Coverage Report

Created: 2025-06-03 16:15

/root/doris/be/src/olap/rowset_builder.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/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/txn_manager.h"
52
#include "runtime/memory/global_memory_arbitrator.h"
53
#include "util/brpc_client_cache.h"
54
#include "util/brpc_closure.h"
55
#include "util/debug_points.h"
56
#include "util/mem_info.h"
57
#include "util/stopwatch.hpp"
58
#include "util/time.h"
59
#include "util/trace.h"
60
#include "vec/common/schema_util.h"
61
#include "vec/core/block.h"
62
63
namespace doris {
64
using namespace ErrorCode;
65
66
BaseRowsetBuilder::BaseRowsetBuilder(const WriteRequest& req, RuntimeProfile* profile)
67
29
        : _req(req), _tablet_schema(std::make_shared<TabletSchema>()) {
68
29
    _init_profile(profile);
69
29
}
70
71
RowsetBuilder::RowsetBuilder(StorageEngine& engine, const WriteRequest& req,
72
                             RuntimeProfile* profile)
73
29
        : BaseRowsetBuilder(req, profile), _engine(engine) {}
74
75
29
void BaseRowsetBuilder::_init_profile(RuntimeProfile* profile) {
76
29
    _profile = profile->create_child(fmt::format("RowsetBuilder {}", _req.tablet_id), true, true);
77
29
    _build_rowset_timer = ADD_TIMER(_profile, "BuildRowsetTime");
78
29
    _submit_delete_bitmap_timer = ADD_TIMER(_profile, "DeleteBitmapSubmitTime");
79
29
    _wait_delete_bitmap_timer = ADD_TIMER(_profile, "DeleteBitmapWaitTime");
80
29
}
81
82
0
void RowsetBuilder::_init_profile(RuntimeProfile* profile) {
83
0
    BaseRowsetBuilder::_init_profile(profile);
84
0
    _commit_txn_timer = ADD_TIMER(_profile, "CommitTxnTime");
85
0
}
86
87
29
BaseRowsetBuilder::~BaseRowsetBuilder() {
88
29
    if (!_is_init) {
89
1
        return;
90
1
    }
91
92
28
    if (_calc_delete_bitmap_token != nullptr) {
93
28
        _calc_delete_bitmap_token->cancel();
94
28
    }
95
28
}
96
97
29
RowsetBuilder::~RowsetBuilder() {
98
29
    if (_is_init && !_is_committed) {
99
5
        _garbage_collection();
100
5
    }
101
29
}
102
103
236
Tablet* RowsetBuilder::tablet() {
104
236
    return static_cast<Tablet*>(_tablet.get());
105
236
}
106
107
0
TabletSharedPtr RowsetBuilder::tablet_sptr() {
108
0
    return std::static_pointer_cast<Tablet>(_tablet);
109
0
}
110
111
5
void RowsetBuilder::_garbage_collection() {
112
5
    Status rollback_status;
113
5
    TxnManager* txn_mgr = _engine.txn_manager();
114
5
    if (tablet() != nullptr) {
115
5
        rollback_status = txn_mgr->rollback_txn(_req.partition_id, *tablet(), _req.txn_id);
116
5
    }
117
    // has to check rollback status, because the rowset maybe committed in this thread and
118
    // published in another thread, then rollback will fail.
119
    // when rollback failed should not delete rowset
120
5
    if (rollback_status.ok()) {
121
5
        _engine.add_unused_rowset(_rowset);
122
5
    }
123
5
}
124
125
3
Status BaseRowsetBuilder::init_mow_context(std::shared_ptr<MowContext>& mow_context) {
126
3
    std::lock_guard<std::shared_mutex> lck(tablet()->get_header_lock());
127
3
    _max_version_in_flush_phase = tablet()->max_version_unlocked();
128
3
    std::vector<RowsetSharedPtr> rowset_ptrs;
129
    // tablet is under alter process. The delete bitmap will be calculated after conversion.
130
3
    if (tablet()->tablet_state() == TABLET_NOTREADY) {
131
        // Disable 'partial_update' when the tablet is undergoing a 'schema changing process'
132
0
        if (_req.table_schema_param->is_partial_update()) {
133
0
            return Status::InternalError(
134
0
                    "Unable to do 'partial_update' when "
135
0
                    "the tablet is undergoing a 'schema changing process'");
136
0
        }
137
0
        _rowset_ids.clear();
138
3
    } else {
139
3
        RETURN_IF_ERROR(
140
3
                tablet()->get_all_rs_id_unlocked(_max_version_in_flush_phase, &_rowset_ids));
141
3
        rowset_ptrs = tablet()->get_rowset_by_ids(&_rowset_ids);
142
3
    }
143
3
    _delete_bitmap = std::make_shared<DeleteBitmap>(tablet()->tablet_id());
144
3
    mow_context = std::make_shared<MowContext>(_max_version_in_flush_phase, _req.txn_id,
145
3
                                               _rowset_ids, rowset_ptrs, _delete_bitmap);
146
3
    return Status::OK();
147
3
}
148
149
28
Status RowsetBuilder::check_tablet_version_count() {
150
28
    bool injection = false;
151
28
    DBUG_EXECUTE_IF("RowsetBuilder.check_tablet_version_count.too_many_version",
152
28
                    { injection = true; });
153
28
    int32_t max_version_config = _tablet->max_version_config();
154
28
    if (injection) {
155
        // do not return if injection
156
28
    } else if (!_tablet->exceed_version_limit(max_version_config - 100) ||
157
28
               GlobalMemoryArbitrator::is_exceed_soft_mem_limit(GB_EXCHANGE_BYTE)) {
158
28
        return Status::OK();
159
28
    }
160
    //trigger compaction
161
0
    auto st = _engine.submit_compaction_task(tablet_sptr(), CompactionType::CUMULATIVE_COMPACTION,
162
0
                                             true);
163
0
    if (!st.ok()) [[unlikely]] {
164
0
        LOG(WARNING) << "failed to trigger compaction, tablet_id=" << _tablet->tablet_id() << " : "
165
0
                     << st;
166
0
    }
167
0
    int version_count = tablet()->version_count();
168
0
    DBUG_EXECUTE_IF("RowsetBuilder.check_tablet_version_count.too_many_version",
169
0
                    { version_count = INT_MAX; });
170
0
    if (version_count > max_version_config) {
171
0
        return Status::Error<TOO_MANY_VERSION>(
172
0
                "failed to init rowset builder. version count: {}, exceed limit: {}, "
173
0
                "tablet: {}. Please reduce the frequency of loading data or adjust the "
174
0
                "max_tablet_version_num or time_series_max_tablet_version_num in be.conf to a "
175
0
                "larger value.",
176
0
                version_count, max_version_config, _tablet->tablet_id());
177
0
    }
178
0
    return Status::OK();
179
0
}
180
181
28
Status RowsetBuilder::prepare_txn() {
182
28
    std::shared_lock base_migration_lock(tablet()->get_migration_lock(), std::defer_lock);
183
28
    if (!base_migration_lock.try_lock_for(
184
28
                std::chrono::milliseconds(config::migration_lock_timeout_ms))) {
185
0
        return Status::Error<TRY_LOCK_FAILED>("try_lock migration lock failed after {}ms",
186
0
                                              config::migration_lock_timeout_ms);
187
0
    }
188
28
    std::lock_guard<std::mutex> push_lock(tablet()->get_push_lock());
189
28
    return _engine.txn_manager()->prepare_txn(_req.partition_id, *tablet(), _req.txn_id,
190
28
                                              _req.load_id);
191
28
}
192
193
29
Status RowsetBuilder::init() {
194
29
    _tablet = DORIS_TRY(_engine.get_tablet(_req.tablet_id));
195
28
    std::shared_ptr<MowContext> mow_context;
196
28
    if (_tablet->enable_unique_key_merge_on_write()) {
197
3
        RETURN_IF_ERROR(init_mow_context(mow_context));
198
3
    }
199
200
28
    if (!config::disable_auto_compaction &&
201
28
        !_tablet->tablet_meta()->tablet_schema()->disable_auto_compaction()) {
202
28
        RETURN_IF_ERROR(check_tablet_version_count());
203
28
    }
204
205
28
    int version_count = tablet()->version_count() + tablet()->stale_version_count();
206
28
    if (tablet()->avg_rs_meta_serialize_size() * version_count >
207
28
        config::tablet_meta_serialize_size_limit) {
208
0
        return Status::Error<TOO_MANY_VERSION>(
209
0
                "failed to init rowset builder. meta serialize size : {}, exceed limit: {}, "
210
0
                "tablet: {}. Please reduce the frequency of loading data or adjust the "
211
0
                "max_tablet_version_num in be.conf to a larger value.",
212
0
                tablet()->avg_rs_meta_serialize_size() * version_count,
213
0
                config::tablet_meta_serialize_size_limit, _tablet->tablet_id());
214
0
    }
215
216
28
    RETURN_IF_ERROR(prepare_txn());
217
218
28
    DBUG_EXECUTE_IF("BaseRowsetBuilder::init.check_partial_update_column_num", {
219
28
        if (_req.table_schema_param->partial_update_input_columns().size() !=
220
28
            dp->param<int>("column_num")) {
221
28
            return Status::InternalError("partial update input column num wrong!");
222
28
        };
223
28
    })
224
    // build tablet schema in request level
225
28
    RETURN_IF_ERROR(_build_current_tablet_schema(_req.index_id, _req.table_schema_param.get(),
226
28
                                                 *_tablet->tablet_schema()));
227
28
    RowsetWriterContext context;
228
28
    context.txn_id = _req.txn_id;
229
28
    context.load_id = _req.load_id;
230
28
    context.rowset_state = PREPARED;
231
28
    context.segments_overlap = OVERLAPPING;
232
28
    context.tablet_schema = _tablet_schema;
233
28
    context.newest_write_timestamp = UnixSeconds();
234
28
    context.tablet_id = _req.tablet_id;
235
28
    context.index_id = _req.index_id;
236
28
    context.tablet = _tablet;
237
28
    context.write_type = DataWriteType::TYPE_DIRECT;
238
28
    context.mow_context = mow_context;
239
28
    context.write_file_cache = _req.write_file_cache;
240
28
    context.partial_update_info = _partial_update_info;
241
28
    _rowset_writer = DORIS_TRY(_tablet->create_rowset_writer(context, false));
242
28
    _pending_rs_guard = _engine.pending_local_rowsets().add(context.rowset_id);
243
244
28
    _calc_delete_bitmap_token = _engine.calc_delete_bitmap_executor()->create_token();
245
246
28
    _is_init = true;
247
28
    return Status::OK();
248
28
}
249
250
23
Status BaseRowsetBuilder::build_rowset() {
251
23
    std::lock_guard<std::mutex> l(_lock);
252
23
    DCHECK(_is_init) << "rowset builder is supposed be to initialized before "
253
0
                        "build_rowset() being called";
254
255
23
    SCOPED_TIMER(_build_rowset_timer);
256
    // use rowset meta manager to save meta
257
23
    RETURN_NOT_OK_STATUS_WITH_WARN(_rowset_writer->build(_rowset), "fail to build rowset");
258
23
    return Status::OK();
259
23
}
260
261
17
Status BaseRowsetBuilder::submit_calc_delete_bitmap_task() {
262
17
    if (!_tablet->enable_unique_key_merge_on_write() || _rowset->num_segments() == 0) {
263
14
        return Status::OK();
264
14
    }
265
3
    std::lock_guard<std::mutex> l(_lock);
266
3
    SCOPED_TIMER(_submit_delete_bitmap_timer);
267
3
    if (_partial_update_info && _partial_update_info->is_flexible_partial_update()) {
268
0
        if (_rowset->num_segments() > 1) {
269
            // in flexible partial update, when there are more one segment in one load,
270
            // we need to do alignment process for same keys between segments, we haven't
271
            // implemented it yet and just report an error when encouter this situation
272
0
            return Status::NotSupported(
273
0
                    "too large input data in flexible partial update, Please "
274
0
                    "reduce the amount of data imported in a single load.");
275
0
        }
276
0
    }
277
278
3
    auto* beta_rowset = reinterpret_cast<BetaRowset*>(_rowset.get());
279
3
    std::vector<segment_v2::SegmentSharedPtr> segments;
280
3
    RETURN_IF_ERROR(beta_rowset->load_segments(&segments));
281
3
    if (segments.size() > 1) {
282
        // calculate delete bitmap between segments
283
0
        if (config::enable_calc_delete_bitmap_between_segments_concurrently) {
284
0
            RETURN_IF_ERROR(_calc_delete_bitmap_token->submit(_tablet, _rowset->rowset_id(),
285
0
                                                              segments, _delete_bitmap));
286
0
        } else {
287
0
            RETURN_IF_ERROR(_tablet->calc_delete_bitmap_between_segments(_rowset->rowset_id(),
288
0
                                                                         segments, _delete_bitmap));
289
0
        }
290
0
    }
291
292
    // For partial update, we need to fill in the entire row of data, during the calculation
293
    // of the delete bitmap. This operation is resource-intensive, and we need to minimize
294
    // the number of times it occurs. Therefore, we skip this operation here.
295
3
    if (_partial_update_info->is_partial_update()) {
296
        // for partial update, the delete bitmap calculation is done while append_block()
297
        // we print it's summarize logs here before commit.
298
0
        LOG(INFO) << fmt::format(
299
0
                "{} calc delete bitmap summary before commit: tablet({}), txn_id({}), "
300
0
                "rowset_ids({}), cur max_version({}), bitmap num({}), bitmap_cardinality({}), num "
301
0
                "rows updated({}), num rows new added({}), num rows deleted({}), total rows({})",
302
0
                _partial_update_info->partial_update_mode_str(), tablet()->tablet_id(), _req.txn_id,
303
0
                _rowset_ids.size(), rowset_writer()->context().mow_context->max_version,
304
0
                _delete_bitmap->get_delete_bitmap_count(), _delete_bitmap->cardinality(),
305
0
                rowset_writer()->num_rows_updated(), rowset_writer()->num_rows_new_added(),
306
0
                rowset_writer()->num_rows_deleted(), rowset_writer()->num_rows());
307
0
        return Status::OK();
308
0
    }
309
310
3
    LOG(INFO) << "submit calc delete bitmap task to executor, tablet_id: " << tablet()->tablet_id()
311
3
              << ", txn_id: " << _req.txn_id;
312
3
    return BaseTablet::commit_phase_update_delete_bitmap(_tablet, _rowset, _rowset_ids,
313
3
                                                         _delete_bitmap, segments, _req.txn_id,
314
3
                                                         _calc_delete_bitmap_token.get(), nullptr);
315
3
}
316
317
17
Status BaseRowsetBuilder::wait_calc_delete_bitmap() {
318
17
    if (!_tablet->enable_unique_key_merge_on_write() || _partial_update_info->is_partial_update()) {
319
14
        return Status::OK();
320
14
    }
321
3
    std::lock_guard<std::mutex> l(_lock);
322
3
    SCOPED_TIMER(_wait_delete_bitmap_timer);
323
3
    RETURN_IF_ERROR(_calc_delete_bitmap_token->wait());
324
3
    return Status::OK();
325
3
}
326
327
23
Status RowsetBuilder::commit_txn() {
328
23
    if (tablet()->enable_unique_key_merge_on_write() &&
329
23
        config::enable_merge_on_write_correctness_check && _rowset->num_rows() != 0 &&
330
23
        tablet()->tablet_state() != TABLET_NOTREADY) {
331
3
        auto st = tablet()->check_delete_bitmap_correctness(
332
3
                _delete_bitmap, _rowset->end_version() - 1, _req.txn_id, _rowset_ids);
333
3
        if (!st.ok()) {
334
0
            LOG(WARNING) << fmt::format(
335
0
                    "[tablet_id:{}][txn_id:{}][load_id:{}][partition_id:{}] "
336
0
                    "delete bitmap correctness check failed in commit phase!",
337
0
                    _req.tablet_id, _req.txn_id, UniqueId(_req.load_id).to_string(),
338
0
                    _req.partition_id);
339
0
            return st;
340
0
        }
341
3
    }
342
23
    std::lock_guard<std::mutex> l(_lock);
343
23
    SCOPED_TIMER(_commit_txn_timer);
344
345
23
    const RowsetWriterContext& rw_ctx = _rowset_writer->context();
346
23
    if (rw_ctx.tablet_schema->num_variant_columns() > 0 && _rowset->num_rows() > 0) {
347
        // Need to merge schema with `rw_ctx.merged_tablet_schema` in prior,
348
        // merged schema keeps the newest merged schema for the rowset, which is updated and merged
349
        // during flushing segments.
350
0
        if (rw_ctx.merged_tablet_schema != nullptr) {
351
0
            RETURN_IF_ERROR(tablet()->update_by_least_common_schema(rw_ctx.merged_tablet_schema));
352
0
        } else {
353
            // We should merge rowset schema further, in case that the merged_tablet_schema maybe null
354
            // when enable_memtable_on_sink_node is true, the merged_tablet_schema will not be passed to
355
            // the destination backend.
356
            // update tablet schema when meet variant columns, before commit_txn
357
            // Eg. rowset schema:       A(int),    B(float),  C(int), D(int)
358
            // _tabelt->tablet_schema:  A(bigint), B(double)
359
            //  => update_schema:       A(bigint), B(double), C(int), D(int)
360
0
            RETURN_IF_ERROR(tablet()->update_by_least_common_schema(rw_ctx.tablet_schema));
361
0
        }
362
0
    }
363
364
    // Transfer ownership of `PendingRowsetGuard` to `TxnManager`
365
23
    Status res = _engine.txn_manager()->commit_txn(
366
23
            _req.partition_id, *tablet(), _req.txn_id, _req.load_id, _rowset,
367
23
            std::move(_pending_rs_guard), false, _partial_update_info);
368
369
23
    if (!res && !res.is<PUSH_TRANSACTION_ALREADY_EXIST>()) {
370
0
        LOG(WARNING) << "Failed to commit txn: " << _req.txn_id
371
0
                     << " for rowset: " << _rowset->rowset_id();
372
0
        return res;
373
0
    }
374
23
    if (_tablet->enable_unique_key_merge_on_write()) {
375
3
        _engine.txn_manager()->set_txn_related_delete_bitmap(
376
3
                _req.partition_id, _req.txn_id, tablet()->tablet_id(), tablet()->tablet_uid(), true,
377
3
                _delete_bitmap, _rowset_ids, _partial_update_info);
378
3
    }
379
380
23
    _is_committed = true;
381
23
    return Status::OK();
382
23
}
383
384
0
Status BaseRowsetBuilder::cancel() {
385
0
    std::lock_guard<std::mutex> l(_lock);
386
0
    if (_is_cancelled) {
387
0
        return Status::OK();
388
0
    }
389
0
    if (_calc_delete_bitmap_token != nullptr) {
390
0
        _calc_delete_bitmap_token->cancel();
391
0
    }
392
0
    _is_cancelled = true;
393
0
    return Status::OK();
394
0
}
395
396
Status BaseRowsetBuilder::_build_current_tablet_schema(
397
        int64_t index_id, const OlapTableSchemaParam* table_schema_param,
398
28
        const TabletSchema& ori_tablet_schema) {
399
    // find the right index id
400
28
    int i = 0;
401
28
    auto indexes = table_schema_param->indexes();
402
28
    for (; i < indexes.size(); i++) {
403
13
        if (indexes[i]->index_id == index_id) {
404
13
            break;
405
13
        }
406
13
    }
407
28
    if (!indexes.empty() && !indexes[i]->columns.empty() &&
408
28
        indexes[i]->columns[0]->unique_id() >= 0) {
409
0
        _tablet_schema->shawdow_copy_without_columns(ori_tablet_schema);
410
0
        _tablet_schema->build_current_tablet_schema(index_id, table_schema_param->version(),
411
0
                                                    indexes[i], ori_tablet_schema);
412
28
    } else {
413
28
        _tablet_schema->copy_from(ori_tablet_schema);
414
28
    }
415
28
    if (_tablet_schema->schema_version() > ori_tablet_schema.schema_version()) {
416
        // After schema change, should include extracted column
417
        // For example: a table has two columns, k and v
418
        // After adding a column v2, the schema version increases, max_version_schema needs to be updated.
419
        // _tablet_schema includes k, v, and v2
420
        // if v is a variant, need to add the columns decomposed from the v to the _tablet_schema.
421
0
        if (_tablet_schema->num_variant_columns() > 0) {
422
0
            TabletSchemaSPtr max_version_schema = std::make_shared<TabletSchema>();
423
0
            max_version_schema->copy_from(*_tablet_schema);
424
0
            max_version_schema->copy_extracted_columns(ori_tablet_schema);
425
0
            _tablet->update_max_version_schema(max_version_schema);
426
0
        } else {
427
0
            _tablet->update_max_version_schema(_tablet_schema);
428
0
        }
429
0
    }
430
431
28
    _tablet_schema->set_table_id(table_schema_param->table_id());
432
28
    _tablet_schema->set_db_id(table_schema_param->db_id());
433
28
    if (table_schema_param->is_partial_update()) {
434
0
        _tablet_schema->set_auto_increment_column(table_schema_param->auto_increment_coulumn());
435
0
    }
436
    // set partial update columns info
437
28
    _partial_update_info = std::make_shared<PartialUpdateInfo>();
438
28
    RETURN_IF_ERROR(_partial_update_info->init(
439
28
            tablet()->tablet_id(), _req.txn_id, *_tablet_schema,
440
28
            table_schema_param->unique_key_update_mode(),
441
28
            table_schema_param->partial_update_input_columns(),
442
28
            table_schema_param->is_strict_mode(), table_schema_param->timestamp_ms(),
443
28
            table_schema_param->nano_seconds(), table_schema_param->timezone(),
444
28
            table_schema_param->auto_increment_coulumn(),
445
28
            table_schema_param->sequence_map_col_uid(), _max_version_in_flush_phase));
446
28
    return Status::OK();
447
28
}
448
449
} // namespace doris