Coverage Report

Created: 2024-11-20 12:06

/root/doris/be/src/olap/tablet.cpp
Line
Count
Source (jump to first uncovered line)
1
// Licensed to the Apache Software Foundation (ASF) under one
2
// or more contributor license agreements.  See the NOTICE file
3
// distributed with this work for additional information
4
// regarding copyright ownership.  The ASF licenses this file
5
// to you under the Apache License, Version 2.0 (the
6
// "License"); you may not use this file except in compliance
7
// with the License.  You may obtain a copy of the License at
8
//
9
//   http://www.apache.org/licenses/LICENSE-2.0
10
//
11
// Unless required by applicable law or agreed to in writing,
12
// software distributed under the License is distributed on an
13
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
14
// KIND, either express or implied.  See the License for the
15
// specific language governing permissions and limitations
16
// under the License.
17
18
#include "olap/tablet.h"
19
20
#include <butil/logging.h>
21
#include <bvar/reducer.h>
22
#include <bvar/window.h>
23
#include <fmt/format.h>
24
#include <gen_cpp/FrontendService_types.h>
25
#include <gen_cpp/MasterService_types.h>
26
#include <gen_cpp/Metrics_types.h>
27
#include <gen_cpp/olap_file.pb.h>
28
#include <gen_cpp/types.pb.h>
29
#include <glog/logging.h>
30
#include <rapidjson/document.h>
31
#include <rapidjson/encodings.h>
32
#include <rapidjson/prettywriter.h>
33
#include <rapidjson/rapidjson.h>
34
#include <rapidjson/stringbuffer.h>
35
36
#include <algorithm>
37
#include <atomic>
38
#include <boost/container/detail/std_fwd.hpp>
39
#include <cstdint>
40
#include <roaring/roaring.hh>
41
42
#include "common/compiler_util.h" // IWYU pragma: keep
43
// IWYU pragma: no_include <bits/chrono.h>
44
#include <chrono> // IWYU pragma: keep
45
#include <filesystem>
46
#include <iterator>
47
#include <limits>
48
#include <map>
49
#include <memory>
50
#include <mutex>
51
#include <set>
52
#include <shared_mutex>
53
#include <string>
54
#include <tuple>
55
#include <type_traits>
56
#include <unordered_map>
57
#include <unordered_set>
58
59
#include "agent/utils.h"
60
#include "common/config.h"
61
#include "common/consts.h"
62
#include "common/logging.h"
63
#include "common/signal_handler.h"
64
#include "common/status.h"
65
#include "exprs/runtime_filter.h"
66
#include "gutil/ref_counted.h"
67
#include "gutil/strings/substitute.h"
68
#include "io/fs/file_reader.h"
69
#include "io/fs/file_reader_writer_fwd.h"
70
#include "io/fs/file_system.h"
71
#include "io/fs/file_writer.h"
72
#include "io/fs/path.h"
73
#include "io/fs/remote_file_system.h"
74
#include "io/io_common.h"
75
#include "olap/base_compaction.h"
76
#include "olap/base_tablet.h"
77
#include "olap/binlog.h"
78
#include "olap/cumulative_compaction.h"
79
#include "olap/cumulative_compaction_policy.h"
80
#include "olap/cumulative_compaction_time_series_policy.h"
81
#include "olap/delete_bitmap_calculator.h"
82
#include "olap/full_compaction.h"
83
#include "olap/memtable.h"
84
#include "olap/olap_common.h"
85
#include "olap/olap_define.h"
86
#include "olap/olap_meta.h"
87
#include "olap/primary_key_index.h"
88
#include "olap/rowset/beta_rowset.h"
89
#include "olap/rowset/rowset.h"
90
#include "olap/rowset/rowset_factory.h"
91
#include "olap/rowset/rowset_fwd.h"
92
#include "olap/rowset/rowset_meta.h"
93
#include "olap/rowset/rowset_meta_manager.h"
94
#include "olap/rowset/rowset_writer.h"
95
#include "olap/rowset/rowset_writer_context.h"
96
#include "olap/rowset/segment_v2/column_reader.h"
97
#include "olap/rowset/segment_v2/common.h"
98
#include "olap/rowset/segment_v2/indexed_column_reader.h"
99
#include "olap/rowset/vertical_beta_rowset_writer.h"
100
#include "olap/schema_change.h"
101
#include "olap/single_replica_compaction.h"
102
#include "olap/storage_engine.h"
103
#include "olap/storage_policy.h"
104
#include "olap/tablet_manager.h"
105
#include "olap/tablet_meta.h"
106
#include "olap/tablet_meta_manager.h"
107
#include "olap/tablet_schema.h"
108
#include "olap/txn_manager.h"
109
#include "olap/types.h"
110
#include "olap/utils.h"
111
#include "segment_loader.h"
112
#include "service/point_query_executor.h"
113
#include "tablet.h"
114
#include "util/bvar_helper.h"
115
#include "util/debug_points.h"
116
#include "util/defer_op.h"
117
#include "util/doris_metrics.h"
118
#include "util/pretty_printer.h"
119
#include "util/scoped_cleanup.h"
120
#include "util/stopwatch.hpp"
121
#include "util/threadpool.h"
122
#include "util/time.h"
123
#include "util/trace.h"
124
#include "util/uid_util.h"
125
#include "util/work_thread_pool.hpp"
126
#include "vec/columns/column.h"
127
#include "vec/columns/column_string.h"
128
#include "vec/common/schema_util.h"
129
#include "vec/common/string_ref.h"
130
#include "vec/data_types/data_type.h"
131
#include "vec/data_types/data_type_factory.hpp"
132
#include "vec/data_types/serde/data_type_serde.h"
133
#include "vec/jsonb/serialize.h"
134
135
namespace doris {
136
class TupleDescriptor;
137
138
namespace vectorized {
139
class Block;
140
} // namespace vectorized
141
142
using namespace ErrorCode;
143
using namespace std::chrono_literals;
144
145
using std::pair;
146
using std::string;
147
using std::vector;
148
using io::FileSystemSPtr;
149
150
namespace {
151
152
bvar::Adder<uint64_t> exceed_version_limit_counter;
153
bvar::Window<bvar::Adder<uint64_t>> exceed_version_limit_counter_minute(
154
        &exceed_version_limit_counter, 60);
155
bvar::Adder<uint64_t> cooldown_pending_task("cooldown_pending_task");
156
bvar::Adder<uint64_t> cooldown_processing_task("cooldown_processing_task");
157
158
0
void set_last_failure_time(Tablet* tablet, const Compaction& compaction, int64_t ms) {
159
0
    switch (compaction.compaction_type()) {
160
0
    case ReaderType::READER_CUMULATIVE_COMPACTION:
161
0
        tablet->set_last_cumu_compaction_failure_time(ms);
162
0
        return;
163
0
    case ReaderType::READER_BASE_COMPACTION:
164
0
        tablet->set_last_base_compaction_failure_time(ms);
165
0
        return;
166
0
    case ReaderType::READER_FULL_COMPACTION:
167
0
        tablet->set_last_full_compaction_failure_time(ms);
168
0
        return;
169
0
    default:
170
0
        LOG(FATAL) << "invalid compaction type " << compaction.compaction_name()
171
0
                   << " tablet_id: " << tablet->tablet_id();
172
0
    }
173
0
};
174
175
} // namespace
176
177
bvar::Adder<uint64_t> unused_remote_rowset_num("unused_remote_rowset_num");
178
179
WriteCooldownMetaExecutors::WriteCooldownMetaExecutors(size_t executor_nums)
180
1
        : _executor_nums(executor_nums) {
181
6
    for (size_t i = 0; i < _executor_nums; i++) {
182
5
        std::unique_ptr<PriorityThreadPool> pool;
183
5
        static_cast<void>(ThreadPoolBuilder("WriteCooldownMetaExecutor")
184
5
                                  .set_min_threads(1)
185
5
                                  .set_max_threads(1)
186
5
                                  .set_max_queue_size(std::numeric_limits<int>::max())
187
5
                                  .build(&pool));
188
5
        _executors.emplace_back(std::move(pool));
189
5
    }
190
1
}
191
192
0
void WriteCooldownMetaExecutors::stop() {
193
0
    for (auto& pool_ptr : _executors) {
194
0
        if (pool_ptr) {
195
0
            pool_ptr->shutdown();
196
0
        }
197
0
    }
198
0
}
199
200
5
void WriteCooldownMetaExecutors::WriteCooldownMetaExecutors::submit(TabletSharedPtr tablet) {
201
5
    auto tablet_id = tablet->tablet_id();
202
203
5
    {
204
5
        std::shared_lock rdlock(tablet->get_header_lock());
205
5
        if (!tablet->tablet_meta()->cooldown_meta_id().initialized()) {
206
0
            VLOG_NOTICE << "tablet " << tablet_id << " is not cooldown replica";
207
0
            return;
208
0
        }
209
5
        if (tablet->tablet_state() == TABLET_SHUTDOWN) [[unlikely]] {
210
0
            LOG_INFO("tablet {} has been dropped, don't do cooldown", tablet_id);
211
0
            return;
212
0
        }
213
5
    }
214
5
    {
215
        // one tablet could at most have one cooldown task to be done
216
5
        std::unique_lock<std::mutex> lck {_latch};
217
5
        if (_pending_tablets.count(tablet_id) > 0) {
218
1
            return;
219
1
        }
220
4
        _pending_tablets.insert(tablet_id);
221
4
    }
222
223
4
    auto async_write_task = [this, t = std::move(tablet)]() {
224
4
        {
225
4
            std::unique_lock<std::mutex> lck {_latch};
226
4
            _pending_tablets.erase(t->tablet_id());
227
4
        }
228
4
        auto s = t->write_cooldown_meta();
229
4
        if (s.ok()) {
230
4
            return;
231
4
        }
232
0
        if (!s.is<ABORTED>()) {
233
0
            LOG_EVERY_SECOND(WARNING)
234
0
                    << "write tablet " << t->tablet_id() << " cooldown meta failed because: " << s;
235
0
            submit(t);
236
0
            return;
237
0
        }
238
0
        VLOG_DEBUG << "tablet " << t->tablet_id() << " is not cooldown replica";
239
0
    };
240
241
4
    cooldown_pending_task << 1;
242
4
    _executors[_get_executor_pos(tablet_id)]->offer([task = std::move(async_write_task)]() {
243
4
        cooldown_pending_task << -1;
244
4
        cooldown_processing_task << 1;
245
4
        task();
246
4
        cooldown_processing_task << -1;
247
4
    });
248
4
}
249
250
Tablet::Tablet(StorageEngine& engine, TabletMetaSharedPtr tablet_meta, DataDir* data_dir,
251
               const std::string_view& cumulative_compaction_type)
252
        : BaseTablet(std::move(tablet_meta)),
253
          _engine(engine),
254
          _data_dir(data_dir),
255
          _is_bad(false),
256
          _last_cumu_compaction_failure_millis(0),
257
          _last_base_compaction_failure_millis(0),
258
          _last_full_compaction_failure_millis(0),
259
          _last_cumu_compaction_success_millis(0),
260
          _last_base_compaction_success_millis(0),
261
          _last_full_compaction_success_millis(0),
262
          _cumulative_point(K_INVALID_CUMULATIVE_POINT),
263
          _newly_created_rowset_num(0),
264
          _last_checkpoint_time(0),
265
          _cumulative_compaction_type(cumulative_compaction_type),
266
          _is_tablet_path_exists(true),
267
          _last_missed_version(-1),
268
275
          _last_missed_time_s(0) {
269
275
    if (_data_dir != nullptr) {
270
139
        _tablet_path = fmt::format("{}/{}/{}/{}/{}", _data_dir->path(), DATA_PREFIX,
271
139
                                   _tablet_meta->shard_id(), tablet_id(), schema_hash());
272
139
    }
273
275
}
274
275
0
bool Tablet::set_tablet_schema_into_rowset_meta() {
276
0
    bool flag = false;
277
0
    for (auto&& rowset_meta : _tablet_meta->all_mutable_rs_metas()) {
278
0
        if (!rowset_meta->tablet_schema()) {
279
0
            rowset_meta->set_tablet_schema(_tablet_meta->tablet_schema());
280
0
            flag = true;
281
0
        }
282
0
    }
283
0
    return flag;
284
0
}
285
286
238
Status Tablet::_init_once_action() {
287
238
    Status res = Status::OK();
288
238
    VLOG_NOTICE << "begin to load tablet. tablet=" << tablet_id()
289
0
                << ", version_size=" << _tablet_meta->version_count();
290
291
238
#ifdef BE_TEST
292
    // init cumulative compaction policy by type
293
238
    _cumulative_compaction_policy =
294
238
            CumulativeCompactionPolicyFactory::create_cumulative_compaction_policy(
295
238
                    _tablet_meta->compaction_policy());
296
238
#endif
297
298
238
    for (const auto& rs_meta : _tablet_meta->all_rs_metas()) {
299
197
        Version version = rs_meta->version();
300
197
        RowsetSharedPtr rowset;
301
197
        res = create_rowset(rs_meta, &rowset);
302
197
        if (!res.ok()) {
303
0
            LOG(WARNING) << "fail to init rowset. tablet_id=" << tablet_id()
304
0
                         << ", schema_hash=" << schema_hash() << ", version=" << version
305
0
                         << ", res=" << res;
306
0
            return res;
307
0
        }
308
197
        _rs_version_map[version] = std::move(rowset);
309
197
    }
310
311
    // init stale rowset
312
238
    for (const auto& stale_rs_meta : _tablet_meta->all_stale_rs_metas()) {
313
0
        Version version = stale_rs_meta->version();
314
0
        RowsetSharedPtr rowset;
315
0
        res = create_rowset(stale_rs_meta, &rowset);
316
0
        if (!res.ok()) {
317
0
            LOG(WARNING) << "fail to init stale rowset. tablet_id:" << tablet_id()
318
0
                         << ", schema_hash:" << schema_hash() << ", version=" << version
319
0
                         << ", res:" << res;
320
0
            return res;
321
0
        }
322
0
        _stale_rs_version_map[version] = std::move(rowset);
323
0
    }
324
325
238
    return res;
326
238
}
327
328
259
Status Tablet::init() {
329
259
    return _init_once.call([this] { return _init_once_action(); });
330
259
}
331
332
// should save tablet meta to remote meta store
333
// if it's a primary replica
334
145
void Tablet::save_meta() {
335
145
    check_table_size_correctness();
336
145
    auto res = _tablet_meta->save_meta(_data_dir);
337
145
    CHECK_EQ(res, Status::OK()) << "fail to save tablet_meta. res=" << res
338
0
                                << ", root=" << _data_dir->path();
339
145
}
340
341
// Caller should hold _meta_lock.
342
Status Tablet::revise_tablet_meta(const std::vector<RowsetSharedPtr>& to_add,
343
                                  const std::vector<RowsetSharedPtr>& to_delete,
344
0
                                  bool is_incremental_clone) {
345
0
    LOG(INFO) << "begin to revise tablet. tablet_id=" << tablet_id();
346
    // 1. for incremental clone, we have to add the rowsets first to make it easy to compute
347
    //    all the delete bitmaps, and it's easy to delete them if we end up with a failure
348
    // 2. for full clone, we can calculate delete bitmaps on the cloned rowsets directly.
349
0
    if (is_incremental_clone) {
350
0
        CHECK(to_delete.empty()); // don't need to delete rowsets
351
0
        add_rowsets(to_add);
352
        // reconstruct from tablet meta
353
0
        _timestamped_version_tracker.construct_versioned_tracker(_tablet_meta->all_rs_metas());
354
0
    }
355
356
0
    Status calc_bm_status;
357
0
    std::vector<RowsetSharedPtr> base_rowsets_for_full_clone = to_add; // copy vector
358
0
    while (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write()) {
359
0
        std::vector<RowsetSharedPtr> calc_delete_bitmap_rowsets;
360
0
        int64_t to_add_min_version = INT64_MAX;
361
0
        int64_t to_add_max_version = INT64_MIN;
362
0
        for (auto& rs : to_add) {
363
0
            if (to_add_min_version > rs->start_version()) {
364
0
                to_add_min_version = rs->start_version();
365
0
            }
366
0
            if (to_add_max_version < rs->end_version()) {
367
0
                to_add_max_version = rs->end_version();
368
0
            }
369
0
        }
370
0
        Version calc_delete_bitmap_ver;
371
0
        if (is_incremental_clone) {
372
            // From the rowset of to_add with smallest version, all other rowsets
373
            // need to recalculate the delete bitmap
374
            // For example:
375
            // local tablet: [0-1] [2-5] [6-6] [9-10]
376
            // clone tablet: [7-7] [8-8]
377
            // new tablet:   [0-1] [2-5] [6-6] [7-7] [8-8] [9-10]
378
            // [7-7] [8-8] [9-10] need to recalculate delete bitmap
379
0
            calc_delete_bitmap_ver = Version(to_add_min_version, max_version_unlocked());
380
0
        } else {
381
            // the delete bitmap of to_add's rowsets has clone from remote when full clone.
382
            // only other rowsets in local need to recalculate the delete bitmap.
383
            // For example:
384
            // local tablet: [0-1]x [2-5]x [6-6]x [7-7]x [9-10]
385
            // clone tablet: [0-1]  [2-4]  [5-6]  [7-8]
386
            // new tablet:   [0-1]  [2-4]  [5-6]  [7-8] [9-10]
387
            // only [9-10] need to recalculate delete bitmap
388
0
            CHECK_EQ(to_add_min_version, 0) << "to_add_min_version is: " << to_add_min_version;
389
0
            calc_delete_bitmap_ver = Version(to_add_max_version + 1, max_version_unlocked());
390
0
        }
391
392
0
        if (calc_delete_bitmap_ver.first <= calc_delete_bitmap_ver.second) {
393
0
            calc_bm_status = capture_consistent_rowsets_unlocked(calc_delete_bitmap_ver,
394
0
                                                                 &calc_delete_bitmap_rowsets);
395
0
            if (!calc_bm_status.ok()) {
396
0
                LOG(WARNING) << "fail to capture_consistent_rowsets, res: " << calc_bm_status;
397
0
                break;
398
0
            }
399
            // FIXME(plat1ko): Use `const TabletSharedPtr&` as parameter
400
0
            auto self = _engine.tablet_manager()->get_tablet(tablet_id());
401
0
            CHECK(self);
402
0
            for (auto rs : calc_delete_bitmap_rowsets) {
403
0
                if (is_incremental_clone) {
404
0
                    calc_bm_status = update_delete_bitmap_without_lock(self, rs);
405
0
                } else {
406
0
                    calc_bm_status = update_delete_bitmap_without_lock(
407
0
                            self, rs, &base_rowsets_for_full_clone);
408
0
                    base_rowsets_for_full_clone.push_back(rs);
409
0
                }
410
0
                if (!calc_bm_status.ok()) {
411
0
                    LOG(WARNING) << "fail to update_delete_bitmap_without_lock, res: "
412
0
                                 << calc_bm_status;
413
0
                    break;
414
0
                }
415
0
            }
416
0
        }
417
0
        break; // while (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write())
418
0
    }
419
420
0
    DBUG_EXECUTE_IF("Tablet.revise_tablet_meta_fail", {
421
0
        auto ptablet_id = dp->param("tablet_id", 0);
422
0
        if (tablet_id() == ptablet_id) {
423
0
            LOG(INFO) << "injected revies_tablet_meta failure for tabelt: " << ptablet_id;
424
0
            calc_bm_status = Status::InternalError("fault injection error");
425
0
        }
426
0
    });
427
428
    // error handling
429
0
    if (!calc_bm_status.ok()) {
430
0
        if (is_incremental_clone) {
431
0
            RETURN_IF_ERROR(delete_rowsets(to_add, false));
432
0
            LOG(WARNING) << "incremental clone on tablet: " << tablet_id() << " failed due to "
433
0
                         << calc_bm_status.msg() << ", revert " << to_add.size()
434
0
                         << " rowsets added before.";
435
0
        } else {
436
0
            LOG(WARNING) << "full clone on tablet: " << tablet_id() << " failed due to "
437
0
                         << calc_bm_status.msg() << ", will not update tablet meta.";
438
0
        }
439
0
        return calc_bm_status;
440
0
    }
441
442
    // full clone, calculate delete bitmap succeeded, update rowset
443
0
    if (!is_incremental_clone) {
444
0
        RETURN_IF_ERROR(delete_rowsets(to_delete, false));
445
0
        add_rowsets(to_add);
446
        // reconstruct from tablet meta
447
0
        _timestamped_version_tracker.construct_versioned_tracker(_tablet_meta->all_rs_metas());
448
449
        // check the rowsets used for delete bitmap calculation is equal to the rowsets
450
        // that we can capture by version
451
0
        if (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write()) {
452
0
            Version full_version = Version(0, max_version_unlocked());
453
0
            std::vector<RowsetSharedPtr> expected_rowsets;
454
0
            auto st = capture_consistent_rowsets_unlocked(full_version, &expected_rowsets);
455
0
            DCHECK(st.ok()) << st;
456
0
            DCHECK_EQ(base_rowsets_for_full_clone.size(), expected_rowsets.size());
457
0
            if (st.ok() && base_rowsets_for_full_clone.size() != expected_rowsets.size())
458
0
                    [[unlikely]] {
459
0
                LOG(WARNING) << "full clone succeeded, but the count("
460
0
                             << base_rowsets_for_full_clone.size()
461
0
                             << ") of base rowsets used for delete bitmap calculation is not match "
462
0
                                "expect count("
463
0
                             << expected_rowsets.size() << ") we capture from tablet meta";
464
0
            }
465
0
        }
466
0
    }
467
468
    // clear stale rowset
469
0
    for (auto& [v, rs] : _stale_rs_version_map) {
470
0
        _engine.add_unused_rowset(rs);
471
0
    }
472
0
    _stale_rs_version_map.clear();
473
0
    _tablet_meta->clear_stale_rowset();
474
0
    save_meta();
475
476
0
    LOG(INFO) << "finish to revise tablet. tablet_id=" << tablet_id();
477
0
    return Status::OK();
478
0
}
479
480
109
Status Tablet::add_rowset(RowsetSharedPtr rowset) {
481
109
    DCHECK(rowset != nullptr);
482
109
    std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
483
109
    SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
484
    // If the rowset already exist, just return directly.  The rowset_id is an unique-id,
485
    // we can use it to check this situation.
486
109
    if (_contains_rowset(rowset->rowset_id())) {
487
0
        return Status::OK();
488
0
    }
489
    // Otherwise, the version should be not contained in any existing rowset.
490
109
    RETURN_IF_ERROR(_contains_version(rowset->version()));
491
492
109
    RETURN_IF_ERROR(_tablet_meta->add_rs_meta(rowset->rowset_meta()));
493
109
    _rs_version_map[rowset->version()] = rowset;
494
109
    _timestamped_version_tracker.add_version(rowset->version());
495
109
    add_compaction_score(rowset->rowset_meta()->get_compaction_score());
496
497
109
    std::vector<RowsetSharedPtr> rowsets_to_delete;
498
    // yiguolei: temp code, should remove the rowset contains by this rowset
499
    // but it should be removed in multi path version
500
135
    for (auto& it : _rs_version_map) {
501
135
        if (rowset->version().contains(it.first) && rowset->version() != it.first) {
502
0
            CHECK(it.second != nullptr)
503
0
                    << "there exist a version=" << it.first
504
0
                    << " contains the input rs with version=" << rowset->version()
505
0
                    << ", but the related rs is null";
506
0
            rowsets_to_delete.push_back(it.second);
507
0
        }
508
135
    }
509
109
    std::vector<RowsetSharedPtr> empty_vec;
510
109
    RETURN_IF_ERROR(modify_rowsets(empty_vec, rowsets_to_delete));
511
109
    ++_newly_created_rowset_num;
512
109
    return Status::OK();
513
109
}
514
515
Status Tablet::modify_rowsets(std::vector<RowsetSharedPtr>& to_add,
516
111
                              std::vector<RowsetSharedPtr>& to_delete, bool check_delete) {
517
    // the compaction process allow to compact the single version, eg: version[4-4].
518
    // this kind of "single version compaction" has same "input version" and "output version".
519
    // which means "to_add->version()" equals to "to_delete->version()".
520
    // So we should delete the "to_delete" before adding the "to_add",
521
    // otherwise, the "to_add" will be deleted from _rs_version_map, eventually.
522
    //
523
    // And if the version of "to_add" and "to_delete" are exactly same. eg:
524
    // to_add:      [7-7]
525
    // to_delete:   [7-7]
526
    // In this case, we no longer need to add the rowset in "to_delete" to
527
    // _stale_rs_version_map, but can delete it directly.
528
529
111
    if (to_add.empty() && to_delete.empty()) {
530
109
        return Status::OK();
531
109
    }
532
533
2
    if (check_delete) {
534
20
        for (auto&& rs : to_delete) {
535
20
            if (auto it = _rs_version_map.find(rs->version()); it == _rs_version_map.end()) {
536
0
                return Status::Error<DELETE_VERSION_ERROR>(
537
0
                        "try to delete not exist version {} from {}", rs->version().to_string(),
538
0
                        tablet_id());
539
20
            } else if (rs->rowset_id() != it->second->rowset_id()) {
540
0
                return Status::Error<DELETE_VERSION_ERROR>(
541
0
                        "try to delete version {} from {}, but rowset id changed, delete rowset id "
542
0
                        "is {}, exists rowsetid is {}",
543
0
                        rs->version().to_string(), tablet_id(), rs->rowset_id().to_string(),
544
0
                        it->second->rowset_id().to_string());
545
0
            }
546
20
        }
547
1
    }
548
549
2
    bool same_version = true;
550
2
    std::sort(to_add.begin(), to_add.end(), Rowset::comparator);
551
2
    std::sort(to_delete.begin(), to_delete.end(), Rowset::comparator);
552
2
    if (to_add.size() == to_delete.size()) {
553
0
        for (int i = 0; i < to_add.size(); ++i) {
554
0
            if (to_add[i]->version() != to_delete[i]->version()) {
555
0
                same_version = false;
556
0
                break;
557
0
            }
558
0
        }
559
2
    } else {
560
2
        same_version = false;
561
2
    }
562
563
2
    std::vector<RowsetMetaSharedPtr> rs_metas_to_delete;
564
20
    for (auto& rs : to_delete) {
565
20
        rs_metas_to_delete.push_back(rs->rowset_meta());
566
20
        _rs_version_map.erase(rs->version());
567
568
20
        if (!same_version) {
569
            // put compaction rowsets in _stale_rs_version_map.
570
20
            _stale_rs_version_map[rs->version()] = rs;
571
20
        }
572
20
    }
573
574
2
    std::vector<RowsetMetaSharedPtr> rs_metas_to_add;
575
2
    for (auto& rs : to_add) {
576
2
        rs_metas_to_add.push_back(rs->rowset_meta());
577
2
        _rs_version_map[rs->version()] = rs;
578
579
2
        if (!same_version) {
580
            // If version are same, then _timestamped_version_tracker
581
            // already has this version, no need to add again.
582
2
            _timestamped_version_tracker.add_version(rs->version());
583
2
        }
584
2
        ++_newly_created_rowset_num;
585
2
    }
586
587
2
    _tablet_meta->modify_rs_metas(rs_metas_to_add, rs_metas_to_delete, same_version);
588
589
2
    if (!same_version) {
590
        // add rs_metas_to_delete to tracker
591
2
        _timestamped_version_tracker.add_stale_path_version(rs_metas_to_delete);
592
2
    } else {
593
        // delete rowset in "to_delete" directly
594
0
        for (auto& rs : to_delete) {
595
0
            LOG(INFO) << "add unused rowset " << rs->rowset_id() << " because of same version";
596
0
            if (rs->is_local()) {
597
0
                _engine.add_unused_rowset(rs);
598
0
            }
599
0
        }
600
0
    }
601
602
2
    int32_t add_score = 0;
603
2
    for (auto rs : to_add) {
604
2
        add_score += rs->rowset_meta()->get_compaction_score();
605
2
    }
606
2
    int32_t sub_score = 0;
607
20
    for (auto rs : to_delete) {
608
20
        sub_score += rs->rowset_meta()->get_compaction_score();
609
20
    }
610
2
    add_compaction_score(add_score - sub_score);
611
612
2
    return Status::OK();
613
2
}
614
615
5
void Tablet::add_rowsets(const std::vector<RowsetSharedPtr>& to_add) {
616
5
    if (to_add.empty()) {
617
0
        return;
618
0
    }
619
5
    std::vector<RowsetMetaSharedPtr> rs_metas;
620
5
    rs_metas.reserve(to_add.size());
621
5
    for (auto& rs : to_add) {
622
5
        _rs_version_map.emplace(rs->version(), rs);
623
5
        _timestamped_version_tracker.add_version(rs->version());
624
5
        rs_metas.push_back(rs->rowset_meta());
625
5
    }
626
5
    _tablet_meta->modify_rs_metas(rs_metas, {});
627
5
}
628
629
5
Status Tablet::delete_rowsets(const std::vector<RowsetSharedPtr>& to_delete, bool move_to_stale) {
630
5
    if (to_delete.empty()) {
631
0
        return Status::OK();
632
0
    }
633
5
    std::vector<RowsetMetaSharedPtr> rs_metas;
634
5
    rs_metas.reserve(to_delete.size());
635
5
    for (const auto& rs : to_delete) {
636
5
        rs_metas.push_back(rs->rowset_meta());
637
5
        _rs_version_map.erase(rs->version());
638
5
    }
639
5
    _tablet_meta->modify_rs_metas({}, rs_metas, !move_to_stale);
640
5
    if (move_to_stale) {
641
0
        for (const auto& rs : to_delete) {
642
0
            _stale_rs_version_map[rs->version()] = rs;
643
0
        }
644
0
        _timestamped_version_tracker.add_stale_path_version(rs_metas);
645
5
    } else {
646
5
        for (const auto& rs : to_delete) {
647
5
            _timestamped_version_tracker.delete_version(rs->version());
648
5
            if (rs->is_local()) {
649
5
                _engine.add_unused_rowset(rs);
650
5
                RETURN_IF_ERROR(RowsetMetaManager::remove(_data_dir->get_meta(), tablet_uid(),
651
5
                                                          rs->rowset_meta()->rowset_id()));
652
5
            }
653
5
        }
654
5
    }
655
5
    return Status::OK();
656
5
}
657
658
0
RowsetSharedPtr Tablet::_rowset_with_largest_size() {
659
0
    RowsetSharedPtr largest_rowset = nullptr;
660
0
    for (auto& it : _rs_version_map) {
661
0
        if (it.second->empty() || it.second->zero_num_rows()) {
662
0
            continue;
663
0
        }
664
0
        if (largest_rowset == nullptr || it.second->rowset_meta()->index_disk_size() >
665
0
                                                 largest_rowset->rowset_meta()->index_disk_size()) {
666
0
            largest_rowset = it.second;
667
0
        }
668
0
    }
669
670
0
    return largest_rowset;
671
0
}
672
673
// add inc rowset should not persist tablet meta, because it will be persisted when publish txn.
674
395
Status Tablet::add_inc_rowset(const RowsetSharedPtr& rowset) {
675
395
    DCHECK(rowset != nullptr);
676
395
    std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
677
395
    SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
678
395
    if (_contains_rowset(rowset->rowset_id())) {
679
0
        return Status::OK();
680
0
    }
681
395
    RETURN_IF_ERROR(_contains_version(rowset->version()));
682
683
395
    RETURN_IF_ERROR(_tablet_meta->add_rs_meta(rowset->rowset_meta()));
684
395
    _rs_version_map[rowset->version()] = rowset;
685
686
395
    _timestamped_version_tracker.add_version(rowset->version());
687
688
395
    ++_newly_created_rowset_num;
689
690
395
    add_compaction_score(rowset->rowset_meta()->get_compaction_score());
691
692
395
    return Status::OK();
693
395
}
694
695
8
void Tablet::_delete_stale_rowset_by_version(const Version& version) {
696
8
    RowsetMetaSharedPtr rowset_meta = _tablet_meta->acquire_stale_rs_meta_by_version(version);
697
8
    if (rowset_meta == nullptr) {
698
8
        return;
699
8
    }
700
0
    _tablet_meta->delete_stale_rs_meta_by_version(version);
701
0
    VLOG_NOTICE << "delete stale rowset. tablet=" << tablet_id() << ", version=" << version;
702
0
}
703
704
2
void Tablet::delete_expired_stale_rowset() {
705
2
    int64_t now = UnixSeconds();
706
    // hold write lock while processing stable rowset
707
2
    {
708
2
        std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
709
2
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
710
        // Compute the end time to delete rowsets, when a expired rowset createtime less then this time, it will be deleted.
711
2
        double expired_stale_sweep_endtime =
712
2
                ::difftime(now, config::tablet_rowset_stale_sweep_time_sec);
713
2
        if (config::tablet_rowset_stale_sweep_by_size) {
714
0
            expired_stale_sweep_endtime = now;
715
0
        }
716
717
2
        std::vector<int64_t> path_id_vec;
718
        // capture the path version to delete
719
2
        _timestamped_version_tracker.capture_expired_paths(
720
2
                static_cast<int64_t>(expired_stale_sweep_endtime), &path_id_vec);
721
722
2
        if (path_id_vec.empty()) {
723
1
            return;
724
1
        }
725
726
1
        const RowsetSharedPtr lastest_delta = get_rowset_with_max_version();
727
1
        if (lastest_delta == nullptr) {
728
0
            LOG(WARNING) << "lastest_delta is null " << tablet_id();
729
0
            return;
730
0
        }
731
732
        // fetch missing version before delete
733
1
        Versions missed_versions = get_missed_versions_unlocked(lastest_delta->end_version());
734
1
        if (!missed_versions.empty()) {
735
0
            LOG(WARNING) << "tablet:" << tablet_id()
736
0
                         << ", missed version for version:" << lastest_delta->end_version();
737
0
            _print_missed_versions(missed_versions);
738
0
            return;
739
0
        }
740
741
        // do check consistent operation
742
1
        auto path_id_iter = path_id_vec.begin();
743
744
1
        std::map<int64_t, PathVersionListSharedPtr> stale_version_path_map;
745
5
        while (path_id_iter != path_id_vec.end()) {
746
4
            PathVersionListSharedPtr version_path =
747
4
                    _timestamped_version_tracker.fetch_and_delete_path_by_id(*path_id_iter);
748
749
4
            Version test_version = Version(0, lastest_delta->end_version());
750
4
            stale_version_path_map[*path_id_iter] = version_path;
751
752
4
            Status status =
753
4
                    capture_consistent_versions_unlocked(test_version, nullptr, false, false);
754
            // 1. When there is no consistent versions, we must reconstruct the tracker.
755
4
            if (!status.ok()) {
756
                // 2. fetch missing version after delete
757
0
                Versions after_missed_versions =
758
0
                        get_missed_versions_unlocked(lastest_delta->end_version());
759
760
                // 2.1 check whether missed_versions and after_missed_versions are the same.
761
                // when they are the same, it means we can delete the path securely.
762
0
                bool is_missing = missed_versions.size() != after_missed_versions.size();
763
764
0
                if (!is_missing) {
765
0
                    for (int ver_index = 0; ver_index < missed_versions.size(); ver_index++) {
766
0
                        if (missed_versions[ver_index] != after_missed_versions[ver_index]) {
767
0
                            is_missing = true;
768
0
                            break;
769
0
                        }
770
0
                    }
771
0
                }
772
773
0
                if (is_missing) {
774
0
                    LOG(WARNING) << "The consistent version check fails, there are bugs. "
775
0
                                 << "Reconstruct the tracker to recover versions in tablet="
776
0
                                 << tablet_id();
777
778
                    // 3. try to recover
779
0
                    _timestamped_version_tracker.recover_versioned_tracker(stale_version_path_map);
780
781
                    // 4. double check the consistent versions
782
                    // fetch missing version after recover
783
0
                    Versions recover_missed_versions =
784
0
                            get_missed_versions_unlocked(lastest_delta->end_version());
785
786
                    // 4.1 check whether missed_versions and recover_missed_versions are the same.
787
                    // when they are the same, it means we recover successfully.
788
0
                    bool is_recover_missing =
789
0
                            missed_versions.size() != recover_missed_versions.size();
790
791
0
                    if (!is_recover_missing) {
792
0
                        for (int ver_index = 0; ver_index < missed_versions.size(); ver_index++) {
793
0
                            if (missed_versions[ver_index] != recover_missed_versions[ver_index]) {
794
0
                                is_recover_missing = true;
795
0
                                break;
796
0
                            }
797
0
                        }
798
0
                    }
799
800
                    // 5. check recover fail, version is mission
801
0
                    if (is_recover_missing) {
802
0
                        if (!config::ignore_rowset_stale_unconsistent_delete) {
803
0
                            LOG(FATAL)
804
0
                                    << "rowset stale unconsistent delete. tablet= " << tablet_id();
805
0
                        } else {
806
0
                            LOG(WARNING)
807
0
                                    << "rowset stale unconsistent delete. tablet= " << tablet_id();
808
0
                        }
809
0
                    }
810
0
                }
811
0
                return;
812
0
            }
813
4
            path_id_iter++;
814
4
        }
815
816
1
        auto old_size = _stale_rs_version_map.size();
817
1
        auto old_meta_size = _tablet_meta->all_stale_rs_metas().size();
818
819
        // do delete operation
820
1
        auto to_delete_iter = stale_version_path_map.begin();
821
5
        while (to_delete_iter != stale_version_path_map.end()) {
822
4
            std::vector<TimestampedVersionSharedPtr>& to_delete_version =
823
4
                    to_delete_iter->second->timestamped_versions();
824
8
            for (auto& timestampedVersion : to_delete_version) {
825
8
                auto it = _stale_rs_version_map.find(timestampedVersion->version());
826
8
                if (it != _stale_rs_version_map.end()) {
827
0
                    it->second->clear_cache();
828
                    // delete rowset
829
0
                    if (it->second->is_local()) {
830
0
                        _engine.add_unused_rowset(it->second);
831
0
                    }
832
0
                    _stale_rs_version_map.erase(it);
833
0
                    VLOG_NOTICE << "delete stale rowset tablet=" << tablet_id() << " version["
834
0
                                << timestampedVersion->version().first << ","
835
0
                                << timestampedVersion->version().second
836
0
                                << "] move to unused_rowset success " << std::fixed
837
0
                                << expired_stale_sweep_endtime;
838
8
                } else {
839
8
                    LOG(WARNING) << "delete stale rowset tablet=" << tablet_id() << " version["
840
8
                                 << timestampedVersion->version().first << ","
841
8
                                 << timestampedVersion->version().second
842
8
                                 << "] not find in stale rs version map";
843
8
                }
844
8
                _delete_stale_rowset_by_version(timestampedVersion->version());
845
8
            }
846
4
            to_delete_iter++;
847
4
        }
848
849
1
        bool reconstructed = _reconstruct_version_tracker_if_necessary();
850
851
1
        VLOG_NOTICE << "delete stale rowset _stale_rs_version_map tablet=" << tablet_id()
852
0
                    << " current_size=" << _stale_rs_version_map.size() << " old_size=" << old_size
853
0
                    << " current_meta_size=" << _tablet_meta->all_stale_rs_metas().size()
854
0
                    << " old_meta_size=" << old_meta_size << " sweep endtime " << std::fixed
855
0
                    << expired_stale_sweep_endtime << ", reconstructed=" << reconstructed;
856
1
    }
857
#ifndef BE_TEST
858
    {
859
        std::shared_lock<std::shared_mutex> rlock(_meta_lock);
860
        save_meta();
861
    }
862
#endif
863
1
}
864
865
Status Tablet::capture_consistent_versions_unlocked(const Version& spec_version,
866
                                                    Versions* version_path,
867
8
                                                    bool skip_missing_version, bool quiet) const {
868
8
    Status status =
869
8
            _timestamped_version_tracker.capture_consistent_versions(spec_version, version_path);
870
8
    if (!status.ok() && !quiet) {
871
1
        Versions missed_versions = get_missed_versions_unlocked(spec_version.second);
872
1
        if (missed_versions.empty()) {
873
            // if version_path is null, it may be a compaction check logic.
874
            // so to avoid print too many logs.
875
0
            if (version_path != nullptr) {
876
0
                LOG(WARNING) << "tablet:" << tablet_id()
877
0
                             << ", version already has been merged. spec_version: " << spec_version
878
0
                             << ", max_version: " << max_version_unlocked();
879
0
            }
880
0
            status = Status::Error<VERSION_ALREADY_MERGED, false>(
881
0
                    "versions are already compacted, spec_version "
882
0
                    "{}, max_version {}, tablet_id {}",
883
0
                    spec_version.second, max_version_unlocked(), tablet_id());
884
1
        } else {
885
1
            if (version_path != nullptr) {
886
1
                LOG(WARNING) << "status:" << status << ", tablet:" << tablet_id()
887
1
                             << ", missed version for version:" << spec_version;
888
1
                _print_missed_versions(missed_versions);
889
1
                if (skip_missing_version) {
890
0
                    LOG(WARNING) << "force skipping missing version for tablet:" << tablet_id();
891
0
                    return Status::OK();
892
0
                }
893
1
            }
894
1
        }
895
1
    }
896
897
8
    DBUG_EXECUTE_IF("TTablet::capture_consistent_versions.inject_failure", {
898
8
        auto tablet_id = dp->param<int64>("tablet_id", -1);
899
8
        if (tablet_id != -1 && tablet_id == _tablet_meta->tablet_id()) {
900
8
            status = Status::Error<VERSION_ALREADY_MERGED>("version already merged");
901
8
        }
902
8
    });
903
904
8
    return status;
905
8
}
906
907
0
Status Tablet::check_version_integrity(const Version& version, bool quiet) {
908
0
    std::shared_lock rdlock(_meta_lock);
909
0
    return capture_consistent_versions_unlocked(version, nullptr, false, quiet);
910
0
}
911
912
25
bool Tablet::exceed_version_limit(int32_t limit) {
913
25
    if (_tablet_meta->version_count() > limit) {
914
0
        exceed_version_limit_counter << 1;
915
0
        return true;
916
0
    }
917
25
    return false;
918
25
}
919
920
// If any rowset contains the specific version, it means the version already exist
921
1
bool Tablet::check_version_exist(const Version& version) const {
922
1
    std::shared_lock rdlock(_meta_lock);
923
3
    for (auto& it : _rs_version_map) {
924
3
        if (it.first.contains(version)) {
925
0
            return true;
926
0
        }
927
3
    }
928
1
    return false;
929
1
}
930
931
// The meta read lock should be held before calling
932
void Tablet::acquire_version_and_rowsets(
933
0
        std::vector<std::pair<Version, RowsetSharedPtr>>* version_rowsets) const {
934
0
    for (const auto& it : _rs_version_map) {
935
0
        version_rowsets->emplace_back(it.first, it.second);
936
0
    }
937
0
}
938
939
Status Tablet::capture_consistent_rowsets_unlocked(const Version& spec_version,
940
2
                                                   std::vector<RowsetSharedPtr>* rowsets) const {
941
2
    std::vector<Version> version_path;
942
2
    RETURN_IF_ERROR(
943
2
            capture_consistent_versions_unlocked(spec_version, &version_path, false, false));
944
2
    RETURN_IF_ERROR(_capture_consistent_rowsets_unlocked(version_path, rowsets));
945
2
    return Status::OK();
946
2
}
947
948
Status Tablet::capture_rs_readers(const Version& spec_version, std::vector<RowSetSplits>* rs_splits,
949
2
                                  bool skip_missing_version) {
950
2
    std::shared_lock rlock(_meta_lock);
951
2
    std::vector<Version> version_path;
952
2
    RETURN_IF_ERROR(capture_consistent_versions_unlocked(spec_version, &version_path,
953
2
                                                         skip_missing_version, false));
954
1
    RETURN_IF_ERROR(capture_rs_readers_unlocked(version_path, rs_splits));
955
1
    return Status::OK();
956
1
}
957
958
2
Versions Tablet::calc_missed_versions(int64_t spec_version, Versions existing_versions) const {
959
2
    DCHECK(spec_version > 0) << "invalid spec_version: " << spec_version;
960
961
    // sort the existing versions in ascending order
962
2
    std::sort(existing_versions.begin(), existing_versions.end(),
963
12
              [](const Version& a, const Version& b) {
964
                  // simple because 2 versions are certainly not overlapping
965
12
                  return a.first < b.first;
966
12
              });
967
968
    // From the first version(=0),  find the missing version until spec_version
969
2
    int64_t last_version = -1;
970
2
    Versions missed_versions;
971
8
    for (const Version& version : existing_versions) {
972
8
        if (version.first > last_version + 1) {
973
4
            for (int64_t i = last_version + 1; i < version.first && i <= spec_version; ++i) {
974
                // Don't merge missed_versions because clone & snapshot use single version.
975
                // For example, if miss 4 ~ 6, clone need [4, 4], [5, 5], [6, 6], but not [4, 6].
976
2
                missed_versions.emplace_back(i, i);
977
2
            }
978
2
        }
979
8
        last_version = version.second;
980
8
        if (last_version >= spec_version) {
981
2
            break;
982
2
        }
983
8
    }
984
2
    for (int64_t i = last_version + 1; i <= spec_version; ++i) {
985
0
        missed_versions.emplace_back(i, i);
986
0
    }
987
988
2
    return missed_versions;
989
2
}
990
991
53
bool Tablet::can_do_compaction(size_t path_hash, CompactionType compaction_type) {
992
53
    if (compaction_type == CompactionType::BASE_COMPACTION && tablet_state() != TABLET_RUNNING) {
993
        // base compaction can only be done for tablet in TABLET_RUNNING state.
994
        // but cumulative compaction can be done for TABLET_NOTREADY, such as tablet under alter process.
995
0
        return false;
996
0
    }
997
998
53
    if (data_dir()->path_hash() != path_hash || !is_used() || !init_succeeded()) {
999
0
        return false;
1000
0
    }
1001
1002
    // In TABLET_NOTREADY, we keep last 10 versions in new tablet so base tablet max_version
1003
    // not merged in new tablet and then we can do compaction
1004
53
    return tablet_state() == TABLET_RUNNING || tablet_state() == TABLET_NOTREADY;
1005
53
}
1006
1007
56
uint32_t Tablet::calc_compaction_score() {
1008
56
    if (_score_check_cnt++ % config::check_score_rounds_num != 0) {
1009
32
        std::shared_lock rdlock(_meta_lock);
1010
32
        if (_compaction_score > 0) {
1011
32
            return _compaction_score;
1012
32
        }
1013
32
    }
1014
1015
24
    {
1016
        // Need meta lock, because it will iterator "all_rs_metas" of tablet meta.
1017
24
        std::shared_lock rdlock(_meta_lock);
1018
24
        int32_t score = get_real_compaction_score();
1019
24
        if (_compaction_score > 0 && _compaction_score != score) {
1020
0
            LOG(WARNING) << "cumu cache score not equal real score, cache score; "
1021
0
                         << _compaction_score << ", real score: " << score
1022
0
                         << ", tablet: " << tablet_id();
1023
0
        }
1024
24
        _compaction_score = score;
1025
24
        return score;
1026
56
    }
1027
56
}
1028
1029
bool Tablet::suitable_for_compaction(
1030
        CompactionType compaction_type,
1031
7
        std::shared_ptr<CumulativeCompactionPolicy> cumulative_compaction_policy) {
1032
    // Need meta lock, because it will iterator "all_rs_metas" of tablet meta.
1033
7
    std::shared_lock rdlock(_meta_lock);
1034
7
    int32_t score = -1;
1035
7
    if (compaction_type == CompactionType::CUMULATIVE_COMPACTION) {
1036
7
        score = _calc_cumulative_compaction_score(cumulative_compaction_policy);
1037
7
    } else {
1038
0
        DCHECK_EQ(compaction_type, CompactionType::BASE_COMPACTION);
1039
0
        score = _calc_base_compaction_score();
1040
0
    }
1041
7
    return score > 0;
1042
7
}
1043
1044
0
uint32_t Tablet::calc_cold_data_compaction_score() const {
1045
0
    uint32_t score = 0;
1046
0
    std::vector<RowsetMetaSharedPtr> cooldowned_rowsets;
1047
0
    int64_t max_delete_version = 0;
1048
0
    {
1049
0
        std::shared_lock rlock(_meta_lock);
1050
0
        for (auto& rs_meta : _tablet_meta->all_rs_metas()) {
1051
0
            if (!rs_meta->is_local()) {
1052
0
                cooldowned_rowsets.push_back(rs_meta);
1053
0
                if (rs_meta->has_delete_predicate() &&
1054
0
                    rs_meta->end_version() > max_delete_version) {
1055
0
                    max_delete_version = rs_meta->end_version();
1056
0
                }
1057
0
            }
1058
0
        }
1059
0
    }
1060
0
    for (auto& rs_meta : cooldowned_rowsets) {
1061
0
        if (rs_meta->end_version() < max_delete_version) {
1062
0
            score += rs_meta->num_segments();
1063
0
        } else {
1064
0
            score += rs_meta->get_compaction_score();
1065
0
        }
1066
0
    }
1067
0
    return (keys_type() != KeysType::DUP_KEYS) ? score * 2 : score;
1068
0
}
1069
1070
uint32_t Tablet::_calc_cumulative_compaction_score(
1071
7
        std::shared_ptr<CumulativeCompactionPolicy> cumulative_compaction_policy) {
1072
7
    if (cumulative_compaction_policy == nullptr) [[unlikely]] {
1073
0
        return 0;
1074
0
    }
1075
#ifndef BE_TEST
1076
    if (_cumulative_compaction_policy == nullptr ||
1077
        _cumulative_compaction_policy->name() != cumulative_compaction_policy->name()) {
1078
        _cumulative_compaction_policy = cumulative_compaction_policy;
1079
    }
1080
#endif
1081
7
    return _cumulative_compaction_policy->calc_cumulative_compaction_score(this);
1082
7
}
1083
1084
0
uint32_t Tablet::_calc_base_compaction_score() const {
1085
0
    uint32_t score = 0;
1086
0
    const int64_t point = cumulative_layer_point();
1087
0
    bool base_rowset_exist = false;
1088
0
    bool has_delete = false;
1089
0
    for (auto& rs_meta : _tablet_meta->all_rs_metas()) {
1090
0
        if (rs_meta->start_version() == 0) {
1091
0
            base_rowset_exist = true;
1092
0
        }
1093
0
        if (rs_meta->start_version() >= point || !rs_meta->is_local()) {
1094
            // all_rs_metas() is not sorted, so we use _continue_ other than _break_ here.
1095
0
            continue;
1096
0
        }
1097
0
        if (rs_meta->has_delete_predicate()) {
1098
0
            has_delete = true;
1099
0
        }
1100
0
        score += rs_meta->get_compaction_score();
1101
0
    }
1102
1103
    // In the time series compaction policy, we want the base compaction to be triggered
1104
    // when there are delete versions present.
1105
0
    if (_tablet_meta->compaction_policy() == CUMULATIVE_TIME_SERIES_POLICY) {
1106
0
        return (base_rowset_exist && has_delete) ? score : 0;
1107
0
    }
1108
1109
    // base不存在可能是tablet正在做alter table,先不选它,设score=0
1110
0
    return base_rowset_exist ? score : 0;
1111
0
}
1112
1113
0
void Tablet::max_continuous_version_from_beginning(Version* version, Version* max_version) {
1114
0
    bool has_version_cross;
1115
0
    std::shared_lock rdlock(_meta_lock);
1116
0
    _max_continuous_version_from_beginning_unlocked(version, max_version, &has_version_cross);
1117
0
}
1118
1119
void Tablet::_max_continuous_version_from_beginning_unlocked(Version* version, Version* max_version,
1120
0
                                                             bool* has_version_cross) const {
1121
0
    std::vector<Version> existing_versions;
1122
0
    *has_version_cross = false;
1123
0
    for (auto& rs : _tablet_meta->all_rs_metas()) {
1124
0
        existing_versions.emplace_back(rs->version());
1125
0
    }
1126
1127
    // sort the existing versions in ascending order
1128
0
    std::sort(existing_versions.begin(), existing_versions.end(),
1129
0
              [](const Version& left, const Version& right) {
1130
                  // simple because 2 versions are certainly not overlapping
1131
0
                  return left.first < right.first;
1132
0
              });
1133
1134
0
    Version max_continuous_version = {-1, -1};
1135
0
    for (int i = 0; i < existing_versions.size(); ++i) {
1136
0
        if (existing_versions[i].first > max_continuous_version.second + 1) {
1137
0
            break;
1138
0
        } else if (existing_versions[i].first <= max_continuous_version.second) {
1139
0
            *has_version_cross = true;
1140
0
        }
1141
0
        max_continuous_version = existing_versions[i];
1142
0
    }
1143
0
    *version = max_continuous_version;
1144
    // tablet may not has rowset, eg, tablet has just been clear for restore.
1145
0
    if (max_version != nullptr && !existing_versions.empty()) {
1146
0
        *max_version = existing_versions.back();
1147
0
    }
1148
0
}
1149
1150
38
void Tablet::calculate_cumulative_point() {
1151
38
    std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
1152
38
    SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
1153
38
    int64_t ret_cumulative_point;
1154
38
    _cumulative_compaction_policy->calculate_cumulative_point(
1155
38
            this, _tablet_meta->all_rs_metas(), _cumulative_point, &ret_cumulative_point);
1156
1157
38
    if (ret_cumulative_point == K_INVALID_CUMULATIVE_POINT) {
1158
10
        return;
1159
10
    }
1160
28
    set_cumulative_layer_point(ret_cumulative_point);
1161
28
}
1162
1163
// NOTE: only used when create_table, so it is sure that there is no concurrent reader and writer.
1164
0
void Tablet::delete_all_files() {
1165
    // Release resources like memory and disk space.
1166
0
    std::shared_lock rdlock(_meta_lock);
1167
0
    for (auto it : _rs_version_map) {
1168
0
        static_cast<void>(it.second->remove());
1169
0
    }
1170
0
    _rs_version_map.clear();
1171
1172
0
    for (auto it : _stale_rs_version_map) {
1173
0
        static_cast<void>(it.second->remove());
1174
0
    }
1175
0
    _stale_rs_version_map.clear();
1176
0
}
1177
1178
0
void Tablet::check_tablet_path_exists() {
1179
0
    if (!tablet_path().empty()) {
1180
0
        std::error_code ec;
1181
0
        if (std::filesystem::is_directory(tablet_path(), ec)) {
1182
0
            _is_tablet_path_exists.store(true, std::memory_order_relaxed);
1183
0
        } else if (ec.value() == ENOENT || ec.value() == 0) {
1184
0
            _is_tablet_path_exists.store(false, std::memory_order_relaxed);
1185
0
        }
1186
0
    }
1187
0
}
1188
1189
504
Status Tablet::_contains_version(const Version& version) {
1190
    // check if there exist a rowset contains the added rowset
1191
3.54k
    for (auto& it : _rs_version_map) {
1192
3.54k
        if (it.first.contains(version)) {
1193
            // TODO(lingbin): Is this check unnecessary?
1194
            // because the value type is std::shared_ptr, when will it be nullptr?
1195
            // In addition, in this class, there are many places that do not make this judgment
1196
            // when access _rs_version_map's value.
1197
0
            CHECK(it.second != nullptr) << "there exist a version=" << it.first
1198
0
                                        << " contains the input rs with version=" << version
1199
0
                                        << ", but the related rs is null";
1200
0
            return Status::Error<PUSH_VERSION_ALREADY_EXIST>("Tablet push duplicate version {}",
1201
0
                                                             version.to_string());
1202
0
        }
1203
3.54k
    }
1204
1205
504
    return Status::OK();
1206
504
}
1207
1208
28
std::vector<RowsetSharedPtr> Tablet::pick_candidate_rowsets_to_cumulative_compaction() {
1209
28
    std::vector<RowsetSharedPtr> candidate_rowsets;
1210
28
    if (_cumulative_point == K_INVALID_CUMULATIVE_POINT) {
1211
0
        return candidate_rowsets;
1212
0
    }
1213
28
    return _pick_visible_rowsets_to_compaction(_cumulative_point,
1214
28
                                               std::numeric_limits<int64_t>::max());
1215
28
}
1216
1217
1
std::vector<RowsetSharedPtr> Tablet::pick_candidate_rowsets_to_base_compaction() {
1218
1
    return _pick_visible_rowsets_to_compaction(std::numeric_limits<int64_t>::min(),
1219
1
                                               _cumulative_point - 1);
1220
1
}
1221
1222
std::vector<RowsetSharedPtr> Tablet::_pick_visible_rowsets_to_compaction(
1223
29
        int64_t min_start_version, int64_t max_start_version) {
1224
29
    auto [visible_version, update_ts] = get_visible_version_and_time();
1225
29
    bool update_time_long = MonotonicMillis() - update_ts >
1226
29
                            config::compaction_keep_invisible_version_timeout_sec * 1000L;
1227
29
    int32_t keep_invisible_version_limit =
1228
29
            update_time_long ? config::compaction_keep_invisible_version_min_count
1229
29
                             : config::compaction_keep_invisible_version_max_count;
1230
1231
29
    std::vector<RowsetSharedPtr> candidate_rowsets;
1232
29
    {
1233
29
        std::shared_lock rlock(_meta_lock);
1234
402
        for (const auto& [version, rs] : _rs_version_map) {
1235
402
            int64_t version_start = version.first;
1236
            // rowset is remote or rowset is not in given range
1237
402
            if (!rs->is_local() || version_start < min_start_version ||
1238
402
                version_start > max_start_version) {
1239
31
                continue;
1240
31
            }
1241
1242
            // can compact, met one of the conditions:
1243
            // 1. had been visible;
1244
            // 2. exceeds the limit of keep invisible versions.
1245
371
            int64_t version_end = version.second;
1246
371
            if (version_end <= visible_version ||
1247
371
                version_end > visible_version + keep_invisible_version_limit) {
1248
371
                candidate_rowsets.push_back(rs);
1249
371
            }
1250
371
        }
1251
29
    }
1252
29
    std::sort(candidate_rowsets.begin(), candidate_rowsets.end(), Rowset::comparator);
1253
29
    return candidate_rowsets;
1254
29
}
1255
1256
0
std::vector<RowsetSharedPtr> Tablet::pick_candidate_rowsets_to_full_compaction() {
1257
0
    std::vector<RowsetSharedPtr> candidate_rowsets;
1258
0
    traverse_rowsets([&candidate_rowsets](const auto& rs) {
1259
        // Do full compaction on all local rowsets.
1260
0
        if (rs->is_local()) {
1261
0
            candidate_rowsets.emplace_back(rs);
1262
0
        }
1263
0
    });
1264
0
    std::sort(candidate_rowsets.begin(), candidate_rowsets.end(), Rowset::comparator);
1265
0
    return candidate_rowsets;
1266
0
}
1267
1268
std::vector<RowsetSharedPtr> Tablet::pick_candidate_rowsets_to_build_inverted_index(
1269
0
        const std::set<int64_t>& alter_index_uids, bool is_drop_op) {
1270
0
    std::vector<RowsetSharedPtr> candidate_rowsets;
1271
0
    {
1272
0
        std::shared_lock rlock(_meta_lock);
1273
0
        auto has_alter_inverted_index = [&](RowsetSharedPtr rowset) -> bool {
1274
0
            for (const auto& index_id : alter_index_uids) {
1275
0
                if (rowset->tablet_schema()->has_inverted_index_with_index_id(index_id)) {
1276
0
                    return true;
1277
0
                }
1278
0
            }
1279
0
            return false;
1280
0
        };
1281
1282
0
        for (const auto& [version, rs] : _rs_version_map) {
1283
0
            if (!has_alter_inverted_index(rs) && is_drop_op) {
1284
0
                continue;
1285
0
            }
1286
0
            if (has_alter_inverted_index(rs) && !is_drop_op) {
1287
0
                continue;
1288
0
            }
1289
1290
0
            if (rs->is_local()) {
1291
0
                candidate_rowsets.push_back(rs);
1292
0
            }
1293
0
        }
1294
0
    }
1295
0
    std::sort(candidate_rowsets.begin(), candidate_rowsets.end(), Rowset::comparator);
1296
0
    return candidate_rowsets;
1297
0
}
1298
1299
29
std::tuple<int64_t, int64_t> Tablet::get_visible_version_and_time() const {
1300
    // some old tablet has bug, its partition_id is 0, fe couldn't update its visible version.
1301
    // so let this tablet's visible version become int64 max.
1302
29
    auto version_info = std::atomic_load_explicit(&_visible_version, std::memory_order_relaxed);
1303
29
    if (version_info != nullptr && partition_id() != 0) {
1304
0
        return std::make_tuple(version_info->version.load(std::memory_order_relaxed),
1305
0
                               version_info->update_ts);
1306
29
    } else {
1307
29
        return std::make_tuple(std::numeric_limits<int64_t>::max(),
1308
29
                               std::numeric_limits<int64_t>::max());
1309
29
    }
1310
29
}
1311
1312
// For http compaction action
1313
0
void Tablet::get_compaction_status(std::string* json_result) {
1314
0
    rapidjson::Document root;
1315
0
    root.SetObject();
1316
1317
0
    rapidjson::Document path_arr;
1318
0
    path_arr.SetArray();
1319
1320
0
    std::vector<RowsetSharedPtr> rowsets;
1321
0
    std::vector<RowsetSharedPtr> stale_rowsets;
1322
0
    std::vector<bool> delete_flags;
1323
0
    {
1324
0
        std::shared_lock rdlock(_meta_lock);
1325
0
        rowsets.reserve(_rs_version_map.size());
1326
0
        for (auto& it : _rs_version_map) {
1327
0
            rowsets.push_back(it.second);
1328
0
        }
1329
0
        std::sort(rowsets.begin(), rowsets.end(), Rowset::comparator);
1330
1331
0
        stale_rowsets.reserve(_stale_rs_version_map.size());
1332
0
        for (auto& it : _stale_rs_version_map) {
1333
0
            stale_rowsets.push_back(it.second);
1334
0
        }
1335
0
        std::sort(stale_rowsets.begin(), stale_rowsets.end(), Rowset::comparator);
1336
1337
0
        delete_flags.reserve(rowsets.size());
1338
0
        for (auto& rs : rowsets) {
1339
0
            delete_flags.push_back(rs->rowset_meta()->has_delete_predicate());
1340
0
        }
1341
        // get snapshot version path json_doc
1342
0
        _timestamped_version_tracker.get_stale_version_path_json_doc(path_arr);
1343
0
    }
1344
0
    rapidjson::Value cumulative_policy_type;
1345
0
    std::string policy_type_str = "cumulative compaction policy not initializied";
1346
0
    if (_cumulative_compaction_policy != nullptr) {
1347
0
        policy_type_str = _cumulative_compaction_policy->name();
1348
0
    }
1349
0
    cumulative_policy_type.SetString(policy_type_str.c_str(), policy_type_str.length(),
1350
0
                                     root.GetAllocator());
1351
0
    root.AddMember("cumulative policy type", cumulative_policy_type, root.GetAllocator());
1352
0
    root.AddMember("cumulative point", _cumulative_point.load(), root.GetAllocator());
1353
0
    rapidjson::Value cumu_value;
1354
0
    std::string format_str = ToStringFromUnixMillis(_last_cumu_compaction_failure_millis.load());
1355
0
    cumu_value.SetString(format_str.c_str(), format_str.length(), root.GetAllocator());
1356
0
    root.AddMember("last cumulative failure time", cumu_value, root.GetAllocator());
1357
0
    rapidjson::Value base_value;
1358
0
    format_str = ToStringFromUnixMillis(_last_base_compaction_failure_millis.load());
1359
0
    base_value.SetString(format_str.c_str(), format_str.length(), root.GetAllocator());
1360
0
    root.AddMember("last base failure time", base_value, root.GetAllocator());
1361
0
    rapidjson::Value full_value;
1362
0
    format_str = ToStringFromUnixMillis(_last_full_compaction_failure_millis.load());
1363
0
    full_value.SetString(format_str.c_str(), format_str.length(), root.GetAllocator());
1364
0
    root.AddMember("last full failure time", full_value, root.GetAllocator());
1365
0
    rapidjson::Value cumu_success_value;
1366
0
    format_str = ToStringFromUnixMillis(_last_cumu_compaction_success_millis.load());
1367
0
    cumu_success_value.SetString(format_str.c_str(), format_str.length(), root.GetAllocator());
1368
0
    root.AddMember("last cumulative success time", cumu_success_value, root.GetAllocator());
1369
0
    rapidjson::Value base_success_value;
1370
0
    format_str = ToStringFromUnixMillis(_last_base_compaction_success_millis.load());
1371
0
    base_success_value.SetString(format_str.c_str(), format_str.length(), root.GetAllocator());
1372
0
    root.AddMember("last base success time", base_success_value, root.GetAllocator());
1373
0
    rapidjson::Value full_success_value;
1374
0
    format_str = ToStringFromUnixMillis(_last_full_compaction_success_millis.load());
1375
0
    full_success_value.SetString(format_str.c_str(), format_str.length(), root.GetAllocator());
1376
0
    root.AddMember("last full success time", full_success_value, root.GetAllocator());
1377
0
    rapidjson::Value base_schedule_value;
1378
0
    format_str = ToStringFromUnixMillis(_last_base_compaction_schedule_millis.load());
1379
0
    base_schedule_value.SetString(format_str.c_str(), format_str.length(), root.GetAllocator());
1380
0
    root.AddMember("last base schedule time", base_schedule_value, root.GetAllocator());
1381
0
    rapidjson::Value base_compaction_status_value;
1382
0
    base_compaction_status_value.SetString(_last_base_compaction_status.c_str(),
1383
0
                                           _last_base_compaction_status.length(),
1384
0
                                           root.GetAllocator());
1385
0
    root.AddMember("last base status", base_compaction_status_value, root.GetAllocator());
1386
1387
    // last single replica compaction status
1388
    // "single replica compaction status": {
1389
    //     "remote peer": "172.100.1.0:10875",
1390
    //     "last failure status": "",
1391
    //     "last fetched rowset": "[8-10]"
1392
    // }
1393
0
    rapidjson::Document status;
1394
0
    status.SetObject();
1395
0
    TReplicaInfo replica_info;
1396
0
    std::string dummp_token;
1397
0
    if (tablet_meta()->tablet_schema()->enable_single_replica_compaction() &&
1398
0
        _engine.get_peer_replica_info(tablet_id(), &replica_info, &dummp_token)) {
1399
        // remote peer
1400
0
        rapidjson::Value peer_addr;
1401
0
        std::string addr = replica_info.host + ":" + std::to_string(replica_info.brpc_port);
1402
0
        peer_addr.SetString(addr.c_str(), addr.length(), status.GetAllocator());
1403
0
        status.AddMember("remote peer", peer_addr, status.GetAllocator());
1404
        // last failure status
1405
0
        rapidjson::Value compaction_status;
1406
0
        compaction_status.SetString(_last_single_compaction_failure_status.c_str(),
1407
0
                                    _last_single_compaction_failure_status.length(),
1408
0
                                    status.GetAllocator());
1409
0
        status.AddMember("last failure status", compaction_status, status.GetAllocator());
1410
        // last fetched rowset
1411
0
        rapidjson::Value version;
1412
0
        std::string fetched_version = _last_fetched_version.to_string();
1413
0
        version.SetString(fetched_version.c_str(), fetched_version.length(), status.GetAllocator());
1414
0
        status.AddMember("last fetched rowset", version, status.GetAllocator());
1415
0
        root.AddMember("single replica compaction status", status, root.GetAllocator());
1416
0
    }
1417
1418
    // print all rowsets' version as an array
1419
0
    rapidjson::Document versions_arr;
1420
0
    rapidjson::Document missing_versions_arr;
1421
0
    versions_arr.SetArray();
1422
0
    missing_versions_arr.SetArray();
1423
0
    int64_t last_version = -1;
1424
0
    for (auto& rowset : rowsets) {
1425
0
        const Version& ver = rowset->version();
1426
0
        if (ver.first != last_version + 1) {
1427
0
            rapidjson::Value miss_value;
1428
0
            miss_value.SetString(fmt::format("[{}-{}]", last_version + 1, ver.first - 1).c_str(),
1429
0
                                 missing_versions_arr.GetAllocator());
1430
0
            missing_versions_arr.PushBack(miss_value, missing_versions_arr.GetAllocator());
1431
0
        }
1432
0
        rapidjson::Value value;
1433
0
        std::string version_str = rowset->get_rowset_info_str();
1434
0
        value.SetString(version_str.c_str(), version_str.length(), versions_arr.GetAllocator());
1435
0
        versions_arr.PushBack(value, versions_arr.GetAllocator());
1436
0
        last_version = ver.second;
1437
0
    }
1438
0
    root.AddMember("rowsets", versions_arr, root.GetAllocator());
1439
0
    root.AddMember("missing_rowsets", missing_versions_arr, root.GetAllocator());
1440
1441
    // print all stale rowsets' version as an array
1442
0
    rapidjson::Document stale_versions_arr;
1443
0
    stale_versions_arr.SetArray();
1444
0
    for (auto& rowset : stale_rowsets) {
1445
0
        rapidjson::Value value;
1446
0
        std::string version_str = rowset->get_rowset_info_str();
1447
0
        value.SetString(version_str.c_str(), version_str.length(),
1448
0
                        stale_versions_arr.GetAllocator());
1449
0
        stale_versions_arr.PushBack(value, stale_versions_arr.GetAllocator());
1450
0
    }
1451
0
    root.AddMember("stale_rowsets", stale_versions_arr, root.GetAllocator());
1452
1453
    // add stale version rowsets
1454
0
    root.AddMember("stale version path", path_arr, root.GetAllocator());
1455
1456
    // to json string
1457
0
    rapidjson::StringBuffer strbuf;
1458
0
    rapidjson::PrettyWriter<rapidjson::StringBuffer> writer(strbuf);
1459
0
    root.Accept(writer);
1460
0
    *json_result = std::string(strbuf.GetString());
1461
0
}
1462
1463
0
bool Tablet::do_tablet_meta_checkpoint() {
1464
0
    std::lock_guard<std::shared_mutex> store_lock(_meta_store_lock);
1465
0
    if (_newly_created_rowset_num == 0) {
1466
0
        return false;
1467
0
    }
1468
0
    if (UnixMillis() - _last_checkpoint_time <
1469
0
                config::tablet_meta_checkpoint_min_interval_secs * 1000 &&
1470
0
        _newly_created_rowset_num < config::tablet_meta_checkpoint_min_new_rowsets_num) {
1471
0
        return false;
1472
0
    }
1473
1474
    // hold read-lock other than write-lock, because it will not modify meta structure
1475
0
    std::shared_lock rdlock(_meta_lock);
1476
0
    if (tablet_state() != TABLET_RUNNING) {
1477
0
        LOG(INFO) << "tablet is under state=" << tablet_state()
1478
0
                  << ", not running, skip do checkpoint"
1479
0
                  << ", tablet=" << tablet_id();
1480
0
        return false;
1481
0
    }
1482
0
    VLOG_NOTICE << "start to do tablet meta checkpoint, tablet=" << tablet_id();
1483
0
    save_meta();
1484
    // if save meta successfully, then should remove the rowset meta existing in tablet
1485
    // meta from rowset meta store
1486
0
    for (auto& rs_meta : _tablet_meta->all_rs_metas()) {
1487
        // If we delete it from rowset manager's meta explicitly in previous checkpoint, just skip.
1488
0
        if (rs_meta->is_remove_from_rowset_meta()) {
1489
0
            continue;
1490
0
        }
1491
0
        if (RowsetMetaManager::check_rowset_meta(_data_dir->get_meta(), tablet_uid(),
1492
0
                                                 rs_meta->rowset_id())) {
1493
0
            RETURN_FALSE_IF_ERROR(RowsetMetaManager::remove(_data_dir->get_meta(), tablet_uid(),
1494
0
                                                            rs_meta->rowset_id()));
1495
0
            VLOG_NOTICE << "remove rowset id from meta store because it is already persistent with "
1496
0
                        << "tablet meta, rowset_id=" << rs_meta->rowset_id();
1497
0
        }
1498
0
        rs_meta->set_remove_from_rowset_meta();
1499
0
    }
1500
1501
    // check _stale_rs_version_map to remove meta from rowset meta store
1502
0
    for (auto& rs_meta : _tablet_meta->all_stale_rs_metas()) {
1503
        // If we delete it from rowset manager's meta explicitly in previous checkpoint, just skip.
1504
0
        if (rs_meta->is_remove_from_rowset_meta()) {
1505
0
            continue;
1506
0
        }
1507
0
        if (RowsetMetaManager::check_rowset_meta(_data_dir->get_meta(), tablet_uid(),
1508
0
                                                 rs_meta->rowset_id())) {
1509
0
            RETURN_FALSE_IF_ERROR(RowsetMetaManager::remove(_data_dir->get_meta(), tablet_uid(),
1510
0
                                                            rs_meta->rowset_id()));
1511
0
            VLOG_NOTICE << "remove rowset id from meta store because it is already persistent with "
1512
0
                        << "tablet meta, rowset_id=" << rs_meta->rowset_id();
1513
0
        }
1514
0
        rs_meta->set_remove_from_rowset_meta();
1515
0
    }
1516
1517
0
    if (keys_type() == UNIQUE_KEYS && enable_unique_key_merge_on_write()) {
1518
0
        RETURN_FALSE_IF_ERROR(TabletMetaManager::remove_old_version_delete_bitmap(
1519
0
                _data_dir, tablet_id(), max_version_unlocked()));
1520
0
    }
1521
1522
0
    _newly_created_rowset_num = 0;
1523
0
    _last_checkpoint_time = UnixMillis();
1524
0
    return true;
1525
0
}
1526
1527
0
bool Tablet::rowset_meta_is_useful(RowsetMetaSharedPtr rowset_meta) {
1528
0
    std::shared_lock rdlock(_meta_lock);
1529
0
    bool find_version = false;
1530
0
    for (auto& version_rowset : _rs_version_map) {
1531
0
        if (version_rowset.second->rowset_id() == rowset_meta->rowset_id()) {
1532
0
            return true;
1533
0
        }
1534
0
        if (version_rowset.second->contains_version(rowset_meta->version())) {
1535
0
            find_version = true;
1536
0
        }
1537
0
    }
1538
0
    for (auto& stale_version_rowset : _stale_rs_version_map) {
1539
0
        if (stale_version_rowset.second->rowset_id() == rowset_meta->rowset_id()) {
1540
0
            return true;
1541
0
        }
1542
0
        if (stale_version_rowset.second->contains_version(rowset_meta->version())) {
1543
0
            find_version = true;
1544
0
        }
1545
0
    }
1546
0
    return !find_version;
1547
0
}
1548
1549
504
bool Tablet::_contains_rowset(const RowsetId rowset_id) {
1550
3.54k
    for (auto& version_rowset : _rs_version_map) {
1551
3.54k
        if (version_rowset.second->rowset_id() == rowset_id) {
1552
0
            return true;
1553
0
        }
1554
3.54k
    }
1555
504
    for (auto& stale_version_rowset : _stale_rs_version_map) {
1556
0
        if (stale_version_rowset.second->rowset_id() == rowset_id) {
1557
0
            return true;
1558
0
        }
1559
0
    }
1560
504
    return false;
1561
504
}
1562
1563
// need check if consecutive version missing in full report
1564
// alter tablet will ignore this check
1565
void Tablet::build_tablet_report_info(TTabletInfo* tablet_info,
1566
                                      bool enable_consecutive_missing_check,
1567
0
                                      bool enable_path_check) {
1568
0
    std::shared_lock rdlock(_meta_lock);
1569
0
    tablet_info->__set_tablet_id(_tablet_meta->tablet_id());
1570
0
    tablet_info->__set_schema_hash(_tablet_meta->schema_hash());
1571
0
    tablet_info->__set_row_count(_tablet_meta->num_rows());
1572
0
    tablet_info->__set_data_size(_tablet_meta->tablet_local_size());
1573
1574
    // Here we need to report to FE if there are any missing versions of tablet.
1575
    // We start from the initial version and traverse backwards until we meet a discontinuous version.
1576
0
    Version cversion;
1577
0
    Version max_version;
1578
0
    bool has_version_cross;
1579
0
    _max_continuous_version_from_beginning_unlocked(&cversion, &max_version, &has_version_cross);
1580
    // cause publish version task runs concurrently, version may be flying
1581
    // so we add a consecutive miss check to solve this problem:
1582
    // if publish version 5 arrives but version 4 flying, we may judge replica miss version
1583
    // and set version miss in tablet_info, which makes fe treat this replica as unhealth
1584
    // and lead to other problems
1585
0
    if (enable_consecutive_missing_check) {
1586
0
        if (cversion.second < max_version.second) {
1587
0
            if (_last_missed_version == cversion.second + 1) {
1588
0
                if (MonotonicSeconds() - _last_missed_time_s >= 60) {
1589
                    // version missed for over 60 seconds
1590
0
                    tablet_info->__set_version_miss(true);
1591
0
                    _last_missed_version = -1;
1592
0
                    _last_missed_time_s = 0;
1593
0
                }
1594
0
            } else {
1595
0
                _last_missed_version = cversion.second + 1;
1596
0
                _last_missed_time_s = MonotonicSeconds();
1597
0
            }
1598
0
        }
1599
0
    } else {
1600
0
        tablet_info->__set_version_miss(cversion.second < max_version.second);
1601
0
    }
1602
1603
0
    DBUG_EXECUTE_IF("Tablet.build_tablet_report_info.version_miss", {
1604
0
        auto tablet_id = dp->param<int64>("tablet_id", -1);
1605
0
        if (tablet_id != -1 && tablet_id == _tablet_meta->tablet_id()) {
1606
0
            auto miss = dp->param<bool>("version_miss", true);
1607
0
            tablet_info->__set_version_miss(miss);
1608
0
        }
1609
0
    });
1610
1611
    // find rowset with max version
1612
0
    auto iter = _rs_version_map.find(max_version);
1613
0
    if (iter == _rs_version_map.end()) {
1614
        // If the tablet is in running state, it must not be doing schema-change. so if we can not
1615
        // access its rowsets, it means that the tablet is bad and needs to be reported to the FE
1616
        // for subsequent repairs (through the cloning task)
1617
0
        if (tablet_state() == TABLET_RUNNING) {
1618
0
            tablet_info->__set_used(false);
1619
0
        }
1620
        // For other states, FE knows that the tablet is in a certain change process, so here
1621
        // still sets the state to normal when reporting. Note that every task has an timeout,
1622
        // so if the task corresponding to this change hangs, when the task timeout, FE will know
1623
        // and perform state modification operations.
1624
0
    }
1625
1626
0
    if (tablet_state() == TABLET_RUNNING) {
1627
0
        if (has_version_cross || is_io_error_too_times() || !data_dir()->is_used()) {
1628
0
            LOG(INFO) << "report " << tablet_id() << " as bad, version_cross=" << has_version_cross
1629
0
                      << ", ioe times=" << get_io_error_times() << ", data_dir used "
1630
0
                      << data_dir()->is_used();
1631
0
            tablet_info->__set_used(false);
1632
0
        }
1633
1634
0
        if (enable_path_check) {
1635
0
            if (!_is_tablet_path_exists.exchange(true, std::memory_order_relaxed)) {
1636
0
                LOG(INFO) << "report " << tablet_id() << " as bad, tablet directory not found";
1637
0
                tablet_info->__set_used(false);
1638
0
            }
1639
0
        }
1640
0
    }
1641
1642
    // There are two cases when tablet state is TABLET_NOTREADY
1643
    // case 1: tablet is doing schema change. Fe knows it's state, doing nothing.
1644
    // case 2: tablet has finished schema change, but failed. Fe will perform recovery.
1645
0
    if (tablet_state() == TABLET_NOTREADY && is_alter_failed()) {
1646
0
        tablet_info->__set_used(false);
1647
0
    }
1648
1649
0
    if (tablet_state() == TABLET_SHUTDOWN) {
1650
0
        tablet_info->__set_used(false);
1651
0
    }
1652
1653
0
    DBUG_EXECUTE_IF("Tablet.build_tablet_report_info.used", {
1654
0
        auto tablet_id = dp->param<int64>("tablet_id", -1);
1655
0
        if (tablet_id != -1 && tablet_id == _tablet_meta->tablet_id()) {
1656
0
            auto used = dp->param<bool>("used", true);
1657
0
            LOG_WARNING("Tablet.build_tablet_report_info.used")
1658
0
                    .tag("tablet id", tablet_id)
1659
0
                    .tag("used", used);
1660
0
            tablet_info->__set_used(used);
1661
0
        } else {
1662
0
            LOG_WARNING("Tablet.build_tablet_report_info.used").tag("tablet id", tablet_id);
1663
0
        }
1664
0
    });
1665
1666
0
    int64_t total_version_count = _tablet_meta->version_count();
1667
1668
    // For compatibility.
1669
    // For old fe, it wouldn't send visible version request to be, then be's visible version is always 0.
1670
    // Let visible_version_count set to total_version_count in be's report.
1671
0
    int64_t visible_version_count = total_version_count;
1672
0
    if (auto [visible_version, _] = get_visible_version_and_time(); visible_version > 0) {
1673
0
        visible_version_count = _tablet_meta->version_count_cross_with_range({0, visible_version});
1674
0
    }
1675
    // the report version is the largest continuous version, same logic as in FE side
1676
0
    tablet_info->__set_version(cversion.second);
1677
    // Useless but it is a required filed in TTabletInfo
1678
0
    tablet_info->__set_version_hash(0);
1679
0
    tablet_info->__set_partition_id(_tablet_meta->partition_id());
1680
0
    tablet_info->__set_storage_medium(_data_dir->storage_medium());
1681
0
    tablet_info->__set_total_version_count(total_version_count);
1682
0
    tablet_info->__set_visible_version_count(visible_version_count);
1683
0
    tablet_info->__set_path_hash(_data_dir->path_hash());
1684
0
    tablet_info->__set_is_in_memory(_tablet_meta->tablet_schema()->is_in_memory());
1685
0
    tablet_info->__set_replica_id(replica_id());
1686
0
    tablet_info->__set_remote_data_size(_tablet_meta->tablet_remote_size());
1687
0
    if (_tablet_meta->cooldown_meta_id().initialized()) { // has cooldowned data
1688
0
        tablet_info->__set_cooldown_term(_cooldown_conf.term);
1689
0
        tablet_info->__set_cooldown_meta_id(_tablet_meta->cooldown_meta_id().to_thrift());
1690
0
    }
1691
0
    if (tablet_state() == TABLET_RUNNING && _tablet_meta->storage_policy_id() > 0) {
1692
        // tablet may not have cooldowned data, but the storage policy is set
1693
0
        tablet_info->__set_cooldown_term(_cooldown_conf.term);
1694
0
    }
1695
0
}
1696
1697
1
void Tablet::report_error(const Status& st) {
1698
1
    if (st.is<ErrorCode::IO_ERROR>()) {
1699
0
        ++_io_error_times;
1700
1
    } else if (st.is<ErrorCode::CORRUPTION>()) {
1701
1
        _io_error_times = config::max_tablet_io_errors + 1;
1702
1
    } else if (st.is<ErrorCode::NOT_FOUND>()) {
1703
0
        check_tablet_path_exists();
1704
0
        if (!_is_tablet_path_exists.load(std::memory_order_relaxed)) {
1705
0
            _io_error_times = config::max_tablet_io_errors + 1;
1706
0
        }
1707
0
    }
1708
1
}
1709
1710
Status Tablet::prepare_compaction_and_calculate_permits(
1711
        CompactionType compaction_type, const TabletSharedPtr& tablet,
1712
10
        std::shared_ptr<CompactionMixin>& compaction, int64_t& permits) {
1713
10
    if (compaction_type == CompactionType::CUMULATIVE_COMPACTION) {
1714
10
        MonotonicStopWatch watch;
1715
10
        watch.start();
1716
1717
10
        compaction = std::make_shared<CumulativeCompaction>(tablet->_engine, tablet);
1718
10
        DorisMetrics::instance()->cumulative_compaction_request_total->increment(1);
1719
10
        Status res = compaction->prepare_compact();
1720
10
        if (!config::disable_compaction_trace_log &&
1721
10
            watch.elapsed_time() / 1e9 > config::cumulative_compaction_trace_threshold) {
1722
0
            std::stringstream ss;
1723
0
            compaction->runtime_profile()->pretty_print(&ss);
1724
0
            LOG(WARNING) << "prepare cumulative compaction cost " << watch.elapsed_time() / 1e9
1725
0
                         << std::endl
1726
0
                         << ss.str();
1727
0
        }
1728
1729
10
        if (!res.ok()) {
1730
0
            tablet->set_last_cumu_compaction_failure_time(UnixMillis());
1731
0
            permits = 0;
1732
0
            if (!res.is<CUMULATIVE_NO_SUITABLE_VERSION>()) {
1733
0
                DorisMetrics::instance()->cumulative_compaction_request_failed->increment(1);
1734
0
                return Status::InternalError("prepare cumulative compaction with err: {}", res);
1735
0
            }
1736
            // return OK if OLAP_ERR_CUMULATIVE_NO_SUITABLE_VERSION, so that we don't need to
1737
            // print too much useless logs.
1738
            // And because we set permits to 0, so even if we return OK here, nothing will be done.
1739
0
            return Status::OK();
1740
0
        }
1741
10
    } else if (compaction_type == CompactionType::BASE_COMPACTION) {
1742
0
        MonotonicStopWatch watch;
1743
0
        watch.start();
1744
1745
0
        compaction = std::make_shared<BaseCompaction>(tablet->_engine, tablet);
1746
0
        DorisMetrics::instance()->base_compaction_request_total->increment(1);
1747
0
        Status res = compaction->prepare_compact();
1748
0
        if (!config::disable_compaction_trace_log &&
1749
0
            watch.elapsed_time() / 1e9 > config::base_compaction_trace_threshold) {
1750
0
            std::stringstream ss;
1751
0
            compaction->runtime_profile()->pretty_print(&ss);
1752
0
            LOG(WARNING) << "prepare base compaction cost " << watch.elapsed_time() / 1e9
1753
0
                         << std::endl
1754
0
                         << ss.str();
1755
0
        }
1756
1757
0
        tablet->set_last_base_compaction_status(res.to_string());
1758
0
        if (!res.ok()) {
1759
0
            tablet->set_last_base_compaction_failure_time(UnixMillis());
1760
0
            permits = 0;
1761
0
            if (!res.is<BE_NO_SUITABLE_VERSION>()) {
1762
0
                DorisMetrics::instance()->base_compaction_request_failed->increment(1);
1763
0
                return Status::InternalError("prepare base compaction with err: {}", res);
1764
0
            }
1765
            // return OK if OLAP_ERR_BE_NO_SUITABLE_VERSION, so that we don't need to
1766
            // print too much useless logs.
1767
            // And because we set permits to 0, so even if we return OK here, nothing will be done.
1768
0
            return Status::OK();
1769
0
        }
1770
0
    } else {
1771
0
        DCHECK_EQ(compaction_type, CompactionType::FULL_COMPACTION);
1772
1773
0
        compaction = std::make_shared<FullCompaction>(tablet->_engine, tablet);
1774
0
        Status res = compaction->prepare_compact();
1775
0
        if (!res.ok()) {
1776
0
            tablet->set_last_full_compaction_failure_time(UnixMillis());
1777
0
            permits = 0;
1778
0
            if (!res.is<FULL_NO_SUITABLE_VERSION>()) {
1779
0
                return Status::InternalError("prepare full compaction with err: {}", res);
1780
0
            }
1781
            // return OK if OLAP_ERR_BE_NO_SUITABLE_VERSION, so that we don't need to
1782
            // print too much useless logs.
1783
            // And because we set permits to 0, so even if we return OK here, nothing will be done.
1784
0
            return Status::OK();
1785
0
        }
1786
0
    }
1787
1788
    // Time series policy does not rely on permits, it uses goal size to control memory
1789
10
    if (tablet->tablet_meta()->compaction_policy() == CUMULATIVE_TIME_SERIES_POLICY) {
1790
        // permits = 0 means that prepare_compaction failed
1791
0
        permits = 1;
1792
10
    } else {
1793
10
        permits = compaction->get_compaction_permits();
1794
10
    }
1795
10
    return Status::OK();
1796
10
}
1797
1798
0
void Tablet::execute_single_replica_compaction(SingleReplicaCompaction& compaction) {
1799
0
    Status res = compaction.execute_compact();
1800
0
    if (!res.ok()) {
1801
0
        set_last_failure_time(this, compaction, UnixMillis());
1802
0
        set_last_single_compaction_failure_status(res.to_string());
1803
0
        if (res.is<CANCELLED>()) {
1804
0
            DorisMetrics::instance()->single_compaction_request_cancelled->increment(1);
1805
            // "CANCELLED" indicates that the peer has not performed compaction,
1806
            // wait for the peer to perform compaction
1807
0
            set_skip_compaction(true, compaction.real_compact_type(), UnixSeconds());
1808
0
            VLOG_CRITICAL << "Cannel fetching from the remote peer. res=" << res
1809
0
                          << ", tablet=" << tablet_id();
1810
0
        } else {
1811
0
            DorisMetrics::instance()->single_compaction_request_failed->increment(1);
1812
0
            LOG(WARNING) << "failed to do single replica compaction. res=" << res
1813
0
                         << ", tablet=" << tablet_id();
1814
0
        }
1815
0
        return;
1816
0
    }
1817
0
    set_last_failure_time(this, compaction, 0);
1818
0
}
1819
1820
17
bool Tablet::should_fetch_from_peer() {
1821
17
    return tablet_meta()->tablet_schema()->enable_single_replica_compaction() &&
1822
17
           _engine.should_fetch_from_peer(tablet_id());
1823
17
}
1824
1825
3
std::vector<Version> Tablet::get_all_local_versions() {
1826
3
    std::vector<Version> local_versions;
1827
3
    {
1828
3
        std::shared_lock rlock(_meta_lock);
1829
63
        for (const auto& [version, rs] : _rs_version_map) {
1830
63
            if (rs->is_local()) {
1831
53
                local_versions.emplace_back(version);
1832
53
            }
1833
63
        }
1834
3
    }
1835
3
    std::sort(local_versions.begin(), local_versions.end(),
1836
318
              [](const Version& left, const Version& right) { return left.first < right.first; });
1837
3
    return local_versions;
1838
3
}
1839
1840
0
void Tablet::execute_compaction(CompactionMixin& compaction) {
1841
0
    signal::tablet_id = tablet_id();
1842
1843
0
    MonotonicStopWatch watch;
1844
0
    watch.start();
1845
1846
0
    Status res = [&]() { RETURN_IF_CATCH_EXCEPTION({ return compaction.execute_compact(); }); }();
1847
1848
0
    if (!res.ok()) [[unlikely]] {
1849
0
        set_last_failure_time(this, compaction, UnixMillis());
1850
0
        LOG(WARNING) << "failed to do " << compaction.compaction_name()
1851
0
                     << ", tablet=" << tablet_id() << " : " << res;
1852
0
    } else {
1853
0
        set_last_failure_time(this, compaction, 0);
1854
0
    }
1855
1856
0
    if (!config::disable_compaction_trace_log) {
1857
0
        auto need_trace = [&compaction, &watch] {
1858
0
            return compaction.compaction_type() == ReaderType::READER_CUMULATIVE_COMPACTION
1859
0
                           ? watch.elapsed_time() / 1e9 >
1860
0
                                     config::cumulative_compaction_trace_threshold
1861
0
                   : compaction.compaction_type() == ReaderType::READER_BASE_COMPACTION
1862
0
                           ? watch.elapsed_time() / 1e9 > config::base_compaction_trace_threshold
1863
0
                           : false;
1864
0
        };
1865
0
        if (need_trace()) {
1866
0
            std::stringstream ss;
1867
0
            compaction.runtime_profile()->pretty_print(&ss);
1868
0
            LOG(WARNING) << "execute " << compaction.compaction_name() << " cost "
1869
0
                         << watch.elapsed_time() / 1e9 << std::endl
1870
0
                         << ss.str();
1871
0
        }
1872
0
    }
1873
0
}
1874
1875
91
Status Tablet::create_initial_rowset(const int64_t req_version) {
1876
91
    if (req_version < 1) {
1877
0
        return Status::Error<CE_CMD_PARAMS_ERROR>(
1878
0
                "init version of tablet should at least 1. req.ver={}", req_version);
1879
0
    }
1880
91
    Version version(0, req_version);
1881
91
    RowsetSharedPtr new_rowset;
1882
    // there is no data in init rowset, so overlapping info is unknown.
1883
91
    RowsetWriterContext context;
1884
91
    context.version = version;
1885
91
    context.rowset_state = VISIBLE;
1886
91
    context.segments_overlap = OVERLAP_UNKNOWN;
1887
91
    context.tablet_schema = tablet_schema();
1888
91
    context.newest_write_timestamp = UnixSeconds();
1889
91
    auto rs_writer = DORIS_TRY(create_rowset_writer(context, false));
1890
91
    RETURN_IF_ERROR(rs_writer->flush());
1891
91
    RETURN_IF_ERROR(rs_writer->build(new_rowset));
1892
91
    RETURN_IF_ERROR(add_rowset(std::move(new_rowset)));
1893
91
    set_cumulative_layer_point(req_version + 1);
1894
91
    return Status::OK();
1895
91
}
1896
1897
Result<std::unique_ptr<RowsetWriter>> Tablet::create_rowset_writer(RowsetWriterContext& context,
1898
121
                                                                   bool vertical) {
1899
121
    context.rowset_id = _engine.next_rowset_id();
1900
121
    _init_context_common_fields(context);
1901
121
    return RowsetFactory::create_rowset_writer(_engine, context, vertical);
1902
121
}
1903
1904
// create a rowset writer with rowset_id and seg_id
1905
// after writer, merge this transient rowset with original rowset
1906
Result<std::unique_ptr<RowsetWriter>> Tablet::create_transient_rowset_writer(
1907
        const Rowset& rowset, std::shared_ptr<PartialUpdateInfo> partial_update_info,
1908
0
        int64_t txn_expiration) {
1909
0
    RowsetWriterContext context;
1910
0
    context.rowset_state = PREPARED;
1911
0
    context.segments_overlap = OVERLAPPING;
1912
0
    context.tablet_schema = std::make_shared<TabletSchema>();
1913
    // During a partial update, the extracted columns of a variant should not be included in the tablet schema.
1914
    // This is because the partial update for a variant needs to ignore the extracted columns.
1915
    // Otherwise, the schema types in different rowsets might be inconsistent. When performing a partial update,
1916
    // the complete variant is constructed by reading all the sub-columns of the variant.
1917
0
    context.tablet_schema = rowset.tablet_schema()->copy_without_variant_extracted_columns();
1918
0
    context.newest_write_timestamp = UnixSeconds();
1919
0
    context.tablet_id = table_id();
1920
0
    context.enable_segcompaction = false;
1921
    // ATTN: context.tablet is a shared_ptr, can't simply set it's value to `this`. We should
1922
    // get the shared_ptr from tablet_manager.
1923
0
    auto tablet = _engine.tablet_manager()->get_tablet(tablet_id());
1924
0
    if (!tablet) {
1925
0
        LOG(WARNING) << "cant find tablet by tablet_id=" << tablet_id();
1926
0
        return ResultError(Status::NotFound("cant find tablet by tablet_id={}", tablet_id()));
1927
0
    }
1928
0
    context.tablet = tablet;
1929
0
    context.write_type = DataWriteType::TYPE_DIRECT;
1930
0
    context.partial_update_info = std::move(partial_update_info);
1931
0
    context.is_transient_rowset_writer = true;
1932
0
    return create_transient_rowset_writer(context, rowset.rowset_id())
1933
0
            .transform([&](auto&& writer) {
1934
0
                writer->set_segment_start_id(rowset.num_segments());
1935
0
                return writer;
1936
0
            });
1937
0
}
1938
1939
Result<std::unique_ptr<RowsetWriter>> Tablet::create_transient_rowset_writer(
1940
0
        RowsetWriterContext& context, const RowsetId& rowset_id) {
1941
0
    context.rowset_id = rowset_id;
1942
0
    _init_context_common_fields(context);
1943
0
    return RowsetFactory::create_rowset_writer(_engine, context, false);
1944
0
}
1945
1946
121
void Tablet::_init_context_common_fields(RowsetWriterContext& context) {
1947
121
    context.tablet_uid = tablet_uid();
1948
121
    context.tablet_id = tablet_id();
1949
121
    context.partition_id = partition_id();
1950
121
    context.tablet_schema_hash = schema_hash();
1951
121
    context.rowset_type = tablet_meta()->preferred_rowset_type();
1952
    // Alpha Rowset will be removed in the future, so that if the tablet's default rowset type is
1953
    // alpha rowset, then set the newly created rowset to storage engine's default rowset.
1954
121
    if (context.rowset_type == ALPHA_ROWSET) {
1955
0
        context.rowset_type = _engine.default_rowset_type();
1956
0
    }
1957
1958
121
    if (context.is_local_rowset()) {
1959
121
        context.tablet_path = _tablet_path;
1960
121
    }
1961
1962
121
    context.data_dir = data_dir();
1963
121
    context.enable_unique_key_merge_on_write = enable_unique_key_merge_on_write();
1964
121
}
1965
1966
197
Status Tablet::create_rowset(const RowsetMetaSharedPtr& rowset_meta, RowsetSharedPtr* rowset) {
1967
197
    return RowsetFactory::create_rowset(_tablet_meta->tablet_schema(),
1968
197
                                        rowset_meta->is_local() ? _tablet_path : "", rowset_meta,
1969
197
                                        rowset);
1970
197
}
1971
1972
7
Status Tablet::cooldown(RowsetSharedPtr rowset) {
1973
7
    std::unique_lock schema_change_lock(_schema_change_lock, std::try_to_lock);
1974
7
    if (!schema_change_lock.owns_lock()) {
1975
0
        return Status::Error<TRY_LOCK_FAILED>("try schema_change_lock failed");
1976
0
    }
1977
    // Check executing serially with compaction task.
1978
7
    std::unique_lock base_compaction_lock(_base_compaction_lock, std::try_to_lock);
1979
7
    if (!base_compaction_lock.owns_lock()) {
1980
0
        return Status::Error<TRY_LOCK_FAILED>("try base_compaction_lock failed");
1981
0
    }
1982
7
    std::unique_lock cumu_compaction_lock(_cumulative_compaction_lock, std::try_to_lock);
1983
7
    if (!cumu_compaction_lock.owns_lock()) {
1984
0
        return Status::Error<TRY_LOCK_FAILED>("try cumu_compaction_lock failed");
1985
0
    }
1986
7
    std::shared_lock cooldown_conf_rlock(_cooldown_conf_lock);
1987
7
    if (_cooldown_conf.cooldown_replica_id <= 0) { // wait for FE to push cooldown conf
1988
2
        return Status::InternalError("invalid cooldown_replica_id");
1989
2
    }
1990
1991
5
    if (_cooldown_conf.cooldown_replica_id == replica_id()) {
1992
        // this replica is cooldown replica
1993
5
        RETURN_IF_ERROR(_cooldown_data(std::move(rowset)));
1994
5
    } else {
1995
0
        Status st = _follow_cooldowned_data();
1996
0
        if (UNLIKELY(!st.ok())) {
1997
0
            _last_failed_follow_cooldown_time = time(nullptr);
1998
0
            return st;
1999
0
        }
2000
0
        _last_failed_follow_cooldown_time = 0;
2001
0
    }
2002
5
    return Status::OK();
2003
5
}
2004
2005
// hold SHARED `cooldown_conf_lock`
2006
5
Status Tablet::_cooldown_data(RowsetSharedPtr rowset) {
2007
5
    DCHECK(_cooldown_conf.cooldown_replica_id == replica_id());
2008
2009
5
    auto storage_resource = DORIS_TRY(get_resource_by_storage_policy_id(storage_policy_id()));
2010
5
    RowsetSharedPtr old_rowset = nullptr;
2011
2012
5
    if (rowset) {
2013
0
        const auto& rowset_id = rowset->rowset_id();
2014
0
        const auto& rowset_version = rowset->version();
2015
0
        std::shared_lock meta_rlock(_meta_lock);
2016
0
        auto iter = _rs_version_map.find(rowset_version);
2017
0
        if (iter != _rs_version_map.end() && iter->second->rowset_id() == rowset_id) {
2018
0
            old_rowset = rowset;
2019
0
        }
2020
0
    }
2021
2022
5
    if (!old_rowset) {
2023
5
        old_rowset = pick_cooldown_rowset();
2024
5
    }
2025
2026
5
    if (!old_rowset) {
2027
0
        LOG(INFO) << "cannot pick cooldown rowset in tablet " << tablet_id();
2028
0
        return Status::OK();
2029
0
    }
2030
2031
5
    RowsetId new_rowset_id = _engine.next_rowset_id();
2032
5
    auto pending_rs_guard = _engine.pending_remote_rowsets().add(new_rowset_id);
2033
5
    Status st;
2034
5
    Defer defer {[&] {
2035
5
        if (!st.ok()) {
2036
            // reclaim the incomplete rowset data in remote storage
2037
0
            record_unused_remote_rowset(new_rowset_id, storage_resource.fs->id(),
2038
0
                                        old_rowset->num_segments());
2039
0
        }
2040
5
    }};
2041
5
    auto start = std::chrono::steady_clock::now();
2042
5
    if (st = old_rowset->upload_to(storage_resource, new_rowset_id); !st.ok()) {
2043
0
        return st;
2044
0
    }
2045
2046
5
    auto duration = std::chrono::duration<float>(std::chrono::steady_clock::now() - start);
2047
5
    LOG(INFO) << "Upload rowset " << old_rowset->version() << " " << new_rowset_id.to_string()
2048
5
              << " to " << storage_resource.fs->root_path().native()
2049
5
              << ", tablet_id=" << tablet_id() << ", duration=" << duration.count()
2050
5
              << ", capacity=" << old_rowset->total_disk_size()
2051
5
              << ", tp=" << old_rowset->total_disk_size() / duration.count()
2052
5
              << ", old rowset_id=" << old_rowset->rowset_id().to_string();
2053
2054
    // gen a new rowset
2055
5
    auto new_rowset_meta = std::make_shared<RowsetMeta>();
2056
5
    new_rowset_meta->init(old_rowset->rowset_meta().get());
2057
5
    new_rowset_meta->set_rowset_id(new_rowset_id);
2058
5
    new_rowset_meta->set_remote_storage_resource(std::move(storage_resource));
2059
5
    new_rowset_meta->set_creation_time(time(nullptr));
2060
5
    UniqueId cooldown_meta_id = UniqueId::gen_uid();
2061
5
    RowsetSharedPtr new_rowset;
2062
5
    RETURN_IF_ERROR(RowsetFactory::create_rowset(_tablet_meta->tablet_schema(), "", new_rowset_meta,
2063
5
                                                 &new_rowset));
2064
2065
5
    {
2066
5
        std::unique_lock meta_wlock(_meta_lock);
2067
5
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2068
5
        if (tablet_state() == TABLET_RUNNING) {
2069
5
            RETURN_IF_ERROR(delete_rowsets({std::move(old_rowset)}, false));
2070
5
            add_rowsets({std::move(new_rowset)});
2071
            // TODO(plat1ko): process primary key
2072
5
            _tablet_meta->set_cooldown_meta_id(cooldown_meta_id);
2073
5
        }
2074
5
    }
2075
5
    {
2076
5
        std::shared_lock meta_rlock(_meta_lock);
2077
5
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2078
5
        save_meta();
2079
5
    }
2080
    // Upload cooldowned rowset meta to remote fs
2081
    // ATTN: Even if it is an empty rowset, in order for the followers to synchronize, the coolown meta must be
2082
    // uploaded, otherwise followers may never completely cooldown.
2083
5
    if (auto t = _engine.tablet_manager()->get_tablet(tablet_id());
2084
5
        t != nullptr) { // `t` can be nullptr if it has been dropped
2085
5
        async_write_cooldown_meta(std::move(t));
2086
5
    }
2087
5
    return Status::OK();
2088
5
}
2089
2090
// hold SHARED `cooldown_conf_lock`
2091
Status Tablet::_read_cooldown_meta(const StorageResource& storage_resource,
2092
0
                                   TabletMetaPB* tablet_meta_pb) {
2093
0
    std::string remote_meta_path = storage_resource.cooldown_tablet_meta_path(
2094
0
            tablet_id(), _cooldown_conf.cooldown_replica_id, _cooldown_conf.term);
2095
0
    io::FileReaderSPtr tablet_meta_reader;
2096
0
    RETURN_IF_ERROR(storage_resource.fs->open_file(remote_meta_path, &tablet_meta_reader));
2097
0
    auto file_size = tablet_meta_reader->size();
2098
0
    size_t bytes_read;
2099
0
    auto buf = std::unique_ptr<uint8_t[]>(new uint8_t[file_size]);
2100
0
    RETURN_IF_ERROR(tablet_meta_reader->read_at(0, {buf.get(), file_size}, &bytes_read));
2101
0
    RETURN_IF_ERROR(tablet_meta_reader->close());
2102
0
    if (!tablet_meta_pb->ParseFromArray(buf.get(), file_size)) {
2103
0
        return Status::InternalError("malformed tablet meta, path={}/{}",
2104
0
                                     storage_resource.fs->root_path().native(), remote_meta_path);
2105
0
    }
2106
0
    return Status::OK();
2107
0
}
2108
2109
// `rs_metas` MUST already be sorted by `RowsetMeta::comparator`
2110
4
Status check_version_continuity(const std::vector<RowsetMetaSharedPtr>& rs_metas) {
2111
4
    if (rs_metas.size() < 2) {
2112
2
        return Status::OK();
2113
2
    }
2114
2
    auto prev = rs_metas.begin();
2115
5
    for (auto it = rs_metas.begin() + 1; it != rs_metas.end(); ++it) {
2116
3
        if ((*prev)->end_version() + 1 != (*it)->start_version()) {
2117
0
            return Status::InternalError("versions are not continuity: prev={} cur={}",
2118
0
                                         (*prev)->version().to_string(),
2119
0
                                         (*it)->version().to_string());
2120
0
        }
2121
3
        prev = it;
2122
3
    }
2123
2
    return Status::OK();
2124
2
}
2125
2126
// It's guaranteed the write cooldown meta task would be invoked at the end unless BE crashes
2127
// one tablet would at most have one async task to be done
2128
5
void Tablet::async_write_cooldown_meta(TabletSharedPtr tablet) {
2129
5
    ExecEnv::GetInstance()->write_cooldown_meta_executors()->submit(std::move(tablet));
2130
5
}
2131
2132
2
bool Tablet::update_cooldown_conf(int64_t cooldown_term, int64_t cooldown_replica_id) {
2133
2
    std::unique_lock wlock(_cooldown_conf_lock, std::try_to_lock);
2134
2
    if (!wlock.owns_lock()) {
2135
0
        LOG(INFO) << "try cooldown_conf_lock failed, tablet_id=" << tablet_id();
2136
0
        return false;
2137
0
    }
2138
2
    if (cooldown_term <= _cooldown_conf.term) {
2139
0
        return false;
2140
0
    }
2141
2
    LOG(INFO) << "update cooldown conf. tablet_id=" << tablet_id()
2142
2
              << " cooldown_replica_id: " << _cooldown_conf.cooldown_replica_id << " -> "
2143
2
              << cooldown_replica_id << ", cooldown_term: " << _cooldown_conf.term << " -> "
2144
2
              << cooldown_term;
2145
2
    _cooldown_conf.cooldown_replica_id = cooldown_replica_id;
2146
2
    _cooldown_conf.term = cooldown_term;
2147
2
    return true;
2148
2
}
2149
2150
4
Status Tablet::write_cooldown_meta() {
2151
4
    std::shared_lock rlock(_cooldown_conf_lock);
2152
4
    if (_cooldown_conf.cooldown_replica_id != _tablet_meta->replica_id()) {
2153
0
        return Status::Aborted<false>("not cooldown replica({} vs {}) tablet_id={}",
2154
0
                                      _tablet_meta->replica_id(),
2155
0
                                      _cooldown_conf.cooldown_replica_id, tablet_id());
2156
0
    }
2157
2158
4
    auto storage_resource = DORIS_TRY(get_resource_by_storage_policy_id(storage_policy_id()));
2159
2160
4
    std::vector<RowsetMetaSharedPtr> cooldowned_rs_metas;
2161
4
    UniqueId cooldown_meta_id;
2162
4
    {
2163
4
        std::shared_lock meta_rlock(_meta_lock);
2164
9
        for (auto& rs_meta : _tablet_meta->all_rs_metas()) {
2165
9
            if (!rs_meta->is_local()) {
2166
7
                cooldowned_rs_metas.push_back(rs_meta);
2167
7
            }
2168
9
        }
2169
4
        cooldown_meta_id = _tablet_meta->cooldown_meta_id();
2170
4
    }
2171
4
    if (cooldowned_rs_metas.empty()) {
2172
0
        LOG(INFO) << "no cooldown meta to write, tablet_id=" << tablet_id();
2173
0
        return Status::OK();
2174
0
    }
2175
4
    std::sort(cooldowned_rs_metas.begin(), cooldowned_rs_metas.end(), RowsetMeta::comparator);
2176
4
    DCHECK(cooldowned_rs_metas.front()->start_version() == 0);
2177
    // If version not continuous, it must be a bug
2178
4
    if (auto st = check_version_continuity(cooldowned_rs_metas); !st.ok()) {
2179
0
        DCHECK(st.ok()) << st << " tablet_id=" << tablet_id();
2180
0
        st.set_code(ABORTED);
2181
0
        return st;
2182
0
    }
2183
2184
4
    TabletMetaPB tablet_meta_pb;
2185
4
    auto* rs_metas = tablet_meta_pb.mutable_rs_metas();
2186
4
    rs_metas->Reserve(cooldowned_rs_metas.size());
2187
7
    for (auto& rs_meta : cooldowned_rs_metas) {
2188
7
        rs_metas->Add(rs_meta->get_rowset_pb());
2189
7
    }
2190
4
    tablet_meta_pb.mutable_cooldown_meta_id()->set_hi(cooldown_meta_id.hi);
2191
4
    tablet_meta_pb.mutable_cooldown_meta_id()->set_lo(cooldown_meta_id.lo);
2192
2193
4
    std::string remote_meta_path = storage_resource.cooldown_tablet_meta_path(
2194
4
            tablet_id(), _cooldown_conf.cooldown_replica_id, _cooldown_conf.term);
2195
4
    io::FileWriterPtr tablet_meta_writer;
2196
    // FIXME(plat1ko): What if object store permanently unavailable?
2197
4
    RETURN_IF_ERROR(storage_resource.fs->create_file(remote_meta_path, &tablet_meta_writer));
2198
4
    auto val = tablet_meta_pb.SerializeAsString();
2199
4
    RETURN_IF_ERROR(tablet_meta_writer->append({val.data(), val.size()}));
2200
4
    return tablet_meta_writer->close();
2201
4
}
2202
2203
// hold SHARED `cooldown_conf_lock`
2204
0
Status Tablet::_follow_cooldowned_data() {
2205
0
    DCHECK(_cooldown_conf.cooldown_replica_id != replica_id());
2206
0
    LOG(INFO) << "try to follow cooldowned data. tablet_id=" << tablet_id()
2207
0
              << " cooldown_replica_id=" << _cooldown_conf.cooldown_replica_id
2208
0
              << " local replica=" << replica_id();
2209
2210
0
    auto storage_resource = DORIS_TRY(get_resource_by_storage_policy_id(storage_policy_id()));
2211
    // MUST executing serially with cold data compaction, because compaction input rowsets may be deleted by this function
2212
0
    std::unique_lock cold_compaction_lock(_cold_compaction_lock, std::try_to_lock);
2213
0
    if (!cold_compaction_lock.owns_lock()) {
2214
0
        return Status::Error<TRY_LOCK_FAILED>("try cold_compaction_lock failed");
2215
0
    }
2216
2217
0
    TabletMetaPB cooldown_meta_pb;
2218
0
    auto st = _read_cooldown_meta(storage_resource, &cooldown_meta_pb);
2219
0
    if (!st.ok()) {
2220
0
        LOG(INFO) << "cannot read cooldown meta: " << st;
2221
0
        return Status::InternalError<false>("cannot read cooldown meta");
2222
0
    }
2223
0
    DCHECK(cooldown_meta_pb.rs_metas_size() > 0);
2224
0
    if (_tablet_meta->cooldown_meta_id() == cooldown_meta_pb.cooldown_meta_id()) {
2225
        // cooldowned rowsets are same, no need to follow
2226
0
        return Status::OK();
2227
0
    }
2228
2229
0
    int64_t cooldowned_version = cooldown_meta_pb.rs_metas().rbegin()->end_version();
2230
2231
0
    std::vector<RowsetSharedPtr> overlap_rowsets;
2232
0
    bool version_aligned = false;
2233
2234
    // Holding these to delete rowsets' shared ptr until save meta can avoid trash sweeping thread
2235
    // deleting these rowsets' files before rowset meta has been removed from disk, which may cause
2236
    // data loss when BE reboot before save meta to disk.
2237
0
    std::vector<RowsetSharedPtr> to_delete;
2238
0
    std::vector<RowsetSharedPtr> to_add;
2239
2240
0
    {
2241
0
        std::lock_guard wlock(_meta_lock);
2242
0
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2243
0
        if (tablet_state() != TABLET_RUNNING) {
2244
0
            return Status::InternalError<false>("tablet not running");
2245
0
        }
2246
2247
0
        for (auto& [v, rs] : _rs_version_map) {
2248
0
            if (v.second <= cooldowned_version) {
2249
0
                overlap_rowsets.push_back(rs);
2250
0
                if (!version_aligned && v.second == cooldowned_version) {
2251
0
                    version_aligned = true;
2252
0
                }
2253
0
            } else if (!rs->is_local()) {
2254
0
                return Status::InternalError<false>(
2255
0
                        "cooldowned version larger than that to follow with cooldown version {}",
2256
0
                        cooldowned_version);
2257
0
            }
2258
0
        }
2259
2260
0
        if (!version_aligned) {
2261
0
            return Status::InternalError<false>("cooldowned version is not aligned with version {}",
2262
0
                                                cooldowned_version);
2263
0
        }
2264
2265
0
        std::sort(overlap_rowsets.begin(), overlap_rowsets.end(), Rowset::comparator);
2266
2267
        // Find different rowset in `overlap_rowsets` and `cooldown_meta_pb.rs_metas`
2268
0
        auto rs_pb_it = cooldown_meta_pb.rs_metas().begin();
2269
0
        auto rs_it = overlap_rowsets.begin();
2270
0
        for (; rs_pb_it != cooldown_meta_pb.rs_metas().end() && rs_it != overlap_rowsets.end();
2271
0
             ++rs_pb_it, ++rs_it) {
2272
0
            if (rs_pb_it->rowset_id_v2() != (*rs_it)->rowset_id().to_string()) {
2273
0
                break;
2274
0
            }
2275
0
        }
2276
2277
0
        to_delete.assign(rs_it, overlap_rowsets.end());
2278
0
        to_add.reserve(cooldown_meta_pb.rs_metas().end() - rs_pb_it);
2279
0
        for (; rs_pb_it != cooldown_meta_pb.rs_metas().end(); ++rs_pb_it) {
2280
0
            auto rs_meta = std::make_shared<RowsetMeta>();
2281
0
            rs_meta->init_from_pb(*rs_pb_it);
2282
0
            RowsetSharedPtr rs;
2283
0
            RETURN_IF_ERROR(
2284
0
                    RowsetFactory::create_rowset(_tablet_meta->tablet_schema(), "", rs_meta, &rs));
2285
0
            to_add.push_back(std::move(rs));
2286
0
        }
2287
        // Note: We CANNOT call `modify_rowsets` here because `modify_rowsets` cannot process version graph correctly.
2288
0
        RETURN_IF_ERROR(delete_rowsets(to_delete, false));
2289
0
        add_rowsets(to_add);
2290
        // TODO(plat1ko): process primary key
2291
0
        _tablet_meta->set_cooldown_meta_id(cooldown_meta_pb.cooldown_meta_id());
2292
0
    }
2293
2294
0
    {
2295
0
        std::lock_guard rlock(_meta_lock);
2296
0
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2297
0
        save_meta();
2298
0
    }
2299
2300
0
    if (!to_add.empty()) {
2301
0
        LOG(INFO) << "modify rowsets when follow cooldowned data, tablet_id=" << tablet_id()
2302
0
                  << [&]() {
2303
0
                         std::stringstream ss;
2304
0
                         ss << " delete rowsets:\n";
2305
0
                         for (auto&& rs : to_delete) {
2306
0
                             ss << rs->version() << ' ' << rs->rowset_id() << '\n';
2307
0
                         }
2308
0
                         ss << "add rowsets:\n";
2309
0
                         for (auto&& rs : to_add) {
2310
0
                             ss << rs->version() << ' ' << rs->rowset_id() << '\n';
2311
0
                         }
2312
0
                         return ss.str();
2313
0
                     }();
2314
0
    }
2315
2316
0
    return Status::OK();
2317
0
}
2318
2319
10
bool Tablet::_has_data_to_cooldown() {
2320
10
    int64_t min_local_version = std::numeric_limits<int64_t>::max();
2321
10
    RowsetSharedPtr rowset;
2322
10
    std::shared_lock meta_rlock(_meta_lock);
2323
    // Ususally once the tablet has done cooldown successfully then the first
2324
    // rowset would always be remote rowset
2325
10
    bool has_cooldowned = false;
2326
29
    for (const auto& [_, rs] : _rs_version_map) {
2327
29
        if (!rs->is_local()) {
2328
3
            has_cooldowned = true;
2329
3
            break;
2330
3
        }
2331
29
    }
2332
32
    for (auto& [v, rs] : _rs_version_map) {
2333
32
        auto predicate = rs->is_local() && v.first < min_local_version;
2334
32
        if (!has_cooldowned) {
2335
25
            predicate = predicate && (rs->data_disk_size() > 0);
2336
25
        }
2337
32
        if (predicate) {
2338
            // this is a local rowset and has data
2339
25
            min_local_version = v.first;
2340
25
            rowset = rs;
2341
25
        }
2342
32
    }
2343
2344
10
    int64_t newest_cooldown_time = 0;
2345
10
    if (rowset != nullptr) {
2346
9
        newest_cooldown_time = _get_newest_cooldown_time(rowset);
2347
9
    }
2348
2349
10
    return (newest_cooldown_time != 0) && (newest_cooldown_time < UnixSeconds());
2350
10
}
2351
2352
9
RowsetSharedPtr Tablet::pick_cooldown_rowset() {
2353
9
    RowsetSharedPtr rowset;
2354
2355
9
    if (!_has_data_to_cooldown()) {
2356
1
        return nullptr;
2357
1
    }
2358
2359
    // TODO(plat1ko): should we maintain `cooldowned_version` in `Tablet`?
2360
8
    int64_t cooldowned_version = -1;
2361
    // We pick the rowset with smallest start version in local.
2362
8
    int64_t min_local_version = std::numeric_limits<int64_t>::max();
2363
8
    {
2364
8
        std::shared_lock meta_rlock(_meta_lock);
2365
26
        for (auto& [v, rs] : _rs_version_map) {
2366
26
            if (!rs->is_local()) {
2367
4
                cooldowned_version = std::max(cooldowned_version, v.second);
2368
22
            } else if (v.first < min_local_version) { // this is a local rowset
2369
22
                min_local_version = v.first;
2370
22
                rowset = rs;
2371
22
            }
2372
26
        }
2373
8
    }
2374
8
    if (!rowset) {
2375
0
        return nullptr;
2376
0
    }
2377
8
    if (tablet_footprint() == 0) {
2378
0
        VLOG_DEBUG << "skip cooldown due to empty tablet_id = " << tablet_id();
2379
0
        return nullptr;
2380
0
    }
2381
8
    if (min_local_version != cooldowned_version + 1) { // ensure version continuity
2382
0
        if (UNLIKELY(cooldowned_version != -1)) {
2383
0
            LOG(WARNING) << "version not continuous. tablet_id=" << tablet_id()
2384
0
                         << " cooldowned_version=" << cooldowned_version
2385
0
                         << " min_local_version=" << min_local_version;
2386
0
        }
2387
0
        return nullptr;
2388
0
    }
2389
8
    return rowset;
2390
8
}
2391
2392
12
int64_t Tablet::_get_newest_cooldown_time(const RowsetSharedPtr& rowset) {
2393
12
    int64_t id = storage_policy_id();
2394
12
    if (id <= 0) {
2395
0
        VLOG_DEBUG << "tablet does not need cooldown, tablet id: " << tablet_id();
2396
0
        return 0;
2397
0
    }
2398
12
    auto storage_policy = get_storage_policy(id);
2399
12
    if (!storage_policy) {
2400
0
        LOG(WARNING) << "Cannot get storage policy: " << id;
2401
0
        return 0;
2402
0
    }
2403
12
    auto cooldown_ttl_sec = storage_policy->cooldown_ttl;
2404
12
    auto cooldown_datetime = storage_policy->cooldown_datetime;
2405
12
    int64_t newest_cooldown_time = std::numeric_limits<int64_t>::max();
2406
2407
12
    if (cooldown_ttl_sec >= 0) {
2408
9
        newest_cooldown_time = rowset->newest_write_timestamp() + cooldown_ttl_sec;
2409
9
    }
2410
12
    if (cooldown_datetime > 0) {
2411
10
        newest_cooldown_time = std::min(newest_cooldown_time, cooldown_datetime);
2412
10
    }
2413
2414
12
    return newest_cooldown_time;
2415
12
}
2416
2417
4
RowsetSharedPtr Tablet::need_cooldown(int64_t* cooldown_timestamp, size_t* file_size) {
2418
4
    RowsetSharedPtr rowset = pick_cooldown_rowset();
2419
4
    if (!rowset) {
2420
1
        VLOG_DEBUG << "pick cooldown rowset, get null, tablet id: " << tablet_id();
2421
1
        return nullptr;
2422
1
    }
2423
2424
3
    auto newest_cooldown_time = _get_newest_cooldown_time(rowset);
2425
2426
    // the rowset should do cooldown job only if it's cooldown ttl plus newest write time is less than
2427
    // current time or it's datatime is less than current time
2428
3
    if (newest_cooldown_time != 0 && newest_cooldown_time < UnixSeconds()) {
2429
3
        *cooldown_timestamp = newest_cooldown_time;
2430
3
        *file_size = rowset->total_disk_size();
2431
3
        VLOG_DEBUG << "tablet need cooldown, tablet id: " << tablet_id()
2432
0
                   << " file_size: " << *file_size;
2433
3
        return rowset;
2434
3
    }
2435
2436
0
    VLOG_DEBUG << "tablet does not need cooldown, tablet id: " << tablet_id()
2437
0
               << " newest write time: " << rowset->newest_write_timestamp();
2438
0
    return nullptr;
2439
3
}
2440
2441
void Tablet::record_unused_remote_rowset(const RowsetId& rowset_id, const std::string& resource,
2442
0
                                         int64_t num_segments) {
2443
0
    auto gc_key = REMOTE_ROWSET_GC_PREFIX + rowset_id.to_string();
2444
0
    RemoteRowsetGcPB gc_pb;
2445
0
    gc_pb.set_resource_id(resource);
2446
0
    gc_pb.set_tablet_id(tablet_id());
2447
0
    gc_pb.set_num_segments(num_segments);
2448
0
    auto st =
2449
0
            _data_dir->get_meta()->put(META_COLUMN_FAMILY_INDEX, gc_key, gc_pb.SerializeAsString());
2450
0
    if (!st.ok()) {
2451
0
        LOG(WARNING) << "failed to record unused remote rowset. tablet_id=" << tablet_id()
2452
0
                     << " rowset_id=" << rowset_id << " resource_id=" << resource;
2453
0
    }
2454
0
    unused_remote_rowset_num << 1;
2455
0
}
2456
2457
0
Status Tablet::remove_all_remote_rowsets() {
2458
0
    DCHECK(tablet_state() == TABLET_SHUTDOWN);
2459
0
    std::set<std::string> resource_ids;
2460
0
    for (auto& rs_meta : _tablet_meta->all_rs_metas()) {
2461
0
        if (!rs_meta->is_local()) {
2462
0
            resource_ids.insert(rs_meta->resource_id());
2463
0
        }
2464
0
    }
2465
0
    if (resource_ids.empty()) {
2466
0
        return Status::OK();
2467
0
    }
2468
0
    auto tablet_gc_key = REMOTE_TABLET_GC_PREFIX + std::to_string(tablet_id());
2469
0
    RemoteTabletGcPB gc_pb;
2470
0
    for (auto& resource_id : resource_ids) {
2471
0
        gc_pb.add_resource_ids(resource_id);
2472
0
    }
2473
0
    return _data_dir->get_meta()->put(META_COLUMN_FAMILY_INDEX, tablet_gc_key,
2474
0
                                      gc_pb.SerializeAsString());
2475
0
}
2476
2477
0
void Tablet::update_max_version_schema(const TabletSchemaSPtr& tablet_schema) {
2478
0
    std::lock_guard wrlock(_meta_lock);
2479
0
    SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2480
    // Double Check for concurrent update
2481
0
    if (!_max_version_schema ||
2482
0
        tablet_schema->schema_version() > _max_version_schema->schema_version()) {
2483
0
        _max_version_schema = tablet_schema;
2484
0
    }
2485
0
}
2486
2487
0
CalcDeleteBitmapExecutor* Tablet::calc_delete_bitmap_executor() {
2488
0
    return _engine.calc_delete_bitmap_executor();
2489
0
}
2490
2491
Status Tablet::save_delete_bitmap(const TabletTxnInfo* txn_info, int64_t txn_id,
2492
                                  DeleteBitmapPtr delete_bitmap, RowsetWriter* rowset_writer,
2493
3
                                  const RowsetIdUnorderedSet& cur_rowset_ids) {
2494
3
    RowsetSharedPtr rowset = txn_info->rowset;
2495
3
    int64_t cur_version = rowset->start_version();
2496
2497
    // update version without write lock, compaction and publish_txn
2498
    // will update delete bitmap, handle compaction with _rowset_update_lock
2499
    // and publish_txn runs sequential so no need to lock here
2500
3
    for (auto& [key, bitmap] : delete_bitmap->delete_bitmap) {
2501
        // skip sentinel mark, which is used for delete bitmap correctness check
2502
3
        if (std::get<1>(key) != DeleteBitmap::INVALID_SEGMENT_ID) {
2503
2
            _tablet_meta->delete_bitmap().merge({std::get<0>(key), std::get<1>(key), cur_version},
2504
2
                                                bitmap);
2505
2
        }
2506
3
    }
2507
2508
3
    return Status::OK();
2509
3
}
2510
2511
0
void Tablet::merge_delete_bitmap(const DeleteBitmap& delete_bitmap) {
2512
0
    _tablet_meta->delete_bitmap().merge(delete_bitmap);
2513
0
}
2514
2515
0
bool Tablet::check_all_rowset_segment() {
2516
0
    std::shared_lock rdlock(_meta_lock);
2517
0
    for (auto& version_rowset : _rs_version_map) {
2518
0
        RowsetSharedPtr rowset = version_rowset.second;
2519
0
        if (!rowset->check_rowset_segment()) {
2520
0
            LOG(WARNING) << "Tablet Segment Check. find a bad tablet, tablet_id=" << tablet_id();
2521
0
            return false;
2522
0
        }
2523
0
    }
2524
0
    return true;
2525
0
}
2526
2527
0
void Tablet::set_skip_compaction(bool skip, CompactionType compaction_type, int64_t start) {
2528
0
    if (!skip) {
2529
0
        _skip_cumu_compaction = false;
2530
0
        _skip_base_compaction = false;
2531
0
        return;
2532
0
    }
2533
0
    if (compaction_type == CompactionType::CUMULATIVE_COMPACTION) {
2534
0
        _skip_cumu_compaction = true;
2535
0
        _skip_cumu_compaction_ts = start;
2536
0
    } else {
2537
0
        DCHECK(compaction_type == CompactionType::BASE_COMPACTION);
2538
0
        _skip_base_compaction = true;
2539
0
        _skip_base_compaction_ts = start;
2540
0
    }
2541
0
}
2542
2543
0
bool Tablet::should_skip_compaction(CompactionType compaction_type, int64_t now) {
2544
0
    if (compaction_type == CompactionType::CUMULATIVE_COMPACTION && _skip_cumu_compaction &&
2545
0
        now < _skip_cumu_compaction_ts + 120) {
2546
0
        return true;
2547
0
    } else if (compaction_type == CompactionType::BASE_COMPACTION && _skip_base_compaction &&
2548
0
               now < _skip_base_compaction_ts + 120) {
2549
0
        return true;
2550
0
    }
2551
0
    return false;
2552
0
}
2553
2554
0
std::pair<std::string, int64_t> Tablet::get_binlog_info(std::string_view binlog_version) const {
2555
0
    return RowsetMetaManager::get_binlog_info(_data_dir->get_meta(), tablet_uid(), binlog_version);
2556
0
}
2557
2558
std::string Tablet::get_rowset_binlog_meta(std::string_view binlog_version,
2559
0
                                           std::string_view rowset_id) const {
2560
0
    return RowsetMetaManager::get_rowset_binlog_meta(_data_dir->get_meta(), tablet_uid(),
2561
0
                                                     binlog_version, rowset_id);
2562
0
}
2563
2564
Status Tablet::get_rowset_binlog_metas(const std::vector<int64_t>& binlog_versions,
2565
0
                                       RowsetBinlogMetasPB* metas_pb) {
2566
0
    return RowsetMetaManager::get_rowset_binlog_metas(_data_dir->get_meta(), tablet_uid(),
2567
0
                                                      binlog_versions, metas_pb);
2568
0
}
2569
2570
4
Status Tablet::get_rowset_binlog_metas(Version binlog_versions, RowsetBinlogMetasPB* metas_pb) {
2571
4
    return RowsetMetaManager::get_rowset_binlog_metas(_data_dir->get_meta(), tablet_uid(),
2572
4
                                                      binlog_versions, metas_pb);
2573
4
}
2574
2575
std::string Tablet::get_segment_filepath(std::string_view rowset_id,
2576
0
                                         std::string_view segment_index) const {
2577
0
    return fmt::format("{}/_binlog/{}_{}.dat", _tablet_path, rowset_id, segment_index);
2578
0
}
2579
2580
0
std::string Tablet::get_segment_filepath(std::string_view rowset_id, int64_t segment_index) const {
2581
0
    return fmt::format("{}/_binlog/{}_{}.dat", _tablet_path, rowset_id, segment_index);
2582
0
}
2583
2584
std::string Tablet::get_segment_index_filepath(std::string_view rowset_id,
2585
                                               std::string_view segment_index,
2586
0
                                               std::string_view index_id) const {
2587
0
    auto format = _tablet_meta->tablet_schema()->get_inverted_index_storage_format();
2588
0
    if (format == doris::InvertedIndexStorageFormatPB::V1) {
2589
0
        return fmt::format("{}/_binlog/{}_{}_{}.idx", _tablet_path, rowset_id, segment_index,
2590
0
                           index_id);
2591
0
    } else {
2592
0
        return fmt::format("{}/_binlog/{}_{}.idx", _tablet_path, rowset_id, segment_index);
2593
0
    }
2594
0
}
2595
2596
std::string Tablet::get_segment_index_filepath(std::string_view rowset_id, int64_t segment_index,
2597
0
                                               int64_t index_id) const {
2598
0
    auto format = _tablet_meta->tablet_schema()->get_inverted_index_storage_format();
2599
0
    if (format == doris::InvertedIndexStorageFormatPB::V1) {
2600
0
        return fmt::format("{}/_binlog/{}_{}_{}.idx", _tablet_path, rowset_id, segment_index,
2601
0
                           index_id);
2602
0
    } else {
2603
0
        DCHECK(index_id == -1);
2604
0
        return fmt::format("{}/_binlog/{}_{}.idx", _tablet_path, rowset_id, segment_index);
2605
0
    }
2606
0
}
2607
2608
0
std::vector<std::string> Tablet::get_binlog_filepath(std::string_view binlog_version) const {
2609
0
    const auto& [rowset_id, num_segments] = get_binlog_info(binlog_version);
2610
0
    std::vector<std::string> binlog_filepath;
2611
0
    for (int i = 0; i < num_segments; ++i) {
2612
        // TODO(Drogon): rewrite by filesystem path
2613
0
        auto segment_file = fmt::format("{}_{}.dat", rowset_id, i);
2614
0
        binlog_filepath.emplace_back(fmt::format("{}/_binlog/{}", _tablet_path, segment_file));
2615
0
    }
2616
0
    return binlog_filepath;
2617
0
}
2618
2619
0
bool Tablet::can_add_binlog(uint64_t total_binlog_size) const {
2620
0
    return !_data_dir->reach_capacity_limit(total_binlog_size);
2621
0
}
2622
2623
11
bool Tablet::is_enable_binlog() {
2624
11
    return config::enable_feature_binlog && tablet_meta()->binlog_config().is_enable();
2625
11
}
2626
2627
0
void Tablet::set_binlog_config(BinlogConfig binlog_config) {
2628
0
    tablet_meta()->set_binlog_config(binlog_config);
2629
0
}
2630
2631
0
void Tablet::gc_binlogs(int64_t version) {
2632
0
    auto meta = _data_dir->get_meta();
2633
0
    DCHECK(meta != nullptr);
2634
2635
0
    const auto& tablet_uid = this->tablet_uid();
2636
0
    const auto tablet_id = this->tablet_id();
2637
0
    std::string begin_key = make_binlog_meta_key_prefix(tablet_uid);
2638
0
    std::string end_key = make_binlog_meta_key_prefix(tablet_uid, version + 1);
2639
0
    LOG(INFO) << fmt::format("gc binlog meta, tablet_id:{}, begin_key:{}, end_key:{}", tablet_id,
2640
0
                             begin_key, end_key);
2641
2642
0
    std::vector<std::string> wait_for_deleted_binlog_keys;
2643
0
    std::vector<std::string> wait_for_deleted_binlog_files;
2644
0
    auto add_to_wait_for_deleted = [&](std::string_view key, std::string_view rowset_id,
2645
0
                                       int64_t num_segments) {
2646
        // add binlog meta key and binlog data key
2647
0
        wait_for_deleted_binlog_keys.emplace_back(key);
2648
0
        wait_for_deleted_binlog_keys.push_back(get_binlog_data_key_from_meta_key(key));
2649
2650
        // add binlog segment files and index files
2651
0
        for (int64_t i = 0; i < num_segments; ++i) {
2652
0
            wait_for_deleted_binlog_files.emplace_back(get_segment_filepath(rowset_id, i));
2653
0
            for (const auto& index : this->tablet_schema()->inverted_indexes()) {
2654
0
                wait_for_deleted_binlog_files.emplace_back(
2655
0
                        get_segment_index_filepath(rowset_id, i, index->index_id()));
2656
0
            }
2657
0
        }
2658
0
    };
2659
2660
0
    auto check_binlog_ttl = [&](std::string_view key, std::string_view value) mutable -> bool {
2661
0
        if (key >= end_key) {
2662
0
            return false;
2663
0
        }
2664
2665
0
        BinlogMetaEntryPB binlog_meta_entry_pb;
2666
0
        if (!binlog_meta_entry_pb.ParseFromArray(value.data(), value.size())) {
2667
0
            LOG(WARNING) << "failed to parse binlog meta entry, key:" << key;
2668
0
            return true;
2669
0
        }
2670
2671
0
        auto num_segments = binlog_meta_entry_pb.num_segments();
2672
0
        std::string rowset_id;
2673
0
        if (binlog_meta_entry_pb.has_rowset_id_v2()) {
2674
0
            rowset_id = binlog_meta_entry_pb.rowset_id_v2();
2675
0
        } else {
2676
            // key is 'binlog_meta_6943f1585fe834b5-e542c2b83a21d0b7_00000000000000000069_020000000000000135449d7cd7eadfe672aa0f928fa99593', extract last part '020000000000000135449d7cd7eadfe672aa0f928fa99593'
2677
0
            auto pos = key.rfind('_');
2678
0
            if (pos == std::string::npos) {
2679
0
                LOG(WARNING) << fmt::format("invalid binlog meta key:{}", key);
2680
0
                return false;
2681
0
            }
2682
0
            rowset_id = key.substr(pos + 1);
2683
0
        }
2684
0
        add_to_wait_for_deleted(key, rowset_id, num_segments);
2685
2686
0
        return true;
2687
0
    };
2688
2689
0
    auto status = meta->iterate(META_COLUMN_FAMILY_INDEX, begin_key, check_binlog_ttl);
2690
0
    if (!status.ok()) {
2691
0
        LOG(WARNING) << "failed to iterate binlog meta, status:" << status;
2692
0
        return;
2693
0
    }
2694
2695
    // first remove binlog files, if failed, just break, then retry next time
2696
    // this keep binlog meta in meta store, so that binlog can be removed next time
2697
0
    bool remove_binlog_files_failed = false;
2698
0
    for (auto& file : wait_for_deleted_binlog_files) {
2699
0
        if (unlink(file.c_str()) != 0) {
2700
            // file not exist, continue
2701
0
            if (errno == ENOENT) {
2702
0
                continue;
2703
0
            }
2704
2705
0
            remove_binlog_files_failed = true;
2706
0
            LOG(WARNING) << "failed to remove binlog file:" << file << ", errno:" << errno;
2707
0
            break;
2708
0
        }
2709
0
    }
2710
0
    if (!remove_binlog_files_failed) {
2711
0
        static_cast<void>(meta->remove(META_COLUMN_FAMILY_INDEX, wait_for_deleted_binlog_keys));
2712
0
    }
2713
0
}
2714
2715
0
Status Tablet::ingest_binlog_metas(RowsetBinlogMetasPB* metas_pb) {
2716
0
    return RowsetMetaManager::ingest_binlog_metas(_data_dir->get_meta(), tablet_uid(), metas_pb);
2717
0
}
2718
2719
69
void Tablet::clear_cache() {
2720
69
    std::vector<RowsetSharedPtr> rowsets;
2721
69
    {
2722
69
        std::shared_lock rlock(get_header_lock());
2723
69
        SCOPED_SIMPLE_TRACE_IF_TIMEOUT(TRACE_TABLET_LOCK_THRESHOLD);
2724
2725
630
        for (auto& [_, rowset] : rowset_map()) {
2726
630
            rowsets.push_back(rowset);
2727
630
        }
2728
69
        for (auto& [_, rowset] : stale_rowset_map()) {
2729
0
            rowsets.push_back(rowset);
2730
0
        }
2731
69
    }
2732
630
    for (auto& rowset : rowsets) {
2733
630
        rowset->clear_cache();
2734
630
    }
2735
69
}
2736
2737
145
void Tablet::check_table_size_correctness() {
2738
145
    if (!config::enable_table_size_correctness_check) {
2739
145
        return;
2740
145
    }
2741
0
    const std::vector<RowsetMetaSharedPtr>& all_rs_metas = _tablet_meta->all_rs_metas();
2742
0
    for (const auto& rs_meta : all_rs_metas) {
2743
0
        int64_t total_segment_size = get_segment_file_size(rs_meta);
2744
0
        int64_t total_inverted_index_size = get_inverted_index_file_szie(rs_meta);
2745
0
        if (rs_meta->data_disk_size() != total_segment_size ||
2746
0
            rs_meta->index_disk_size() != total_inverted_index_size ||
2747
0
            rs_meta->data_disk_size() + rs_meta->index_disk_size() != rs_meta->total_disk_size()) {
2748
0
            LOG(WARNING) << "[Local table table size check failed]:"
2749
0
                         << " tablet id: " << rs_meta->tablet_id()
2750
0
                         << ", rowset id:" << rs_meta->rowset_id()
2751
0
                         << ", rowset data disk size:" << rs_meta->data_disk_size()
2752
0
                         << ", rowset real data disk size:" << total_segment_size
2753
0
                         << ", rowset index disk size:" << rs_meta->index_disk_size()
2754
0
                         << ", rowset real index disk size:" << total_inverted_index_size
2755
0
                         << ", rowset total disk size:" << rs_meta->total_disk_size()
2756
0
                         << ", rowset segment path:"
2757
0
                         << StorageResource().remote_segment_path(
2758
0
                                    rs_meta->tablet_id(), rs_meta->rowset_id().to_string(), 0);
2759
0
            DCHECK(false);
2760
0
        }
2761
0
    }
2762
0
}
2763
2764
0
std::string Tablet::get_segment_path(const RowsetMetaSharedPtr& rs_meta, int64_t seg_id) {
2765
0
    std::string segment_path;
2766
0
    if (rs_meta->is_local()) {
2767
0
        segment_path = local_segment_path(_tablet_path, rs_meta->rowset_id().to_string(), seg_id);
2768
0
    } else {
2769
0
        segment_path = rs_meta->remote_storage_resource().value()->remote_segment_path(
2770
0
                rs_meta->tablet_id(), rs_meta->rowset_id().to_string(), seg_id);
2771
0
    }
2772
0
    return segment_path;
2773
0
}
2774
2775
0
int64_t Tablet::get_segment_file_size(const RowsetMetaSharedPtr& rs_meta) {
2776
0
    const auto& fs = rs_meta->fs();
2777
0
    if (!fs) {
2778
0
        LOG(WARNING) << "get fs failed, resource_id={}" << rs_meta->resource_id();
2779
0
    }
2780
0
    int64_t total_segment_size = 0;
2781
0
    for (int64_t seg_id = 0; seg_id < rs_meta->num_segments(); seg_id++) {
2782
0
        std::string segment_path = get_segment_path(rs_meta, seg_id);
2783
0
        int64_t segment_file_size = 0;
2784
0
        auto st = fs->file_size(segment_path, &segment_file_size);
2785
0
        if (!st.ok()) {
2786
0
            segment_file_size = 0;
2787
0
            LOG(WARNING) << "table size correctness check get segment size failed! msg:"
2788
0
                         << st.to_string() << ", segment path:" << segment_path;
2789
0
        }
2790
0
        total_segment_size += segment_file_size;
2791
0
    }
2792
0
    return total_segment_size;
2793
0
}
2794
2795
0
int64_t Tablet::get_inverted_index_file_szie(const RowsetMetaSharedPtr& rs_meta) {
2796
0
    const auto& fs = rs_meta->fs();
2797
0
    if (!fs) {
2798
0
        LOG(WARNING) << "get fs failed, resource_id={}" << rs_meta->resource_id();
2799
0
    }
2800
0
    int64_t total_inverted_index_size = 0;
2801
2802
0
    if (rs_meta->tablet_schema()->get_inverted_index_storage_format() ==
2803
0
        InvertedIndexStorageFormatPB::V1) {
2804
0
        const auto& indices = rs_meta->tablet_schema()->inverted_indexes();
2805
0
        for (auto& index : indices) {
2806
0
            for (int seg_id = 0; seg_id < rs_meta->num_segments(); ++seg_id) {
2807
0
                std::string segment_path = get_segment_path(rs_meta, seg_id);
2808
0
                int64_t file_size = 0;
2809
2810
0
                std::string inverted_index_file_path =
2811
0
                        InvertedIndexDescriptor::get_index_file_path_v1(
2812
0
                                InvertedIndexDescriptor::get_index_file_path_prefix(segment_path),
2813
0
                                index->index_id(), index->get_index_suffix());
2814
0
                auto st = fs->file_size(inverted_index_file_path, &file_size);
2815
0
                if (!st.ok()) {
2816
0
                    file_size = 0;
2817
0
                    LOG(WARNING) << " tablet id: " << get_tablet_info().tablet_id
2818
0
                                 << ", rowset id:" << rs_meta->rowset_id()
2819
0
                                 << ", table size correctness check get inverted index v1 "
2820
0
                                    "size failed! msg:"
2821
0
                                 << st.to_string()
2822
0
                                 << ", inverted index path:" << inverted_index_file_path;
2823
0
                }
2824
0
                total_inverted_index_size += file_size;
2825
0
            }
2826
0
        }
2827
0
    } else {
2828
0
        for (int seg_id = 0; seg_id < rs_meta->num_segments(); ++seg_id) {
2829
0
            int64_t file_size = 0;
2830
0
            std::string segment_path = get_segment_path(rs_meta, seg_id);
2831
0
            std::string inverted_index_file_path = InvertedIndexDescriptor::get_index_file_path_v2(
2832
0
                    InvertedIndexDescriptor::get_index_file_path_prefix(segment_path));
2833
0
            auto st = fs->file_size(inverted_index_file_path, &file_size);
2834
0
            if (!st.ok()) {
2835
0
                file_size = 0;
2836
0
                if (st.is<NOT_FOUND>()) {
2837
0
                    LOG(INFO) << " tablet id: " << get_tablet_info().tablet_id
2838
0
                              << ", rowset id:" << rs_meta->rowset_id()
2839
0
                              << ", table size correctness check get inverted index v2 failed "
2840
0
                                 "because file not exist:"
2841
0
                              << inverted_index_file_path;
2842
0
                } else {
2843
0
                    LOG(WARNING) << " tablet id: " << get_tablet_info().tablet_id
2844
0
                                 << ", rowset id:" << rs_meta->rowset_id()
2845
0
                                 << ", table size correctness check get inverted index v2 "
2846
0
                                    "size failed! msg:"
2847
0
                                 << st.to_string()
2848
0
                                 << ", inverted index path:" << inverted_index_file_path;
2849
0
                }
2850
0
            }
2851
0
            total_inverted_index_size += file_size;
2852
0
        }
2853
0
    }
2854
0
    return total_inverted_index_size;
2855
0
}
2856
2857
} // namespace doris