Coverage Report

Created: 2025-10-16 19:56

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
/root/doris/be/src/olap/memtable_writer.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/memtable_writer.h"
19
20
#include <fmt/format.h>
21
22
#include <filesystem>
23
#include <ostream>
24
#include <string>
25
#include <utility>
26
27
#include "common/compiler_util.h" // IWYU pragma: keep
28
#include "common/config.h"
29
#include "common/logging.h"
30
#include "common/status.h"
31
#include "exec/tablet_info.h"
32
#include "io/fs/file_writer.h" // IWYU pragma: keep
33
#include "olap/memtable.h"
34
#include "olap/memtable_flush_executor.h"
35
#include "olap/memtable_memory_limiter.h"
36
#include "olap/rowset/beta_rowset_writer.h"
37
#include "olap/rowset/rowset_writer.h"
38
#include "olap/schema_change.h"
39
#include "olap/storage_engine.h"
40
#include "olap/tablet_schema.h"
41
#include "runtime/exec_env.h"
42
#include "runtime/memory/mem_tracker.h"
43
#include "service/backend_options.h"
44
#include "util/mem_info.h"
45
#include "util/stopwatch.hpp"
46
#include "vec/core/block.h"
47
48
namespace doris {
49
using namespace ErrorCode;
50
51
17
MemTableWriter::MemTableWriter(const WriteRequest& req) : _req(req) {}
52
53
17
MemTableWriter::~MemTableWriter() {
54
17
    if (!_is_init) {
55
2
        return;
56
2
    }
57
15
    if (_flush_token != nullptr) {
58
        // cancel and wait all memtables in flush queue to be finished
59
15
        _flush_token->cancel();
60
15
    }
61
15
    _mem_table.reset();
62
15
}
63
64
Status MemTableWriter::init(std::shared_ptr<RowsetWriter> rowset_writer,
65
                            TabletSchemaSPtr tablet_schema,
66
                            std::shared_ptr<PartialUpdateInfo> partial_update_info,
67
15
                            std::shared_ptr<WorkloadGroup> wg_sptr, bool unique_key_mow) {
68
15
    _rowset_writer = rowset_writer;
69
15
    _tablet_schema = tablet_schema;
70
15
    _unique_key_mow = unique_key_mow;
71
15
    _partial_update_info = partial_update_info;
72
15
    _resource_ctx = thread_context()->resource_ctx();
73
74
15
    _reset_mem_table();
75
76
    // create flush handler
77
    // by assigning segment_id to memtable before submiting to flush executor,
78
    // we can make sure same keys sort in the same order in all replicas.
79
15
    RETURN_IF_ERROR(
80
15
            ExecEnv::GetInstance()->storage_engine().memtable_flush_executor()->create_flush_token(
81
15
                    _flush_token, _rowset_writer, _req.is_high_priority, wg_sptr));
82
83
15
    _is_init = true;
84
15
    return Status::OK();
85
15
}
86
87
Status MemTableWriter::write(const vectorized::Block* block,
88
20
                             const DorisVector<uint32_t>& row_idxs) {
89
20
    if (UNLIKELY(row_idxs.empty())) {
90
0
        return Status::OK();
91
0
    }
92
20
    _lock_watch.start();
93
20
    std::lock_guard<std::mutex> l(_lock);
94
20
    _lock_watch.stop();
95
20
    if (_is_cancelled) {
96
0
        return _cancel_status;
97
0
    }
98
20
    if (!_is_init) {
99
0
        return Status::Error<NOT_INITIALIZED>("delta segment writer has not been initialized");
100
0
    }
101
20
    if (_is_closed) {
102
0
        return Status::Error<ALREADY_CLOSED>("write block after closed tablet_id={}, load_id={}-{}",
103
0
                                             _req.tablet_id, _req.load_id.hi(), _req.load_id.lo());
104
0
    }
105
106
    // Flush and reset memtable if it is raw rows great than int32_t.
107
20
    int64_t raw_rows = _mem_table->raw_rows();
108
20
    DBUG_EXECUTE_IF("MemTableWriter.too_many_raws",
109
20
                    { raw_rows = std::numeric_limits<int32_t>::max(); });
110
20
    if (raw_rows + row_idxs.size() > std::numeric_limits<int32_t>::max()) {
111
0
        RETURN_IF_ERROR(_flush_memtable());
112
0
    }
113
114
20
    _total_received_rows += row_idxs.size();
115
20
    auto st = _mem_table->insert(block, row_idxs);
116
117
    // Reset memtable immediately after insert failure to prevent potential flush operations.
118
    // This is a defensive measure because:
119
    // 1. When insert fails (e.g., memory allocation failure during add_rows),
120
    //    the memtable is in an inconsistent state and should not be flushed
121
    // 2. However, memory pressure might trigger a flush operation on this failed memtable
122
    // 3. By resetting here, we ensure the failed memtable won't be included in any subsequent flush,
123
    //    thus preventing potential crashes
124
20
    DBUG_EXECUTE_IF("MemTableWriter.write.random_insert_error", {
125
20
        if (rand() % 100 < (100 * dp->param("percent", 0.3))) {
126
20
            st = Status::InternalError<false>("write memtable random failed for debug");
127
20
        }
128
20
    });
129
20
    if (!st.ok()) [[unlikely]] {
130
0
        _reset_mem_table();
131
0
        return st;
132
0
    }
133
134
20
    if (UNLIKELY(_mem_table->need_agg() && config::enable_shrink_memory)) {
135
0
        _mem_table->shrink_memtable_by_agg();
136
0
    }
137
20
    if (UNLIKELY(_mem_table->need_flush())) {
138
0
        RETURN_IF_ERROR(_flush_memtable());
139
0
    }
140
141
20
    return Status::OK();
142
20
}
143
144
0
Status MemTableWriter::_flush_memtable() {
145
0
    auto s = _flush_memtable_async();
146
0
    _reset_mem_table();
147
0
    if (UNLIKELY(!s.ok())) {
148
0
        return s;
149
0
    }
150
0
    return Status::OK();
151
0
}
152
153
15
Status MemTableWriter::_flush_memtable_async() {
154
15
    DCHECK(_flush_token != nullptr);
155
15
    std::shared_ptr<MemTable> memtable;
156
15
    {
157
15
        std::lock_guard<std::mutex> l(_mem_table_ptr_lock);
158
15
        memtable = _mem_table;
159
15
        _mem_table = nullptr;
160
15
    }
161
15
    {
162
15
        std::lock_guard<std::mutex> l(_mem_table_ptr_lock);
163
15
        memtable->update_mem_type(MemType::WRITE_FINISHED);
164
15
        _freezed_mem_tables.push_back(memtable);
165
15
    }
166
15
    return _flush_token->submit(memtable);
167
15
}
168
169
0
Status MemTableWriter::flush_async() {
170
0
    std::lock_guard<std::mutex> l(_lock);
171
    // Three calling paths:
172
    // 1. call by local, from `VTabletWriterV2::_write_memtable`.
173
    // 2. call by remote, from `LoadChannelMgr::_get_load_channel`.
174
    // 3. call by daemon thread, from `handle_paused_queries` -> `flush_workload_group_memtables`.
175
0
    SCOPED_SWITCH_THREAD_MEM_TRACKER_LIMITER(_resource_ctx->memory_context()->mem_tracker());
176
0
    if (!_is_init || _is_closed) {
177
        // This writer is uninitialized or closed before flushing, do nothing.
178
        // We return OK instead of NOT_INITIALIZED or ALREADY_CLOSED.
179
        // Because this method maybe called when trying to reduce mem consumption,
180
        // and at that time, the writer may not be initialized yet and that is a normal case.
181
0
        return Status::OK();
182
0
    }
183
184
0
    if (_is_cancelled) {
185
0
        return _cancel_status;
186
0
    }
187
188
0
    VLOG_NOTICE << "flush memtable to reduce mem consumption. memtable size: "
189
0
                << PrettyPrinter::print_bytes(_mem_table->memory_usage())
190
0
                << ", tablet: " << _req.tablet_id << ", load id: " << print_id(_req.load_id);
191
0
    auto s = _flush_memtable_async();
192
0
    _reset_mem_table();
193
0
    return s;
194
0
}
195
196
9
Status MemTableWriter::wait_flush() {
197
9
    {
198
9
        std::lock_guard<std::mutex> l(_lock);
199
9
        if (!_is_init || _is_closed) {
200
            // return OK instead of NOT_INITIALIZED or ALREADY_CLOSED for same reason
201
            // as described in flush_async()
202
9
            return Status::OK();
203
9
        }
204
0
        if (_is_cancelled) {
205
0
            return _cancel_status;
206
0
        }
207
0
    }
208
0
    SCOPED_RAW_TIMER(&_wait_flush_time_ns);
209
0
    RETURN_IF_ERROR(_flush_token->wait());
210
0
    return Status::OK();
211
0
}
212
213
15
void MemTableWriter::_reset_mem_table() {
214
15
    {
215
15
        std::lock_guard<std::mutex> l(_mem_table_ptr_lock);
216
15
        _mem_table.reset(new MemTable(_req.tablet_id, _tablet_schema, _req.slots, _req.tuple_desc,
217
15
                                      _unique_key_mow, _partial_update_info.get(), _resource_ctx));
218
15
    }
219
220
15
    _segment_num++;
221
15
}
222
223
15
Status MemTableWriter::close() {
224
15
    _lock_watch.start();
225
15
    std::lock_guard<std::mutex> l(_lock);
226
15
    _lock_watch.stop();
227
15
    if (_is_cancelled) {
228
0
        return _cancel_status;
229
0
    }
230
15
    if (!_is_init) {
231
0
        return Status::Error<NOT_INITIALIZED>("delta segment writer has not been initialized");
232
0
    }
233
15
    if (_is_closed) {
234
0
        LOG(WARNING) << "close after closed tablet_id=" << _req.tablet_id
235
0
                     << " load_id=" << _req.load_id;
236
0
        return Status::OK();
237
0
    }
238
239
15
    auto s = _flush_memtable_async();
240
15
    {
241
15
        std::lock_guard<std::mutex> l(_mem_table_ptr_lock);
242
15
        _mem_table.reset();
243
15
    }
244
15
    _is_closed = true;
245
15
    if (UNLIKELY(!s.ok())) {
246
0
        return s;
247
15
    } else {
248
15
        return Status::OK();
249
15
    }
250
15
}
251
252
15
Status MemTableWriter::_do_close_wait() {
253
15
    SCOPED_RAW_TIMER(&_close_wait_time_ns);
254
15
    std::lock_guard<std::mutex> l(_lock);
255
15
    DCHECK(_is_init)
256
0
            << "delta writer is supposed be to initialized before close_wait() being called";
257
258
15
    if (_is_cancelled) {
259
0
        return _cancel_status;
260
0
    }
261
262
15
    Status st;
263
    // return error if previous flush failed
264
15
    {
265
15
        SCOPED_RAW_TIMER(&_wait_flush_time_ns);
266
15
        st = _flush_token->wait();
267
15
    }
268
15
    if (UNLIKELY(!st.ok())) {
269
0
        LOG(WARNING) << "previous flush failed tablet " << _req.tablet_id;
270
0
        return st;
271
0
    }
272
273
15
    if (_rowset_writer->num_rows() + _flush_token->memtable_stat().merged_rows !=
274
15
        _total_received_rows) {
275
0
        LOG(WARNING) << "the rows number written doesn't match, rowset num rows written to file: "
276
0
                     << _rowset_writer->num_rows()
277
0
                     << ", merged_rows: " << _flush_token->memtable_stat().merged_rows
278
0
                     << ", total received rows: " << _total_received_rows;
279
0
        return Status::InternalError("rows number written by delta writer dosen't match");
280
0
    }
281
282
    // print slow log if wait more than 1s
283
15
    if (_wait_flush_time_ns > 1000UL * 1000 * 1000) {
284
0
        LOG(INFO) << "close delta writer for tablet: " << _req.tablet_id
285
0
                  << ", load id: " << print_id(_req.load_id) << ", wait close for "
286
0
                  << _wait_flush_time_ns << "(ns), stats: " << _flush_token->get_stats();
287
0
    }
288
289
15
    return Status::OK();
290
15
}
291
292
15
void MemTableWriter::_update_profile(RuntimeProfile* profile) {
293
    // NOTE: MemTableWriter may be accessed when profile is out of scope, in MemTableMemoryLimiter.
294
    // To avoid accessing dangling pointers, we cannot make profile as a member of MemTableWriter.
295
15
    auto child =
296
15
            profile->create_child(fmt::format("MemTableWriter {}", _req.tablet_id), true, true);
297
15
    auto lock_timer = ADD_TIMER(child, "LockTime");
298
15
    auto sort_timer = ADD_TIMER(child, "MemTableSortTime");
299
15
    auto agg_timer = ADD_TIMER(child, "MemTableAggTime");
300
15
    auto memtable_duration_timer = ADD_TIMER(child, "MemTableDurationTime");
301
15
    auto segment_writer_timer = ADD_TIMER(child, "SegmentWriterTime");
302
15
    auto wait_flush_timer = ADD_TIMER(child, "MemTableWaitFlushTime");
303
15
    auto put_into_output_timer = ADD_TIMER(child, "MemTablePutIntoOutputTime");
304
15
    auto delete_bitmap_timer = ADD_TIMER(child, "DeleteBitmapTime");
305
15
    auto close_wait_timer = ADD_TIMER(child, "CloseWaitTime");
306
15
    auto sort_times = ADD_COUNTER(child, "MemTableSortTimes", TUnit::UNIT);
307
15
    auto agg_times = ADD_COUNTER(child, "MemTableAggTimes", TUnit::UNIT);
308
15
    auto segment_num = ADD_COUNTER(child, "SegmentNum", TUnit::UNIT);
309
15
    auto raw_rows_num = ADD_COUNTER(child, "RawRowNum", TUnit::UNIT);
310
15
    auto merged_rows_num = ADD_COUNTER(child, "MergedRowNum", TUnit::UNIT);
311
312
15
    COUNTER_UPDATE(lock_timer, _lock_watch.elapsed_time());
313
15
    COUNTER_SET(delete_bitmap_timer, _rowset_writer->delete_bitmap_ns());
314
15
    COUNTER_SET(segment_writer_timer, _rowset_writer->segment_writer_ns());
315
15
    COUNTER_SET(wait_flush_timer, _wait_flush_time_ns);
316
15
    COUNTER_SET(close_wait_timer, _close_wait_time_ns);
317
15
    COUNTER_SET(segment_num, _segment_num);
318
15
    const auto& memtable_stat = _flush_token->memtable_stat();
319
15
    COUNTER_SET(sort_timer, memtable_stat.sort_ns);
320
15
    COUNTER_SET(agg_timer, memtable_stat.agg_ns);
321
15
    COUNTER_SET(memtable_duration_timer, memtable_stat.duration_ns);
322
15
    COUNTER_SET(put_into_output_timer, memtable_stat.put_into_output_ns);
323
15
    COUNTER_SET(sort_times, memtable_stat.sort_times);
324
15
    COUNTER_SET(agg_times, memtable_stat.agg_times);
325
15
    COUNTER_SET(raw_rows_num, memtable_stat.raw_rows);
326
15
    COUNTER_SET(merged_rows_num, memtable_stat.merged_rows);
327
15
}
328
329
15
Status MemTableWriter::cancel() {
330
15
    return cancel_with_status(Status::Cancelled("already cancelled"));
331
15
}
332
333
15
Status MemTableWriter::cancel_with_status(const Status& st) {
334
15
    std::lock_guard<std::mutex> l(_lock);
335
15
    if (_is_cancelled) {
336
0
        return Status::OK();
337
0
    }
338
15
    {
339
15
        std::lock_guard<std::mutex> l(_mem_table_ptr_lock);
340
15
        _mem_table.reset();
341
15
    }
342
15
    if (_flush_token != nullptr) {
343
        // cancel and wait all memtables in flush queue to be finished
344
15
        _flush_token->cancel();
345
15
    }
346
15
    _is_cancelled = true;
347
15
    _cancel_status = st;
348
15
    return Status::OK();
349
15
}
350
351
15
const FlushStatistic& MemTableWriter::get_flush_token_stats() {
352
15
    return _flush_token->get_stats();
353
15
}
354
355
20
uint64_t MemTableWriter::flush_running_count() const {
356
20
    return _flush_token == nullptr ? 0 : _flush_token->get_stats().flush_running_count.load();
357
20
}
358
359
0
int64_t MemTableWriter::mem_consumption(MemType mem) {
360
0
    if (!_is_init) {
361
        // This method may be called before this writer is initialized.
362
        // So _flush_token may be null.
363
0
        return 0;
364
0
    }
365
0
    int64_t mem_usage = 0;
366
0
    {
367
0
        std::lock_guard<std::mutex> l(_mem_table_ptr_lock);
368
0
        for (const auto& mem_table : _freezed_mem_tables) {
369
0
            auto mem_table_sptr = mem_table.lock();
370
0
            if (mem_table_sptr != nullptr && mem_table_sptr->get_mem_type() == mem) {
371
0
                mem_usage += mem_table_sptr->memory_usage();
372
0
            }
373
0
        }
374
0
    }
375
0
    return mem_usage;
376
0
}
377
378
0
int64_t MemTableWriter::active_memtable_mem_consumption() {
379
0
    std::lock_guard<std::mutex> l(_mem_table_ptr_lock);
380
0
    return _mem_table != nullptr ? _mem_table->memory_usage() : 0;
381
0
}
382
383
} // namespace doris