13.2k
                if (start_version < 0) {
876
2.38k
                    start_version = timestampedVersion->version().first;
877
2.38k
                }
878
13.2k
                end_version = timestampedVersion->version().second;
879
13.2k
                _delete_stale_rowset_by_version(timestampedVersion->version());
880
13.2k
            }
881
2.38k
            Version version(start_version, end_version);
882
2.38k
            to_delete_iter++;
883
2.38k
            if (!remove_rowset_ids.empty()) {
884
0
                deleted_stale_rowsets.emplace_back(version, remove_rowset_ids);
885
0
            }
886
2.38k
        }
887
888
1.49k
        bool reconstructed = _reconstruct_version_tracker_if_necessary();
889
890
1.49k
        VLOG_NOTICE << "delete stale rowset _stale_rs_version_map tablet=" << tablet_id()
891
1
                    << " current_size=" << _stale_rs_version_map.size() << " old_size=" << old_size
892
1
                    << " current_meta_size=" << _tablet_meta->all_stale_rs_metas().size()
893
1
                    << " old_meta_size=" << old_meta_size << " sweep endtime " << std::fixed
894
1
                    << expired_stale_sweep_endtime << ", reconstructed=" << reconstructed;
895
1.49k
    }
896
1.49k
    if (config::enable_agg_and_remove_pre_rowsets_delete_bitmap && !deleted_stale_rowsets.empty()) {
897
        // agg delete bitmap for pre rowsets; record unused delete bitmap key ranges
898
0
        OlapStopWatch watch;
899
0
        for (const auto& [version, remove_rowset_ids] : deleted_stale_rowsets) {
900
            // agg delete bitmap for pre rowset
901
0
            DeleteBitmapKeyRanges remove_delete_bitmap_key_ranges;
902
0
            agg_delete_bitmap_for_stale_rowsets(version, remove_delete_bitmap_key_ranges);
903
            // add remove delete bitmap
904
0
            if (!remove_delete_bitmap_key_ranges.empty()) {
905
0
                _engine.add_unused_delete_bitmap_key_ranges(tablet_id(), remove_rowset_ids,
906
0
                                                            remove_delete_bitmap_key_ranges);
907
0
            }
908
0
        }
909
0
        LOG(INFO) << "agg pre rowsets delete bitmap. tablet_id=" << tablet_id()
910
0
                  << ", size=" << deleted_stale_rowsets.size()
911
0
                  << ", cost(us)=" << watch.get_elapse_time_us();
912
0
    }
913
1.49k
#ifndef BE_TEST
914
1.49k
    {
915
1.49k
        std::shared_lock<std::shared_mutex> rlock(_meta_lock);
916
1.49k
        save_meta();
917
1.49k
    }
918
1.49k
#endif
919
1.49k
    if (config::enable_mow_verbose_log) {
920
0
        LOG_INFO("finish delete_expired_stale_rowset for tablet={}", tablet_id());
921
0
    }
922
1.49k
    DBUG_EXECUTE_IF("Tablet.delete_expired_stale_rowset.start_delete_unused_rowset", {
923
1.49k
        _engine.start_delete_unused_rowset();
924
1.49k
        [[maybe_unused]] auto st = _engine.start_trash_sweep(nullptr);
925
1.49k
    });
926
1.49k
}
927
928
0
Status Tablet::check_version_integrity(const Version& version, bool quiet) {
929
0
    std::shared_lock rdlock(_meta_lock);
930
0
    [[maybe_unused]] auto _versions = DORIS_TRY(
931
0
            capture_consistent_versions_unlocked(version, CaptureRowsetOps {.quiet = quiet}));
932
0
    return Status::OK();
933
0
}
934
935
96
bool Tablet::exceed_version_limit(int32_t limit) {
936
96
    if (_tablet_meta->version_count() > limit) {
937
0
        exceed_version_limit_counter << 1;
938
0
        return true;
939
0
    }
940
96
    return false;
941
96
}
942
943
// If any rowset contains the specific version, it means the version already exist
944
97
bool Tablet::check_version_exist(const Version& version) const {
945
97
    std::shared_lock rdlock(_meta_lock);
946
119
    for (auto& it : _rs_version_map) {
947
119
        if (it.first.contains(version)) {
948
96
            return true;
949
96
        }
950
119
    }
951
1
    return false;
952
97
}
953
954
// The meta read lock should be held before calling
955
void Tablet::acquire_version_and_rowsets(
956
0
        std::vector<std::pair<Version, RowsetSharedPtr>>* version_rowsets) const {
957
0
    for (const auto& it : _rs_version_map) {
958
0
        version_rowsets->emplace_back(it.first, it.second);
959
0
    }
960
0
}
961
962
Status Tablet::capture_rs_readers(const Version& spec_version, std::vector<RowSetSplits>* rs_splits,
963
2
                                  const CaptureRowsetOps& opts) {
964
2
    std::shared_lock rlock(_meta_lock);
965
2
    std::vector<Version> version_path;
966
2
    *rs_splits = DORIS_TRY(capture_rs_readers_unlocked(
967
1
            spec_version, CaptureRowsetOps {.skip_missing_versions = opts.skip_missing_versions}));
968
1
    return Status::OK();
969
2
}
970
971
1.50k
Versions Tablet::calc_missed_versions(int64_t spec_version, Versions existing_versions) const {
972
1.50k
    DCHECK(spec_version > 0) << "invalid spec_version: " << spec_version;
973
974
    // sort the existing versions in ascending order
975
1.50k
    std::sort(existing_versions.begin(), existing_versions.end(),
976
7.44k
              [](const Version& a, const Version& b) {
977
                  // simple because 2 versions are certainly not overlapping
978
7.44k
                  return a.first < b.first;
979
7.44k
              });
980
981
    // From the first version(=0),  find the missing version until spec_version
982
1.50k
    int64_t last_version = -1;
983
1.50k
    Versions missed_versions;
984
4.56k
    for (const Version& version : existing_versions) {
985
4.56k
        if (version.first > last_version + 1) {
986
4
            for (int64_t i = last_version + 1; i < version.first && i <= spec_version; ++i) {
987
                // Don't merge missed_versions because clone & snapshot use single version.
988
                // For example, if miss 4 ~ 6, clone need [4, 4], [5, 5], [6, 6], but not [4, 6].
989
2
                missed_versions.emplace_back(i, i);
990
2
            }
991
2
        }
992
4.56k
        last_version = version.second;
993
4.56k
        if (last_version >= spec_version) {
994
1.50k
            break;
995
1.50k
        }
996
4.56k
    }
997
1.50k
    for (int64_t i = last_version + 1; i <= spec_version; ++i) {
998
0
        missed_versions.emplace_back(i, i);
999
0
    }
1000
1001
1.50k
    return missed_versions;
1002
1.50k
}
1003
1004
25.1M
bool Tablet::can_do_compaction(size_t path_hash, CompactionType compaction_type) {
1005
25.1M
    if (compaction_type == CompactionType::BASE_COMPACTION && tablet_state() != TABLET_RUNNING) {
1006
        // base compaction can only be done for tablet in TABLET_RUNNING state.
1007
        // but cumulative compaction can be done for TABLET_NOTREADY, such as tablet under alter process.
1008
4
        return false;
1009
4
    }
1010
1011
25.1M
    if (data_dir()->path_hash() != path_hash || !is_used() || !init_succeeded()) {
1012
9.94M
        return false;
1013
9.94M
    }
1014
1015
    // In TABLET_NOTREADY, we keep last 10 versions in new tablet so base tablet max_version
1016
    // not merged in new tablet and then we can do compaction
1017
15.2M
    return tablet_state() == TABLET_RUNNING || tablet_state() == TABLET_NOTREADY;
1018
25.1M
}
1019
1020
14.6M
uint32_t Tablet::calc_compaction_score() {
1021
14.6M
    if (_score_check_cnt++ % config::check_score_rounds_num != 0) {
1022
14.3M
        std::shared_lock rdlock(_meta_lock);
1023
14.3M
        if (_compaction_score > 0) {
1024
14.3M
            return _compaction_score;
1025
14.3M
        }
1026
14.3M
    }
1027
1028
333k
    {
1029
        // Need meta lock, because it will iterator "all_rs_metas" of tablet meta.
1030
333k
        std::shared_lock rdlock(_meta_lock);
1031
333k
        int32_t score = get_real_compaction_score();
1032
333k
        if (_compaction_score > 0 && _compaction_score != score) {
1033
0
            LOG(WARNING) << "cumu cache score not equal real score, cache score; "
1034
0
                         << _compaction_score << ", real score: " << score
1035
0
                         << ", tablet: " << tablet_id();
1036
0
        }
1037
333k
        _compaction_score = score;
1038
333k
        return score;
1039
14.6M
    }
1040
14.6M
}
1041
1042
bool Tablet::suitable_for_compaction(
1043
        CompactionType compaction_type,
1044
541k
        std::shared_ptr<CumulativeCompactionPolicy> cumulative_compaction_policy) {
1045
    // Need meta lock, because it will iterator "all_rs_metas" of tablet meta.
1046
541k
    std::shared_lock rdlock(_meta_lock);
1047
541k
    int32_t score = -1;
1048
541k
    if (compaction_type == CompactionType::CUMULATIVE_COMPACTION) {
1049
88.7k
        score = _calc_cumulative_compaction_score(cumulative_compaction_policy);
1050
453k
    } else {
1051
453k
        DCHECK_EQ(compaction_type, CompactionType::BASE_COMPACTION);
1052
453k
        score = _calc_base_compaction_score();
1053
453k
    }
1054
541k
    return score > 0;
1055
541k
}
1056
1057
7
uint32_t Tablet::calc_cold_data_compaction_score() const {
1058
7
    uint32_t score = 0;
1059
7
    std::vector<RowsetMetaSharedPtr> cooldowned_rowsets;
1060
7
    int64_t max_delete_version = 0;
1061
7
    {
1062
7
        std::shared_lock rlock(_meta_lock);
1063
14
        for (const auto& [_, rs_meta] : _tablet_meta->all_rs_metas()) {
1064
14
            if (!rs_meta->is_local()) {
1065
14
                cooldowned_rowsets.push_back(rs_meta);
1066
14
                if (rs_meta->has_delete_predicate() &&
1067
14
                    rs_meta->end_version() > max_delete_version) {
1068
0
                    max_delete_version = rs_meta->end_version();
1069
0
                }
1070
14
            }
1071
14
        }
1072
7
    }
1073
14
    for (auto& rs_meta : cooldowned_rowsets) {
1074
14
        if (rs_meta->end_version() < max_delete_version) {
1075
0
            score += rs_meta->num_segments();
1076
14
        } else {
1077
14
            score += rs_meta->get_compaction_score();
1078
14
        }
1079
14
    }
1080
7
    return (keys_type() != KeysType::DUP_KEYS) ? score * 2 : score;
1081
7
}
1082
1083
uint32_t Tablet::_calc_cumulative_compaction_score(
1084
88.7k
        std::shared_ptr<CumulativeCompactionPolicy> cumulative_compaction_policy) {
1085
88.7k
    if (cumulative_compaction_policy == nullptr) [[unlikely]] {
1086
0
        return 0;
1087
0
    }
1088
88.7k
#ifndef BE_TEST
1089
88.7k
    if (_cumulative_compaction_policy == nullptr ||
1090
88.7k
        _cumulative_compaction_policy->name() != cumulative_compaction_policy->name()) {
1091
3.92k
        _cumulative_compaction_policy = cumulative_compaction_policy;
1092
3.92k
    }
1093
88.7k
#endif
1094
88.7k
    DBUG_EXECUTE_IF("Tablet._calc_cumulative_compaction_score.return", {
1095
88.7k
        LOG_WARNING("Tablet._calc_cumulative_compaction_score.return")
1096
88.7k
                .tag("tablet id", tablet_id());
1097
88.7k
        return 0;
1098
88.7k
    });
1099
88.7k
    return _cumulative_compaction_policy->calc_cumulative_compaction_score(this);
1100
88.7k
}
1101
1102
453k
uint32_t Tablet::_calc_base_compaction_score() const {
1103
453k
    uint32_t score = 0;
1104
453k
    const int64_t point = cumulative_layer_point();
1105
453k
    bool base_rowset_exist = false;
1106
453k
    bool has_delete = false;
1107
785k
    for (const auto& [_, rs_meta] : _tablet_meta->all_rs_metas()) {
1108
785k
        if (rs_meta->start_version() == 0) {
1109
453k
            base_rowset_exist = true;
1110
453k
        }
1111
785k
        if (rs_meta->start_version() >= point || !rs_meta->is_local()) {
1112
            // all_rs_metas() is not sorted, so we use _continue_ other than _break_ here.
1113
769k
            continue;
1114
769k
        }
1115
15.2k
        if (rs_meta->has_delete_predicate()) {
1116
432
            has_delete = true;
1117
432
        }
1118
15.2k
        score += rs_meta->get_compaction_score();
1119
15.2k
    }
1120
1121
    // In the time series compaction policy, we want the base compaction to be triggered
1122
    // when there are delete versions present.
1123
453k
    if (_tablet_meta->compaction_policy() == CUMULATIVE_TIME_SERIES_POLICY) {
1124
49
        return (base_rowset_exist && has_delete) ? score : 0;
1125
49
    }
1126
1127
    // base不存在可能是tablet正在做alter table,先不选它,设score=0
1128
453k
    return base_rowset_exist ? score : 0;
1129
453k
}
1130
1131
0
void Tablet::max_continuous_version_from_beginning(Version* version, Version* max_version) {
1132
0
    bool has_version_cross;
1133
0
    std::shared_lock rdlock(_meta_lock);
1134
0
    _max_continuous_version_from_beginning_unlocked(version, max_version, &has_version_cross);
1135
0
}
1136
1137
void Tablet::_max_continuous_version_from_beginning_unlocked(Version* version, Version* max_version,
1138
1.30M
                                                             bool* has_version_cross) const {
1139
1.30M
    std::vector<Version> existing_versions;
1140
1.30M
    *has_version_cross = false;
1141
2.33M
    for (const auto& [ver, _] : _tablet_meta->all_rs_metas()) {
1142
2.33M
        existing_versions.emplace_back(ver);
1143
2.33M
    }
1144
1145
    // sort the existing versions in ascending order
1146
1.30M
    std::sort(existing_versions.begin(), existing_versions.end(),
1147
2.26M
              [](const Version& left, const Version& right) {
1148
                  // simple because 2 versions are certainly not overlapping
1149
2.26M
                  return left.first < right.first;
1150
2.26M
              });
1151
1152
1.30M
    Version max_continuous_version = {-1, -1};
1153
3.63M
    for (int i = 0; i < existing_versions.size(); ++i) {
1154
2.33M
        if (existing_versions[i].first > max_continuous_version.second + 1) {
1155
0
            break;
1156
2.33M
        } else if (existing_versions[i].first <= max_continuous_version.second) {
1157
0
            *has_version_cross = true;
1158
0
        }
1159
2.33M
        max_continuous_version = existing_versions[i];
1160
2.33M
    }
1161
1.30M
    *version = max_continuous_version;
1162
    // tablet may not has rowset, eg, tablet has just been clear for restore.
1163
1.30M
    if (max_version != nullptr && !existing_versions.empty()) {
1164
1.30M
        *max_version = existing_versions.back();
1165
1.30M
    }
1166
1.30M
}
1167
1168
65.0k
void Tablet::calculate_cumulative_point() {
1169
65.0k
    std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
1170
65.0k
    SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
1171
65.0k
    int64_t ret_cumulative_point;
1172
65.0k
    _cumulative_compaction_policy->calculate_cumulative_point(
1173
65.0k
            this, _tablet_meta->all_rs_metas(), _cumulative_point, &ret_cumulative_point);
1174
1175
65.0k
    if (ret_cumulative_point == K_INVALID_CUMULATIVE_POINT) {
1176
62.7k
        return;
1177
62.7k
    }
1178
2.27k
    set_cumulative_layer_point(ret_cumulative_point);
1179
2.27k
}
1180
1181
// NOTE: only used when create_table, so it is sure that there is no concurrent reader and writer.
1182
0
void Tablet::delete_all_files() {
1183
    // Release resources like memory and disk space.
1184
0
    std::shared_lock rdlock(_meta_lock);
1185
0
    for (auto it : _rs_version_map) {
1186
0
        static_cast<void>(it.second->remove());
1187
0
    }
1188
0
    _rs_version_map.clear();
1189
1190
0
    for (auto it : _stale_rs_version_map) {
1191
0
        static_cast<void>(it.second->remove());
1192
0
    }
1193
0
    _stale_rs_version_map.clear();
1194
0
}
1195
1196
0
void Tablet::check_tablet_path_exists() {
1197
0
    if (!tablet_path().empty()) {
1198
0
        std::error_code ec;
1199
0
        if (std::filesystem::is_directory(tablet_path(), ec)) {
1200
0
            _is_tablet_path_exists.store(true, std::memory_order_relaxed);
1201
0
        } else if (ec.value() == ENOENT || ec.value() == 0) {
1202
0
            _is_tablet_path_exists.store(false, std::memory_order_relaxed);
1203
0
        }
1204
0
    }
1205
0
}
1206
1207
12.9k
Status Tablet::_contains_version(const Version& version) {
1208
    // check if there exist a rowset contains the added rowset
1209
359k
    for (auto& it : _rs_version_map) {
1210
359k
        if (it.first.contains(version)) {
1211
            // TODO(lingbin): Is this check unnecessary?
1212
            // because the value type is std::shared_ptr, when will it be nullptr?
1213
            // In addition, in this class, there are many places that do not make this judgment
1214
            // when access _rs_version_map's value.
1215
808
            CHECK(it.second != nullptr) << "there exist a version=" << it.first
1216
0
                                        << " contains the input rs with version=" << version
1217
0
                                        << ", but the related rs is null";
1218
808
            return Status::Error<PUSH_VERSION_ALREADY_EXIST>("Tablet push duplicate version {}",
1219
808
                                                             version.to_string());
1220
808
        }
1221
359k
    }
1222
1223
12.1k
    return Status::OK();
1224
12.9k
}
1225
1226
65.4k
std::vector<RowsetSharedPtr> Tablet::pick_candidate_rowsets_to_cumulative_compaction() {
1227
65.4k
    std::vector<RowsetSharedPtr> candidate_rowsets;
1228
65.4k
    if (_cumulative_point == K_INVALID_CUMULATIVE_POINT) {
1229
437
        return candidate_rowsets;
1230
437
    }
1231
64.9k
    return _pick_visible_rowsets_to_compaction(_cumulative_point,
1232
64.9k
                                               std::numeric_limits<int64_t>::max());
1233
65.4k
}
1234
1235
12.3k
std::vector<RowsetSharedPtr> Tablet::pick_candidate_rowsets_to_base_compaction() {
1236
12.3k
    return _pick_visible_rowsets_to_compaction(std::numeric_limits<int64_t>::min(),
1237
12.3k
                                               _cumulative_point - 1);
1238
12.3k
}
1239
1240
std::vector<RowsetSharedPtr> Tablet::_pick_visible_rowsets_to_compaction(
1241
77.3k
        int64_t min_start_version, int64_t max_start_version) {
1242
77.3k
    auto [visible_version, update_ts] = get_visible_version_and_time();
1243
77.3k
    bool update_time_long = MonotonicMillis() - update_ts >
1244
77.3k
                            config::compaction_keep_invisible_version_timeout_sec * 1000L;
1245
77.3k
    int32_t keep_invisible_version_limit =
1246
77.3k
            update_time_long ? config::compaction_keep_invisible_version_min_count
1247
77.3k
                             : config::compaction_keep_invisible_version_max_count;
1248
1249
77.3k
    std::vector<RowsetSharedPtr> candidate_rowsets;
1250
77.3k
    {
1251
77.3k
        std::shared_lock rlock(_meta_lock);
1252
375k
        for (const auto& [version, rs] : _rs_version_map) {
1253
375k
            int64_t version_start = version.first;
1254
            // rowset is remote or rowset is not in given range
1255
375k
            if (!rs->is_local() || version_start < min_start_version ||
1256
375k
                version_start > max_start_version) {
1257
107k
                continue;
1258
107k
            }
1259
1260
            // can compact, met one of the conditions:
1261
            // 1. had been visible;
1262
            // 2. exceeds the limit of keep invisible versions.
1263
267k
            int64_t version_end = version.second;
1264
267k
            if (version_end <= visible_version ||
1265
267k
                version_end > visible_version + keep_invisible_version_limit) {
1266
254k
                candidate_rowsets.push_back(rs);
1267
254k
            }
1268
267k
        }
1269
77.3k
    }
1270
77.3k
    std::sort(candidate_rowsets.begin(), candidate_rowsets.end(), Rowset::comparator);
1271
77.3k
    return candidate_rowsets;
1272
77.3k
}
1273
1274
0
std::vector<RowsetSharedPtr> Tablet::pick_candidate_rowsets_to_full_compaction() {
1275
0
    std::vector<RowsetSharedPtr> candidate_rowsets;
1276
0
    traverse_rowsets([&candidate_rowsets](const auto& rs) {
1277
        // Do full compaction on all local rowsets.
1278
0
        if (rs->is_local()) {
1279
0
            candidate_rowsets.emplace_back(rs);
1280
0
        }
1281
0
    });
1282
0
    std::sort(candidate_rowsets.begin(), candidate_rowsets.end(), Rowset::comparator);
1283
0
    return candidate_rowsets;
1284
0
}
1285
1286
std::vector<RowsetSharedPtr> Tablet::pick_candidate_rowsets_to_build_inverted_index(
1287
21
        const std::set<int64_t>& alter_index_uids, bool is_drop_op) {
1288
21
    std::vector<RowsetSharedPtr> candidate_rowsets;
1289
21
    {
1290
21
        std::shared_lock rlock(_meta_lock);
1291
41
        auto has_alter_inverted_index = [&](RowsetSharedPtr rowset) -> bool {
1292
45
            for (const auto& index_id : alter_index_uids) {
1293
45
                if (rowset->tablet_schema()->has_inverted_index_with_index_id(index_id)) {
1294
8
                    return true;
1295
8
                }
1296
45
            }
1297
33
            return false;
1298
41
        };
1299
1300
21
        for (const auto& [version, rs] : _rs_version_map) {
1301
21
            if (!has_alter_inverted_index(rs) && is_drop_op) {
1302
1
                continue;
1303
1
            }
1304
20
            if (has_alter_inverted_index(rs) && !is_drop_op) {
1305
0
                continue;
1306
0
            }
1307
1308
20
            if (rs->is_local()) {
1309
20
                candidate_rowsets.push_back(rs);
1310
20
            }
1311
20
        }
1312
21
    }
1313
21
    std::sort(candidate_rowsets.begin(), candidate_rowsets.end(), Rowset::comparator);
1314
21
    return candidate_rowsets;
1315
21
}
1316
1317
1.37M
std::tuple<int64_t, int64_t> Tablet::get_visible_version_and_time() const {
1318
    // some old tablet has bug, its partition_id is 0, fe couldn't update its visible version.
1319
    // so let this tablet's visible version become int64 max.
1320
1.37M
    auto version_info = _visible_version.load();
1321
1.37M
    if (version_info != nullptr && partition_id() != 0) {
1322
1.37M
        return std::make_tuple(version_info->version.load(std::memory_order_relaxed),
1323
1.37M
                               version_info->update_ts);
1324
1.37M
    } else {
1325
72
        return std::make_tuple(std::numeric_limits<int64_t>::max(),
1326
72
                               std::numeric_limits<int64_t>::max());
1327
72
    }
1328
1.37M
}
1329
1330
// For http compaction action
1331
0
void Tablet::get_compaction_status(std::string* json_result) {
1332
0
    rapidjson::Document root;
1333
0
    root.SetObject();
1334
1335
0
    rapidjson::Document path_arr;
1336
0
    path_arr.SetArray();
1337
1338
0
    std::vector<RowsetSharedPtr> rowsets;
1339
0
    std::vector<RowsetSharedPtr> stale_rowsets;
1340
0
    std::vector<bool> delete_flags;
1341
0
    {
1342
0
        std::shared_lock rdlock(_meta_lock);
1343
0
        rowsets.reserve(_rs_version_map.size());
1344
0
        for (auto& it : _rs_version_map) {
1345
0
            rowsets.push_back(it.second);
1346
0
        }
1347
0
        std::sort(rowsets.begin(), rowsets.end(), Rowset::comparator);
1348
1349
0
        stale_rowsets.reserve(_stale_rs_version_map.size());
1350
0
        for (auto& it : _stale_rs_version_map) {
1351
0
            stale_rowsets.push_back(it.second);
1352
0
        }
1353
0
        std::sort(stale_rowsets.begin(), stale_rowsets.end(), Rowset::comparator);
1354
1355
0
        delete_flags.reserve(rowsets.size());
1356
0
        for (auto& rs : rowsets) {
1357
0
            delete_flags.push_back(rs->rowset_meta()->has_delete_predicate());
1358
0
        }
1359
        // get snapshot version path json_doc
1360
0
        _timestamped_version_tracker.get_stale_version_path_json_doc(path_arr);
1361
0
    }
1362
0
    rapidjson::Value cumulative_policy_type;
1363
0
    std::string policy_type_str = "cumulative compaction policy not initializied";
1364
0
    if (_cumulative_compaction_policy != nullptr) {
1365
0
        policy_type_str = _cumulative_compaction_policy->name();
1366
0
    }
1367
0
    cumulative_policy_type.SetString(policy_type_str.c_str(),
1368
0
                                     cast_set<uint32_t>(policy_type_str.length()),
1369
0
                                     root.GetAllocator());
1370
0
    root.AddMember("cumulative policy type", cumulative_policy_type, root.GetAllocator());
1371
0
    root.AddMember("cumulative point", _cumulative_point.load(), root.GetAllocator());
1372
1373
0
#define FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, key, unixmillis_value)                 \
1374
0
    {                                                                                \
1375
0
        rapidjson::Value value;                                                      \
1376
0
        std::string format_str = ToStringFromUnixMillis(unixmillis_value.load());    \
1377
0
        value.SetString(format_str.c_str(), cast_set<uint32_t>(format_str.length()), \
1378
0
                        root.GetAllocator());                                        \
1379
0
        root.AddMember(key, value, root.GetAllocator());                             \
1380
0
    }
1381
0
#define FORMAT_STRING_ADD_JSON_NODE(root, key, str_value)                          \
1382
0
    {                                                                              \
1383
0
        rapidjson::Value value;                                                    \
1384
0
        value.SetString(str_value.c_str(), cast_set<uint32_t>(str_value.length()), \
1385
0
                        root.GetAllocator());                                      \
1386
0
        root.AddMember(key, value, root.GetAllocator());                           \
1387
0
    }
1388
1389
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last cumulative failure time",
1390
0
                                    _last_cumu_compaction_failure_millis)
1391
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last base failure time",
1392
0
                                    _last_base_compaction_failure_millis)
1393
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last full failure time",
1394
0
                                    _last_full_compaction_failure_millis)
1395
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last cumulative success time",
1396
0
                                    _last_cumu_compaction_success_millis)
1397
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last base success time",
1398
0
                                    _last_base_compaction_success_millis)
1399
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last full success time",
1400
0
                                    _last_full_compaction_success_millis)
1401
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last cumulative schedule time",
1402
0
                                    _last_cumu_compaction_schedule_millis)
1403
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last base schedule time",
1404
0
                                    _last_base_compaction_schedule_millis)
1405
0
    FORMAT_UNIXMILLIS_ADD_JSON_NODE(root, "last full schedule time",
1406
0
                                    _last_full_compaction_schedule_millis)
1407
0
    FORMAT_STRING_ADD_JSON_NODE(root, "last cumulative status", _last_cumu_compaction_status)
1408
0
    FORMAT_STRING_ADD_JSON_NODE(root, "last base status", _last_base_compaction_status)
1409
0
    FORMAT_STRING_ADD_JSON_NODE(root, "last full status", _last_full_compaction_status)
1410
1411
    // last single replica compaction status
1412
    // "single replica compaction status": {
1413
    //     "remote peer": "172.100.1.0:10875",
1414
    //     "last failure status": "",
1415
    //     "last fetched rowset": "[8-10]"
1416
    // }
1417
0
    rapidjson::Document status;
1418
0
    status.SetObject();
1419
0
    TReplicaInfo replica_info;
1420
0
    std::string dummp_token;
1421
0
    if (tablet_meta()->tablet_schema()->enable_single_replica_compaction() &&
1422
0
        _engine.get_peer_replica_info(tablet_id(), &replica_info, &dummp_token)) {
1423
        // remote peer
1424
0
        rapidjson::Value peer_addr;
1425
0
        std::string addr = replica_info.host + ":" + std::to_string(replica_info.brpc_port);
1426
0
        peer_addr.SetString(addr.c_str(), cast_set<uint32_t>(addr.length()), status.GetAllocator());
1427
0
        status.AddMember("remote peer", peer_addr, status.GetAllocator());
1428
        // last failure status
1429
0
        rapidjson::Value compaction_status;
1430
0
        compaction_status.SetString(
1431
0
                _last_single_compaction_failure_status.c_str(),
1432
0
                cast_set<uint32_t>(_last_single_compaction_failure_status.length()),
1433
0
                status.GetAllocator());
1434
0
        status.AddMember("last failure status", compaction_status, status.GetAllocator());
1435
        // last fetched rowset
1436
0
        rapidjson::Value version;
1437
0
        std::string fetched_version = _last_fetched_version.to_string();
1438
0
        version.SetString(fetched_version.c_str(), cast_set<uint32_t>(fetched_version.length()),
1439
0
                          status.GetAllocator());
1440
0
        status.AddMember("last fetched rowset", version, status.GetAllocator());
1441
0
        root.AddMember("single replica compaction status", status, root.GetAllocator());
1442
0
    }
1443
1444
    // print all rowsets' version as an array
1445
0
    rapidjson::Document versions_arr;
1446
0
    rapidjson::Document missing_versions_arr;
1447
0
    versions_arr.SetArray();
1448
0
    missing_versions_arr.SetArray();
1449
0
    int64_t last_version = -1;
1450
0
    for (auto& rowset : rowsets) {
1451
0
        const Version& ver = rowset->version();
1452
0
        if (ver.first != last_version + 1) {
1453
0
            rapidjson::Value miss_value;
1454
0
            miss_value.SetString(fmt::format("[{}-{}]", last_version + 1, ver.first - 1).c_str(),
1455
0
                                 missing_versions_arr.GetAllocator());
1456
0
            missing_versions_arr.PushBack(miss_value, missing_versions_arr.GetAllocator());
1457
0
        }
1458
0
        rapidjson::Value value;
1459
0
        std::string version_str = rowset->get_rowset_info_str();
1460
0
        value.SetString(version_str.c_str(), cast_set<uint32_t>(version_str.length()),
1461
0
                        versions_arr.GetAllocator());
1462
0
        versions_arr.PushBack(value, versions_arr.GetAllocator());
1463
0
        last_version = ver.second;
1464
0
    }
1465
0
    root.AddMember("rowsets", versions_arr, root.GetAllocator());
1466
0
    root.AddMember("missing_rowsets", missing_versions_arr, root.GetAllocator());
1467
1468
    // print all stale rowsets' version as an array
1469
0
    rapidjson::Document stale_versions_arr;
1470
0
    stale_versions_arr.SetArray();
1471
0
    for (auto& rowset : stale_rowsets) {
1472
0
        rapidjson::Value value;
1473
0
        std::string version_str = rowset->get_rowset_info_str();
1474
0
        value.SetString(version_str.c_str(), cast_set<uint32_t>(version_str.length()),
1475
0
                        stale_versions_arr.GetAllocator());
1476
0
        stale_versions_arr.PushBack(value, stale_versions_arr.GetAllocator());
1477
0
    }
1478
0
    root.AddMember("stale_rowsets", stale_versions_arr, root.GetAllocator());
1479
1480
    // add stale version rowsets
1481
0
    root.AddMember("stale version path", path_arr, root.GetAllocator());
1482
1483
    // to json string
1484
0
    rapidjson::StringBuffer strbuf;
1485
0
    rapidjson::PrettyWriter<rapidjson::StringBuffer> writer(strbuf);
1486
0
    root.Accept(writer);
1487
0
    *json_result = std::string(strbuf.GetString());
1488
0
}
1489
1490
334k
bool Tablet::do_tablet_meta_checkpoint() {
1491
334k
    std::lock_guard<std::shared_mutex> store_lock(_meta_store_lock);
1492
334k
    if (_newly_created_rowset_num == 0) {
1493
333k
        return false;
1494
333k
    }
1495
951
    if (UnixMillis() - _last_checkpoint_time <
1496
951
                config::tablet_meta_checkpoint_min_interval_secs * 1000 &&
1497
951
        _newly_created_rowset_num < config::tablet_meta_checkpoint_min_new_rowsets_num) {
1498
0
        return false;
1499
0
    }
1500
    // hold read-lock other than write-lock, because it will not modify meta structure
1501
951
    std::shared_lock rdlock(_meta_lock);
1502
951
    if (tablet_state() != TABLET_RUNNING) {
1503
1
        LOG(INFO) << "tablet is under state=" << tablet_state()
1504
1
                  << ", not running, skip do checkpoint"
1505
1
                  << ", tablet=" << tablet_id();
1506
1
        return false;
1507
1
    }
1508
950
    VLOG_NOTICE << "start to do tablet meta checkpoint, tablet=" << tablet_id();
1509
950
    save_meta();
1510
    // if save meta successfully, then should remove the rowset meta existing in tablet
1511
    // meta from rowset meta store
1512
2.10k
    for (const auto& [_, rs_meta] : _tablet_meta->all_rs_metas()) {
1513
        // If we delete it from rowset manager's meta explicitly in previous checkpoint, just skip.
1514
2.10k
        if (rs_meta->is_remove_from_rowset_meta()) {
1515
0
            continue;
1516
0
        }
1517
2.10k
        if (RowsetMetaManager::check_rowset_meta(_data_dir->get_meta(), tablet_uid(),
1518
2.10k
                                                 rs_meta->rowset_id())) {
1519
2.07k
            RETURN_FALSE_IF_ERROR(RowsetMetaManager::remove(_data_dir->get_meta(), tablet_uid(),
1520
2.07k
                                                            rs_meta->rowset_id()));
1521
2.07k
            VLOG_NOTICE << "remove rowset id from meta store because it is already persistent with "
1522
0
                        << "tablet meta, rowset_id=" << rs_meta->rowset_id();
1523
2.07k
        }
1524
2.10k
        rs_meta->set_remove_from_rowset_meta();
1525
2.10k
    }
1526
1527
    // check _stale_rs_version_map to remove meta from rowset meta store
1528
950
    for (const auto& [_, rs_meta] : _tablet_meta->all_stale_rs_metas()) {
1529
        // If we delete it from rowset manager's meta explicitly in previous checkpoint, just skip.
1530
24
        if (rs_meta->is_remove_from_rowset_meta()) {
1531
0
            continue;
1532
0
        }
1533
24
        if (RowsetMetaManager::check_rowset_meta(_data_dir->get_meta(), tablet_uid(),
1534
24
                                                 rs_meta->rowset_id())) {
1535
24
            RETURN_FALSE_IF_ERROR(RowsetMetaManager::remove(_data_dir->get_meta(), tablet_uid(),
1536
24
                                                            rs_meta->rowset_id()));
1537
24
            VLOG_NOTICE << "remove rowset id from meta store because it is already persistent with "
1538
0
                        << "tablet meta, rowset_id=" << rs_meta->rowset_id();
1539
24
        }
1540
24
        rs_meta->set_remove_from_rowset_meta();
1541
24
    }
1542
1543
950
    if (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write()) {
1544
60
        RETURN_FALSE_IF_ERROR(TabletMetaManager::remove_old_version_delete_bitmap(
1545
60
                _data_dir, tablet_id(), max_version_unlocked()));
1546
60
    }
1547
1548
950
    _newly_created_rowset_num = 0;
1549
950
    _last_checkpoint_time = UnixMillis();
1550
950
    return true;
1551
950
}
1552
1553
1.43k
bool Tablet::rowset_meta_is_useful(RowsetMetaSharedPtr rowset_meta) {
1554
1.43k
    std::shared_lock rdlock(_meta_lock);
1555
1.43k
    bool find_version = false;
1556
3.16k
    for (auto& version_rowset : _rs_version_map) {
1557
3.16k
        if (version_rowset.second->rowset_id() == rowset_meta->rowset_id()) {
1558
158
            return true;
1559
158
        }
1560
3.00k
        if (version_rowset.second->contains_version(rowset_meta->version())) {
1561
1.28k
            find_version = true;
1562
1.28k
        }
1563
3.00k
    }
1564
1.51k
    for (auto& stale_version_rowset : _stale_rs_version_map) {
1565
1.51k
        if (stale_version_rowset.second->rowset_id() == rowset_meta->rowset_id()) {
1566
468
            return true;
1567
468
        }
1568
1.04k
        if (stale_version_rowset.second->contains_version(rowset_meta->version())) {
1569
80
            find_version = true;
1570
80
        }
1571
1.04k
    }
1572
812
    return !find_version;
1573
1.28k
}
1574
1575
13.2k
bool Tablet::_contains_rowset(const RowsetId rowset_id) {
1576
361k
    for (auto& version_rowset : _rs_version_map) {
1577
361k
        if (version_rowset.second->rowset_id() == rowset_id) {
1578
67
            return true;
1579
67
        }
1580
361k
    }
1581
13.2k
    for (auto& stale_version_rowset : _stale_rs_version_map) {
1582
894
        if (stale_version_rowset.second->rowset_id() == rowset_id) {
1583
242
            return true;
1584
242
        }
1585
894
    }
1586
12.9k
    return false;
1587
13.2k
}
1588
1589
// need check if consecutive version missing in full report
1590
// alter tablet will ignore this check
1591
void Tablet::build_tablet_report_info(TTabletInfo* tablet_info,
1592
                                      bool enable_consecutive_missing_check,
1593
1.30M
                                      bool enable_path_check) {
1594
1.30M
    std::shared_lock rdlock(_meta_lock);
1595
1.30M
    tablet_info->__set_tablet_id(_tablet_meta->tablet_id());
1596
1.30M
    tablet_info->__set_schema_hash(_tablet_meta->schema_hash());
1597
1.30M
    tablet_info->__set_row_count(_tablet_meta->num_rows());
1598
1.30M
    tablet_info->__set_data_size(_tablet_meta->tablet_local_size());
1599
1600
    // Here we need to report to FE if there are any missing versions of tablet.
1601
    // We start from the initial version and traverse backwards until we meet a discontinuous version.
1602
1.30M
    Version cversion;
1603
1.30M
    Version max_version;
1604
1.30M
    bool has_version_cross;
1605
1.30M
    _max_continuous_version_from_beginning_unlocked(&cversion, &max_version, &has_version_cross);
1606
    // cause publish version task runs concurrently, version may be flying
1607
    // so we add a consecutive miss check to solve this problem:
1608
    // if publish version 5 arrives but version 4 flying, we may judge replica miss version
1609
    // and set version miss in tablet_info, which makes fe treat this replica as unhealth
1610
    // and lead to other problems
1611
1.30M
    if (enable_consecutive_missing_check) {
1612
1.30M
        if (cversion.second < max_version.second) {
1613
0
            if (_last_missed_version == cversion.second + 1) {
1614
0
                if (MonotonicSeconds() - _last_missed_time_s >= 60) {
1615
                    // version missed for over 60 seconds
1616
0
                    tablet_info->__set_version_miss(true);
1617
0
                    _last_missed_version = -1;
1618
0
                    _last_missed_time_s = 0;
1619
0
                }
1620
0
            } else {
1621
0
                _last_missed_version = cversion.second + 1;
1622
0
                _last_missed_time_s = MonotonicSeconds();
1623
0
            }
1624
0
        }
1625
1.30M
    } else {
1626
0
        tablet_info->__set_version_miss(cversion.second < max_version.second);
1627
0
    }
1628
1629
1.30M
    DBUG_EXECUTE_IF("Tablet.build_tablet_report_info.version_miss", {
1630
1.30M
        auto tablet_id = dp->param<int64_t>("tablet_id", -1);
1631
1.30M
        if (tablet_id != -1 && tablet_id == _tablet_meta->tablet_id()) {
1632
1.30M
            auto miss = dp->param<bool>("version_miss", true);
1633
1.30M
            tablet_info->__set_version_miss(miss);
1634
1.30M
        }
1635
1.30M
    });
1636
1637
    // find rowset with max version
1638
1.30M
    auto iter = _rs_version_map.find(max_version);
1639
1.30M
    if (iter == _rs_version_map.end()) {
1640
        // If the tablet is in running state, it must not be doing schema-change. so if we can not
1641
        // access its rowsets, it means that the tablet is bad and needs to be reported to the FE
1642
        // for subsequent repairs (through the cloning task)
1643
0
        if (tablet_state() == TABLET_RUNNING) {
1644
0
            tablet_info->__set_used(false);
1645
0
        }
1646
        // For other states, FE knows that the tablet is in a certain change process, so here
1647
        // still sets the state to normal when reporting. Note that every task has an timeout,
1648
        // so if the task corresponding to this change hangs, when the task timeout, FE will know
1649
        // and perform state modification operations.
1650
0
    }
1651
1652
1.30M
    if (tablet_state() == TABLET_RUNNING) {
1653
1.30M
        if (has_version_cross || is_io_error_too_times() || !data_dir()->is_used()) {
1654
0
            LOG(INFO) << "report " << tablet_id() << " as bad, version_cross=" << has_version_cross
1655
0
                      << ", ioe times=" << get_io_error_times() << ", data_dir used "
1656
0
                      << data_dir()->is_used();
1657
0
            tablet_info->__set_used(false);
1658
0
        }
1659
1660
1.30M
        if (enable_path_check) {
1661
1.30M
            if (!_is_tablet_path_exists.exchange(true, std::memory_order_relaxed)) {
1662
0
                LOG(INFO) << "report " << tablet_id() << " as bad, tablet directory not found";
1663
0
                tablet_info->__set_used(false);
1664
0
            }
1665
1.30M
        }
1666
1.30M
    }
1667
1668
    // There are two cases when tablet state is TABLET_NOTREADY
1669
    // case 1: tablet is doing schema change. Fe knows it's state, doing nothing.
1670
    // case 2: tablet has finished schema change, but failed. Fe will perform recovery.
1671
1.30M
    if (tablet_state() == TABLET_NOTREADY && is_alter_failed()) {
1672
0
        tablet_info->__set_used(false);
1673
0
    }
1674
1675
1.30M
    if (tablet_state() == TABLET_SHUTDOWN) {
1676
0
        tablet_info->__set_used(false);
1677
0
    }
1678
1679
1.30M
    DBUG_EXECUTE_IF("Tablet.build_tablet_report_info.used", {
1680
1.30M
        auto tablet_id = dp->param<int64_t>("tablet_id", -1);
1681
1.30M
        if (tablet_id != -1 && tablet_id == _tablet_meta->tablet_id()) {
1682
1.30M
            auto used = dp->param<bool>("used", true);
1683
1.30M
            LOG_WARNING("Tablet.build_tablet_report_info.used")
1684
1.30M
                    .tag("tablet id", tablet_id)
1685
1.30M
                    .tag("used", used);
1686
1.30M
            tablet_info->__set_used(used);
1687
1.30M
        } else {
1688
1.30M
            LOG_WARNING("Tablet.build_tablet_report_info.used").tag("tablet id", tablet_id);
1689
1.30M
        }
1690
1.30M
    });
1691
1692
1.30M
    int64_t total_version_count = _tablet_meta->version_count();
1693
1694
    // For compatibility.
1695
    // For old fe, it wouldn't send visible version request to be, then be's visible version is always 0.
1696
    // Let visible_version_count set to total_version_count in be's report.
1697
1.30M
    int64_t visible_version_count = total_version_count;
1698
1.30M
    if (auto [visible_version, _] = get_visible_version_and_time(); visible_version > 0) {
1699
955k
        visible_version_count = _tablet_meta->version_count_cross_with_range({0, visible_version});
1700
955k
    }
1701
    // the report version is the largest continuous version, same logic as in FE side
1702
1.30M
    tablet_info->__set_version(cversion.second);
1703
    // Useless but it is a required filed in TTabletInfo
1704
1.30M
    tablet_info->__set_version_hash(0);
1705
1.30M
    tablet_info->__set_partition_id(_tablet_meta->partition_id());
1706
1.30M
    tablet_info->__set_storage_medium(_data_dir->storage_medium());
1707
1.30M
    tablet_info->__set_total_version_count(total_version_count);
1708
1.30M
    tablet_info->__set_visible_version_count(visible_version_count);
1709
1.30M
    tablet_info->__set_path_hash(_data_dir->path_hash());
1710
1.30M
    tablet_info->__set_is_in_memory(_tablet_meta->tablet_schema()->is_in_memory());
1711
1.30M
    tablet_info->__set_replica_id(replica_id());
1712
1.30M
    tablet_info->__set_remote_data_size(_tablet_meta->tablet_remote_size());
1713
1.30M
    if (_tablet_meta->cooldown_meta_id().initialized()) { // has cooldowned data
1714
8
        tablet_info->__set_cooldown_term(_cooldown_conf.term);
1715
8
        tablet_info->__set_cooldown_meta_id(_tablet_meta->cooldown_meta_id().to_thrift());
1716
8
    }
1717
1.30M
    if (tablet_state() == TABLET_RUNNING && _tablet_meta->storage_policy_id() > 0) {
1718
        // tablet may not have cooldowned data, but the storage policy is set
1719
1.28k
        tablet_info->__set_cooldown_term(_cooldown_conf.term);
1720
1.28k
    }
1721
1.30M
    tablet_info->__set_local_index_size(_tablet_meta->tablet_local_index_size());
1722
1.30M
    tablet_info->__set_local_segment_size(_tablet_meta->tablet_local_segment_size());
1723
1.30M
    tablet_info->__set_remote_index_size(_tablet_meta->tablet_remote_index_size());
1724
1.30M
    tablet_info->__set_remote_segment_size(_tablet_meta->tablet_remote_segment_size());
1725
1.30M
}
1726
1727
1
void Tablet::report_error(const Status& st) {
1728
1
    if (st.is<ErrorCode::IO_ERROR>()) {
1729
0
        ++_io_error_times;
1730
1
    } else if (st.is<ErrorCode::CORRUPTION>()) {
1731
1
        _io_error_times = config::max_tablet_io_errors + 1;
1732
1
    } else if (st.is<ErrorCode::NOT_FOUND>()) {
1733
0
        check_tablet_path_exists();
1734
0
        if (!_is_tablet_path_exists.load(std::memory_order_relaxed)) {
1735
0
            _io_error_times = config::max_tablet_io_errors + 1;
1736
0
        }
1737
0
    }
1738
1
}
1739
1740
Status Tablet::prepare_compaction_and_calculate_permits(
1741
        CompactionType compaction_type, const TabletSharedPtr& tablet,
1742
77.3k
        std::shared_ptr<CompactionMixin>& compaction, int64_t& permits) {
1743
77.3k
    if (compaction_type == CompactionType::CUMULATIVE_COMPACTION) {
1744
64.9k
        MonotonicStopWatch watch;
1745
64.9k
        watch.start();
1746
1747
64.9k
        compaction = std::make_shared<CumulativeCompaction>(tablet->_engine, tablet);
1748
64.9k
        DorisMetrics::instance()->cumulative_compaction_request_total->increment(1);
1749
64.9k
        Status res = compaction->prepare_compact();
1750
64.9k
        if (!config::disable_compaction_trace_log &&
1751
64.9k
            static_cast<double>(watch.elapsed_time()) / 1e9 >
1752
0
                    config::cumulative_compaction_trace_threshold) {
1753
0
            std::stringstream ss;
1754
0
            compaction->runtime_profile()->pretty_print(&ss);
1755
0
            LOG(WARNING) << "prepare cumulative compaction cost "
1756
0
                         << static_cast<double>(watch.elapsed_time()) / 1e9 << std::endl
1757
0
                         << ss.str();
1758
0
        }
1759
1760
64.9k
        if (!res.ok()) {
1761
64.8k
            permits = 0;
1762
            // if we meet a delete version, should increase the cumulative point to let base compaction handle the delete version.
1763
            // no need to wait 5s.
1764
64.8k
            if (!res.is<ErrorCode::CUMULATIVE_MEET_DELETE_VERSION>() ||
1765
64.8k
                config::enable_sleep_between_delete_cumu_compaction) {
1766
64.8k
                tablet->set_last_cumu_compaction_failure_time(UnixMillis());
1767
64.8k
            }
1768
64.8k
            if (!res.is<CUMULATIVE_NO_SUITABLE_VERSION>() &&
1769
64.8k
                !res.is<ErrorCode::CUMULATIVE_MEET_DELETE_VERSION>()) {
1770
0
                DorisMetrics::instance()->cumulative_compaction_request_failed->increment(1);
1771
0
                return Status::InternalError("prepare cumulative compaction with err: {}",
1772
0
                                             res.to_string());
1773
0
            }
1774
            // return OK if OLAP_ERR_CUMULATIVE_NO_SUITABLE_VERSION, so that we don't need to
1775
            // print too much useless logs.
1776
            // And because we set permits to 0, so even if we return OK here, nothing will be done.
1777
64.8k
            return Status::OK();
1778
64.8k
        }
1779
64.9k
    } else if (compaction_type == CompactionType::BASE_COMPACTION) {
1780
12.3k
        MonotonicStopWatch watch;
1781
12.3k
        watch.start();
1782
1783
12.3k
        compaction = std::make_shared<BaseCompaction>(tablet->_engine, tablet);
1784
12.3k
        DorisMetrics::instance()->base_compaction_request_total->increment(1);
1785
12.3k
        Status res = compaction->prepare_compact();
1786
12.3k
        if (!config::disable_compaction_trace_log &&
1787
12.3k
            static_cast<double>(watch.elapsed_time()) / 1e9 >
1788
0
                    config::base_compaction_trace_threshold) {
1789
0
            std::stringstream ss;
1790
0
            compaction->runtime_profile()->pretty_print(&ss);
1791
0
            LOG(WARNING) << "prepare base compaction cost "
1792
0
                         << static_cast<double>(watch.elapsed_time()) / 1e9 << std::endl
1793
0
                         << ss.str();
1794
0
        }
1795
1796
12.3k
        tablet->set_last_base_compaction_status(res.to_string());
1797
12.3k
        if (!res.ok()) {
1798
12.3k
            tablet->set_last_base_compaction_failure_time(UnixMillis());
1799
12.3k
            permits = 0;
1800
12.3k
            if (!res.is<BE_NO_SUITABLE_VERSION>()) {
1801
0
                DorisMetrics::instance()->base_compaction_request_failed->increment(1);
1802
0
                return Status::InternalError("prepare base compaction with err: {}",
1803
0
                                             res.to_string());
1804
0
            }
1805
            // return OK if OLAP_ERR_BE_NO_SUITABLE_VERSION, so that we don't need to
1806
            // print too much useless logs.
1807
            // And because we set permits to 0, so even if we return OK here, nothing will be done.
1808
12.3k
            return Status::OK();
1809
12.3k
        }
1810
12.3k
    } else {
1811
0
        DCHECK_EQ(compaction_type, CompactionType::FULL_COMPACTION);
1812
1813
0
        compaction = std::make_shared<FullCompaction>(tablet->_engine, tablet);
1814
0
        Status res = compaction->prepare_compact();
1815
0
        if (!res.ok()) {
1816
0
            tablet->set_last_full_compaction_failure_time(UnixMillis());
1817
0
            permits = 0;
1818
0
            if (!res.is<FULL_NO_SUITABLE_VERSION>()) {
1819
0
                return Status::InternalError("prepare full compaction with err: {}",
1820
0
                                             res.to_string());
1821
0
            }
1822
            // return OK if OLAP_ERR_BE_NO_SUITABLE_VERSION, so that we don't need to
1823
            // print too much useless logs.
1824
            // And because we set permits to 0, so even if we return OK here, nothing will be done.
1825
0
            return Status::OK();
1826
0
        }
1827
0
    }
1828
1829
    // Time series policy does not rely on permits, it uses goal size to control memory
1830
71
    if (tablet->tablet_meta()->compaction_policy() == CUMULATIVE_TIME_SERIES_POLICY) {
1831
        // permits = 0 means that prepare_compaction failed
1832
0
        permits = 1;
1833
71
    } else {
1834
71
        permits = compaction->get_compaction_permits();
1835
71
    }
1836
71
    return Status::OK();
1837
77.3k
}
1838
1839
0
void Tablet::execute_single_replica_compaction(SingleReplicaCompaction& compaction) {
1840
0
    Status res = compaction.execute_compact();
1841
0
    if (!res.ok()) {
1842
0
        set_last_failure_time(this, compaction, UnixMillis());
1843
0
        set_last_single_compaction_failure_status(res.to_string());
1844
0
        if (res.is<CANCELLED>()) {
1845
0
            DorisMetrics::instance()->single_compaction_request_cancelled->increment(1);
1846
            // "CANCELLED" indicates that the peer has not performed compaction,
1847
            // wait for the peer to perform compaction
1848
0
            set_skip_compaction(true, compaction.real_compact_type(), UnixSeconds());
1849
0
            VLOG_CRITICAL << "Cannel fetching from the remote peer. res=" << res
1850
0
                          << ", tablet=" << tablet_id();
1851
0
        } else {
1852
0
            DorisMetrics::instance()->single_compaction_request_failed->increment(1);
1853
0
            LOG(WARNING) << "failed to do single replica compaction. res=" << res
1854
0
                         << ", tablet=" << tablet_id();
1855
0
        }
1856
0
        return;
1857
0
    }
1858
0
    set_last_failure_time(this, compaction, 0);
1859
0
}
1860
1861
29.3M
bool Tablet::should_fetch_from_peer() {
1862
29.3M
    return tablet_meta()->tablet_schema()->enable_single_replica_compaction() &&
1863
29.3M
           _engine.should_fetch_from_peer(tablet_id());
1864
29.3M
}
1865
1866
3
std::vector<Version> Tablet::get_all_local_versions() {
1867
3
    std::vector<Version> local_versions;
1868
3
    {
1869
3
        std::shared_lock rlock(_meta_lock);
1870
63
        for (const auto& [version, rs] : _rs_version_map) {
1871
63
            if (rs->is_local()) {
1872
53
                local_versions.emplace_back(version);
1873
53
            }
1874
63
        }
1875
3
    }
1876
3
    std::sort(local_versions.begin(), local_versions.end(),
1877
324
              [](const Version& left, const Version& right) { return left.first < right.first; });
1878
3
    return local_versions;
1879
3
}
1880
1881
51
void Tablet::execute_compaction(CompactionMixin& compaction) {
1882
51
    signal::tablet_id = tablet_id();
1883
1884
51
    MonotonicStopWatch watch;
1885
51
    watch.start();
1886
1887
51
    Status res = [&]() { RETURN_IF_CATCH_EXCEPTION({ return compaction.execute_compact(); }); }();
1888
1889
51
    if (!res.ok()) [[unlikely]] {
1890
0
        set_last_failure_time(this, compaction, UnixMillis());
1891
0
        LOG(WARNING) << "failed to do " << compaction.compaction_name()
1892
0
                     << ", tablet=" << tablet_id() << " : " << res;
1893
51
    } else {
1894
51
        set_last_failure_time(this, compaction, 0);
1895
51
    }
1896
1897
51
    if (!config::disable_compaction_trace_log) {
1898
0
        auto need_trace = [&compaction, &watch] {
1899
0
            return compaction.compaction_type() == ReaderType::READER_CUMULATIVE_COMPACTION
1900
0
                           ? static_cast<double>(watch.elapsed_time()) / 1e9 >
1901
0
                                     config::cumulative_compaction_trace_threshold
1902
0
                   : compaction.compaction_type() == ReaderType::READER_BASE_COMPACTION
1903
0
                           ? static_cast<double>(watch.elapsed_time()) / 1e9 >
1904
0
                                     config::base_compaction_trace_threshold
1905
0
                           : false;
1906
0
        };
1907
0
        if (need_trace()) {
1908
0
            std::stringstream ss;
1909
0
            compaction.runtime_profile()->pretty_print(&ss);
1910
0
            LOG(WARNING) << "execute " << compaction.compaction_name() << " cost "
1911
0
                         << static_cast<double>(watch.elapsed_time()) / 1e9 << std::endl
1912
0
                         << ss.str();
1913
0
        }
1914
0
    }
1915
51
}
1916
1917
304
Status Tablet::create_initial_rowset(const int64_t req_version) {
1918
304
    if (req_version < 1) {
1919
0
        return Status::Error<CE_CMD_PARAMS_ERROR>(
1920
0
                "init version of tablet should at least 1. req.ver={}", req_version);
1921
0
    }
1922
304
    Version version(0, req_version);
1923
304
    RowsetSharedPtr new_rowset;
1924
    // there is no data in init rowset, so overlapping info is unknown.
1925
304
    RowsetWriterContext context;
1926
304
    context.version = version;
1927
304
    context.rowset_state = VISIBLE;
1928
304
    context.segments_overlap = OVERLAP_UNKNOWN;
1929
304
    context.tablet_schema = tablet_schema();
1930
304
    context.newest_write_timestamp = UnixSeconds();
1931
304
    auto rs_writer = DORIS_TRY(create_rowset_writer(context, false));
1932
304
    RETURN_IF_ERROR(rs_writer->flush());
1933
304
    RETURN_IF_ERROR(rs_writer->build(new_rowset));
1934
304
    RETURN_IF_ERROR(add_rowset(std::move(new_rowset)));
1935
304
    set_cumulative_layer_point(req_version + 1);
1936
304
    return Status::OK();
1937
304
}
1938
1939
Result<std::unique_ptr<RowsetWriter>> Tablet::create_rowset_writer(RowsetWriterContext& context,
1940
540
                                                                   bool vertical) {
1941
540
    context.rowset_id = _engine.next_rowset_id();
1942
540
    _init_context_common_fields(context);
1943
540
    return RowsetFactory::create_rowset_writer(_engine, context, vertical);
1944
540
}
1945
1946
// create a rowset writer with rowset_id and seg_id
1947
// after writer, merge this transient rowset with original rowset
1948
Result<std::unique_ptr<RowsetWriter>> Tablet::create_transient_rowset_writer(
1949
        const Rowset& rowset, std::shared_ptr<PartialUpdateInfo> partial_update_info,
1950
42
        int64_t txn_expiration) {
1951
42
    RowsetWriterContext context;
1952
42
    context.rowset_state = PREPARED;
1953
42
    context.segments_overlap = OVERLAPPING;
1954
42
    context.tablet_schema = std::make_shared<TabletSchema>();
1955
    // During a partial update, the extracted columns of a variant should not be included in the tablet schema.
1956
    // This is because the partial update for a variant needs to ignore the extracted columns.
1957
    // Otherwise, the schema types in different rowsets might be inconsistent. When performing a partial update,
1958
    // the complete variant is constructed by reading all the sub-columns of the variant.
1959
42
    context.tablet_schema = rowset.tablet_schema()->copy_without_variant_extracted_columns();
1960
42
    context.newest_write_timestamp = UnixSeconds();
1961
42
    context.tablet_id = table_id();
1962
42
    context.enable_segcompaction = false;
1963
    // ATTN: context.tablet is a shared_ptr, can't simply set it's value to `this`. We should
1964
    // get the shared_ptr from tablet_manager.
1965
42
    auto tablet = _engine.tablet_manager()->get_tablet(tablet_id());
1966
42
    if (!tablet) {
1967
0
        LOG(WARNING) << "cant find tablet by tablet_id=" << tablet_id();
1968
0
        return ResultError(Status::NotFound("cant find tablet by tablet_id={}", tablet_id()));
1969
0
    }
1970
42
    context.tablet = tablet;
1971
42
    context.write_type = DataWriteType::TYPE_DIRECT;
1972
42
    context.partial_update_info = std::move(partial_update_info);
1973
42
    context.is_transient_rowset_writer = true;
1974
42
    return create_transient_rowset_writer(context, rowset.rowset_id())
1975
42
            .transform([&](auto&& writer) {
1976
42
                writer->set_segment_start_id(cast_set<int32_t>(rowset.num_segments()));
1977
42
                return writer;
1978
42
            });
1979
42
}
1980
1981
Result<std::unique_ptr<RowsetWriter>> Tablet::create_transient_rowset_writer(
1982
42
        RowsetWriterContext& context, const RowsetId& rowset_id) {
1983
42
    context.rowset_id = rowset_id;
1984
42
    _init_context_common_fields(context);
1985
42
    return RowsetFactory::create_rowset_writer(_engine, context, false);
1986
42
}
1987
1988
582
void Tablet::_init_context_common_fields(RowsetWriterContext& context) {
1989
582
    context.tablet_uid = tablet_uid();
1990
582
    context.tablet_id = tablet_id();
1991
582
    context.partition_id = partition_id();
1992
582
    context.tablet_schema_hash = schema_hash();
1993
582
    context.rowset_type = tablet_meta()->preferred_rowset_type();
1994
    // Alpha Rowset will be removed in the future, so that if the tablet's default rowset type is
1995
    // alpha rowset, then set the newly created rowset to storage engine's default rowset.
1996
582
    if (context.rowset_type == ALPHA_ROWSET) {
1997
0
        context.rowset_type = _engine.default_rowset_type();
1998
0
    }
1999
2000
582
    if (context.is_local_rowset()) {
2001
582
        context.tablet_path = _tablet_path;
2002
582
    }
2003
2004
582
    context.data_dir = data_dir();
2005
582
    context.enable_unique_key_merge_on_write = enable_unique_key_merge_on_write();
2006
2007
582
    context.encrypt_algorithm = tablet_meta()->encryption_algorithm();
2008
582
}
2009
2010
621k
Status Tablet::create_rowset(const RowsetMetaSharedPtr& rowset_meta, RowsetSharedPtr* rowset) {
2011
621k
    return RowsetFactory::create_rowset(_tablet_meta->tablet_schema(),
2012
621k
                                        rowset_meta->is_local() ? _tablet_path : "", rowset_meta,
2013
621k
                                        rowset);
2014
621k
}
2015
2016
7
Status Tablet::cooldown(RowsetSharedPtr rowset) {
2017
7
    std::unique_lock schema_change_lock(_schema_change_lock, std::try_to_lock);
2018
7
    if (!schema_change_lock.owns_lock()) {
2019
0
        return Status::Error<TRY_LOCK_FAILED>(
2020
0
                "try schema_change_lock failed, schema change running or inverted index built "
2021
0
                "on "
2022
0
                "this tablet={}",
2023
0
                tablet_id());
2024
0
    }
2025
    // Check executing serially with compaction task.
2026
7
    std::unique_lock base_compaction_lock(_base_compaction_lock, std::try_to_lock);
2027
7
    if (!base_compaction_lock.owns_lock()) {
2028
0
        return Status::Error<TRY_LOCK_FAILED>("try base_compaction_lock failed");
2029
0
    }
2030
7
    std::unique_lock cumu_compaction_lock(_cumulative_compaction_lock, std::try_to_lock);
2031
7
    if (!cumu_compaction_lock.owns_lock()) {
2032
0
        return Status::Error<TRY_LOCK_FAILED>("try cumu_compaction_lock failed");
2033
0
    }
2034
7
    std::shared_lock cooldown_conf_rlock(_cooldown_conf_lock);
2035
7
    if (_cooldown_conf.cooldown_replica_id <= 0) { // wait for FE to push cooldown conf
2036
2
        return Status::InternalError("invalid cooldown_replica_id");
2037
2
    }
2038
2039
5
    auto mem_tracker = MemTrackerLimiter::create_shared(
2040
5
            MemTrackerLimiter::Type::OTHER,
2041
5
            fmt::format("Tablet::cooldown#tableId={}:replicaId={}", std::to_string(tablet_id()),
2042
5
                        std::to_string(replica_id())));
2043
5
    SCOPED_ATTACH_TASK(mem_tracker);
2044
2045
5
    if (_cooldown_conf.cooldown_replica_id == replica_id()) {
2046
        // this replica is cooldown replica
2047
5
        RETURN_IF_ERROR(_cooldown_data(std::move(rowset)));
2048
5
    } else {
2049
0
        Status st = _follow_cooldowned_data();
2050
0
        if (UNLIKELY(!st.ok())) {
2051
0
            _last_failed_follow_cooldown_time = time(nullptr);
2052
0
            return st;
2053
0
        }
2054
0
        _last_failed_follow_cooldown_time = 0;
2055
0
    }
2056
5
    return Status::OK();
2057
5
}
2058
2059
// hold SHARED `cooldown_conf_lock`
2060
5
Status Tablet::_cooldown_data(RowsetSharedPtr rowset) {
2061
5
    DCHECK(_cooldown_conf.cooldown_replica_id == replica_id());
2062
2063
5
    auto storage_resource = DORIS_TRY(get_resource_by_storage_policy_id(storage_policy_id()));
2064
5
    RowsetSharedPtr old_rowset = nullptr;
2065
2066
5
    if (rowset) {
2067
0
        const auto& rowset_id = rowset->rowset_id();
2068
0
        const auto& rowset_version = rowset->version();
2069
0
        std::shared_lock meta_rlock(_meta_lock);
2070
0
        auto iter = _rs_version_map.find(rowset_version);
2071
0
        if (iter != _rs_version_map.end() && iter->second->rowset_id() == rowset_id) {
2072
0
            old_rowset = rowset;
2073
0
        }
2074
0
    }
2075
2076
5
    if (!old_rowset) {
2077
5
        old_rowset = pick_cooldown_rowset();
2078
5
    }
2079
2080
5
    if (!old_rowset) {
2081
0
        LOG(INFO) << "cannot pick cooldown rowset in tablet " << tablet_id();
2082
0
        return Status::OK();
2083
0
    }
2084
2085
5
    RowsetId new_rowset_id = _engine.next_rowset_id();
2086
5
    auto pending_rs_guard = _engine.pending_remote_rowsets().add(new_rowset_id);
2087
5
    Status st;
2088
5
    Defer defer {[&] {
2089
5
        if (!st.ok()) {
2090
            // reclaim the incomplete rowset data in remote storage
2091
0
            record_unused_remote_rowset(new_rowset_id, storage_resource.fs->id(),
2092
0
                                        old_rowset->num_segments());
2093
0
        }
2094
5
    }};
2095
5
    auto start = std::chrono::steady_clock::now();
2096
5
    if (st = old_rowset->upload_to(storage_resource, new_rowset_id); !st.ok()) {
2097
0
        return st;
2098
0
    }
2099
2100
5
    auto duration = std::chrono::duration<double>(std::chrono::steady_clock::now() - start);
2101
5
    LOG(INFO) << "Upload rowset " << old_rowset->version() << " " << new_rowset_id.to_string()
2102
5
              << " to " << storage_resource.fs->root_path().native()
2103
5
              << ", tablet_id=" << tablet_id() << ", duration=" << duration.count()
2104
5
              << ", capacity=" << old_rowset->total_disk_size()
2105
5
              << ", tp=" << static_cast<double>(old_rowset->total_disk_size()) / duration.count()
2106
5
              << ", old rowset_id=" << old_rowset->rowset_id().to_string();
2107
2108
    // gen a new rowset
2109
5
    auto new_rowset_meta = std::make_shared<RowsetMeta>();
2110
5
    new_rowset_meta->init(old_rowset->rowset_meta().get());
2111
5
    new_rowset_meta->set_rowset_id(new_rowset_id);
2112
5
    new_rowset_meta->set_remote_storage_resource(std::move(storage_resource));
2113
5
    new_rowset_meta->set_creation_time(time(nullptr));
2114
5
    UniqueId cooldown_meta_id = UniqueId::gen_uid();
2115
5
    RowsetSharedPtr new_rowset;
2116
5
    RETURN_IF_ERROR(RowsetFactory::create_rowset(_tablet_meta->tablet_schema(), "", new_rowset_meta,
2117
5
                                                 &new_rowset));
2118
2119
5
    {
2120
5
        std::unique_lock meta_wlock(_meta_lock);
2121
5
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2122
5
        if (tablet_state() == TABLET_RUNNING) {
2123
5
            RETURN_IF_ERROR(delete_rowsets({std::move(old_rowset)}, false));
2124
5
            add_rowsets({std::move(new_rowset)});
2125
            // TODO(plat1ko): process primary key
2126
5
            _tablet_meta->set_cooldown_meta_id(cooldown_meta_id);
2127
5
        }
2128
5
    }
2129
5
    {
2130
5
        std::shared_lock meta_rlock(_meta_lock);
2131
5
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2132
5
        save_meta();
2133
5
    }
2134
    // Upload cooldowned rowset meta to remote fs
2135
    // ATTN: Even if it is an empty rowset, in order for the followers to synchronize, the coolown meta must be
2136
    // uploaded, otherwise followers may never completely cooldown.
2137
5
    if (auto t = _engine.tablet_manager()->get_tablet(tablet_id());
2138
5
        t != nullptr) { // `t` can be nullptr if it has been dropped
2139
5
        async_write_cooldown_meta(std::move(t));
2140
5
    }
2141
5
    return Status::OK();
2142
5
}
2143
2144
// hold SHARED `cooldown_conf_lock`
2145
Status Tablet::_read_cooldown_meta(const StorageResource& storage_resource,
2146
0
                                   TabletMetaPB* tablet_meta_pb) {
2147
0
    std::string remote_meta_path = storage_resource.cooldown_tablet_meta_path(
2148
0
            tablet_id(), _cooldown_conf.cooldown_replica_id, _cooldown_conf.term);
2149
0
    io::FileReaderSPtr tablet_meta_reader;
2150
0
    RETURN_IF_ERROR(storage_resource.fs->open_file(remote_meta_path, &tablet_meta_reader));
2151
0
    auto file_size = tablet_meta_reader->size();
2152
0
    size_t bytes_read;
2153
0
    auto buf = make_unique_buffer<uint8_t>(file_size);
2154
0
    RETURN_IF_ERROR(tablet_meta_reader->read_at(0, {buf.get(), file_size}, &bytes_read));
2155
0
    RETURN_IF_ERROR(tablet_meta_reader->close());
2156
0
    if (!tablet_meta_pb->ParseFromArray(buf.get(), cast_set<int>(file_size))) {
2157
0
        return Status::InternalError("malformed tablet meta, path={}/{}",
2158
0
                                     storage_resource.fs->root_path().native(), remote_meta_path);
2159
0
    }
2160
0
    return Status::OK();
2161
0
}
2162
2163
// `rs_metas` MUST already be sorted by `RowsetMeta::comparator`
2164
7
Status check_version_continuity(const std::vector<RowsetMetaSharedPtr>& rs_metas) {
2165
7
    if (rs_metas.size() < 2) {
2166
2
        return Status::OK();
2167
2
    }
2168
5
    auto prev = rs_metas.begin();
2169
11
    for (auto it = rs_metas.begin() + 1; it != rs_metas.end(); ++it) {
2170
6
        if ((*prev)->end_version() + 1 != (*it)->start_version()) {
2171
0
            return Status::InternalError("versions are not continuity: prev={} cur={}",
2172
0
                                         (*prev)->version().to_string(),
2173
0
                                         (*it)->version().to_string());
2174
0
        }
2175
6
        prev = it;
2176
6
    }
2177
5
    return Status::OK();
2178
5
}
2179
2180
// It's guaranteed the write cooldown meta task would be invoked at the end unless BE crashes
2181
// one tablet would at most have one async task to be done
2182
7
void Tablet::async_write_cooldown_meta(TabletSharedPtr tablet) {
2183
7
    ExecEnv::GetInstance()->write_cooldown_meta_executors()->submit(std::move(tablet));
2184
7
}
2185
2186
328
bool Tablet::update_cooldown_conf(int64_t cooldown_term, int64_t cooldown_replica_id) {
2187
328
    std::unique_lock wlock(_cooldown_conf_lock, std::try_to_lock);
2188
328
    if (!wlock.owns_lock()) {
2189
0
        LOG(INFO) << "try cooldown_conf_lock failed, tablet_id=" << tablet_id();
2190
0
        return false;
2191
0
    }
2192
328
    if (cooldown_term <= _cooldown_conf.term) {
2193
0
        return false;
2194
0
    }
2195
328
    LOG(INFO) << "update cooldown conf. tablet_id=" << tablet_id()
2196
328
              << " cooldown_replica_id: " << _cooldown_conf.cooldown_replica_id << " -> "
2197
328
              << cooldown_replica_id << ", cooldown_term: " << _cooldown_conf.term << " -> "
2198
328
              << cooldown_term;
2199
328
    _cooldown_conf.cooldown_replica_id = cooldown_replica_id;
2200
328
    _cooldown_conf.term = cooldown_term;
2201
328
    return true;
2202
328
}
2203
2204
116k
Status Tablet::write_cooldown_meta() {
2205
116k
    std::shared_lock rlock(_cooldown_conf_lock);
2206
116k
    if (_cooldown_conf.cooldown_replica_id != _tablet_meta->replica_id()) {
2207
0
        return Status::Aborted<false>("not cooldown replica({} vs {}) tablet_id={}",
2208
0
                                      _tablet_meta->replica_id(),
2209
0
                                      _cooldown_conf.cooldown_replica_id, tablet_id());
2210
0
    }
2211
2212
116k
    auto storage_resource = DORIS_TRY(get_resource_by_storage_policy_id(storage_policy_id()));
2213
2214
7
    std::vector<RowsetMetaSharedPtr> cooldowned_rs_metas;
2215
7
    UniqueId cooldown_meta_id;
2216
7
    {
2217
7
        std::shared_lock meta_rlock(_meta_lock);
2218
15
        for (const auto& [_, rs_meta] : _tablet_meta->all_rs_metas()) {
2219
15
            if (!rs_meta->is_local()) {
2220
13
                cooldowned_rs_metas.push_back(rs_meta);
2221
13
            }
2222
15
        }
2223
7
        cooldown_meta_id = _tablet_meta->cooldown_meta_id();
2224
7
    }
2225
7
    if (cooldowned_rs_metas.empty()) {
2226
0
        LOG(INFO) << "no cooldown meta to write, tablet_id=" << tablet_id();
2227
0
        return Status::OK();
2228
0
    }
2229
7
    std::sort(cooldowned_rs_metas.begin(), cooldowned_rs_metas.end(), RowsetMeta::comparator);
2230
7
    DCHECK(cooldowned_rs_metas.front()->start_version() == 0);
2231
    // If version not continuous, it must be a bug
2232
7
    if (auto st = check_version_continuity(cooldowned_rs_metas); !st.ok()) {
2233
0
        DCHECK(st.ok()) << st << " tablet_id=" << tablet_id();
2234
0
        st.set_code(ABORTED);
2235
0
        return st;
2236
0
    }
2237
2238
7
    TabletMetaPB tablet_meta_pb;
2239
7
    auto* rs_metas = tablet_meta_pb.mutable_rs_metas();
2240
7
    rs_metas->Reserve(cast_set<int>(cooldowned_rs_metas.size()));
2241
13
    for (auto& rs_meta : cooldowned_rs_metas) {
2242
13
        rs_metas->Add(rs_meta->get_rowset_pb());
2243
13
    }
2244
7
    tablet_meta_pb.mutable_cooldown_meta_id()->set_hi(cooldown_meta_id.hi);
2245
7
    tablet_meta_pb.mutable_cooldown_meta_id()->set_lo(cooldown_meta_id.lo);
2246
2247
7
    std::string remote_meta_path = storage_resource.cooldown_tablet_meta_path(
2248
7
            tablet_id(), _cooldown_conf.cooldown_replica_id, _cooldown_conf.term);
2249
7
    io::FileWriterPtr tablet_meta_writer;
2250
    // FIXME(plat1ko): What if object store permanently unavailable?
2251
7
    RETURN_IF_ERROR(storage_resource.fs->create_file(remote_meta_path, &tablet_meta_writer));
2252
7
    auto val = tablet_meta_pb.SerializeAsString();
2253
7
    RETURN_IF_ERROR(tablet_meta_writer->append({val.data(), val.size()}));
2254
7
    return tablet_meta_writer->close();
2255
7
}
2256
2257
// hold SHARED `cooldown_conf_lock`
2258
0
Status Tablet::_follow_cooldowned_data() {
2259
0
    DCHECK(_cooldown_conf.cooldown_replica_id != replica_id());
2260
0
    LOG(INFO) << "try to follow cooldowned data. tablet_id=" << tablet_id()
2261
0
              << " cooldown_replica_id=" << _cooldown_conf.cooldown_replica_id
2262
0
              << " local replica=" << replica_id();
2263
2264
0
    auto storage_resource = DORIS_TRY(get_resource_by_storage_policy_id(storage_policy_id()));
2265
    // MUST executing serially with cold data compaction, because compaction input rowsets may be deleted by this function
2266
0
    std::unique_lock cold_compaction_lock(_cold_compaction_lock, std::try_to_lock);
2267
0
    if (!cold_compaction_lock.owns_lock()) {
2268
0
        return Status::Error<TRY_LOCK_FAILED>("try cold_compaction_lock failed");
2269
0
    }
2270
2271
0
    TabletMetaPB cooldown_meta_pb;
2272
0
    auto st = _read_cooldown_meta(storage_resource, &cooldown_meta_pb);
2273
0
    if (!st.ok()) {
2274
0
        LOG(INFO) << "cannot read cooldown meta: " << st;
2275
0
        return Status::InternalError<false>("cannot read cooldown meta");
2276
0
    }
2277
2278
0
    if (cooldown_meta_pb.rs_metas_size() <= 0) {
2279
0
        LOG(WARNING)
2280
0
                << "Cooldown meta file exists but rs_metas is empty for tablet " << tablet_id()
2281
0
                << ". Cooldown meta id: " << cooldown_meta_pb.cooldown_meta_id()
2282
0
                << ". This may indicate a cooldown meta synchronization issue or an invalid file.";
2283
0
        return Status::InternalError<false>("Cooldown meta rs_metas is empty");
2284
0
    }
2285
2286
0
    if (_tablet_meta->cooldown_meta_id() == cooldown_meta_pb.cooldown_meta_id()) {
2287
        // cooldowned rowsets are same, no need to follow
2288
0
        return Status::OK();
2289
0
    }
2290
2291
0
    int64_t cooldowned_version = cooldown_meta_pb.rs_metas().rbegin()->end_version();
2292
2293
0
    std::vector<RowsetSharedPtr> overlap_rowsets;
2294
0
    bool version_aligned = false;
2295
2296
    // Holding these to delete rowsets' shared ptr until save meta can avoid trash sweeping thread
2297
    // deleting these rowsets' files before rowset meta has been removed from disk, which may cause
2298
    // data loss when BE reboot before save meta to disk.
2299
0
    std::vector<RowsetSharedPtr> to_delete;
2300
0
    std::vector<RowsetSharedPtr> to_add;
2301
2302
0
    {
2303
0
        std::lock_guard wlock(_meta_lock);
2304
0
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2305
0
        if (tablet_state() != TABLET_RUNNING) {
2306
0
            return Status::InternalError<false>("tablet not running");
2307
0
        }
2308
2309
0
        for (auto& [v, rs] : _rs_version_map) {
2310
0
            if (v.second <= cooldowned_version) {
2311
0
                overlap_rowsets.push_back(rs);
2312
0
                if (!version_aligned && v.second == cooldowned_version) {
2313
0
                    version_aligned = true;
2314
0
                }
2315
0
            } else if (!rs->is_local()) {
2316
0
                return Status::InternalError<false>(
2317
0
                        "cooldowned version larger than that to follow with cooldown version "
2318
0
                        "{}",
2319
0
                        cooldowned_version);
2320
0
            }
2321
0
        }
2322
2323
0
        if (!version_aligned) {
2324
0
            return Status::InternalError<false>("cooldowned version is not aligned with version {}",
2325
0
                                                cooldowned_version);
2326
0
        }
2327
2328
0
        std::sort(overlap_rowsets.begin(), overlap_rowsets.end(), Rowset::comparator);
2329
2330
        // Find different rowset in `overlap_rowsets` and `cooldown_meta_pb.rs_metas`
2331
0
        auto rs_pb_it = cooldown_meta_pb.rs_metas().begin();
2332
0
        auto rs_it = overlap_rowsets.begin();
2333
0
        for (; rs_pb_it != cooldown_meta_pb.rs_metas().end() && rs_it != overlap_rowsets.end();
2334
0
             ++rs_pb_it, ++rs_it) {
2335
0
            if (rs_pb_it->rowset_id_v2() != (*rs_it)->rowset_id().to_string()) {
2336
0
                break;
2337
0
            }
2338
0
        }
2339
2340
0
        to_delete.assign(rs_it, overlap_rowsets.end());
2341
0
        to_add.reserve(cooldown_meta_pb.rs_metas().end() - rs_pb_it);
2342
0
        for (; rs_pb_it != cooldown_meta_pb.rs_metas().end(); ++rs_pb_it) {
2343
0
            auto rs_meta = std::make_shared<RowsetMeta>();
2344
0
            rs_meta->init_from_pb(*rs_pb_it);
2345
0
            RowsetSharedPtr rs;
2346
0
            RETURN_IF_ERROR(
2347
0
                    RowsetFactory::create_rowset(_tablet_meta->tablet_schema(), "", rs_meta, &rs));
2348
0
            to_add.push_back(std::move(rs));
2349
0
        }
2350
        // Note: We CANNOT call `modify_rowsets` here because `modify_rowsets` cannot process version graph correctly.
2351
0
        RETURN_IF_ERROR(delete_rowsets(to_delete, false));
2352
0
        add_rowsets(to_add);
2353
        // TODO(plat1ko): process primary key
2354
0
        _tablet_meta->set_cooldown_meta_id(cooldown_meta_pb.cooldown_meta_id());
2355
0
    }
2356
2357
0
    {
2358
0
        std::lock_guard rlock(_meta_lock);
2359
0
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2360
0
        save_meta();
2361
0
    }
2362
2363
0
    if (!to_add.empty()) {
2364
0
        LOG(INFO) << "modify rowsets when follow cooldowned data, tablet_id=" << tablet_id()
2365
0
                  << [&]() {
2366
0
                         std::stringstream ss;
2367
0
                         ss << " delete rowsets:\n";
2368
0
                         for (auto&& rs : to_delete) {
2369
0
                             ss << rs->version() << ' ' << rs->rowset_id() << '\n';
2370
0
                         }
2371
0
                         ss << "add rowsets:\n";
2372
0
                         for (auto&& rs : to_add) {
2373
0
                             ss << rs->version() << ' ' << rs->rowset_id() << '\n';
2374
0
                         }
2375
0
                         return ss.str();
2376
0
                     }();
2377
0
    }
2378
2379
0
    return Status::OK();
2380
0
}
2381
2382
3.96M
bool Tablet::_has_data_to_cooldown() {
2383
3.96M
    int64_t min_local_version = std::numeric_limits<int64_t>::max();
2384
3.96M
    RowsetSharedPtr rowset;
2385
3.96M
    std::shared_lock meta_rlock(_meta_lock);
2386
    // Ususally once the tablet has done cooldown successfully then the first
2387
    // rowset would always be remote rowset
2388
3.96M
    bool has_cooldowned = false;
2389
7.09M
    for (const auto& [_, rs] : _rs_version_map) {
2390
7.09M
        if (!rs->is_local()) {
2391
27
            has_cooldowned = true;
2392
27
            break;
2393
27
        }
2394
7.09M
    }
2395
7.09M
    for (auto& [v, rs] : _rs_version_map) {
2396
7.09M
        auto predicate = rs->is_local() && v.first < min_local_version;
2397
7.09M
        if (!has_cooldowned) {
2398
7.09M
            predicate = predicate && (rs->data_disk_size() > 0);
2399
7.09M
        }
2400
7.09M
        if (predicate) {
2401
            // this is a local rowset and has data
2402
696k
            min_local_version = v.first;
2403
696k
            rowset = rs;
2404
696k
        }
2405
7.09M
    }
2406
2407
3.96M
    int64_t newest_cooldown_time = 0;
2408
3.96M
    if (rowset != nullptr) {
2409
634k
        newest_cooldown_time = _get_newest_cooldown_time(rowset);
2410
634k
    }
2411
2412
3.96M
    return (newest_cooldown_time != 0) && (newest_cooldown_time < UnixSeconds());
2413
3.96M
}
2414
2415
3.96M
RowsetSharedPtr Tablet::pick_cooldown_rowset() {
2416
3.96M
    RowsetSharedPtr rowset;
2417
2418
3.96M
    if (!_has_data_to_cooldown()) {
2419
3.96M
        return nullptr;
2420
3.96M
    }
2421
2422
    // TODO(plat1ko): should we maintain `cooldowned_version` in `Tablet`?
2423
8
    int64_t cooldowned_version = -1;
2424
    // We pick the rowset with smallest start version in local.
2425
8
    int64_t min_local_version = std::numeric_limits<int64_t>::max();
2426
8
    {
2427
8
        std::shared_lock meta_rlock(_meta_lock);
2428
26
        for (auto& [v, rs] : _rs_version_map) {
2429
26
            if (!rs->is_local()) {
2430
4
                cooldowned_version = std::max(cooldowned_version, v.second);
2431
22
            } else if (v.first < min_local_version) { // this is a local rowset
2432
10
                min_local_version = v.first;
2433
10
                rowset = rs;
2434
10
            }
2435
26
        }
2436
8
    }
2437
8
    if (!rowset) {
2438
0
        return nullptr;
2439
0
    }
2440
8
    if (tablet_footprint() == 0) {
2441
0
        VLOG_DEBUG << "skip cooldown due to empty tablet_id = " << tablet_id();
2442
0
        return nullptr;
2443
0
    }
2444
8
    if (min_local_version != cooldowned_version + 1) { // ensure version continuity
2445
0
        if (UNLIKELY(cooldowned_version != -1)) {
2446
0
            LOG(WARNING) << "version not continuous. tablet_id=" << tablet_id()
2447
0
                         << " cooldowned_version=" << cooldowned_version
2448
0
                         << " min_local_version=" << min_local_version;
2449
0
        }
2450
0
        return nullptr;
2451
0
    }
2452
8
    return rowset;
2453
8
}
2454
2455
634k
int64_t Tablet::_get_newest_cooldown_time(const RowsetSharedPtr& rowset) {
2456
634k
    int64_t id = storage_policy_id();
2457
634k
    if (id <= 0) {
2458
634k
        VLOG_DEBUG << "tablet does not need cooldown, tablet id: " << tablet_id();
2459
634k
        return 0;
2460
634k
    }
2461
12
    auto storage_policy = get_storage_policy(id);
2462
12
    if (!storage_policy) {
2463
0
        LOG(WARNING) << "Cannot get storage policy: " << id;
2464
0
        return 0;
2465
0
    }
2466
12
    auto cooldown_ttl_sec = storage_policy->cooldown_ttl;
2467
12
    auto cooldown_datetime = storage_policy->cooldown_datetime;
2468
12
    int64_t newest_cooldown_time = std::numeric_limits<int64_t>::max();
2469
2470
12
    if (cooldown_ttl_sec >= 0) {
2471
9
        newest_cooldown_time = rowset->newest_write_timestamp() + cooldown_ttl_sec;
2472
9
    }
2473
12
    if (cooldown_datetime > 0) {
2474
10
        newest_cooldown_time = std::min(newest_cooldown_time, cooldown_datetime);
2475
10
    }
2476
2477
12
    return newest_cooldown_time;
2478
12
}
2479
2480
3.96M
RowsetSharedPtr Tablet::need_cooldown(int64_t* cooldown_timestamp, size_t* file_size) {
2481
3.96M
    RowsetSharedPtr rowset = pick_cooldown_rowset();
2482
3.96M
    if (!rowset) {
2483
3.96M
        VLOG_DEBUG << "pick cooldown rowset, get null, tablet id: " << tablet_id();
2484
3.96M
        return nullptr;
2485
3.96M
    }
2486
2487
3
    auto newest_cooldown_time = _get_newest_cooldown_time(rowset);
2488
2489
    // the rowset should do cooldown job only if it's cooldown ttl plus newest write time is less than
2490
    // current time or it's datatime is less than current time
2491
3
    if (newest_cooldown_time != 0 && newest_cooldown_time < UnixSeconds()) {
2492
3
        *cooldown_timestamp = newest_cooldown_time;
2493
3
        *file_size = rowset->total_disk_size();
2494
3
        VLOG_DEBUG << "tablet need cooldown, tablet id: " << tablet_id()
2495
0
                   << " file_size: " << *file_size;
2496
3
        return rowset;
2497
3
    }
2498
2499
0
    VLOG_DEBUG << "tablet does not need cooldown, tablet id: " << tablet_id()
2500
0
               << " newest write time: " << rowset->newest_write_timestamp();
2501
0
    return nullptr;
2502
3
}
2503
2504
void Tablet::record_unused_remote_rowset(const RowsetId& rowset_id, const std::string& resource,
2505
0
                                         int64_t num_segments) {
2506
0
    auto gc_key = REMOTE_ROWSET_GC_PREFIX + rowset_id.to_string();
2507
0
    RemoteRowsetGcPB gc_pb;
2508
0
    gc_pb.set_resource_id(resource);
2509
0
    gc_pb.set_tablet_id(tablet_id());
2510
0
    gc_pb.set_num_segments(num_segments);
2511
0
    auto st =
2512
0
            _data_dir->get_meta()->put(META_COLUMN_FAMILY_INDEX, gc_key, gc_pb.SerializeAsString());
2513
0
    if (!st.ok()) {
2514
0
        LOG(WARNING) << "failed to record unused remote rowset. tablet_id=" << tablet_id()
2515
0
                     << " rowset_id=" << rowset_id << " resource_id=" << resource;
2516
0
    }
2517
0
    unused_remote_rowset_num << 1;
2518
0
}
2519
2520
8.32k
Status Tablet::remove_all_remote_rowsets() {
2521
8.32k
    DCHECK(tablet_state() == TABLET_SHUTDOWN);
2522
8.32k
    std::set<std::string> resource_ids;
2523
19.1k
    for (const auto& [_, rs_meta] : _tablet_meta->all_rs_metas()) {
2524
19.1k
        if (!rs_meta->is_local()) {
2525
0
            resource_ids.insert(rs_meta->resource_id());
2526
0
        }
2527
19.1k
    }
2528
8.32k
    if (resource_ids.empty()) {
2529
8.32k
        return Status::OK();
2530
8.32k
    }
2531
2
    auto tablet_gc_key = REMOTE_TABLET_GC_PREFIX + std::to_string(tablet_id());
2532
2
    RemoteTabletGcPB gc_pb;
2533
2
    for (auto& resource_id : resource_ids) {
2534
0
        gc_pb.add_resource_ids(resource_id);
2535
0
    }
2536
2
    return _data_dir->get_meta()->put(META_COLUMN_FAMILY_INDEX, tablet_gc_key,
2537
2
                                      gc_pb.SerializeAsString());
2538
8.32k
}
2539
2540
0
void Tablet::update_max_version_schema(const TabletSchemaSPtr& tablet_schema) {
2541
0
    std::lock_guard wrlock(_meta_lock);
2542
0
    SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2543
    // Double Check for concurrent update
2544
0
    if (!_max_version_schema ||
2545
0
        tablet_schema->schema_version() > _max_version_schema->schema_version()) {
2546
0
        _max_version_schema = tablet_schema;
2547
0
    }
2548
0
}
2549
2550
0
CalcDeleteBitmapExecutor* Tablet::calc_delete_bitmap_executor() {
2551
0
    return _engine.calc_delete_bitmap_executor();
2552
0
}
2553
2554
Status Tablet::save_delete_bitmap(const TabletTxnInfo* txn_info, int64_t txn_id,
2555
                                  DeleteBitmapPtr delete_bitmap, RowsetWriter* rowset_writer,
2556
                                  const RowsetIdUnorderedSet& cur_rowset_ids, int64_t lock_id,
2557
87
                                  int64_t next_visible_version) {
2558
87
    RowsetSharedPtr rowset = txn_info->rowset;
2559
87
    int64_t cur_version = rowset->start_version();
2560
2561
    // update version without write lock, compaction and publish_txn
2562
    // will update delete bitmap, handle compaction with _rowset_update_lock
2563
    // and publish_txn runs sequential so no need to lock here
2564
289
    for (auto& [key, bitmap] : delete_bitmap->delete_bitmap) {
2565
        // skip sentinel mark, which is used for delete bitmap correctness check
2566
289
        if (std::get<1>(key) != DeleteBitmap::INVALID_SEGMENT_ID) {
2567
38
            _tablet_meta->delete_bitmap().merge({std::get<0>(key), std::get<1>(key), cur_version},
2568
38
                                                bitmap);
2569
38
        }
2570
289
    }
2571
2572
87
    return Status::OK();
2573
87
}
2574
2575
22
void Tablet::merge_delete_bitmap(const DeleteBitmap& delete_bitmap) {
2576
22
    _tablet_meta->delete_bitmap().merge(delete_bitmap);
2577
22
}
2578
2579
0
bool Tablet::check_all_rowset_segment() {
2580
0
    std::shared_lock rdlock(_meta_lock);
2581
0
    for (auto& version_rowset : _rs_version_map) {
2582
0
        RowsetSharedPtr rowset = version_rowset.second;
2583
0
        if (!rowset->check_rowset_segment()) {
2584
0
            LOG(WARNING) << "Tablet Segment Check. find a bad tablet, tablet_id=" << tablet_id();
2585
0
            return false;
2586
0
        }
2587
0
    }
2588
0
    return true;
2589
0
}
2590
2591
13.4M
void Tablet::set_skip_compaction(bool skip, CompactionType compaction_type, int64_t start) {
2592
13.4M
    if (!skip) {
2593
0
        _skip_cumu_compaction = false;
2594
0
        _skip_base_compaction = false;
2595
0
        return;
2596
0
    }
2597
13.4M
    if (compaction_type == CompactionType::CUMULATIVE_COMPACTION) {
2598
7.59M
        _skip_cumu_compaction = true;
2599
7.59M
        _skip_cumu_compaction_ts = start;
2600
7.59M
    } else {
2601
5.84M
        DCHECK(compaction_type == CompactionType::BASE_COMPACTION);
2602
5.84M
        _skip_base_compaction = true;
2603
5.84M
        _skip_base_compaction_ts = start;
2604
5.84M
    }
2605
13.4M
}
2606
2607
318M
bool Tablet::should_skip_compaction(CompactionType compaction_type, int64_t now) {
2608
318M
    if (compaction_type == CompactionType::CUMULATIVE_COMPACTION && _skip_cumu_compaction &&
2609
318M
        now < _skip_cumu_compaction_ts + config::skip_tablet_compaction_second) {
2610
271M
        return true;
2611
271M
    } else if (compaction_type == CompactionType::BASE_COMPACTION && _skip_base_compaction &&
2612
47.4M
               now < _skip_base_compaction_ts + config::skip_tablet_compaction_second) {
2613
22.2M
        return true;
2614
22.2M
    }
2615
25.1M
    return false;
2616
318M
}
2617
2618
0
std::pair<std::string, int64_t> Tablet::get_binlog_info(std::string_view binlog_version) const {
2619
0
    return RowsetMetaManager::get_binlog_info(_data_dir->get_meta(), tablet_uid(), binlog_version);
2620
0
}
2621
2622
std::string Tablet::get_rowset_binlog_meta(std::string_view binlog_version,
2623
0
                                           std::string_view rowset_id) const {
2624
0
    return RowsetMetaManager::get_rowset_binlog_meta(_data_dir->get_meta(), tablet_uid(),
2625
0
                                                     binlog_version, rowset_id);
2626
0
}
2627
2628
Status Tablet::get_rowset_binlog_metas(const std::vector<int64_t>& binlog_versions,
2629
0
                                       RowsetBinlogMetasPB* metas_pb) {
2630
0
    return RowsetMetaManager::get_rowset_binlog_metas(_data_dir->get_meta(), tablet_uid(),
2631
0
                                                      binlog_versions, metas_pb);
2632
0
}
2633
2634
4
Status Tablet::get_rowset_binlog_metas(Version binlog_versions, RowsetBinlogMetasPB* metas_pb) {
2635
4
    return RowsetMetaManager::get_rowset_binlog_metas(_data_dir->get_meta(), tablet_uid(),
2636
4
                                                      binlog_versions, metas_pb);
2637
4
}
2638
2639
std::string Tablet::get_segment_filepath(std::string_view rowset_id,
2640
0
                                         std::string_view segment_index) const {
2641
0
    return fmt::format("{}/_binlog/{}_{}.dat", _tablet_path, rowset_id, segment_index);
2642
0
}
2643
2644
1
std::string Tablet::get_segment_filepath(std::string_view rowset_id, int64_t segment_index) const {
2645
1
    return fmt::format("{}/_binlog/{}_{}.dat", _tablet_path, rowset_id, segment_index);
2646
1
}
2647
2648
0
std::vector<std::string> Tablet::get_binlog_filepath(std::string_view binlog_version) const {
2649
0
    const auto& [rowset_id, num_segments] = get_binlog_info(binlog_version);
2650
0
    std::vector<std::string> binlog_filepath;
2651
0
    for (int i = 0; i < num_segments; ++i) {
2652
        // TODO(Drogon): rewrite by filesystem path
2653
0
        auto segment_file = fmt::format("{}_{}.dat", rowset_id, i);
2654
0
        binlog_filepath.emplace_back(fmt::format("{}/_binlog/{}", _tablet_path, segment_file));
2655
0
    }
2656
0
    return binlog_filepath;
2657
0
}
2658
2659
0
bool Tablet::can_add_binlog(uint64_t total_binlog_size) const {
2660
0
    return !_data_dir->reach_capacity_limit(total_binlog_size);
2661
0
}
2662
2663
110
bool Tablet::is_enable_binlog() {
2664
110
    return config::enable_feature_binlog && tablet_meta()->binlog_config().is_enable();
2665
110
}
2666
2667
0
void Tablet::set_binlog_config(BinlogConfig binlog_config) {
2668
0
    tablet_meta()->set_binlog_config(binlog_config);
2669
0
}
2670
2671
2
void Tablet::gc_binlogs(int64_t version) {
2672
2
    auto meta = _data_dir->get_meta();
2673
2
    DCHECK(meta != nullptr);
2674
2675
2
    const auto& tablet_uid = this->tablet_uid();
2676
2
    const auto tablet_id = this->tablet_id();
2677
2
    std::string begin_key = make_binlog_meta_key_prefix(tablet_uid);
2678
2
    std::string end_key = make_binlog_meta_key_prefix(tablet_uid, version + 1);
2679
2
    LOG(INFO) << fmt::format("gc binlog meta, tablet_id:{}, begin_key:{}, end_key:{}", tablet_id,
2680
2
                             begin_key, end_key);
2681
2682
2
    std::vector<std::string> wait_for_deleted_binlog_keys;
2683
2
    std::vector<std::string> wait_for_deleted_binlog_files;
2684
2
    auto add_to_wait_for_deleted = [&](std::string_view key, std::string_view rowset_id,
2685
2
                                       int64_t num_segments) {
2686
        // add binlog meta key and binlog data key
2687
1
        wait_for_deleted_binlog_keys.emplace_back(key);
2688
1
        wait_for_deleted_binlog_keys.push_back(get_binlog_data_key_from_meta_key(key));
2689
2690
        // add binlog segment files and index files
2691
2
        for (int64_t i = 0; i < num_segments; ++i) {
2692
1
            auto segment_file_path = get_segment_filepath(rowset_id, i);
2693
1
            wait_for_deleted_binlog_files.emplace_back(segment_file_path);
2694
2695
            // index files
2696
1
            if (tablet_schema()->has_inverted_index()) {
2697
1
                if (tablet_schema()->get_inverted_index_storage_format() ==
2698
1
                    InvertedIndexStorageFormatPB::V1) {
2699
0
                    for (const auto& index : tablet_schema()->inverted_indexes()) {
2700
0
                        auto index_file = InvertedIndexDescriptor::get_index_file_path_v1(
2701
0
                                InvertedIndexDescriptor::get_index_file_path_prefix(
2702
0
                                        segment_file_path),
2703
0
                                index->index_id(), index->get_index_suffix());
2704
0
                        wait_for_deleted_binlog_files.emplace_back(index_file);
2705
0
                    }
2706
1
                } else {
2707
1
                    auto index_file = InvertedIndexDescriptor::get_index_file_path_v2(
2708
1
                            InvertedIndexDescriptor::get_index_file_path_prefix(segment_file_path));
2709
1
                    wait_for_deleted_binlog_files.emplace_back(index_file);
2710
1
                }
2711
1
            }
2712
1
        }
2713
1
    };
2714
2715
2
    auto check_binlog_ttl = [&](std::string_view key, std::string_view value) mutable -> bool {
2716
2
        if (key >= end_key) {
2717
1
            return false;
2718
1
        }
2719
2720
1
        BinlogMetaEntryPB binlog_meta_entry_pb;
2721
1
        if (!binlog_meta_entry_pb.ParseFromArray(value.data(), cast_set<int>(value.size()))) {
2722
0
            LOG(WARNING) << "failed to parse binlog meta entry, key:" << key;
2723
0
            return true;
2724
0
        }
2725
2726
1
        auto num_segments = binlog_meta_entry_pb.num_segments();
2727
1
        std::string rowset_id;
2728
1
        if (binlog_meta_entry_pb.has_rowset_id_v2()) {
2729
1
            rowset_id = binlog_meta_entry_pb.rowset_id_v2();
2730
1
        } else {
2731
            // key is 'binlog_meta_6943f1585fe834b5-e542c2b83a21d0b7_00000000000000000069_020000000000000135449d7cd7eadfe672aa0f928fa99593', extract last part '020000000000000135449d7cd7eadfe672aa0f928fa99593'
2732
0
            auto pos = key.rfind('_');
2733
0
            if (pos == std::string::npos) {
2734
0
                LOG(WARNING) << fmt::format("invalid binlog meta key:{}", key);
2735
0
                return false;
2736
0
            }
2737
0
            rowset_id = key.substr(pos + 1);
2738
0
        }
2739
1
        add_to_wait_for_deleted(key, rowset_id, num_segments);
2740
2741
1
        return true;
2742
1
    };
2743
2744
2
    auto status = meta->iterate(META_COLUMN_FAMILY_INDEX, begin_key, check_binlog_ttl);
2745
2
    if (!status.ok()) {
2746
0
        LOG(WARNING) << "failed to iterate binlog meta, status:" << status;
2747
0
        return;
2748
0
    }
2749
2750
    // first remove binlog files, if failed, just break, then retry next time
2751
    // this keep binlog meta in meta store, so that binlog can be removed next time
2752
2
    bool remove_binlog_files_failed = false;
2753
2
    for (auto& file : wait_for_deleted_binlog_files) {
2754
2
        if (unlink(file.c_str()) != 0) {
2755
            // file not exist, continue
2756
0
            if (errno == ENOENT) {
2757
0
                continue;
2758
0
            }
2759
2760
0
            remove_binlog_files_failed = true;
2761
0
            LOG(WARNING) << "failed to remove binlog file:" << file << ", errno:" << errno;
2762
0
            break;
2763
0
        }
2764
2
    }
2765
2
    if (!remove_binlog_files_failed) {
2766
2
        static_cast<void>(meta->remove(META_COLUMN_FAMILY_INDEX, wait_for_deleted_binlog_keys));
2767
2
    }
2768
2
}
2769
2770
0
Status Tablet::ingest_binlog_metas(RowsetBinlogMetasPB* metas_pb) {
2771
0
    return RowsetMetaManager::ingest_binlog_metas(_data_dir->get_meta(), tablet_uid(), metas_pb);
2772
0
}
2773
2774
16.8k
void Tablet::clear_cache() {
2775
16.8k
    std::vector<RowsetSharedPtr> rowsets;
2776
16.8k
    {
2777
16.8k
        std::shared_lock rlock(get_header_lock());
2778
16.8k
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2779
2780
59.2k
        for (auto& [_, rowset] : rowset_map()) {
2781
59.2k
            rowsets.push_back(rowset);
2782
59.2k
        }
2783
16.8k
        for (auto& [_, rowset] : stale_rowset_map()) {
2784
640
            rowsets.push_back(rowset);
2785
640
        }
2786
16.8k
    }
2787
59.8k
    for (auto& rowset : rowsets) {
2788
59.8k
        rowset->clear_cache();
2789
59.8k
    }
2790
16.8k
}
2791
2792
11.4k
void Tablet::check_table_size_correctness() {
2793
11.4k
    if (!config::enable_table_size_correctness_check) {
2794
11.4k
        return;
2795
11.4k
    }
2796
0
    for (const auto& [_, rs_meta] : _tablet_meta->all_rs_metas()) {
2797
0
        int64_t total_segment_size = get_segment_file_size(rs_meta);
2798
0
        int64_t total_inverted_index_size = get_inverted_index_file_size(rs_meta);
2799
0
        if (rs_meta->data_disk_size() != total_segment_size ||
2800
0
            rs_meta->index_disk_size() != total_inverted_index_size ||
2801
0
            rs_meta->data_disk_size() + rs_meta->index_disk_size() != rs_meta->total_disk_size()) {
2802
0
            LOG(WARNING) << "[Local table table size check failed]:"
2803
0
                         << " tablet id: " << rs_meta->tablet_id()
2804
0
                         << ", rowset id:" << rs_meta->rowset_id()
2805
0
                         << ", rowset data disk size:" << rs_meta->data_disk_size()
2806
0
                         << ", rowset real data disk size:" << total_segment_size
2807
0
                         << ", rowset index disk size:" << rs_meta->index_disk_size()
2808
0
                         << ", rowset real index disk size:" << total_inverted_index_size
2809
0
                         << ", rowset total disk size:" << rs_meta->total_disk_size()
2810
0
                         << ", rowset segment path:"
2811
0
                         << StorageResource().remote_segment_path(
2812
0
                                    rs_meta->tablet_id(), rs_meta->rowset_id().to_string(), 0);
2813
0
            DCHECK(false);
2814
0
        }
2815
0
    }
2816
0
}
2817
2818
0
std::string Tablet::get_segment_path(const RowsetMetaSharedPtr& rs_meta, int64_t seg_id) {
2819
0
    std::string segment_path;
2820
0
    if (rs_meta->is_local()) {
2821
0
        segment_path = local_segment_path(_tablet_path, rs_meta->rowset_id().to_string(), seg_id);
2822
0
    } else {
2823
0
        segment_path = rs_meta->remote_storage_resource().value()->remote_segment_path(
2824
0
                rs_meta->tablet_id(), rs_meta->rowset_id().to_string(), seg_id);
2825
0
    }
2826
0
    return segment_path;
2827
0
}
2828
2829
0
int64_t Tablet::get_segment_file_size(const RowsetMetaSharedPtr& rs_meta) {
2830
0
    const auto& fs = rs_meta->fs();
2831
0
    if (!fs) {
2832
0
        LOG(WARNING) << "get fs failed, resource_id={}" << rs_meta->resource_id();
2833
0
    }
2834
0
    int64_t total_segment_size = 0;
2835
0
    for (int64_t seg_id = 0; seg_id < rs_meta->num_segments(); seg_id++) {
2836
0
        std::string segment_path = get_segment_path(rs_meta, seg_id);
2837
0
        int64_t segment_file_size = 0;
2838
0
        auto st = fs->file_size(segment_path, &segment_file_size);
2839
0
        if (!st.ok()) {
2840
0
            segment_file_size = 0;
2841
0
            LOG(WARNING) << "table size correctness check get segment size failed! msg:"
2842
0
                         << st.to_string() << ", segment path:" << segment_path;
2843
0
        }
2844
0
        total_segment_size += segment_file_size;
2845
0
    }
2846
0
    return total_segment_size;
2847
0
}
2848
2849
0
int64_t Tablet::get_inverted_index_file_size(const RowsetMetaSharedPtr& rs_meta) {
2850
0
    const auto& fs = rs_meta->fs();
2851
0
    if (!fs) {
2852
0
        LOG(WARNING) << "get fs failed, resource_id={}" << rs_meta->resource_id();
2853
0
    }
2854
0
    int64_t total_inverted_index_size = 0;
2855
2856
0
    if (rs_meta->tablet_schema()->get_inverted_index_storage_format() ==
2857
0
        InvertedIndexStorageFormatPB::V1) {
2858
0
        const auto& indices = rs_meta->tablet_schema()->inverted_indexes();
2859
0
        for (auto& index : indices) {
2860
0
            for (int seg_id = 0; seg_id < rs_meta->num_segments(); ++seg_id) {
2861
0
                std::string segment_path = get_segment_path(rs_meta, seg_id);
2862
0
                int64_t file_size = 0;
2863
2864
0
                std::string inverted_index_file_path =
2865
0
                        InvertedIndexDescriptor::get_index_file_path_v1(
2866
0
                                InvertedIndexDescriptor::get_index_file_path_prefix(segment_path),
2867
0
                                index->index_id(), index->get_index_suffix());
2868
0
                auto st = fs->file_size(inverted_index_file_path, &file_size);
2869
0
                if (!st.ok()) {
2870
0
                    file_size = 0;
2871
0
                    LOG(WARNING) << " tablet id: " << get_tablet_info().tablet_id
2872
0
                                 << ", rowset id:" << rs_meta->rowset_id()
2873
0
                                 << ", table size correctness check get inverted index v1 "
2874
0
                                    "size failed! msg:"
2875
0
                                 << st.to_string()
2876
0
                                 << ", inverted index path:" << inverted_index_file_path;
2877
0
                }
2878
0
                total_inverted_index_size += file_size;
2879
0
            }
2880
0
        }
2881
0
    } else {
2882
0
        for (int seg_id = 0; seg_id < rs_meta->num_segments(); ++seg_id) {
2883
0
            int64_t file_size = 0;
2884
0
            std::string segment_path = get_segment_path(rs_meta, seg_id);
2885
0
            std::string inverted_index_file_path = InvertedIndexDescriptor::get_index_file_path_v2(
2886
0
                    InvertedIndexDescriptor::get_index_file_path_prefix(segment_path));
2887
0
            auto st = fs->file_size(inverted_index_file_path, &file_size);
2888
0
            if (!st.ok()) {
2889
0
                file_size = 0;
2890
0
                if (st.is<NOT_FOUND>()) {
2891
0
                    LOG(INFO) << " tablet id: " << get_tablet_info().tablet_id
2892
0
                              << ", rowset id:" << rs_meta->rowset_id()
2893
0
                              << ", table size correctness check get inverted index v2 failed "
2894
0
                                 "because file not exist:"
2895
0
                              << inverted_index_file_path;
2896
0
                } else {
2897
0
                    LOG(WARNING) << " tablet id: " << get_tablet_info().tablet_id
2898
0
                                 << ", rowset id:" << rs_meta->rowset_id()
2899
0
                                 << ", table size correctness check get inverted index v2 "
2900
0
                                    "size failed! msg:"
2901
0
                                 << st.to_string()
2902
0
                                 << ", inverted index path:" << inverted_index_file_path;
2903
0
                }
2904
0
            }
2905
0
            total_inverted_index_size += file_size;
2906
0
        }
2907
0
    }
2908
0
    return total_inverted_index_size;
2909
0
}
2910
2911
Status Tablet::prepare_txn(TPartitionId partition_id, TTransactionId transaction_id,
2912
124
                           const PUniqueId& load_id, bool ingest) {
2913
124
    std::shared_lock base_migration_lock(get_migration_lock(), std::defer_lock);
2914
2915
    // TODO: rename debugpoint.
2916
124
    DBUG_EXECUTE_IF("PushHandler::_do_streaming_ingestion.try_lock_fail", {
2917
124
        return Status::ObtainLockFailed("PushHandler::_do_streaming_ingestion get lock failed");
2918
124
    })
2919
2920
124
    if (!base_migration_lock.try_lock_for(
2921
124
                std::chrono::milliseconds(config::migration_lock_timeout_ms))) {
2922
0
        return Status::ObtainLockFailed("try_lock migration lock failed after {}ms",
2923
0
                                        config::migration_lock_timeout_ms);
2924
0
    }
2925
2926
124
    std::lock_guard<std::mutex> push_lock(get_push_lock());
2927
124
    return _engine.txn_manager()->prepare_txn(partition_id, transaction_id, tablet_id(),
2928
124
                                              tablet_uid(), load_id, ingest);
2929
124
}
2930
2931
#include "common/compile_check_end.h"
2932
} // namespace doris