Coverage Report

Created: 2026-09-30 16:20

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
/root/doris/be/src/olap/rowset_builder.cpp
Line
Count
Source
1
// Licensed to the Apache Software Foundation (ASF) under one
2
// or more contributor license agreements.  See the NOTICE file
3
// distributed with this work for additional information
4
// regarding copyright ownership.  The ASF licenses this file
5
// to you under the Apache License, Version 2.0 (the
6
// "License"); you may not use this file except in compliance
7
// with the License.  You may obtain a copy of the License at
8
//
9
//   http://www.apache.org/licenses/LICENSE-2.0
10
//
11
// Unless required by applicable law or agreed to in writing,
12
// software distributed under the License is distributed on an
13
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
14
// KIND, either express or implied.  See the License for the
15
// specific language governing permissions and limitations
16
// under the License.
17
18
#include "olap/rowset_builder.h"
19
20
#include <brpc/controller.h>
21
#include <fmt/format.h>
22
23
#include <filesystem>
24
#include <memory>
25
#include <ostream>
26
#include <string>
27
#include <utility>
28
29
// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
30
#include "common/compiler_util.h" // IWYU pragma: keep
31
#include "common/config.h"
32
#include "common/status.h"
33
#include "exec/tablet_info.h"
34
#include "io/fs/file_system.h"
35
#include "io/fs/file_writer.h" // IWYU pragma: keep
36
#include "olap/calc_delete_bitmap_executor.h"
37
#include "olap/olap_define.h"
38
#include "olap/partial_update_info.h"
39
#include "olap/rowset/beta_rowset.h"
40
#include "olap/rowset/beta_rowset_writer.h"
41
#include "olap/rowset/pending_rowset_helper.h"
42
#include "olap/rowset/rowset_meta.h"
43
#include "olap/rowset/rowset_meta_manager.h"
44
#include "olap/rowset/rowset_writer.h"
45
#include "olap/rowset/rowset_writer_context.h"
46
#include "olap/schema_change.h"
47
#include "olap/storage_engine.h"
48
#include "olap/tablet_manager.h"
49
#include "olap/tablet_meta.h"
50
#include "olap/tablet_schema.h"
51
#include "olap/tablet_schema_cache.h"
52
#include "olap/txn_manager.h"
53
#include "runtime/memory/global_memory_arbitrator.h"
54
#include "util/brpc_client_cache.h"
55
#include "util/brpc_closure.h"
56
#include "util/debug_points.h"
57
#include "util/mem_info.h"
58
#include "util/stopwatch.hpp"
59
#include "util/time.h"
60
#include "util/trace.h"
61
#include "vec/common/schema_util.h"
62
#include "vec/core/block.h"
63
64
namespace doris {
65
#include "common/compile_check_begin.h"
66
using namespace ErrorCode;
67
68
BaseRowsetBuilder::BaseRowsetBuilder(const WriteRequest& req, RuntimeProfile* profile)
69
29
        : _req(req), _tablet_schema(std::make_shared<TabletSchema>()) {
70
29
    if (profile != nullptr) {
71
29
        _init_profile(profile);
72
29
    }
73
29
}
74
75
RowsetBuilder::RowsetBuilder(StorageEngine& engine, const WriteRequest& req,
76
                             RuntimeProfile* profile)
77
29
        : BaseRowsetBuilder(req, profile), _engine(engine) {}
78
79
29
void BaseRowsetBuilder::_init_profile(RuntimeProfile* profile) {
80
29
    DCHECK(profile != nullptr);
81
29
    _profile = profile->create_child(fmt::format("RowsetBuilder {}", _req.tablet_id), true, true);
82
29
    _build_rowset_timer = ADD_TIMER(_profile, "BuildRowsetTime");
83
29
    _submit_delete_bitmap_timer = ADD_TIMER(_profile, "DeleteBitmapSubmitTime");
84
29
    _wait_delete_bitmap_timer = ADD_TIMER(_profile, "DeleteBitmapWaitTime");
85
29
}
86
87
0
void RowsetBuilder::_init_profile(RuntimeProfile* profile) {
88
0
    DCHECK(profile != nullptr);
89
0
    BaseRowsetBuilder::_init_profile(profile);
90
0
    _commit_txn_timer = ADD_TIMER(_profile, "CommitTxnTime");
91
0
}
92
93
29
BaseRowsetBuilder::~BaseRowsetBuilder() {
94
29
    if (!_is_init) {
95
1
        return;
96
1
    }
97
98
28
    if (_calc_delete_bitmap_token != nullptr) {
99
28
        _calc_delete_bitmap_token->cancel();
100
28
    }
101
28
}
102
103
29
RowsetBuilder::~RowsetBuilder() {
104
29
    if (_is_init && !_is_committed) {
105
5
        _garbage_collection();
106
5
    }
107
29
}
108
109
208
Tablet* RowsetBuilder::tablet() {
110
208
    return static_cast<Tablet*>(_tablet.get());
111
208
}
112
113
0
TabletSharedPtr RowsetBuilder::tablet_sptr() {
114
0
    return std::static_pointer_cast<Tablet>(_tablet);
115
0
}
116
117
5
void RowsetBuilder::_garbage_collection() {
118
5
    Status rollback_status;
119
5
    TxnManager* txn_mgr = _engine.txn_manager();
120
5
    if (tablet() != nullptr) {
121
5
        rollback_status = txn_mgr->rollback_txn(_req.partition_id, *tablet(), _req.txn_id);
122
5
    }
123
    // has to check rollback status, because the rowset maybe committed in this thread and
124
    // published in another thread, then rollback will fail.
125
    // when rollback failed should not delete rowset
126
5
    if (rollback_status.ok()) {
127
5
        _engine.add_unused_rowset(_rowset);
128
5
    }
129
5
}
130
131
3
Status BaseRowsetBuilder::init_mow_context(std::shared_ptr<MowContext>& mow_context) {
132
3
    std::lock_guard lck(tablet()->get_header_lock());
133
3
    _max_version_in_flush_phase = tablet()->max_version_unlocked();
134
3
    std::vector<RowsetSharedPtr> rowset_ptrs;
135
    // tablet is under alter process. The delete bitmap will be calculated after conversion.
136
3
    if (tablet()->tablet_state() == TABLET_NOTREADY) {
137
        // Disable 'partial_update' when the tablet is undergoing a 'schema changing process'
138
0
        if (_req.table_schema_param->is_partial_update()) {
139
0
            return Status::InternalError(
140
0
                    "Unable to do 'partial_update' when "
141
0
                    "the tablet is undergoing a 'schema changing process'");
142
0
        }
143
0
        _rowset_ids->clear();
144
3
    } else {
145
3
        RETURN_IF_ERROR(
146
3
                tablet()->get_all_rs_id_unlocked(_max_version_in_flush_phase, _rowset_ids.get()));
147
3
        rowset_ptrs = tablet()->get_rowset_by_ids(_rowset_ids.get());
148
3
    }
149
3
    _delete_bitmap = std::make_shared<DeleteBitmap>(tablet()->tablet_id());
150
3
    mow_context = std::make_shared<MowContext>(_max_version_in_flush_phase, _req.txn_id,
151
3
                                               _rowset_ids, rowset_ptrs, _delete_bitmap);
152
3
    return Status::OK();
153
3
}
154
155
28
Status RowsetBuilder::check_tablet_version_count() {
156
28
    auto max_version_config = _tablet->max_version_config();
157
28
    auto version_count = tablet()->version_count();
158
28
    DBUG_EXECUTE_IF("RowsetBuilder.check_tablet_version_count.too_many_version",
159
28
                    { version_count = INT_MAX; });
160
    // Trigger TOO MANY VERSION error first
161
28
    if (version_count > max_version_config) {
162
0
        return Status::Error<TOO_MANY_VERSION>(
163
0
                "failed to init rowset builder. version count: {}, exceed limit: {}, "
164
0
                "tablet: {}. Please reduce the frequency of loading data or adjust the "
165
0
                "max_tablet_version_num or time_series_max_tablet_version_num in be.conf to a "
166
0
                "larger value.",
167
0
                version_count, max_version_config, _tablet->tablet_id());
168
0
    }
169
    // (TODO Refrain) Maybe we can use a configurable param instead of hardcoded values '100'.
170
    // max_version_config must > 100, otherwise silent errors will occur.
171
28
    if ((!config::disable_auto_compaction &&
172
28
         !_tablet->tablet_meta()->tablet_schema()->disable_auto_compaction()) &&
173
28
        (version_count > max_version_config - 100) &&
174
28
        !GlobalMemoryArbitrator::is_exceed_soft_mem_limit(GB_EXCHANGE_BYTE)) {
175
        // Trigger compaction
176
0
        auto st = _engine.submit_compaction_task(
177
0
                tablet_sptr(), CompactionType::CUMULATIVE_COMPACTION, true, true, 2);
178
0
        if (!st.ok()) [[unlikely]] {
179
0
            LOG(WARNING) << "failed to trigger compaction, tablet_id=" << _tablet->tablet_id()
180
0
                         << " : " << st;
181
0
        }
182
0
    }
183
28
    return Status::OK();
184
28
}
185
186
28
Status RowsetBuilder::prepare_txn() {
187
28
    return tablet()->prepare_txn(_req.partition_id, _req.txn_id, _req.load_id, false);
188
28
}
189
190
29
Status RowsetBuilder::init() {
191
29
    _tablet = DORIS_TRY(_engine.get_tablet(_req.tablet_id));
192
28
    std::shared_ptr<MowContext> mow_context;
193
28
    if (_tablet->enable_unique_key_merge_on_write()) {
194
3
        RETURN_IF_ERROR(init_mow_context(mow_context));
195
3
    }
196
197
28
    RETURN_IF_ERROR(check_tablet_version_count());
198
199
28
    auto version_count = tablet()->version_count() + tablet()->stale_version_count();
200
28
    if (tablet()->avg_rs_meta_serialize_size() * version_count >
201
28
        config::tablet_meta_serialize_size_limit) {
202
0
        return Status::Error<TOO_MANY_VERSION>(
203
0
                "failed to init rowset builder. meta serialize size : {}, exceed limit: {}, "
204
0
                "tablet: {}. Please reduce the frequency of loading data or adjust the "
205
0
                "max_tablet_version_num in be.conf to a larger value.",
206
0
                tablet()->avg_rs_meta_serialize_size() * version_count,
207
0
                config::tablet_meta_serialize_size_limit, _tablet->tablet_id());
208
0
    }
209
210
28
    RETURN_IF_ERROR(prepare_txn());
211
212
28
    DBUG_EXECUTE_IF("BaseRowsetBuilder::init.check_partial_update_column_num", {
213
28
        if (_req.table_schema_param->partial_update_input_columns().size() !=
214
28
            dp->param<int>("column_num")) {
215
28
            return Status::InternalError("partial update input column num wrong!");
216
28
        };
217
28
    })
218
    // build tablet schema in request level
219
28
    RETURN_IF_ERROR(_build_current_tablet_schema(_req.index_id, _req.table_schema_param.get(),
220
28
                                                 *_tablet->tablet_schema()));
221
28
    RowsetWriterContext context;
222
28
    context.txn_id = _req.txn_id;
223
28
    context.load_id = _req.load_id;
224
28
    context.rowset_state = PREPARED;
225
28
    context.segments_overlap = OVERLAPPING;
226
28
    context.tablet_schema = _tablet_schema;
227
28
    context.newest_write_timestamp = UnixSeconds();
228
28
    context.tablet_id = _req.tablet_id;
229
28
    context.index_id = _req.index_id;
230
28
    context.tablet = _tablet;
231
28
    context.enable_segcompaction = true;
232
28
    context.write_type = DataWriteType::TYPE_DIRECT;
233
28
    context.mow_context = mow_context;
234
28
    context.write_file_cache = _req.write_file_cache;
235
28
    context.partial_update_info = _partial_update_info;
236
28
    _rowset_writer = DORIS_TRY(_tablet->create_rowset_writer(context, false));
237
28
    _pending_rs_guard = _engine.pending_local_rowsets().add(context.rowset_id);
238
239
28
    _calc_delete_bitmap_token = _engine.calc_delete_bitmap_executor()->create_token();
240
241
28
    _is_init = true;
242
28
    return Status::OK();
243
28
}
244
245
23
Status BaseRowsetBuilder::build_rowset() {
246
23
    std::lock_guard<std::mutex> l(_lock);
247
23
    DCHECK(_is_init) << "rowset builder is supposed be to initialized before "
248
0
                        "build_rowset() being called";
249
250
23
    SCOPED_TIMER(_build_rowset_timer);
251
    // use rowset meta manager to save meta
252
23
    RETURN_NOT_OK_STATUS_WITH_WARN(_rowset_writer->build(_rowset), "fail to build rowset");
253
23
    return Status::OK();
254
23
}
255
256
17
Status BaseRowsetBuilder::submit_calc_delete_bitmap_task() {
257
17
    if (!_tablet->enable_unique_key_merge_on_write() || _rowset->num_segments() == 0) {
258
14
        return Status::OK();
259
14
    }
260
3
    std::lock_guard<std::mutex> l(_lock);
261
3
    SCOPED_TIMER(_submit_delete_bitmap_timer);
262
3
    if (_partial_update_info && _partial_update_info->is_flexible_partial_update()) {
263
0
        if (_rowset->num_segments() > 1) {
264
            // in flexible partial update, when there are more one segment in one load,
265
            // we need to do alignment process for same keys between segments, we haven't
266
            // implemented it yet and just report an error when encouter this situation
267
0
            return Status::NotSupported(
268
0
                    "too large input data in flexible partial update, Please "
269
0
                    "reduce the amount of data imported in a single load.");
270
0
        }
271
0
    }
272
273
3
    auto* beta_rowset = reinterpret_cast<BetaRowset*>(_rowset.get());
274
3
    std::vector<segment_v2::SegmentSharedPtr> segments;
275
3
    RETURN_IF_ERROR(beta_rowset->load_segments(&segments));
276
3
    if (segments.size() > 1) {
277
        // calculate delete bitmap between segments
278
0
        if (config::enable_calc_delete_bitmap_between_segments_concurrently) {
279
0
            RETURN_IF_ERROR(_calc_delete_bitmap_token->submit(
280
0
                    _tablet, _tablet_schema, _rowset->rowset_id(), segments, _delete_bitmap));
281
0
        } else {
282
0
            RETURN_IF_ERROR(_tablet->calc_delete_bitmap_between_segments(
283
0
                    _tablet_schema, _rowset->rowset_id(), segments, _delete_bitmap));
284
0
        }
285
0
    }
286
287
    // For partial update, we need to fill in the entire row of data, during the calculation
288
    // of the delete bitmap. This operation is resource-intensive, and we need to minimize
289
    // the number of times it occurs. Therefore, we skip this operation here.
290
3
    if (_partial_update_info->is_partial_update()) {
291
        // for partial update, the delete bitmap calculation is done while append_block()
292
        // we print it's summarize logs here before commit.
293
0
        LOG(INFO) << fmt::format(
294
0
                "{} calc delete bitmap summary before commit: tablet({}), txn_id({}), "
295
0
                "rowset_ids({}), cur max_version({}), bitmap num({}), bitmap_cardinality({}), num "
296
0
                "rows updated({}), num rows new added({}), num rows deleted({}), total rows({})",
297
0
                _partial_update_info->partial_update_mode_str(), tablet()->tablet_id(), _req.txn_id,
298
0
                _rowset_ids->size(), rowset_writer()->context().mow_context->max_version,
299
0
                _delete_bitmap->get_delete_bitmap_count(), _delete_bitmap->cardinality(),
300
0
                rowset_writer()->num_rows_updated(), rowset_writer()->num_rows_new_added(),
301
0
                rowset_writer()->num_rows_deleted(), rowset_writer()->num_rows());
302
0
        return Status::OK();
303
0
    }
304
305
3
    LOG(INFO) << "submit calc delete bitmap task to executor, tablet_id: " << tablet()->tablet_id()
306
3
              << ", txn_id: " << _req.txn_id;
307
3
    return BaseTablet::commit_phase_update_delete_bitmap(_tablet, _rowset, *_rowset_ids,
308
3
                                                         _delete_bitmap, segments, _req.txn_id,
309
3
                                                         _calc_delete_bitmap_token.get(), nullptr);
310
3
}
311
312
17
Status BaseRowsetBuilder::wait_calc_delete_bitmap() {
313
17
    if (!_tablet->enable_unique_key_merge_on_write() || _partial_update_info->is_partial_update()) {
314
14
        return Status::OK();
315
14
    }
316
3
    std::lock_guard<std::mutex> l(_lock);
317
3
    SCOPED_TIMER(_wait_delete_bitmap_timer);
318
3
    RETURN_IF_ERROR(_calc_delete_bitmap_token->wait());
319
3
    return Status::OK();
320
3
}
321
322
23
Status RowsetBuilder::commit_txn() {
323
23
    if (tablet()->enable_unique_key_merge_on_write() &&
324
23
        config::enable_merge_on_write_correctness_check && _rowset->num_rows() != 0 &&
325
23
        tablet()->tablet_state() != TABLET_NOTREADY) {
326
3
        auto st = tablet()->check_delete_bitmap_correctness(
327
3
                _delete_bitmap, _rowset->end_version() - 1, _req.txn_id, *_rowset_ids);
328
3
        if (!st.ok()) {
329
0
            LOG(WARNING) << fmt::format(
330
0
                    "[tablet_id:{}][txn_id:{}][load_id:{}][partition_id:{}] "
331
0
                    "delete bitmap correctness check failed in commit phase!",
332
0
                    _req.tablet_id, _req.txn_id, UniqueId(_req.load_id).to_string(),
333
0
                    _req.partition_id);
334
0
            return st;
335
0
        }
336
3
    }
337
23
    std::lock_guard<std::mutex> l(_lock);
338
23
    SCOPED_TIMER(_commit_txn_timer);
339
340
    // Transfer ownership of `PendingRowsetGuard` to `TxnManager`
341
23
    Status res = _engine.txn_manager()->commit_txn(
342
23
            _req.partition_id, *tablet(), _req.txn_id, _req.load_id, _rowset,
343
23
            std::move(_pending_rs_guard), false, _partial_update_info);
344
345
23
    if (!res && !res.is<PUSH_TRANSACTION_ALREADY_EXIST>()) {
346
0
        LOG(WARNING) << "Failed to commit txn: " << _req.txn_id
347
0
                     << " for rowset: " << _rowset->rowset_id();
348
0
        return res;
349
0
    }
350
23
    if (_tablet->enable_unique_key_merge_on_write()) {
351
3
        _engine.txn_manager()->set_txn_related_delete_bitmap(
352
3
                _req.partition_id, _req.txn_id, tablet()->tablet_id(), tablet()->tablet_uid(), true,
353
3
                _delete_bitmap, *_rowset_ids, _partial_update_info);
354
3
    }
355
356
23
    _is_committed = true;
357
23
    return Status::OK();
358
23
}
359
360
0
Status BaseRowsetBuilder::cancel() {
361
0
    std::lock_guard<std::mutex> l(_lock);
362
0
    if (_is_cancelled) {
363
0
        return Status::OK();
364
0
    }
365
0
    if (_calc_delete_bitmap_token != nullptr) {
366
0
        _calc_delete_bitmap_token->cancel();
367
0
    }
368
0
    _is_cancelled = true;
369
0
    return Status::OK();
370
0
}
371
372
Status BaseRowsetBuilder::_build_current_tablet_schema(
373
        int64_t index_id, const OlapTableSchemaParam* table_schema_param,
374
28
        const TabletSchema& ori_tablet_schema) {
375
    // find the right index id
376
28
    const OlapTableIndexSchema* index_schema = nullptr;
377
28
    for (const auto* schema : table_schema_param->indexes()) {
378
13
        if (schema->index_id == index_id) {
379
13
            index_schema = schema;
380
13
            break;
381
13
        }
382
13
    }
383
384
28
    auto cache_key = TabletSchemaCache::build_load_schema_cache_key(
385
28
            index_id, table_schema_param, ori_tablet_schema, index_schema);
386
28
    auto cached_schema = TabletSchemaCache::instance()->lookup_schema(cache_key);
387
28
    if (cached_schema.first != nullptr) {
388
21
        _tablet_schema = cached_schema.second;
389
21
        TabletSchemaCache::instance()->release(cached_schema.first);
390
21
    } else {
391
7
        if (index_schema != nullptr && !index_schema->columns.empty() &&
392
7
            index_schema->columns[0]->unique_id() >= 0) {
393
0
            _tablet_schema->shawdow_copy_without_columns(ori_tablet_schema);
394
0
            _tablet_schema->build_current_tablet_schema(
395
0
                    index_id, cast_set<int32_t>(table_schema_param->version()), index_schema,
396
0
                    ori_tablet_schema);
397
7
        } else {
398
7
            _tablet_schema->copy_from(ori_tablet_schema);
399
7
        }
400
7
        _tablet_schema->set_table_id(table_schema_param->table_id());
401
7
        _tablet_schema->set_db_id(table_schema_param->db_id());
402
7
        if (table_schema_param->is_partial_update()) {
403
0
            _tablet_schema->set_auto_increment_column(table_schema_param->auto_increment_coulumn());
404
0
        }
405
7
        auto inserted_schema = TabletSchemaCache::instance()->insert(cache_key, _tablet_schema);
406
7
        _tablet_schema = inserted_schema.second;
407
7
        TabletSchemaCache::instance()->release(inserted_schema.first);
408
7
    }
409
28
    if (_tablet_schema->schema_version() > ori_tablet_schema.schema_version()) {
410
        // After schema change, should include extracted column
411
        // For example: a table has two columns, k and v
412
        // After adding a column v2, the schema version increases, max_version_schema needs to be updated.
413
        // _tablet_schema includes k, v, and v2
414
        // if v is a variant, need to add the columns decomposed from the v to the _tablet_schema.
415
0
        if (_tablet_schema->num_variant_columns() > 0) {
416
0
            TabletSchemaSPtr max_version_schema = std::make_shared<TabletSchema>();
417
0
            max_version_schema->copy_from(*_tablet_schema);
418
0
            max_version_schema->copy_extracted_columns(ori_tablet_schema);
419
0
            _tablet->update_max_version_schema(max_version_schema);
420
0
        } else {
421
0
            _tablet->update_max_version_schema(_tablet_schema);
422
0
        }
423
0
    }
424
425
    // set partial update columns info
426
28
    _partial_update_info = std::make_shared<PartialUpdateInfo>();
427
28
    RETURN_IF_ERROR(_partial_update_info->init(
428
28
            tablet()->tablet_id(), _req.txn_id, *_tablet_schema,
429
28
            table_schema_param->unique_key_update_mode(),
430
28
            table_schema_param->partial_update_new_key_policy(),
431
28
            table_schema_param->partial_update_input_columns(),
432
28
            table_schema_param->is_strict_mode(), table_schema_param->timestamp_ms(),
433
28
            table_schema_param->nano_seconds(), table_schema_param->timezone(),
434
28
            table_schema_param->auto_increment_coulumn(),
435
28
            table_schema_param->sequence_map_col_uid(), _max_version_in_flush_phase));
436
28
    return Status::OK();
437
28
}
438
#include "common/compile_check_end.h"
439
} // namespace doris