Coverage Report

Created: 2024-11-20 15:53

/root/doris/be/src/olap/tablet_meta.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_meta.h"
19
20
#include <gen_cpp/Descriptors_types.h>
21
#include <gen_cpp/Types_types.h>
22
#include <gen_cpp/olap_common.pb.h>
23
#include <gen_cpp/olap_file.pb.h>
24
#include <gen_cpp/segment_v2.pb.h>
25
#include <gen_cpp/types.pb.h>
26
#include <json2pb/pb_to_json.h>
27
#include <time.h>
28
29
#include <cstdint>
30
#include <set>
31
#include <utility>
32
33
#include "common/config.h"
34
#include "gutil/integral_types.h"
35
#include "io/fs/file_writer.h"
36
#include "olap/data_dir.h"
37
#include "olap/file_header.h"
38
#include "olap/olap_common.h"
39
#include "olap/olap_define.h"
40
#include "olap/rowset/rowset.h"
41
#include "olap/rowset/rowset_meta_manager.h"
42
#include "olap/tablet_meta_manager.h"
43
#include "olap/utils.h"
44
#include "util/debug_points.h"
45
#include "util/mem_info.h"
46
#include "util/parse_util.h"
47
#include "util/string_util.h"
48
#include "util/time.h"
49
#include "util/uid_util.h"
50
51
using std::string;
52
using std::unordered_map;
53
using std::vector;
54
55
namespace doris {
56
using namespace ErrorCode;
57
58
TabletMetaSharedPtr TabletMeta::create(
59
        const TCreateTabletReq& request, const TabletUid& tablet_uid, uint64_t shard_id,
60
        uint32_t next_unique_id,
61
69
        const unordered_map<uint32_t, uint32_t>& col_ordinal_to_unique_id) {
62
69
    std::optional<TBinlogConfig> binlog_config;
63
69
    if (request.__isset.binlog_config) {
64
0
        binlog_config = request.binlog_config;
65
0
    }
66
69
    return std::make_shared<TabletMeta>(
67
69
            request.table_id, request.partition_id, request.tablet_id, request.replica_id,
68
69
            request.tablet_schema.schema_hash, shard_id, request.tablet_schema, next_unique_id,
69
69
            col_ordinal_to_unique_id, tablet_uid,
70
69
            request.__isset.tablet_type ? request.tablet_type : TTabletType::TABLET_TYPE_DISK,
71
69
            request.compression_type, request.storage_policy_id,
72
69
            request.__isset.enable_unique_key_merge_on_write
73
69
                    ? request.enable_unique_key_merge_on_write
74
69
                    : false,
75
69
            std::move(binlog_config), request.compaction_policy,
76
69
            request.time_series_compaction_goal_size_mbytes,
77
69
            request.time_series_compaction_file_count_threshold,
78
69
            request.time_series_compaction_time_threshold_seconds,
79
69
            request.time_series_compaction_empty_rowsets_threshold,
80
69
            request.inverted_index_storage_format, request.time_series_compaction_level_threshold);
81
69
}
82
83
TabletMeta::TabletMeta()
84
        : _tablet_uid(0, 0),
85
          _schema(new TabletSchema),
86
67
          _delete_bitmap(new DeleteBitmap(_tablet_id)) {}
87
88
TabletMeta::TabletMeta(int64_t table_id, int64_t partition_id, int64_t tablet_id,
89
                       int64_t replica_id, int32_t schema_hash, uint64_t shard_id,
90
                       const TTabletSchema& tablet_schema, uint32_t next_unique_id,
91
                       const std::unordered_map<uint32_t, uint32_t>& col_ordinal_to_unique_id,
92
                       TabletUid tablet_uid, TTabletType::type tabletType,
93
                       TCompressionType::type compression_type, int64_t storage_policy_id,
94
                       bool enable_unique_key_merge_on_write,
95
                       std::optional<TBinlogConfig> binlog_config, std::string compaction_policy,
96
                       int64_t time_series_compaction_goal_size_mbytes,
97
                       int64_t time_series_compaction_file_count_threshold,
98
                       int64_t time_series_compaction_time_threshold_seconds,
99
                       int64_t time_series_compaction_empty_rowsets_threshold,
100
                       TInvertedIndexStorageFormat::type inverted_index_storage_format,
101
                       int64_t time_series_compaction_level_threshold)
102
        : _tablet_uid(0, 0),
103
          _schema(new TabletSchema),
104
212
          _delete_bitmap(new DeleteBitmap(tablet_id)) {
105
212
    TabletMetaPB tablet_meta_pb;
106
212
    tablet_meta_pb.set_table_id(table_id);
107
212
    tablet_meta_pb.set_partition_id(partition_id);
108
212
    tablet_meta_pb.set_tablet_id(tablet_id);
109
212
    tablet_meta_pb.set_replica_id(replica_id);
110
212
    tablet_meta_pb.set_schema_hash(schema_hash);
111
212
    tablet_meta_pb.set_shard_id(shard_id);
112
    // Persist the creation time, but it is not used
113
212
    tablet_meta_pb.set_creation_time(time(nullptr));
114
212
    tablet_meta_pb.set_cumulative_layer_point(-1);
115
212
    tablet_meta_pb.set_tablet_state(PB_RUNNING);
116
212
    *(tablet_meta_pb.mutable_tablet_uid()) = tablet_uid.to_proto();
117
212
    tablet_meta_pb.set_tablet_type(tabletType == TTabletType::TABLET_TYPE_DISK
118
212
                                           ? TabletTypePB::TABLET_TYPE_DISK
119
212
                                           : TabletTypePB::TABLET_TYPE_MEMORY);
120
212
    tablet_meta_pb.set_enable_unique_key_merge_on_write(enable_unique_key_merge_on_write);
121
212
    tablet_meta_pb.set_storage_policy_id(storage_policy_id);
122
212
    tablet_meta_pb.set_compaction_policy(compaction_policy);
123
212
    tablet_meta_pb.set_time_series_compaction_goal_size_mbytes(
124
212
            time_series_compaction_goal_size_mbytes);
125
212
    tablet_meta_pb.set_time_series_compaction_file_count_threshold(
126
212
            time_series_compaction_file_count_threshold);
127
212
    tablet_meta_pb.set_time_series_compaction_time_threshold_seconds(
128
212
            time_series_compaction_time_threshold_seconds);
129
212
    tablet_meta_pb.set_time_series_compaction_empty_rowsets_threshold(
130
212
            time_series_compaction_empty_rowsets_threshold);
131
212
    tablet_meta_pb.set_time_series_compaction_level_threshold(
132
212
            time_series_compaction_level_threshold);
133
212
    TabletSchemaPB* schema = tablet_meta_pb.mutable_schema();
134
212
    schema->set_num_short_key_columns(tablet_schema.short_key_column_count);
135
212
    schema->set_num_rows_per_row_block(config::default_num_rows_per_column_file_block);
136
212
    schema->set_sequence_col_idx(tablet_schema.sequence_col_idx);
137
212
    switch (tablet_schema.keys_type) {
138
30
    case TKeysType::DUP_KEYS:
139
30
        schema->set_keys_type(KeysType::DUP_KEYS);
140
30
        break;
141
73
    case TKeysType::UNIQUE_KEYS:
142
73
        schema->set_keys_type(KeysType::UNIQUE_KEYS);
143
73
        break;
144
61
    case TKeysType::AGG_KEYS:
145
61
        schema->set_keys_type(KeysType::AGG_KEYS);
146
61
        break;
147
48
    default:
148
48
        LOG(WARNING) << "unknown tablet keys type";
149
48
        break;
150
212
    }
151
    // compress_kind used to compress segment files
152
212
    schema->set_compress_kind(COMPRESS_LZ4);
153
154
    // compression_type used to compress segment page
155
212
    switch (compression_type) {
156
0
    case TCompressionType::NO_COMPRESSION:
157
0
        schema->set_compression_type(segment_v2::NO_COMPRESSION);
158
0
        break;
159
0
    case TCompressionType::SNAPPY:
160
0
        schema->set_compression_type(segment_v2::SNAPPY);
161
0
        break;
162
0
    case TCompressionType::LZ4:
163
0
        schema->set_compression_type(segment_v2::LZ4);
164
0
        break;
165
212
    case TCompressionType::LZ4F:
166
212
        schema->set_compression_type(segment_v2::LZ4F);
167
212
        break;
168
0
    case TCompressionType::ZLIB:
169
0
        schema->set_compression_type(segment_v2::ZLIB);
170
0
        break;
171
0
    case TCompressionType::ZSTD:
172
0
        schema->set_compression_type(segment_v2::ZSTD);
173
0
        break;
174
0
    default:
175
0
        schema->set_compression_type(segment_v2::LZ4F);
176
0
        break;
177
212
    }
178
179
212
    switch (inverted_index_storage_format) {
180
212
    case TInvertedIndexStorageFormat::V1:
181
212
        schema->set_inverted_index_storage_format(InvertedIndexStorageFormatPB::V1);
182
212
        break;
183
0
    case TInvertedIndexStorageFormat::V2:
184
0
        schema->set_inverted_index_storage_format(InvertedIndexStorageFormatPB::V2);
185
0
        break;
186
0
    default:
187
0
        schema->set_inverted_index_storage_format(InvertedIndexStorageFormatPB::V1);
188
0
        break;
189
212
    }
190
191
212
    switch (tablet_schema.sort_type) {
192
0
    case TSortType::type::ZORDER:
193
0
        schema->set_sort_type(SortType::ZORDER);
194
0
        break;
195
212
    default:
196
212
        schema->set_sort_type(SortType::LEXICAL);
197
212
    }
198
212
    schema->set_sort_col_num(tablet_schema.sort_col_num);
199
212
    for (const auto& i : tablet_schema.cluster_key_idxes) {
200
0
        schema->add_cluster_key_idxes(i);
201
0
    }
202
212
    tablet_meta_pb.set_in_restore_mode(false);
203
204
    // set column information
205
212
    uint32_t col_ordinal = 0;
206
212
    bool has_bf_columns = false;
207
1.41k
    for (TColumn tcolumn : tablet_schema.columns) {
208
1.41k
        ColumnPB* column = schema->add_column();
209
1.41k
        uint32_t unique_id = -1;
210
1.41k
        if (tcolumn.col_unique_id >= 0) {
211
0
            unique_id = tcolumn.col_unique_id;
212
1.41k
        } else {
213
1.41k
            unique_id = col_ordinal_to_unique_id.at(col_ordinal);
214
1.41k
        }
215
1.41k
        col_ordinal++;
216
1.41k
        init_column_from_tcolumn(unique_id, tcolumn, column);
217
218
1.41k
        if (column->is_bf_column()) {
219
0
            has_bf_columns = true;
220
0
        }
221
222
1.41k
        if (tablet_schema.__isset.indexes) {
223
0
            for (auto& index : tablet_schema.indexes) {
224
0
                if (index.index_type == TIndexType::type::BITMAP) {
225
0
                    DCHECK_EQ(index.columns.size(), 1);
226
0
                    if (iequal(tcolumn.column_name, index.columns[0])) {
227
0
                        column->set_has_bitmap_index(true);
228
0
                        break;
229
0
                    }
230
0
                } else if (index.index_type == TIndexType::type::BLOOMFILTER ||
231
0
                           index.index_type == TIndexType::type::NGRAM_BF) {
232
0
                    DCHECK_EQ(index.columns.size(), 1);
233
0
                    if (iequal(tcolumn.column_name, index.columns[0])) {
234
0
                        column->set_is_bf_column(true);
235
0
                        break;
236
0
                    }
237
0
                }
238
0
            }
239
0
        }
240
1.41k
    }
241
242
    // copy index meta
243
212
    if (tablet_schema.__isset.indexes) {
244
0
        for (auto& index : tablet_schema.indexes) {
245
0
            TabletIndexPB* index_pb = schema->add_index();
246
0
            index_pb->set_index_id(index.index_id);
247
0
            index_pb->set_index_name(index.index_name);
248
            // init col_unique_id in index at be side, since col_unique_id may be -1 at fe side
249
            // get column unique id by name
250
0
            for (auto column_name : index.columns) {
251
0
                for (auto column : schema->column()) {
252
0
                    if (iequal(column.name(), column_name)) {
253
0
                        index_pb->add_col_unique_id(column.unique_id());
254
0
                    }
255
0
                }
256
0
            }
257
0
            switch (index.index_type) {
258
0
            case TIndexType::BITMAP:
259
0
                index_pb->set_index_type(IndexType::BITMAP);
260
0
                break;
261
0
            case TIndexType::INVERTED:
262
0
                index_pb->set_index_type(IndexType::INVERTED);
263
0
                break;
264
0
            case TIndexType::BLOOMFILTER:
265
0
                index_pb->set_index_type(IndexType::BLOOMFILTER);
266
0
                break;
267
0
            case TIndexType::NGRAM_BF:
268
0
                index_pb->set_index_type(IndexType::NGRAM_BF);
269
0
                break;
270
0
            }
271
272
0
            if (index.__isset.properties) {
273
0
                auto properties = index_pb->mutable_properties();
274
0
                for (auto kv : index.properties) {
275
0
                    (*properties)[kv.first] = kv.second;
276
0
                }
277
0
            }
278
0
        }
279
0
    }
280
281
212
    schema->set_next_column_unique_id(next_unique_id);
282
212
    if (has_bf_columns && tablet_schema.__isset.bloom_filter_fpp) {
283
0
        schema->set_bf_fpp(tablet_schema.bloom_filter_fpp);
284
0
    }
285
286
212
    if (tablet_schema.__isset.is_in_memory) {
287
0
        schema->set_is_in_memory(tablet_schema.is_in_memory);
288
0
    }
289
290
212
    if (tablet_schema.__isset.disable_auto_compaction) {
291
0
        schema->set_disable_auto_compaction(tablet_schema.disable_auto_compaction);
292
0
    }
293
294
212
    if (tablet_schema.__isset.enable_single_replica_compaction) {
295
212
        schema->set_enable_single_replica_compaction(
296
212
                tablet_schema.enable_single_replica_compaction);
297
212
    }
298
299
212
    if (tablet_schema.__isset.delete_sign_idx) {
300
212
        schema->set_delete_sign_idx(tablet_schema.delete_sign_idx);
301
212
    }
302
212
    if (tablet_schema.__isset.store_row_column) {
303
212
        schema->set_store_row_column(tablet_schema.store_row_column);
304
212
    }
305
212
    if (tablet_schema.__isset.row_store_page_size) {
306
212
        schema->set_row_store_page_size(tablet_schema.row_store_page_size);
307
212
    }
308
212
    if (tablet_schema.__isset.skip_write_index_on_load) {
309
212
        schema->set_skip_write_index_on_load(tablet_schema.skip_write_index_on_load);
310
212
    }
311
212
    if (binlog_config.has_value()) {
312
0
        BinlogConfig tmp_binlog_config;
313
0
        tmp_binlog_config = binlog_config.value();
314
0
        tmp_binlog_config.to_pb(tablet_meta_pb.mutable_binlog_config());
315
0
    }
316
317
212
    init_from_pb(tablet_meta_pb);
318
212
}
319
320
TabletMeta::TabletMeta(const TabletMeta& b)
321
        : _table_id(b._table_id),
322
          _partition_id(b._partition_id),
323
          _tablet_id(b._tablet_id),
324
          _replica_id(b._replica_id),
325
          _schema_hash(b._schema_hash),
326
          _shard_id(b._shard_id),
327
          _creation_time(b._creation_time),
328
          _cumulative_layer_point(b._cumulative_layer_point),
329
          _tablet_uid(b._tablet_uid),
330
          _tablet_type(b._tablet_type),
331
          _tablet_state(b._tablet_state),
332
          _schema(b._schema),
333
          _rs_metas(b._rs_metas),
334
          _stale_rs_metas(b._stale_rs_metas),
335
          _in_restore_mode(b._in_restore_mode),
336
          _preferred_rowset_type(b._preferred_rowset_type),
337
          _storage_policy_id(b._storage_policy_id),
338
          _cooldown_meta_id(b._cooldown_meta_id),
339
          _enable_unique_key_merge_on_write(b._enable_unique_key_merge_on_write),
340
          _delete_bitmap(b._delete_bitmap),
341
          _binlog_config(b._binlog_config),
342
          _compaction_policy(b._compaction_policy),
343
          _time_series_compaction_goal_size_mbytes(b._time_series_compaction_goal_size_mbytes),
344
          _time_series_compaction_file_count_threshold(
345
                  b._time_series_compaction_file_count_threshold),
346
          _time_series_compaction_time_threshold_seconds(
347
                  b._time_series_compaction_time_threshold_seconds),
348
          _time_series_compaction_empty_rowsets_threshold(
349
                  b._time_series_compaction_empty_rowsets_threshold),
350
0
          _time_series_compaction_level_threshold(b._time_series_compaction_level_threshold) {};
351
352
void TabletMeta::init_column_from_tcolumn(uint32_t unique_id, const TColumn& tcolumn,
353
1.41k
                                          ColumnPB* column) {
354
1.41k
    column->set_unique_id(unique_id);
355
1.41k
    column->set_name(tcolumn.column_name);
356
1.41k
    column->set_has_bitmap_index(tcolumn.has_bitmap_index);
357
1.41k
    column->set_is_auto_increment(tcolumn.is_auto_increment);
358
1.41k
    string data_type;
359
1.41k
    EnumToString(TPrimitiveType, tcolumn.column_type.type, data_type);
360
1.41k
    column->set_type(data_type);
361
362
1.41k
    uint32_t length = TabletColumn::get_field_length_by_type(tcolumn.column_type.type,
363
1.41k
                                                             tcolumn.column_type.len);
364
1.41k
    column->set_length(length);
365
1.41k
    column->set_index_length(length);
366
1.41k
    column->set_precision(tcolumn.column_type.precision);
367
1.41k
    column->set_frac(tcolumn.column_type.scale);
368
369
1.41k
    if (tcolumn.__isset.result_is_nullable) {
370
0
        column->set_result_is_nullable(tcolumn.result_is_nullable);
371
0
    }
372
373
1.41k
    if (tcolumn.column_type.type == TPrimitiveType::VARCHAR ||
374
1.41k
        tcolumn.column_type.type == TPrimitiveType::STRING) {
375
101
        if (!tcolumn.column_type.__isset.index_len) {
376
101
            column->set_index_length(10);
377
101
        } else {
378
0
            column->set_index_length(tcolumn.column_type.index_len);
379
0
        }
380
101
    }
381
1.41k
    if (!tcolumn.is_key) {
382
678
        column->set_is_key(false);
383
678
        if (tcolumn.__isset.aggregation) {
384
0
            column->set_aggregation(tcolumn.aggregation);
385
678
        } else {
386
678
            string aggregation_type;
387
678
            EnumToString(TAggregationType, tcolumn.aggregation_type, aggregation_type);
388
678
            column->set_aggregation(aggregation_type);
389
678
        }
390
741
    } else {
391
741
        column->set_is_key(true);
392
741
        column->set_aggregation("NONE");
393
741
    }
394
1.41k
    column->set_is_nullable(tcolumn.is_allow_null);
395
1.41k
    if (tcolumn.__isset.default_value) {
396
0
        column->set_default_value(tcolumn.default_value);
397
0
    }
398
1.41k
    if (tcolumn.__isset.is_bloom_filter_column) {
399
0
        column->set_is_bf_column(tcolumn.is_bloom_filter_column);
400
0
    }
401
1.41k
    for (size_t i = 0; i < tcolumn.children_column.size(); i++) {
402
0
        ColumnPB* children_column = column->add_children_columns();
403
0
        init_column_from_tcolumn(tcolumn.children_column[i].col_unique_id,
404
0
                                 tcolumn.children_column[i], children_column);
405
0
    }
406
1.41k
}
407
408
5
Status TabletMeta::create_from_file(const string& file_path) {
409
5
    FileHeader<TabletMetaPB> file_header(file_path);
410
    // In file_header.deserialize(), it validates file length, signature, checksum of protobuf.
411
5
    RETURN_IF_ERROR(file_header.deserialize());
412
5
    TabletMetaPB tablet_meta_pb;
413
5
    try {
414
5
        tablet_meta_pb.CopyFrom(file_header.message());
415
5
    } catch (...) {
416
0
        return Status::Error<PARSE_PROTOBUF_ERROR>("fail to copy protocol buffer object. file={}",
417
0
                                                   file_path);
418
0
    }
419
420
5
    init_from_pb(tablet_meta_pb);
421
5
    return Status::OK();
422
5
}
423
424
std::string TabletMeta::construct_header_file_path(const string& schema_hash_path,
425
2
                                                   int64_t tablet_id) {
426
2
    std::stringstream header_name_stream;
427
2
    header_name_stream << schema_hash_path << "/" << tablet_id << ".hdr";
428
2
    return header_name_stream.str();
429
2
}
430
431
0
Status TabletMeta::save_as_json(const string& file_path, DataDir* dir) {
432
0
    std::string json_meta;
433
0
    json2pb::Pb2JsonOptions json_options;
434
0
    json_options.pretty_json = true;
435
0
    json_options.bytes_to_base64 = true;
436
0
    to_json(&json_meta, json_options);
437
    // save to file
438
0
    io::FileWriterPtr file_writer;
439
0
    RETURN_IF_ERROR(dir->fs()->create_file(file_path, &file_writer));
440
0
    RETURN_IF_ERROR(file_writer->append(json_meta));
441
0
    RETURN_IF_ERROR(file_writer->close());
442
0
    return Status::OK();
443
0
}
444
445
6
Status TabletMeta::save(const string& file_path) {
446
6
    TabletMetaPB tablet_meta_pb;
447
6
    to_meta_pb(&tablet_meta_pb);
448
6
    return TabletMeta::save(file_path, tablet_meta_pb);
449
6
}
450
451
8
Status TabletMeta::save(const string& file_path, const TabletMetaPB& tablet_meta_pb) {
452
8
    DCHECK(!file_path.empty());
453
8
    FileHeader<TabletMetaPB> file_header(file_path);
454
8
    try {
455
8
        file_header.mutable_message()->CopyFrom(tablet_meta_pb);
456
8
    } catch (...) {
457
0
        LOG(WARNING) << "fail to copy protocol buffer object. file='" << file_path;
458
0
        return Status::Error<ErrorCode::INTERNAL_ERROR>(
459
0
                "fail to copy protocol buffer object. file={}", file_path);
460
0
    }
461
8
    RETURN_IF_ERROR(file_header.prepare());
462
8
    RETURN_IF_ERROR(file_header.serialize());
463
8
    return Status::OK();
464
8
}
465
466
102
Status TabletMeta::save_meta(DataDir* data_dir) {
467
102
    std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
468
102
    return _save_meta(data_dir);
469
102
}
470
471
102
Status TabletMeta::_save_meta(DataDir* data_dir) {
472
    // check if tablet uid is valid
473
102
    if (_tablet_uid.hi == 0 && _tablet_uid.lo == 0) {
474
0
        LOG(FATAL) << "tablet_uid is invalid"
475
0
                   << " tablet=" << tablet_id() << " _tablet_uid=" << _tablet_uid.to_string();
476
0
    }
477
102
    string meta_binary;
478
479
102
    auto t1 = MonotonicMicros();
480
102
    serialize(&meta_binary);
481
102
    auto t2 = MonotonicMicros();
482
102
    Status status = TabletMetaManager::save(data_dir, tablet_id(), schema_hash(), meta_binary);
483
102
    if (!status.ok()) {
484
0
        LOG(FATAL) << "fail to save tablet_meta. status=" << status << ", tablet_id=" << tablet_id()
485
0
                   << ", schema_hash=" << schema_hash();
486
0
    }
487
102
    auto t3 = MonotonicMicros();
488
102
    auto cost = t3 - t1;
489
102
    if (cost > 1 * 1000 * 1000) {
490
0
        LOG(INFO) << "save tablet(" << tablet_id() << ") meta too slow. serialize cost " << t2 - t1
491
0
                  << "(us), serialized binary size: " << meta_binary.length()
492
0
                  << "(bytes), write rocksdb cost " << t3 - t2 << "(us)";
493
0
    }
494
102
    return status;
495
102
}
496
497
105
void TabletMeta::serialize(string* meta_binary) {
498
105
    TabletMetaPB tablet_meta_pb;
499
105
    to_meta_pb(&tablet_meta_pb);
500
105
    if (tablet_meta_pb.partition_id() <= 0) {
501
9
        LOG(WARNING) << "invalid partition id " << tablet_meta_pb.partition_id() << " tablet "
502
9
                     << tablet_meta_pb.tablet_id();
503
9
    }
504
105
    DBUG_EXECUTE_IF("TabletMeta::serialize::zero_partition_id", {
505
105
        long partition_id = tablet_meta_pb.partition_id();
506
105
        tablet_meta_pb.set_partition_id(0);
507
105
        LOG(WARNING) << "set debug point TabletMeta::serialize::zero_partition_id old="
508
105
                     << partition_id << " new=" << tablet_meta_pb.DebugString();
509
105
    });
510
105
    bool serialize_success = tablet_meta_pb.SerializeToString(meta_binary);
511
105
    if (!_rs_metas.empty() || !_stale_rs_metas.empty()) {
512
105
        _avg_rs_meta_serialize_size =
513
105
                meta_binary->length() / (_rs_metas.size() + _stale_rs_metas.size());
514
105
        if (meta_binary->length() > config::tablet_meta_serialize_size_limit ||
515
105
            !serialize_success) {
516
0
            int64_t origin_meta_size = meta_binary->length();
517
0
            int64_t stale_rowsets_num = tablet_meta_pb.stale_rs_metas().size();
518
0
            tablet_meta_pb.clear_stale_rs_metas();
519
0
            meta_binary->clear();
520
0
            serialize_success = tablet_meta_pb.SerializeToString(meta_binary);
521
0
            LOG(WARNING) << "tablet meta serialization size exceeds limit: "
522
0
                         << config::tablet_meta_serialize_size_limit
523
0
                         << " clean up stale rowsets, tablet id: " << tablet_id()
524
0
                         << " stale rowset num: " << stale_rowsets_num
525
0
                         << " serialization size before clean " << origin_meta_size
526
0
                         << " serialization size after clean " << meta_binary->length();
527
0
        }
528
105
    }
529
530
105
    if (!serialize_success) {
531
0
        LOG(FATAL) << "failed to serialize meta " << tablet_id();
532
0
    }
533
105
}
534
535
10
Status TabletMeta::deserialize(const string& meta_binary) {
536
10
    TabletMetaPB tablet_meta_pb;
537
10
    bool parsed = tablet_meta_pb.ParseFromString(meta_binary);
538
10
    if (!parsed) {
539
0
        return Status::Error<INIT_FAILED>("parse tablet meta failed");
540
0
    }
541
10
    init_from_pb(tablet_meta_pb);
542
10
    return Status::OK();
543
10
}
544
545
2
void TabletMeta::init_rs_metas_fs(const io::FileSystemSPtr& fs) {
546
4
    for (auto& rs_meta : _rs_metas) {
547
4
        if (rs_meta->is_local()) {
548
4
            rs_meta->set_fs(fs);
549
4
        }
550
4
    }
551
2
    for (auto& rs_meta : _stale_rs_metas) {
552
0
        if (rs_meta->is_local()) {
553
0
            rs_meta->set_fs(fs);
554
0
        }
555
0
    }
556
2
}
557
558
229
void TabletMeta::init_from_pb(const TabletMetaPB& tablet_meta_pb) {
559
229
    _table_id = tablet_meta_pb.table_id();
560
229
    _partition_id = tablet_meta_pb.partition_id();
561
229
    _tablet_id = tablet_meta_pb.tablet_id();
562
229
    _replica_id = tablet_meta_pb.replica_id();
563
229
    _schema_hash = tablet_meta_pb.schema_hash();
564
229
    _shard_id = tablet_meta_pb.shard_id();
565
229
    _creation_time = tablet_meta_pb.creation_time();
566
229
    _cumulative_layer_point = tablet_meta_pb.cumulative_layer_point();
567
229
    _tablet_uid = TabletUid(tablet_meta_pb.tablet_uid());
568
229
    if (tablet_meta_pb.has_tablet_type()) {
569
229
        _tablet_type = tablet_meta_pb.tablet_type();
570
229
    } else {
571
0
        _tablet_type = TabletTypePB::TABLET_TYPE_DISK;
572
0
    }
573
574
    // init _tablet_state
575
229
    switch (tablet_meta_pb.tablet_state()) {
576
3
    case PB_NOTREADY:
577
3
        _tablet_state = TabletState::TABLET_NOTREADY;
578
3
        break;
579
223
    case PB_RUNNING:
580
223
        _tablet_state = TabletState::TABLET_RUNNING;
581
223
        break;
582
0
    case PB_TOMBSTONED:
583
0
        _tablet_state = TabletState::TABLET_TOMBSTONED;
584
0
        break;
585
0
    case PB_STOPPED:
586
0
        _tablet_state = TabletState::TABLET_STOPPED;
587
0
        break;
588
3
    case PB_SHUTDOWN:
589
3
        _tablet_state = TabletState::TABLET_SHUTDOWN;
590
3
        break;
591
0
    default:
592
0
        LOG(WARNING) << "tablet has no state. tablet=" << tablet_id()
593
0
                     << ", schema_hash=" << schema_hash();
594
229
    }
595
596
    // init _schema
597
229
    _schema->init_from_pb(tablet_meta_pb.schema());
598
599
229
    if (tablet_meta_pb.has_enable_unique_key_merge_on_write()) {
600
229
        _enable_unique_key_merge_on_write = tablet_meta_pb.enable_unique_key_merge_on_write();
601
229
    }
602
603
    // init _rs_metas
604
229
    for (auto& it : tablet_meta_pb.rs_metas()) {
605
27
        RowsetMetaSharedPtr rs_meta(new RowsetMeta());
606
27
        rs_meta->init_from_pb(it);
607
27
        _rs_metas.push_back(std::move(rs_meta));
608
27
    }
609
610
    // For mow table, delete bitmap of stale rowsets has not been persisted.
611
    // When be restart, query should not read the stale rowset, otherwise duplicate keys
612
    // will be read out. Therefore, we don't add them to _stale_rs_meta for mow table.
613
229
    if (!config::skip_loading_stale_rowset_meta && !_enable_unique_key_merge_on_write) {
614
195
        for (auto& it : tablet_meta_pb.stale_rs_metas()) {
615
0
            RowsetMetaSharedPtr rs_meta(new RowsetMeta());
616
0
            rs_meta->init_from_pb(it);
617
0
            _stale_rs_metas.push_back(std::move(rs_meta));
618
0
        }
619
195
    }
620
621
229
    if (tablet_meta_pb.has_in_restore_mode()) {
622
229
        _in_restore_mode = tablet_meta_pb.in_restore_mode();
623
229
    }
624
625
229
    if (tablet_meta_pb.has_preferred_rowset_type()) {
626
17
        _preferred_rowset_type = tablet_meta_pb.preferred_rowset_type();
627
17
    }
628
629
229
    _storage_policy_id = tablet_meta_pb.storage_policy_id();
630
229
    if (tablet_meta_pb.has_cooldown_meta_id()) {
631
0
        _cooldown_meta_id = tablet_meta_pb.cooldown_meta_id();
632
0
    }
633
634
229
    if (tablet_meta_pb.has_delete_bitmap()) {
635
0
        int rst_ids_size = tablet_meta_pb.delete_bitmap().rowset_ids_size();
636
0
        int seg_ids_size = tablet_meta_pb.delete_bitmap().segment_ids_size();
637
0
        int versions_size = tablet_meta_pb.delete_bitmap().versions_size();
638
0
        int seg_maps_size = tablet_meta_pb.delete_bitmap().segment_delete_bitmaps_size();
639
0
        CHECK(rst_ids_size == seg_ids_size && seg_ids_size == seg_maps_size &&
640
0
              seg_maps_size == versions_size);
641
0
        for (size_t i = 0; i < rst_ids_size; ++i) {
642
0
            RowsetId rst_id;
643
0
            rst_id.init(tablet_meta_pb.delete_bitmap().rowset_ids(i));
644
0
            auto seg_id = tablet_meta_pb.delete_bitmap().segment_ids(i);
645
0
            uint32_t ver = tablet_meta_pb.delete_bitmap().versions(i);
646
0
            auto bitmap = tablet_meta_pb.delete_bitmap().segment_delete_bitmaps(i).data();
647
0
            delete_bitmap().delete_bitmap[{rst_id, seg_id, ver}] = roaring::Roaring::read(bitmap);
648
0
        }
649
0
    }
650
651
229
    if (tablet_meta_pb.has_binlog_config()) {
652
15
        _binlog_config = tablet_meta_pb.binlog_config();
653
15
    }
654
229
    _compaction_policy = tablet_meta_pb.compaction_policy();
655
229
    _time_series_compaction_goal_size_mbytes =
656
229
            tablet_meta_pb.time_series_compaction_goal_size_mbytes();
657
229
    _time_series_compaction_file_count_threshold =
658
229
            tablet_meta_pb.time_series_compaction_file_count_threshold();
659
229
    _time_series_compaction_time_threshold_seconds =
660
229
            tablet_meta_pb.time_series_compaction_time_threshold_seconds();
661
229
    _time_series_compaction_empty_rowsets_threshold =
662
229
            tablet_meta_pb.time_series_compaction_empty_rowsets_threshold();
663
229
    _time_series_compaction_level_threshold =
664
229
            tablet_meta_pb.time_series_compaction_level_threshold();
665
229
}
666
667
117
void TabletMeta::to_meta_pb(TabletMetaPB* tablet_meta_pb) {
668
117
    tablet_meta_pb->set_table_id(table_id());
669
117
    tablet_meta_pb->set_partition_id(partition_id());
670
117
    tablet_meta_pb->set_tablet_id(tablet_id());
671
117
    tablet_meta_pb->set_replica_id(replica_id());
672
117
    tablet_meta_pb->set_schema_hash(schema_hash());
673
117
    tablet_meta_pb->set_shard_id(shard_id());
674
117
    tablet_meta_pb->set_creation_time(creation_time());
675
117
    tablet_meta_pb->set_cumulative_layer_point(cumulative_layer_point());
676
117
    *(tablet_meta_pb->mutable_tablet_uid()) = tablet_uid().to_proto();
677
117
    tablet_meta_pb->set_tablet_type(_tablet_type);
678
117
    switch (tablet_state()) {
679
3
    case TABLET_NOTREADY:
680
3
        tablet_meta_pb->set_tablet_state(PB_NOTREADY);
681
3
        break;
682
86
    case TABLET_RUNNING:
683
86
        tablet_meta_pb->set_tablet_state(PB_RUNNING);
684
86
        break;
685
0
    case TABLET_TOMBSTONED:
686
0
        tablet_meta_pb->set_tablet_state(PB_TOMBSTONED);
687
0
        break;
688
0
    case TABLET_STOPPED:
689
0
        tablet_meta_pb->set_tablet_state(PB_STOPPED);
690
0
        break;
691
28
    case TABLET_SHUTDOWN:
692
28
        tablet_meta_pb->set_tablet_state(PB_SHUTDOWN);
693
28
        break;
694
117
    }
695
696
158
    for (auto& rs : _rs_metas) {
697
158
        rs->to_rowset_pb(tablet_meta_pb->add_rs_metas());
698
158
    }
699
117
    for (auto rs : _stale_rs_metas) {
700
0
        rs->to_rowset_pb(tablet_meta_pb->add_stale_rs_metas());
701
0
    }
702
117
    _schema->to_schema_pb(tablet_meta_pb->mutable_schema());
703
704
117
    tablet_meta_pb->set_in_restore_mode(in_restore_mode());
705
706
    // to avoid modify tablet meta to the greatest extend
707
117
    if (_preferred_rowset_type == BETA_ROWSET) {
708
117
        tablet_meta_pb->set_preferred_rowset_type(_preferred_rowset_type);
709
117
    }
710
117
    if (_storage_policy_id > 0) {
711
5
        tablet_meta_pb->set_storage_policy_id(_storage_policy_id);
712
5
    }
713
117
    if (_cooldown_meta_id.initialized()) {
714
5
        tablet_meta_pb->mutable_cooldown_meta_id()->CopyFrom(_cooldown_meta_id.to_proto());
715
5
    }
716
717
117
    tablet_meta_pb->set_enable_unique_key_merge_on_write(_enable_unique_key_merge_on_write);
718
719
117
    if (_enable_unique_key_merge_on_write) {
720
2
        std::set<RowsetId> stale_rs_ids;
721
2
        for (const auto& rowset : _stale_rs_metas) {
722
0
            stale_rs_ids.insert(rowset->rowset_id());
723
0
        }
724
2
        DeleteBitmapPB* delete_bitmap_pb = tablet_meta_pb->mutable_delete_bitmap();
725
2
        for (auto& [id, bitmap] : delete_bitmap().snapshot().delete_bitmap) {
726
2
            auto& [rowset_id, segment_id, ver] = id;
727
2
            if (stale_rs_ids.count(rowset_id) != 0) {
728
0
                continue;
729
0
            }
730
2
            delete_bitmap_pb->add_rowset_ids(rowset_id.to_string());
731
2
            delete_bitmap_pb->add_segment_ids(segment_id);
732
2
            delete_bitmap_pb->add_versions(ver);
733
2
            std::string bitmap_data(bitmap.getSizeInBytes(), '\0');
734
2
            bitmap.write(bitmap_data.data());
735
2
            *(delete_bitmap_pb->add_segment_delete_bitmaps()) = std::move(bitmap_data);
736
2
        }
737
2
    }
738
117
    _binlog_config.to_pb(tablet_meta_pb->mutable_binlog_config());
739
117
    tablet_meta_pb->set_compaction_policy(compaction_policy());
740
117
    tablet_meta_pb->set_time_series_compaction_goal_size_mbytes(
741
117
            time_series_compaction_goal_size_mbytes());
742
117
    tablet_meta_pb->set_time_series_compaction_file_count_threshold(
743
117
            time_series_compaction_file_count_threshold());
744
117
    tablet_meta_pb->set_time_series_compaction_time_threshold_seconds(
745
117
            time_series_compaction_time_threshold_seconds());
746
117
    tablet_meta_pb->set_time_series_compaction_empty_rowsets_threshold(
747
117
            time_series_compaction_empty_rowsets_threshold());
748
117
    tablet_meta_pb->set_time_series_compaction_level_threshold(
749
117
            time_series_compaction_level_threshold());
750
117
}
751
752
96
int64_t TabletMeta::mem_size() const {
753
96
    auto size = sizeof(TabletMeta);
754
96
    size += _schema->mem_size();
755
96
    return size;
756
96
}
757
758
2
void TabletMeta::to_json(string* json_string, json2pb::Pb2JsonOptions& options) {
759
2
    TabletMetaPB tablet_meta_pb;
760
2
    to_meta_pb(&tablet_meta_pb);
761
2
    json2pb::ProtoMessageToJson(tablet_meta_pb, json_string, options);
762
2
}
763
764
91
Version TabletMeta::max_version() const {
765
91
    Version max_version = {-1, 0};
766
145
    for (auto& rs_meta : _rs_metas) {
767
145
        if (rs_meta->end_version() > max_version.second) {
768
143
            max_version = rs_meta->version();
769
143
        }
770
145
    }
771
91
    return max_version;
772
91
}
773
774
0
size_t TabletMeta::version_count_cross_with_range(const Version& range) const {
775
0
    size_t count = 0;
776
0
    for (const auto& rs_meta : _rs_metas) {
777
0
        if (!(range.first > rs_meta->version().second || range.second < rs_meta->version().first)) {
778
0
            count++;
779
0
        }
780
0
    }
781
0
    return count;
782
0
}
783
784
362
Status TabletMeta::add_rs_meta(const RowsetMetaSharedPtr& rs_meta) {
785
    // check RowsetMeta is valid
786
1.36k
    for (auto& rs : _rs_metas) {
787
1.36k
        if (rs->version() == rs_meta->version()) {
788
0
            if (rs->rowset_id() != rs_meta->rowset_id()) {
789
0
                return Status::Error<PUSH_VERSION_ALREADY_EXIST>(
790
0
                        "version already exist. rowset_id={}, version={}, tablet={}",
791
0
                        rs->rowset_id().to_string(), rs->version().to_string(), tablet_id());
792
0
            } else {
793
                // rowsetid,version is equal, it is a duplicate req, skip it
794
0
                return Status::OK();
795
0
            }
796
0
        }
797
1.36k
    }
798
362
    _rs_metas.push_back(rs_meta);
799
362
    return Status::OK();
800
362
}
801
802
void TabletMeta::delete_rs_meta_by_version(const Version& version,
803
0
                                           std::vector<RowsetMetaSharedPtr>* deleted_rs_metas) {
804
0
    auto it = _rs_metas.begin();
805
0
    while (it != _rs_metas.end()) {
806
0
        if ((*it)->version() == version) {
807
0
            if (deleted_rs_metas != nullptr) {
808
0
                deleted_rs_metas->push_back(*it);
809
0
            }
810
0
            _rs_metas.erase(it);
811
0
            return;
812
0
        } else {
813
0
            ++it;
814
0
        }
815
0
    }
816
0
}
817
818
void TabletMeta::modify_rs_metas(const std::vector<RowsetMetaSharedPtr>& to_add,
819
                                 const std::vector<RowsetMetaSharedPtr>& to_delete,
820
11
                                 bool same_version) {
821
    // Remove to_delete rowsets from _rs_metas
822
11
    for (auto rs_to_del : to_delete) {
823
5
        auto it = _rs_metas.begin();
824
7
        while (it != _rs_metas.end()) {
825
7
            if (rs_to_del->version() == (*it)->version()) {
826
5
                _rs_metas.erase(it);
827
                // there should be only one rowset match the version
828
5
                break;
829
5
            } else {
830
2
                ++it;
831
2
            }
832
7
        }
833
5
    }
834
11
    if (!same_version) {
835
        // put to_delete rowsets in _stale_rs_metas.
836
6
        _stale_rs_metas.insert(_stale_rs_metas.end(), to_delete.begin(), to_delete.end());
837
6
    }
838
    // put to_add rowsets in _rs_metas.
839
11
    _rs_metas.insert(_rs_metas.end(), to_add.begin(), to_add.end());
840
11
}
841
842
// Use the passing "rs_metas" to replace the rs meta in this tablet meta
843
// Also clear the _stale_rs_metas because this tablet meta maybe copyied from
844
// an existing tablet before. Add after revise, only the passing "rs_metas"
845
// is needed.
846
3
void TabletMeta::revise_rs_metas(std::vector<RowsetMetaSharedPtr>&& rs_metas) {
847
3
    std::lock_guard<std::shared_mutex> wrlock(_meta_lock);
848
3
    _rs_metas = std::move(rs_metas);
849
3
    _stale_rs_metas.clear();
850
3
}
851
852
// This method should call after revise_rs_metas, since new rs_metas might be a subset
853
// of original tablet, we should revise the delete_bitmap according to current rowset.
854
//
855
// Delete bitmap is protected by Tablet::_meta_lock, we don't need to acquire the
856
// TabletMeta's _meta_lock
857
1
void TabletMeta::revise_delete_bitmap_unlocked(const DeleteBitmap& delete_bitmap) {
858
1
    _delete_bitmap = std::make_unique<DeleteBitmap>(tablet_id());
859
2
    for (auto rs : _rs_metas) {
860
2
        DeleteBitmap rs_bm(tablet_id());
861
2
        delete_bitmap.subset({rs->rowset_id(), 0, 0}, {rs->rowset_id(), UINT32_MAX, INT64_MAX},
862
2
                             &rs_bm);
863
2
        _delete_bitmap->merge(rs_bm);
864
2
    }
865
1
    for (auto rs : _stale_rs_metas) {
866
0
        DeleteBitmap rs_bm(tablet_id());
867
0
        delete_bitmap.subset({rs->rowset_id(), 0, 0}, {rs->rowset_id(), UINT32_MAX, INT64_MAX},
868
0
                             &rs_bm);
869
0
        _delete_bitmap->merge(rs_bm);
870
0
    }
871
1
}
872
873
0
void TabletMeta::delete_stale_rs_meta_by_version(const Version& version) {
874
0
    auto it = _stale_rs_metas.begin();
875
0
    while (it != _stale_rs_metas.end()) {
876
0
        if ((*it)->version() == version) {
877
0
            if (_enable_unique_key_merge_on_write) {
878
                // remove rowset delete bitmap
879
0
                delete_bitmap().remove({(*it)->rowset_id(), 0, 0},
880
0
                                       {(*it)->rowset_id(), UINT32_MAX, 0});
881
0
            }
882
0
            it = _stale_rs_metas.erase(it);
883
0
        } else {
884
0
            it++;
885
0
        }
886
0
    }
887
0
}
888
889
0
RowsetMetaSharedPtr TabletMeta::acquire_rs_meta_by_version(const Version& version) const {
890
0
    for (auto it : _rs_metas) {
891
0
        if (it->version() == version) {
892
0
            return it;
893
0
        }
894
0
    }
895
0
    return nullptr;
896
0
}
897
898
8
RowsetMetaSharedPtr TabletMeta::acquire_stale_rs_meta_by_version(const Version& version) const {
899
8
    for (auto it : _stale_rs_metas) {
900
0
        if (it->version() == version) {
901
0
            return it;
902
0
        }
903
0
    }
904
8
    return nullptr;
905
8
}
906
907
22
Status TabletMeta::set_partition_id(int64_t partition_id) {
908
22
    if ((_partition_id > 0 && _partition_id != partition_id) || partition_id < 1) {
909
0
        LOG(WARNING) << "cur partition id=" << _partition_id << " new partition id=" << partition_id
910
0
                     << " not equal";
911
0
    }
912
22
    _partition_id = partition_id;
913
22
    return Status::OK();
914
22
}
915
916
1
bool operator==(const TabletMeta& a, const TabletMeta& b) {
917
1
    if (a._table_id != b._table_id) return false;
918
1
    if (a._partition_id != b._partition_id) return false;
919
1
    if (a._tablet_id != b._tablet_id) return false;
920
1
    if (a._replica_id != b._replica_id) return false;
921
1
    if (a._schema_hash != b._schema_hash) return false;
922
1
    if (a._shard_id != b._shard_id) return false;
923
1
    if (a._creation_time != b._creation_time) return false;
924
1
    if (a._cumulative_layer_point != b._cumulative_layer_point) return false;
925
1
    if (a._tablet_uid != b._tablet_uid) return false;
926
1
    if (a._tablet_type != b._tablet_type) return false;
927
1
    if (a._tablet_state != b._tablet_state) return false;
928
1
    if (*a._schema != *b._schema) return false;
929
1
    if (a._rs_metas.size() != b._rs_metas.size()) return false;
930
1
    for (int i = 0; i < a._rs_metas.size(); ++i) {
931
0
        if (a._rs_metas[i] != b._rs_metas[i]) return false;
932
0
    }
933
1
    if (a._in_restore_mode != b._in_restore_mode) return false;
934
1
    if (a._preferred_rowset_type != b._preferred_rowset_type) return false;
935
1
    if (a._storage_policy_id != b._storage_policy_id) return false;
936
1
    if (a._compaction_policy != b._compaction_policy) return false;
937
1
    if (a._time_series_compaction_goal_size_mbytes != b._time_series_compaction_goal_size_mbytes)
938
0
        return false;
939
1
    if (a._time_series_compaction_file_count_threshold !=
940
1
        b._time_series_compaction_file_count_threshold)
941
0
        return false;
942
1
    if (a._time_series_compaction_time_threshold_seconds !=
943
1
        b._time_series_compaction_time_threshold_seconds)
944
0
        return false;
945
1
    if (a._time_series_compaction_empty_rowsets_threshold !=
946
1
        b._time_series_compaction_empty_rowsets_threshold)
947
0
        return false;
948
1
    if (a._time_series_compaction_level_threshold != b._time_series_compaction_level_threshold)
949
0
        return false;
950
1
    return true;
951
1
}
952
953
0
bool operator!=(const TabletMeta& a, const TabletMeta& b) {
954
0
    return !(a == b);
955
0
}
956
957
307
DeleteBitmap::DeleteBitmap(int64_t tablet_id) : _tablet_id(tablet_id) {
958
    // The default delete bitmap cache is set to 100MB,
959
    // which can be insufficient and cause performance issues when the amount of user data is large.
960
    // To mitigate the problem of an inadequate cache,
961
    // we will take the larger of 0.5% of the total memory and 100MB as the delete bitmap cache size.
962
307
    bool is_percent = false;
963
307
    int64_t delete_bitmap_agg_cache_cache_limit =
964
307
            ParseUtil::parse_mem_spec(config::delete_bitmap_dynamic_agg_cache_limit,
965
307
                                      MemInfo::mem_limit(), MemInfo::physical_mem(), &is_percent);
966
307
    _agg_cache.reset(new AggCache(delete_bitmap_agg_cache_cache_limit >
967
307
                                                  config::delete_bitmap_agg_cache_capacity
968
307
                                          ? delete_bitmap_agg_cache_cache_limit
969
307
                                          : config::delete_bitmap_agg_cache_capacity));
970
307
}
971
972
5
DeleteBitmap::DeleteBitmap(const DeleteBitmap& o) {
973
5
    delete_bitmap = o.delete_bitmap; // just copy data
974
5
    _tablet_id = o._tablet_id;
975
5
}
976
977
0
DeleteBitmap& DeleteBitmap::operator=(const DeleteBitmap& o) {
978
0
    delete_bitmap = o.delete_bitmap; // just copy data
979
0
    _tablet_id = o._tablet_id;
980
0
    return *this;
981
0
}
982
983
0
DeleteBitmap::DeleteBitmap(DeleteBitmap&& o) {
984
0
    delete_bitmap = std::move(o.delete_bitmap);
985
0
    _tablet_id = o._tablet_id;
986
0
}
987
988
0
DeleteBitmap& DeleteBitmap::operator=(DeleteBitmap&& o) {
989
0
    delete_bitmap = std::move(o.delete_bitmap);
990
0
    _tablet_id = o._tablet_id;
991
0
    return *this;
992
0
}
993
994
5
DeleteBitmap DeleteBitmap::snapshot() const {
995
5
    std::shared_lock l(lock);
996
5
    return DeleteBitmap(*this);
997
5
}
998
999
3
DeleteBitmap DeleteBitmap::snapshot(Version version) const {
1000
    // Take snapshot first, then remove keys greater than given version.
1001
3
    DeleteBitmap snapshot = this->snapshot();
1002
3
    auto it = snapshot.delete_bitmap.begin();
1003
412
    while (it != snapshot.delete_bitmap.end()) {
1004
409
        if (std::get<2>(it->first) > version) {
1005
4
            it = snapshot.delete_bitmap.erase(it);
1006
405
        } else {
1007
405
            it++;
1008
405
        }
1009
409
    }
1010
3
    return snapshot;
1011
3
}
1012
1013
459k
void DeleteBitmap::add(const BitmapKey& bmk, uint32_t row_id) {
1014
459k
    std::lock_guard l(lock);
1015
459k
    delete_bitmap[bmk].add(row_id);
1016
459k
}
1017
1018
0
int DeleteBitmap::remove(const BitmapKey& bmk, uint32_t row_id) {
1019
0
    std::lock_guard l(lock);
1020
0
    auto it = delete_bitmap.find(bmk);
1021
0
    if (it == delete_bitmap.end()) return -1;
1022
0
    it->second.remove(row_id);
1023
0
    return 0;
1024
0
}
1025
1026
8
void DeleteBitmap::remove(const BitmapKey& start, const BitmapKey& end) {
1027
8
    std::lock_guard l(lock);
1028
107
    for (auto it = delete_bitmap.lower_bound(start); it != delete_bitmap.end();) {
1029
101
        auto& [k, _] = *it;
1030
101
        if (k >= end) {
1031
2
            break;
1032
2
        }
1033
99
        it = delete_bitmap.erase(it);
1034
99
    }
1035
8
}
1036
1037
6
bool DeleteBitmap::contains(const BitmapKey& bmk, uint32_t row_id) const {
1038
6
    std::shared_lock l(lock);
1039
6
    auto it = delete_bitmap.find(bmk);
1040
6
    return it != delete_bitmap.end() && it->second.contains(row_id);
1041
6
}
1042
1043
2
bool DeleteBitmap::contains_agg(const BitmapKey& bmk, uint32_t row_id) const {
1044
2
    return get_agg(bmk)->contains(row_id);
1045
2
}
1046
1047
0
bool DeleteBitmap::empty() const {
1048
0
    std::shared_lock l(lock);
1049
0
    return delete_bitmap.empty();
1050
0
}
1051
1052
0
uint64_t DeleteBitmap::cardinality() const {
1053
0
    uint64_t res = 0;
1054
0
    for (auto entry : delete_bitmap) {
1055
0
        res += entry.second.cardinality();
1056
0
    }
1057
0
    return res;
1058
0
}
1059
1060
1
bool DeleteBitmap::contains_agg_without_cache(const BitmapKey& bmk, uint32_t row_id) const {
1061
1
    std::shared_lock l(lock);
1062
1
    DeleteBitmap::BitmapKey start {std::get<0>(bmk), std::get<1>(bmk), 0};
1063
1
    for (auto it = delete_bitmap.lower_bound(start); it != delete_bitmap.end(); ++it) {
1064
0
        auto& [k, bm] = *it;
1065
0
        if (std::get<0>(k) != std::get<0>(bmk) || std::get<1>(k) != std::get<1>(bmk) ||
1066
0
            std::get<2>(k) > std::get<2>(bmk)) {
1067
0
            break;
1068
0
        }
1069
0
        if (bm.contains(row_id)) {
1070
0
            return true;
1071
0
        }
1072
0
    }
1073
1
    return false;
1074
1
}
1075
1076
38
int DeleteBitmap::set(const BitmapKey& bmk, const roaring::Roaring& segment_delete_bitmap) {
1077
38
    std::lock_guard l(lock);
1078
38
    auto [_, inserted] = delete_bitmap.insert_or_assign(bmk, segment_delete_bitmap);
1079
38
    return inserted;
1080
38
}
1081
1082
3
int DeleteBitmap::get(const BitmapKey& bmk, roaring::Roaring* segment_delete_bitmap) const {
1083
3
    std::shared_lock l(lock);
1084
3
    auto it = delete_bitmap.find(bmk);
1085
3
    if (it == delete_bitmap.end()) return -1;
1086
3
    *segment_delete_bitmap = it->second; // copy
1087
3
    return 0;
1088
3
}
1089
1090
54
const roaring::Roaring* DeleteBitmap::get(const BitmapKey& bmk) const {
1091
54
    std::shared_lock l(lock);
1092
54
    auto it = delete_bitmap.find(bmk);
1093
54
    if (it == delete_bitmap.end()) return nullptr;
1094
41
    return &(it->second); // get address
1095
54
}
1096
1097
void DeleteBitmap::subset(const BitmapKey& start, const BitmapKey& end,
1098
3
                          DeleteBitmap* subset_rowset_map) const {
1099
3
    roaring::Roaring roaring;
1100
3
    DCHECK(start < end);
1101
3
    std::shared_lock l(lock);
1102
26
    for (auto it = delete_bitmap.lower_bound(start); it != delete_bitmap.end(); ++it) {
1103
25
        auto& [k, bm] = *it;
1104
25
        if (k >= end) {
1105
2
            break;
1106
2
        }
1107
23
        subset_rowset_map->set(k, bm);
1108
23
    }
1109
3
}
1110
1111
2
void DeleteBitmap::merge(const BitmapKey& bmk, const roaring::Roaring& segment_delete_bitmap) {
1112
2
    std::lock_guard l(lock);
1113
2
    auto [iter, succ] = delete_bitmap.emplace(bmk, segment_delete_bitmap);
1114
2
    if (!succ) {
1115
0
        iter->second |= segment_delete_bitmap;
1116
0
    }
1117
2
}
1118
1119
8
void DeleteBitmap::merge(const DeleteBitmap& other) {
1120
8
    std::lock_guard l(lock);
1121
29
    for (auto& i : other.delete_bitmap) {
1122
29
        auto [j, succ] = this->delete_bitmap.insert(i);
1123
29
        if (!succ) j->second |= i.second;
1124
29
    }
1125
8
}
1126
1127
// We cannot just copy the underlying memory to construct a string
1128
// due to equivalent objects may have different padding bytes.
1129
// Reading padding bytes is undefined behavior, neither copy nor
1130
// placement new will help simplify the code.
1131
// Refer to C11 standards §6.2.6.1/6 and §6.7.9/21 for more info.
1132
44
static std::string agg_cache_key(int64_t tablet_id, const DeleteBitmap::BitmapKey& bmk) {
1133
44
    std::string ret(sizeof(tablet_id) + sizeof(bmk), '\0');
1134
44
    *reinterpret_cast<int64_t*>(ret.data()) = tablet_id;
1135
44
    auto t = reinterpret_cast<DeleteBitmap::BitmapKey*>(ret.data() + sizeof(tablet_id));
1136
44
    std::get<RowsetId>(*t).version = std::get<RowsetId>(bmk).version;
1137
44
    std::get<RowsetId>(*t).hi = std::get<RowsetId>(bmk).hi;
1138
44
    std::get<RowsetId>(*t).mi = std::get<RowsetId>(bmk).mi;
1139
44
    std::get<RowsetId>(*t).lo = std::get<RowsetId>(bmk).lo;
1140
44
    std::get<1>(*t) = std::get<1>(bmk);
1141
44
    std::get<2>(*t) = std::get<2>(bmk);
1142
44
    return ret;
1143
44
}
1144
1145
44
std::shared_ptr<roaring::Roaring> DeleteBitmap::get_agg(const BitmapKey& bmk) const {
1146
44
    std::string key_str = agg_cache_key(_tablet_id, bmk); // Cache key container
1147
44
    CacheKey key(key_str);
1148
44
    Cache::Handle* handle = _agg_cache->repr()->lookup(key);
1149
1150
44
    AggCache::Value* val =
1151
44
            handle == nullptr
1152
44
                    ? nullptr
1153
44
                    : reinterpret_cast<AggCache::Value*>(_agg_cache->repr()->value(handle));
1154
    // FIXME: do we need a mutex here to get rid of duplicated initializations
1155
    //        of cache entries in some cases?
1156
44
    if (val == nullptr) { // Renew if needed, put a new Value to cache
1157
38
        val = new AggCache::Value();
1158
38
        {
1159
38
            std::shared_lock l(lock);
1160
38
            DeleteBitmap::BitmapKey start {std::get<0>(bmk), std::get<1>(bmk), 0};
1161
69
            for (auto it = delete_bitmap.lower_bound(start); it != delete_bitmap.end(); ++it) {
1162
66
                auto& [k, bm] = *it;
1163
66
                if (std::get<0>(k) != std::get<0>(bmk) || std::get<1>(k) != std::get<1>(bmk) ||
1164
66
                    std::get<2>(k) > std::get<2>(bmk)) {
1165
35
                    break;
1166
35
                }
1167
31
                val->bitmap |= bm;
1168
31
            }
1169
38
        }
1170
38
        size_t charge = val->bitmap.getSizeInBytes() + sizeof(AggCache::Value);
1171
38
        handle = _agg_cache->repr()->insert(key, val, charge, charge, CachePriority::NORMAL);
1172
38
    }
1173
1174
    // It is natural for the cache to reclaim the underlying memory
1175
44
    return std::shared_ptr<roaring::Roaring>(
1176
44
            &val->bitmap, [this, handle](...) { _agg_cache->repr()->release(handle); });
1177
44
}
1178
1179
std::atomic<DeleteBitmap::AggCachePolicy*> DeleteBitmap::AggCache::s_repr {nullptr};
1180
1181
0
std::string tablet_state_name(TabletState state) {
1182
0
    switch (state) {
1183
0
    case TABLET_NOTREADY:
1184
0
        return "TABLET_NOTREADY";
1185
1186
0
    case TABLET_RUNNING:
1187
0
        return "TABLET_RUNNING";
1188
1189
0
    case TABLET_TOMBSTONED:
1190
0
        return "TABLET_TOMBSTONED";
1191
1192
0
    case TABLET_STOPPED:
1193
0
        return "TABLET_STOPPED";
1194
1195
0
    case TABLET_SHUTDOWN:
1196
0
        return "TABLET_SHUTDOWN";
1197
1198
0
    default:
1199
0
        return "TabletState(" + std::to_string(state) + ")";
1200
0
    }
1201
0
}
1202
1203
} // namespace doris