be/src/exec/spill/spill_file_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 "exec/spill/spill_file_writer.h" |
19 | | |
20 | | #include "agent/be_exec_version_manager.h" |
21 | | #include "common/config.h" |
22 | | #include "common/status.h" |
23 | | #include "exec/spill/spill_file.h" |
24 | | #include "exec/spill/spill_file_manager.h" |
25 | | #include "io/fs/local_file_system.h" |
26 | | #include "io/fs/local_file_writer.h" |
27 | | #include "runtime/exec_env.h" |
28 | | #include "runtime/query_context.h" |
29 | | #include "runtime/runtime_state.h" |
30 | | #include "runtime/thread_context.h" |
31 | | |
32 | | namespace doris { |
33 | | |
34 | | SpillFileWriter::SpillFileWriter(const std::shared_ptr<SpillFile>& spill_file, RuntimeState* state, |
35 | | RuntimeProfile* profile, SpillDataDir* data_dir, |
36 | | const std::string& spill_dir) |
37 | 299 | : _spill_file_wptr(spill_file), |
38 | 299 | _data_dir(data_dir), |
39 | 299 | _spill_dir(spill_dir), |
40 | 299 | _max_part_size(config::spill_file_part_size_bytes), |
41 | 299 | _resource_ctx(state->get_query_ctx()->resource_ctx()) { |
42 | | // Common counters |
43 | 299 | RuntimeProfile* common_profile = profile->get_child("CommonCounters"); |
44 | 299 | DCHECK(common_profile != nullptr); |
45 | 299 | _memory_used_counter = common_profile->get_counter("MemoryUsage"); |
46 | | |
47 | | // Register this writer as the active writer for the SpillFile. |
48 | 299 | spill_file->_active_writer = this; |
49 | | |
50 | | // Custom (spill-specific) counters |
51 | 299 | RuntimeProfile* custom_profile = profile->get_child("CustomCounters"); |
52 | 299 | _write_file_timer = custom_profile->get_counter("SpillWriteFileTime"); |
53 | 299 | _serialize_timer = custom_profile->get_counter("SpillWriteSerializeBlockTime"); |
54 | 299 | _write_block_counter = custom_profile->get_counter("SpillWriteBlockCount"); |
55 | 299 | _write_block_bytes_counter = custom_profile->get_counter("SpillWriteBlockBytes"); |
56 | 299 | _write_file_total_size = custom_profile->get_counter("SpillWriteFileBytes"); |
57 | 299 | _write_file_current_size = custom_profile->get_counter("SpillWriteFileCurrentBytes"); |
58 | 299 | _write_rows_counter = custom_profile->get_counter("SpillWriteRows"); |
59 | 299 | _total_file_count = custom_profile->get_counter("SpillWriteFileTotalCount"); |
60 | 299 | } |
61 | | |
62 | 299 | SpillFileWriter::~SpillFileWriter() { |
63 | 299 | if (_closed) { |
64 | 272 | return; |
65 | 272 | } |
66 | 27 | Status st = close(); |
67 | 27 | if (!st.ok()) { |
68 | 15 | LOG(WARNING) << "SpillFileWriter::~SpillFileWriter() failed: " << st.to_string() |
69 | 15 | << ", spill_dir=" << _spill_dir; |
70 | 15 | } |
71 | 27 | } |
72 | | |
73 | 263 | Status SpillFileWriter::_open_next_part() { |
74 | 263 | _current_part_path = _spill_dir + "/" + std::to_string(_current_part_index); |
75 | | // Create the spill directory lazily on first part |
76 | 263 | if (_current_part_index == 0) { |
77 | 253 | RETURN_IF_ERROR(io::global_local_filesystem()->create_directory(_spill_dir)); |
78 | 253 | } |
79 | 263 | RETURN_IF_ERROR(io::global_local_filesystem()->create_file(_current_part_path, &_file_writer)); |
80 | 263 | COUNTER_UPDATE(_total_file_count, 1); |
81 | 263 | return Status::OK(); |
82 | 263 | } |
83 | | |
84 | 308 | Status SpillFileWriter::_close_current_part(const std::shared_ptr<SpillFile>& spill_file) { |
85 | 308 | if (!_file_writer) { |
86 | 45 | return Status::OK(); |
87 | 45 | } |
88 | | |
89 | | // Write footer: block offsets + max_sub_block_size + block_count |
90 | 263 | _part_meta.append((const char*)&_part_max_sub_block_size, sizeof(_part_max_sub_block_size)); |
91 | 263 | _part_meta.append((const char*)&_part_written_blocks, sizeof(_part_written_blocks)); |
92 | | |
93 | 263 | { |
94 | 263 | SCOPED_TIMER(_write_file_timer); |
95 | 263 | RETURN_IF_ERROR(_file_writer->append(_part_meta)); |
96 | 263 | } |
97 | | |
98 | 263 | int64_t meta_size = _part_meta.size(); |
99 | 263 | _part_written_bytes += meta_size; |
100 | 263 | _total_written_bytes += meta_size; |
101 | 263 | COUNTER_UPDATE(_write_file_total_size, meta_size); |
102 | 263 | if (_resource_ctx) { |
103 | 263 | _resource_ctx->io_context()->update_spill_write_bytes_to_local_storage(meta_size); |
104 | 263 | } |
105 | 263 | if (_write_file_current_size) { |
106 | 179 | COUNTER_UPDATE(_write_file_current_size, meta_size); |
107 | 179 | } |
108 | 263 | _data_dir->update_spill_data_usage(meta_size); |
109 | 263 | ExecEnv::GetInstance()->spill_file_mgr()->update_spill_write_bytes(meta_size); |
110 | | // Incrementally update SpillFile's accounting so gc() can always |
111 | | // decrement the correct amount, even if close() is never called. |
112 | 263 | if (spill_file) { |
113 | 248 | spill_file->update_written_bytes(meta_size); |
114 | 248 | } |
115 | | |
116 | 263 | RETURN_IF_ERROR(_file_writer->close()); |
117 | 248 | _file_writer.reset(); |
118 | | |
119 | | // Advance to next part |
120 | 248 | ++_current_part_index; |
121 | 248 | ++_total_parts; |
122 | 248 | if (spill_file) { |
123 | 248 | spill_file->increment_part_count(); |
124 | 248 | } |
125 | 248 | _part_written_blocks = 0; |
126 | 248 | _part_written_bytes = 0; |
127 | 248 | _part_max_sub_block_size = 0; |
128 | 248 | _part_meta.clear(); |
129 | | |
130 | 248 | return Status::OK(); |
131 | 263 | } |
132 | | |
133 | 470 | Status SpillFileWriter::_rotate_if_needed(const std::shared_ptr<SpillFile>& spill_file) { |
134 | 470 | if (_file_writer && _part_written_bytes >= _max_part_size) { |
135 | 10 | RETURN_IF_ERROR(_close_current_part(spill_file)); |
136 | 10 | } |
137 | 470 | return Status::OK(); |
138 | 470 | } |
139 | | |
140 | 473 | Status SpillFileWriter::write_block(RuntimeState* state, const Block& block) { |
141 | 473 | DCHECK(!_closed); |
142 | | |
143 | | // Lock the SpillFile to ensure it is still alive. If it has already been |
144 | | // destroyed (gc'd), we must not write any more data because the disk |
145 | | // accounting would be out of sync. |
146 | 473 | auto spill_file = _spill_file_wptr.lock(); |
147 | 473 | if (!spill_file) { |
148 | 0 | return Status::Error<INTERNAL_ERROR>( |
149 | 0 | "SpillFile has been destroyed, cannot write more data, spill_dir={}", _spill_dir); |
150 | 0 | } |
151 | | |
152 | | // Lazily open the first part |
153 | 473 | if (!_file_writer) { |
154 | 263 | RETURN_IF_ERROR(_open_next_part()); |
155 | 263 | } |
156 | | |
157 | 473 | DBUG_EXECUTE_IF("fault_inject::spill_file::spill_block", { |
158 | 473 | return Status::Error<INTERNAL_ERROR>("fault_inject spill_file spill_block failed"); |
159 | 473 | }); |
160 | | |
161 | 470 | auto rows = block.rows(); |
162 | 470 | COUNTER_UPDATE(_write_rows_counter, rows); |
163 | 470 | COUNTER_UPDATE(_write_block_bytes_counter, block.bytes()); |
164 | | |
165 | 470 | RETURN_IF_ERROR(_write_internal(block, spill_file)); |
166 | | |
167 | | // Auto-rotate if current part is full |
168 | 470 | return _rotate_if_needed(spill_file); |
169 | 470 | } |
170 | | |
171 | 317 | Status SpillFileWriter::close() { |
172 | 317 | if (_closed) { |
173 | 18 | return Status::OK(); |
174 | 18 | } |
175 | 299 | _closed = true; |
176 | | |
177 | 299 | DBUG_EXECUTE_IF("fault_inject::spill_file::spill_eof", { |
178 | 299 | return Status::Error<INTERNAL_ERROR>("fault_inject spill_file spill_eof failed"); |
179 | 299 | }); |
180 | | |
181 | 298 | auto spill_file = _spill_file_wptr.lock(); |
182 | 298 | RETURN_IF_ERROR(_close_current_part(spill_file)); |
183 | | |
184 | 283 | if (spill_file) { |
185 | 283 | if (spill_file->_active_writer != this) { |
186 | 0 | return Status::Error<INTERNAL_ERROR>( |
187 | 0 | "SpillFileWriter close() called but not registered as active writer, possible " |
188 | 0 | "double close or logic error"); |
189 | 0 | } |
190 | 283 | spill_file->finish_writing(); |
191 | 283 | } |
192 | | |
193 | 283 | return Status::OK(); |
194 | 283 | } |
195 | | |
196 | | Status SpillFileWriter::_write_internal(const Block& block, |
197 | 470 | const std::shared_ptr<SpillFile>& spill_file) { |
198 | 470 | size_t uncompressed_bytes = 0, compressed_bytes = 0; |
199 | | |
200 | 470 | Status status; |
201 | 470 | std::string buff; |
202 | 470 | int64_t buff_size {0}; |
203 | | |
204 | 470 | if (block.rows() > 0) { |
205 | 449 | { |
206 | 449 | PBlock pblock; |
207 | 449 | SCOPED_TIMER(_serialize_timer); |
208 | 449 | int64_t compressed_time = 0; |
209 | 449 | status = block.serialize( |
210 | 449 | BeExecVersionManager::get_newest_version(), &pblock, &uncompressed_bytes, |
211 | 449 | &compressed_bytes, &compressed_time, |
212 | 449 | segment_v2::CompressionTypePB::ZSTD); // ZSTD for better compression ratio |
213 | 449 | RETURN_IF_ERROR(status); |
214 | 449 | int64_t pblock_mem = pblock.ByteSizeLong(); |
215 | 449 | COUNTER_UPDATE(_memory_used_counter, pblock_mem); |
216 | 449 | Defer defer {[&]() { COUNTER_UPDATE(_memory_used_counter, -pblock_mem); }}; |
217 | 449 | if (!pblock.SerializeToString(&buff)) { |
218 | 0 | return Status::Error<ErrorCode::SERIALIZE_PROTOBUF_ERROR>( |
219 | 0 | "serialize spill data error. [path={}]", _current_part_path); |
220 | 0 | } |
221 | 449 | buff_size = buff.size(); |
222 | 449 | COUNTER_UPDATE(_memory_used_counter, buff_size); |
223 | 449 | Defer defer2 {[&]() { COUNTER_UPDATE(_memory_used_counter, -buff_size); }}; |
224 | 449 | } |
225 | 449 | if (_data_dir->reach_capacity_limit(buff_size)) { |
226 | 0 | return Status::Error<ErrorCode::DISK_REACH_CAPACITY_LIMIT>( |
227 | 0 | "spill data total size exceed limit, path: {}, size limit: {}, spill data " |
228 | 0 | "size: {}", |
229 | 0 | _data_dir->path(), |
230 | 0 | PrettyPrinter::print_bytes(_data_dir->get_spill_data_limit()), |
231 | 0 | PrettyPrinter::print_bytes(_data_dir->get_spill_data_bytes())); |
232 | 0 | } |
233 | | |
234 | 449 | { |
235 | 449 | Defer defer {[&]() { |
236 | 449 | if (status.ok()) { |
237 | 449 | _data_dir->update_spill_data_usage(buff_size); |
238 | 449 | ExecEnv::GetInstance()->spill_file_mgr()->update_spill_write_bytes(buff_size); |
239 | | |
240 | 449 | _part_max_sub_block_size = |
241 | 449 | std::max(_part_max_sub_block_size, (size_t)buff_size); |
242 | | |
243 | 449 | _part_meta.append((const char*)&_part_written_bytes, sizeof(size_t)); |
244 | 449 | COUNTER_UPDATE(_write_file_total_size, buff_size); |
245 | 449 | if (_resource_ctx) { |
246 | 449 | _resource_ctx->io_context()->update_spill_write_bytes_to_local_storage( |
247 | 449 | buff_size); |
248 | 449 | } |
249 | 449 | if (_write_file_current_size) { |
250 | 313 | COUNTER_UPDATE(_write_file_current_size, buff_size); |
251 | 313 | } |
252 | 449 | COUNTER_UPDATE(_write_block_counter, 1); |
253 | 449 | _part_written_bytes += buff_size; |
254 | 449 | _total_written_bytes += buff_size; |
255 | 449 | ++_part_written_blocks; |
256 | | // Incrementally update SpillFile so gc() can always |
257 | | // decrement the correct amount from _data_dir. |
258 | 449 | spill_file->update_written_bytes(buff_size); |
259 | 449 | } |
260 | 449 | }}; |
261 | 449 | { |
262 | 449 | SCOPED_TIMER(_write_file_timer); |
263 | 449 | status = _file_writer->append(buff); |
264 | 449 | RETURN_IF_ERROR(status); |
265 | 449 | } |
266 | 449 | } |
267 | 449 | } |
268 | | |
269 | 470 | return status; |
270 | 470 | } |
271 | | |
272 | | } // namespace doris |