Coverage Report

Created: 2026-01-18 13:22

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