be/src/storage/tablet/tablet.h
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 | | #pragma once |
19 | | |
20 | | #include <butil/macros.h> |
21 | | #include <glog/logging.h> |
22 | | |
23 | | #include <array> |
24 | | #include <atomic> |
25 | | #include <cstddef> |
26 | | #include <cstdint> |
27 | | #include <functional> |
28 | | #include <memory> |
29 | | #include <mutex> |
30 | | #include <ostream> |
31 | | #include <set> |
32 | | #include <shared_mutex> |
33 | | #include <string> |
34 | | #include <string_view> |
35 | | #include <utility> |
36 | | #include <vector> |
37 | | |
38 | | #include "common/atomic_shared_ptr.h" |
39 | | #include "common/config.h" |
40 | | #include "common/metrics/metrics.h" |
41 | | #include "common/status.h" |
42 | | #include "gen_cpp/AgentService_types.h" |
43 | | #include "storage/binlog.h" |
44 | | #include "storage/binlog_config.h" |
45 | | #include "storage/compaction/binlog_compaction_policy.h" |
46 | | #include "storage/data_dir.h" |
47 | | #include "storage/olap_common.h" |
48 | | #include "storage/partial_update_info.h" |
49 | | #include "storage/rowset/rowset.h" |
50 | | #include "storage/rowset/rowset_meta.h" |
51 | | #include "storage/rowset/rowset_reader.h" |
52 | | #include "storage/segment/segment.h" |
53 | | #include "storage/segment/segment_loader.h" |
54 | | #include "storage/tablet/base_tablet.h" |
55 | | #include "storage/version_graph.h" |
56 | | #include "util/once.h" |
57 | | #include "util/slice.h" |
58 | | |
59 | | namespace bvar { |
60 | | template <typename T> |
61 | | class Adder; |
62 | | } |
63 | | |
64 | | namespace doris { |
65 | | |
66 | | class Tablet; |
67 | | class CumulativeCompactionPolicy; |
68 | | class CompactionMixin; |
69 | | class RowsetWriter; |
70 | | struct RowsetWriterContext; |
71 | | class TTabletInfo; |
72 | | class TabletMetaPB; |
73 | | class TupleDescriptor; |
74 | | class CalcDeleteBitmapToken; |
75 | | enum CompressKind : int; |
76 | | class RowsetBinlogMetasPB; |
77 | | |
78 | | namespace io { |
79 | | class RemoteFileSystem; |
80 | | } // namespace io |
81 | | class Block; |
82 | | struct RowLocation; |
83 | | enum KeysType : int; |
84 | | enum SortType : int; |
85 | | |
86 | | enum TabletStorageType { STORAGE_TYPE_LOCAL, STORAGE_TYPE_REMOTE, STORAGE_TYPE_REMOTE_AND_LOCAL }; |
87 | | |
88 | | extern bvar::Adder<uint64_t> unused_remote_rowset_num; |
89 | | |
90 | | static inline constexpr auto TRACE_TABLET_LOCK_THRESHOLD = std::chrono::seconds(1); |
91 | | |
92 | | struct WriteCooldownMetaExecutors { |
93 | | WriteCooldownMetaExecutors(size_t executor_nums = 5); |
94 | | |
95 | | void stop(); |
96 | | |
97 | | void submit(TabletSharedPtr tablet); |
98 | 59.0k | size_t _get_executor_pos(int64_t tablet_id) const { |
99 | 59.0k | return std::hash<int64_t>()(tablet_id) % _executor_nums; |
100 | 59.0k | }; |
101 | | // Each executor is a mpsc to ensure uploads of the same tablet meta are not concurrent |
102 | | // FIXME(AlexYue): Use mpsc instead of `ThreadPool` with 1 thread |
103 | | // We use PriorityThreadPool since it would call status inside it's `shutdown` function. |
104 | | // Consider one situation where the StackTraceCache's singleton is detructed before |
105 | | // this WriteCooldownMetaExecutors's singleton, then invoking the status would also call |
106 | | // StackTraceCache which would then result in heap use after free like #23834 |
107 | | std::vector<std::unique_ptr<PriorityThreadPool>> _executors; |
108 | | std::unordered_set<int64_t> _pending_tablets; |
109 | | std::mutex _latch; |
110 | | size_t _executor_nums; |
111 | | }; |
112 | | |
113 | | class Tablet final : public BaseTablet { |
114 | | public: |
115 | | Tablet(StorageEngine& engine, TabletMetaSharedPtr tablet_meta, DataDir* data_dir, |
116 | | const std::string_view& cumulative_compaction_type = ""); |
117 | | |
118 | 14.5M | DataDir* data_dir() const { return _data_dir; } |
119 | 562k | int64_t replica_id() const { return _tablet_meta->replica_id(); } |
120 | | |
121 | 256k | std::string tablet_path() const override { return _tablet_path; } |
122 | 50 | std::string row_binlog_path() const { |
123 | 50 | return fmt::format("{}/{}", _tablet_path, FDRowBinlogSuffix); |
124 | 50 | } |
125 | 247k | std::string get_rowset_path(const RowsetMetaSharedPtr& rowset_meta) const { |
126 | 247k | if (!rowset_meta->is_local()) { |
127 | 15 | return ""; |
128 | 15 | } |
129 | 247k | return rowset_meta->is_row_binlog() ? row_binlog_path() : tablet_path(); |
130 | 247k | } |
131 | | |
132 | | bool set_tablet_schema_into_rowset_meta(); |
133 | | Status init(); |
134 | | bool init_succeeded(); |
135 | | |
136 | | bool is_used(); |
137 | | |
138 | | void register_tablet_into_dir(); |
139 | | void deregister_tablet_from_dir(); |
140 | | |
141 | | void save_meta(); |
142 | | // Used in clone task, to update local meta when finishing a clone job |
143 | | Status revise_tablet_meta(const std::vector<RowsetSharedPtr>& to_add, |
144 | | const std::vector<RowsetSharedPtr>& to_delete, |
145 | | bool is_incremental_clone, bool copy_row_binlog = false); |
146 | | |
147 | | int64_t cumulative_layer_point() const; |
148 | | void set_cumulative_layer_point(int64_t new_point); |
149 | | inline int64_t cumulative_promotion_size() const; |
150 | | inline void set_cumulative_promotion_size(int64_t new_size); |
151 | | |
152 | | // Disk space occupied by tablet, contain local and remote. |
153 | | size_t tablet_footprint() override; |
154 | | // Local disk space occupied by tablet. |
155 | | size_t tablet_local_size(); |
156 | | // Remote disk space occupied by tablet. |
157 | | size_t tablet_remote_size(); |
158 | | |
159 | | size_t num_rows(); |
160 | | size_t version_count() const; |
161 | | size_t stale_version_count() const; |
162 | | bool exceed_version_limit(int32_t limit) override; |
163 | | uint64_t segment_count() const; |
164 | | Version max_version() const; |
165 | | CumulativeCompactionPolicy* cumulative_compaction_policy(); |
166 | | BinlogCompactionPolicy* binlog_compaction_policy(); |
167 | | |
168 | | // properties encapsulated in TabletSchema |
169 | | SortType sort_type() const; |
170 | | size_t sort_col_num() const; |
171 | | size_t num_columns() const; |
172 | | size_t num_null_columns() const; |
173 | | size_t num_short_key_columns() const; |
174 | | size_t num_rows_per_row_block() const; |
175 | | double bloom_filter_fpp() const; |
176 | | size_t next_unique_id() const; |
177 | | int64_t avg_rs_meta_serialize_size() const; |
178 | | |
179 | | // operation in rowsets |
180 | | Status add_rowset(RowsetSharedPtr rowset, RowsetSharedPtr row_binlog_rowset = nullptr); |
181 | | Status create_initial_rowset(const int64_t version); |
182 | | |
183 | | // MUST hold EXCLUSIVE `_meta_lock`. |
184 | | Status modify_rowsets(std::vector<RowsetSharedPtr>& to_add, |
185 | | std::vector<RowsetSharedPtr>& to_delete, bool check_delete = false); |
186 | | Status modify_row_binlog_rowsets(std::vector<RowsetSharedPtr>& to_add, |
187 | | std::vector<RowsetSharedPtr>& to_delete); |
188 | | bool rowset_exists_unlocked(const RowsetSharedPtr& rowset); |
189 | | |
190 | | // Add a committed data rowset and its row binlog rowset |
191 | | Status add_inc_rowset(const RowsetSharedPtr& rowset, |
192 | | const RowsetSharedPtr& row_binlog_rowset = nullptr); |
193 | | /// Delete stale rowset by timing. This delete policy uses now() minutes |
194 | | /// config::tablet_rowset_expired_stale_sweep_time_sec to compute the deadline of expired rowset |
195 | | /// to delete. When rowset is deleted, it will be added to StorageEngine unused map and record |
196 | | /// need to delete flag. |
197 | | void delete_expired_stale_rowset(); |
198 | | |
199 | | // if quiet is true, no error log will be printed if there are missing versions |
200 | | Status check_version_integrity(const Version& version, bool quiet = false); |
201 | | bool check_version_exist(const Version& version) const; |
202 | | void acquire_version_and_rowsets( |
203 | | std::vector<std::pair<Version, RowsetSharedPtr>>* version_rowsets) const; |
204 | | |
205 | | // If skip_missing_version is true, skip versions if they are missing. |
206 | | Status capture_rs_readers(const Version& spec_version, std::vector<RowSetSplits>* rs_splits, |
207 | | const CaptureRowsetOps& opts) override; |
208 | | |
209 | | // Find the missed versions until the spec_version. |
210 | | // |
211 | | // for example: |
212 | | // [0-4][5-5][8-8][9-9][14-14] |
213 | | // if spec_version = 12, it will return [6, 6], [7, 7], [10, 10], [11, 11], [12, 12] |
214 | | Versions calc_missed_versions(int64_t spec_version, Versions existing_versions) const override; |
215 | | |
216 | | // meta lock |
217 | 62.9k | std::shared_mutex& get_header_lock() { return _meta_lock; } |
218 | 123 | std::mutex& get_rowset_update_lock() { return _rowset_update_lock; } |
219 | 140 | std::mutex& get_push_lock() { return _ingest_lock; } |
220 | 2.93M | std::mutex& get_base_compaction_lock() { return _base_compaction_lock; } |
221 | 3.36M | std::mutex& get_cumulative_compaction_lock() { return _cumulative_compaction_lock; } |
222 | 294 | std::mutex& get_binlog_compaction_lock() { return _binlog_compaction_lock; } |
223 | 1 | std::shared_mutex& get_meta_store_lock() { return _meta_store_lock; } |
224 | | |
225 | 271 | std::shared_timed_mutex& get_migration_lock() { return _migration_lock; } |
226 | | |
227 | 22 | std::mutex& get_build_inverted_index_lock() { return _build_inverted_index_lock; } |
228 | | |
229 | | // operation for compaction |
230 | | bool can_do_compaction(size_t path_hash, CompactionType compaction_type); |
231 | | bool suitable_for_compaction( |
232 | | CompactionType compaction_type, |
233 | | std::shared_ptr<CumulativeCompactionPolicy> cumulative_compaction_policy); |
234 | | |
235 | | uint32_t calc_compaction_score(CompactionType compaction_type, |
236 | | int8_t* prefer_compaction_level = nullptr); |
237 | | |
238 | | // This function to find max continuous version from the beginning. |
239 | | // For example: If there are 1, 2, 3, 5, 6, 7 versions belongs tablet, then 3 is target. |
240 | | // 3 will be saved in "version", and 7 will be saved in "max_version", if max_version != nullptr |
241 | | void max_continuous_version_from_beginning(Version* version, Version* max_version = nullptr); |
242 | | |
243 | 0 | void set_bad(bool is_bad) { _is_bad = is_bad; } |
244 | | |
245 | 6.48M | int64_t last_cumu_compaction_failure_time() { return _last_cumu_compaction_failure_millis; } |
246 | 30.2k | void set_last_cumu_compaction_failure_time(int64_t millis) { |
247 | 30.2k | _last_cumu_compaction_failure_millis = millis; |
248 | 30.2k | } |
249 | | |
250 | 2.93M | int64_t last_base_compaction_failure_time() { return _last_base_compaction_failure_millis; } |
251 | 11.5k | void set_last_base_compaction_failure_time(int64_t millis) { |
252 | 11.5k | _last_base_compaction_failure_millis = millis; |
253 | 11.5k | } |
254 | | |
255 | 0 | int64_t last_full_compaction_failure_time() { return _last_full_compaction_failure_millis; } |
256 | 0 | void set_last_full_compaction_failure_time(int64_t millis) { |
257 | 0 | _last_full_compaction_failure_millis = millis; |
258 | 0 | } |
259 | | |
260 | 28.6k | int64_t last_cumu_compaction_success_time() { return _last_cumu_compaction_success_millis; } |
261 | 931 | void set_last_cumu_compaction_success_time(int64_t millis) { |
262 | 931 | _last_cumu_compaction_success_millis = millis; |
263 | 931 | } |
264 | | |
265 | 28.6k | int64_t last_base_compaction_success_time() { return _last_base_compaction_success_millis; } |
266 | 910 | void set_last_base_compaction_success_time(int64_t millis) { |
267 | 910 | _last_base_compaction_success_millis = millis; |
268 | 910 | } |
269 | | |
270 | 0 | int64_t last_full_compaction_success_time() { return _last_full_compaction_success_millis; } |
271 | 0 | void set_last_full_compaction_success_time(int64_t millis) { |
272 | 0 | _last_full_compaction_success_millis = millis; |
273 | 0 | } |
274 | | |
275 | 146 | int64_t last_binlog_compaction_success_time(int8_t compaction_level) { |
276 | 146 | DCHECK(compaction_level >= 0 && |
277 | 146 | compaction_level < BinlogCompactionPolicy::kBinlogCompactionMaxLevel); |
278 | 146 | return _last_binlog_compaction_success_millis[compaction_level].load(); |
279 | 146 | } |
280 | 2.91k | int64_t last_binlog_compaction_failure_time() { |
281 | 2.91k | return _last_binlog_compaction_failure_millis.load(); |
282 | 2.91k | } |
283 | | void init_last_binlog_compaction_success_and_failure_time(); |
284 | 0 | void set_last_binlog_compaction_success_time(int8_t compaction_level, int64_t millis) { |
285 | 0 | DCHECK(compaction_level >= 0 && |
286 | 0 | compaction_level < BinlogCompactionPolicy::kBinlogCompactionMaxLevel); |
287 | 0 | _last_binlog_compaction_success_millis[compaction_level] = millis; |
288 | 0 | } |
289 | 146 | void set_last_binlog_compaction_failure_time(int64_t millis) { |
290 | 146 | _last_binlog_compaction_failure_millis = millis; |
291 | 146 | } |
292 | | |
293 | 0 | int64_t last_cumu_compaction_schedule_time() { return _last_cumu_compaction_schedule_millis; } |
294 | 30.2k | void set_last_cumu_compaction_schedule_time(int64_t millis) { |
295 | 30.2k | _last_cumu_compaction_schedule_millis = millis; |
296 | 30.2k | } |
297 | | |
298 | 0 | int64_t last_base_compaction_schedule_time() { return _last_base_compaction_schedule_millis; } |
299 | 5.78k | void set_last_base_compaction_schedule_time(int64_t millis) { |
300 | 5.78k | _last_base_compaction_schedule_millis = millis; |
301 | 5.78k | } |
302 | | |
303 | 0 | int64_t last_full_compaction_schedule_time() { return _last_full_compaction_schedule_millis; } |
304 | 0 | void set_last_full_compaction_schedule_time(int64_t millis) { |
305 | 0 | _last_full_compaction_schedule_millis = millis; |
306 | 0 | } |
307 | | |
308 | | void delete_all_files(); |
309 | | |
310 | | void check_tablet_path_exists(); |
311 | | |
312 | | std::vector<RowsetSharedPtr> pick_candidate_rowsets_to_cumulative_compaction(); |
313 | | std::vector<RowsetSharedPtr> pick_candidate_rowsets_to_base_compaction(); |
314 | | std::vector<RowsetSharedPtr> pick_candidate_rowsets_to_full_compaction(); |
315 | | std::vector<RowsetSharedPtr> pick_candidate_rowsets_to_binlog_compaction(); |
316 | | std::vector<RowsetSharedPtr> pick_candidate_rowsets_to_build_inverted_index( |
317 | | const std::set<int64_t>& alter_index_uids, bool is_drop_op); |
318 | | |
319 | | // used for single compaction to get the local versions |
320 | | // Single compaction does not require remote rowsets and cannot violate the cooldown semantics |
321 | | std::vector<Version> get_all_local_versions(); |
322 | | |
323 | | void calculate_cumulative_point(); |
324 | | // TODO(ygl): |
325 | 0 | bool is_primary_replica() { return false; } |
326 | | |
327 | | // return true if the checkpoint is actually done |
328 | | bool do_tablet_meta_checkpoint(); |
329 | | |
330 | | // Check whether the rowset is useful or not, unuseful rowset can be swept up then. |
331 | | // Rowset which is under tablet's management is useful, i.e. rowset is in |
332 | | // _rs_version_map, or _stale_rs_version_map. |
333 | | // Rowset whose version range is not covered by this tablet is also useful. |
334 | | bool rowset_meta_is_useful(RowsetMetaSharedPtr rowset_meta); |
335 | | |
336 | | void build_tablet_report_info(TTabletInfo* tablet_info, |
337 | | bool enable_consecutive_missing_check = false, |
338 | | bool enable_path_check = false); |
339 | | |
340 | | // return a json string to show the compaction status of this tablet |
341 | | void get_compaction_status(std::string* json_result); |
342 | | |
343 | | static Status prepare_compaction_and_calculate_permits( |
344 | | CompactionType compaction_type, const TabletSharedPtr& tablet, |
345 | | std::shared_ptr<CompactionMixin>& compaction, int64_t& permits, |
346 | | int8_t prefer_compaction_level = -1); |
347 | | |
348 | | void execute_compaction(CompactionMixin& compaction); |
349 | | |
350 | | void set_cumulative_compaction_policy( |
351 | 0 | std::shared_ptr<CumulativeCompactionPolicy> cumulative_compaction_policy) { |
352 | 0 | _cumulative_compaction_policy = cumulative_compaction_policy; |
353 | 0 | } |
354 | | |
355 | 0 | std::shared_ptr<CumulativeCompactionPolicy> get_cumulative_compaction_policy() { |
356 | 0 | return _cumulative_compaction_policy; |
357 | 0 | } |
358 | | |
359 | 30.2k | void set_last_cumu_compaction_status(std::string status) { |
360 | 30.2k | _last_cumu_compaction_status = std::move(status); |
361 | 30.2k | } |
362 | | |
363 | 0 | std::string get_last_cumu_compaction_status() { return _last_cumu_compaction_status; } |
364 | | |
365 | 11.5k | void set_last_base_compaction_status(std::string status) { |
366 | 11.5k | _last_base_compaction_status = std::move(status); |
367 | 11.5k | } |
368 | | |
369 | 0 | std::string get_last_base_compaction_status() { return _last_base_compaction_status; } |
370 | | |
371 | 0 | void set_last_full_compaction_status(std::string status) { |
372 | 0 | _last_full_compaction_status = std::move(status); |
373 | 0 | } |
374 | | |
375 | 0 | std::string get_last_full_compaction_status() { return _last_full_compaction_status; } |
376 | | |
377 | 146 | void set_last_binlog_compaction_status(std::string status) { |
378 | 146 | _last_binlog_compaction_status = std::move(status); |
379 | 146 | } |
380 | | |
381 | 0 | std::string get_last_binlog_compaction_status() { return _last_binlog_compaction_status; } |
382 | | |
383 | | std::tuple<int64_t, int64_t> get_visible_version_and_time() const; |
384 | | |
385 | 141k | void set_visible_version(const std::shared_ptr<const VersionWithTime>& visible_version) { |
386 | 141k | _visible_version.store(visible_version); |
387 | 141k | } |
388 | | |
389 | 0 | inline bool all_beta() const { |
390 | 0 | std::shared_lock rdlock(_meta_lock); |
391 | 0 | return _tablet_meta->all_beta(); |
392 | 0 | } |
393 | | |
394 | 0 | const TabletSchemaSPtr& tablet_schema_unlocked() const { return _max_version_schema; } |
395 | | |
396 | | Result<std::unique_ptr<RowsetWriter>> create_rowset_writer(RowsetWriterContext& context, |
397 | | bool vertical) override; |
398 | | |
399 | | Result<std::unique_ptr<RowsetWriter>> create_transient_rowset_writer( |
400 | | const Rowset& rowset, std::shared_ptr<PartialUpdateInfo> partial_update_info, |
401 | | int64_t txn_expiration = 0) override; |
402 | | Result<std::unique_ptr<RowsetWriter>> create_transient_rowset_writer( |
403 | | RowsetWriterContext& context, const RowsetId& rowset_id); |
404 | | |
405 | | Status create_rowset(const RowsetMetaSharedPtr& rowset_meta, RowsetSharedPtr* rowset); |
406 | | |
407 | | // MUST hold EXCLUSIVE `_meta_lock` |
408 | | void add_rowsets(const std::vector<RowsetSharedPtr>& to_add, bool copy_row_binlog = false); |
409 | | // MUST hold EXCLUSIVE `_meta_lock` |
410 | | Status delete_rowsets(const std::vector<RowsetSharedPtr>& to_delete, bool move_to_stale, |
411 | | bool copy_row_binlog = false); |
412 | | |
413 | | // MUST hold SHARED `_meta_lock` |
414 | 2.64k | const auto& rowset_map() const { return _rs_version_map; } |
415 | | // MUST hold SHARED `_meta_lock` |
416 | 2.63k | const auto& stale_rowset_map() const { return _stale_rs_version_map; } |
417 | | // MUST hold SHARED `_meta_lock` |
418 | 0 | const auto& row_binlog_rowset_map() const { return _row_binlog_rs_version_map; } |
419 | | |
420 | | //////////////////////////////////////////////////////////////////////////// |
421 | | // begin cooldown functions |
422 | | //////////////////////////////////////////////////////////////////////////// |
423 | 359k | int64_t storage_policy_id() const { return _tablet_meta->storage_policy_id(); } |
424 | | void set_storage_policy_id(int64_t id) { _tablet_meta->set_storage_policy_id(id); } |
425 | | |
426 | 1.54M | int64_t last_failed_follow_cooldown_time() const { return _last_failed_follow_cooldown_time; } |
427 | | |
428 | | // Cooldown to remote fs. |
429 | | Status cooldown(RowsetSharedPtr rowset = nullptr); |
430 | | |
431 | | RowsetSharedPtr pick_cooldown_rowset(); |
432 | | |
433 | | RowsetSharedPtr need_cooldown(int64_t* cooldown_timestamp, size_t* file_size); |
434 | | |
435 | | struct CooldownConf { |
436 | | int64_t term = -1; |
437 | | int64_t cooldown_replica_id = -1; |
438 | | }; |
439 | | |
440 | 3 | CooldownConf cooldown_conf() const { |
441 | 3 | std::shared_lock rlock(_cooldown_conf_lock); |
442 | 3 | return _cooldown_conf; |
443 | 3 | } |
444 | | |
445 | 9 | CooldownConf cooldown_conf_unlocked() const { return _cooldown_conf; } |
446 | | |
447 | | // Return `true` if update success |
448 | | bool update_cooldown_conf(int64_t cooldown_term, int64_t cooldown_replica_id); |
449 | | |
450 | | Status remove_all_remote_rowsets(); |
451 | | |
452 | | void record_unused_remote_rowset(const RowsetId& rowset_id, const std::string& resource, |
453 | | int64_t num_segments); |
454 | | |
455 | | uint32_t calc_cold_data_compaction_score() const; |
456 | | |
457 | 22 | std::mutex& get_cold_compaction_lock() { return _cold_compaction_lock; } |
458 | | |
459 | 0 | std::shared_mutex& get_cooldown_conf_lock() { return _cooldown_conf_lock; } |
460 | | |
461 | | static void async_write_cooldown_meta(TabletSharedPtr tablet); |
462 | | // Return `ABORTED` if should not to retry again |
463 | | Status write_cooldown_meta(); |
464 | | //////////////////////////////////////////////////////////////////////////// |
465 | | // end cooldown functions |
466 | | //////////////////////////////////////////////////////////////////////////// |
467 | | |
468 | | CalcDeleteBitmapExecutor* calc_delete_bitmap_executor() override; |
469 | | Status save_delete_bitmap(const TabletTxnInfo* txn_info, int64_t txn_id, |
470 | | DeleteBitmapPtr delete_bitmap, RowsetWriter* rowset_writer, |
471 | | const RowsetIdUnorderedSet& cur_rowset_ids, int64_t lock_id = -1, |
472 | | int64_t next_visible_version = -1) override; |
473 | | |
474 | | void merge_delete_bitmap(const DeleteBitmap& delete_bitmap); |
475 | | bool check_all_rowset_segment(); |
476 | | |
477 | | void update_max_version_schema(const TabletSchemaSPtr& tablet_schema); |
478 | | |
479 | | void set_skip_compaction(bool skip, |
480 | | CompactionType compaction_type = CompactionType::CUMULATIVE_COMPACTION, |
481 | | int64_t start = -1); |
482 | | bool should_skip_compaction(CompactionType compaction_type, int64_t now); |
483 | | |
484 | | std::vector<std::string> get_binlog_filepath(std::string_view binlog_version) const; |
485 | | std::pair<std::string, int64_t> get_binlog_info(std::string_view binlog_version) const; |
486 | | std::string get_rowset_binlog_meta(std::string_view binlog_version, |
487 | | std::string_view rowset_id) const; |
488 | | Status get_rowset_binlog_metas(const std::vector<int64_t>& binlog_versions, |
489 | | RowsetBinlogMetasPB* metas_pb); |
490 | | Status get_rowset_binlog_metas(Version binlog_versions, RowsetBinlogMetasPB* metas_pb); |
491 | | std::string get_segment_filepath(std::string_view rowset_id, |
492 | | std::string_view segment_index) const; |
493 | | std::string get_segment_filepath(std::string_view rowset_id, int64_t segment_index) const; |
494 | | bool can_add_binlog(uint64_t total_binlog_size) const; |
495 | | void gc_binlogs(int64_t version); |
496 | | Status ingest_binlog_metas(RowsetBinlogMetasPB* metas_pb); |
497 | | |
498 | | void report_error(const Status& st); |
499 | | |
500 | 0 | inline int64_t get_io_error_times() const { return _io_error_times; } |
501 | | |
502 | 560k | inline bool is_io_error_too_times() const { |
503 | 560k | return config::max_tablet_io_errors > 0 && _io_error_times >= config::max_tablet_io_errors; |
504 | 560k | } |
505 | | |
506 | 148 | int64_t get_table_id() { return _tablet_meta->table_id(); } |
507 | | |
508 | | // binlog related functions |
509 | 245 | bool enable_binlog() const { |
510 | 245 | return config::enable_feature_binlog && _tablet_meta->binlog_config().is_enable(); |
511 | 245 | } |
512 | 245 | bool enable_ccr_binlog() const { |
513 | 245 | return enable_binlog() && _tablet_meta->binlog_config().is_ccr_binlog_format(); |
514 | 245 | } |
515 | 129M | bool enable_row_binlog() const { |
516 | 129M | return _tablet_meta->binlog_config().is_enable() && |
517 | 129M | _tablet_meta->binlog_config().is_row_binlog_format(); |
518 | 129M | } |
519 | | |
520 | 0 | int64_t binlog_ttl_ms() const { return _tablet_meta->binlog_config().ttl_seconds(); } |
521 | 0 | int64_t binlog_max_bytes() const { return _tablet_meta->binlog_config().max_bytes(); } |
522 | | |
523 | | void set_binlog_config(BinlogConfig binlog_config); |
524 | | |
525 | | // row_binlog |
526 | 3 | int32_t row_binlog_schema_hash() const { return _tablet_meta->row_binlog_schema_hash(); } |
527 | | |
528 | 0 | void set_is_full_compaction_running(bool is_full_compaction_running) { |
529 | 0 | _is_full_compaction_running = is_full_compaction_running; |
530 | 0 | } |
531 | 0 | inline bool is_full_compaction_running() const { return _is_full_compaction_running; } |
532 | | void clear_cache() override; |
533 | | |
534 | | int32_t get_compaction_score() const { return _compaction_score; } |
535 | | |
536 | 0 | void set_compaction_score(int32_t compaction_score) { _compaction_score = compaction_score; } |
537 | | |
538 | 6.71k | void add_compaction_score(int32_t score) { |
539 | 6.71k | if (_compaction_score < 0) { |
540 | 6.56k | return; |
541 | 6.56k | } |
542 | 147 | _compaction_score += score; |
543 | 147 | } |
544 | | |
545 | 0 | void minus_compaction_score(int32_t score) { |
546 | 0 | if (_compaction_score < 0) { |
547 | 0 | return; |
548 | 0 | } |
549 | 0 | _compaction_score -= score; |
550 | 0 | } |
551 | | |
552 | | Status prepare_txn(TPartitionId partition_id, TTransactionId transaction_id, |
553 | | const PUniqueId& load_id, bool ingest); |
554 | | // TODO: commit_txn |
555 | | |
556 | | private: |
557 | | Status _init_once_action(); |
558 | | bool _contains_rowset(const RowsetId rowset_id); |
559 | | Status _contains_version(const Version& version); |
560 | | Status _add_row_binlog_rowset_unlocked(const RowsetSharedPtr& rowset, |
561 | | const RowsetSharedPtr& row_binlog_rowset); |
562 | | |
563 | | // Returns: |
564 | | // version: the max continuous version from beginning |
565 | | // max_version: the max version of this tablet |
566 | | void _max_continuous_version_from_beginning_unlocked(Version* version, Version* max_version, |
567 | | bool* has_version_cross) const; |
568 | | RowsetSharedPtr _rowset_with_largest_size(); |
569 | | /// Delete stale rowset by version. This method not only delete the version in expired rowset map, |
570 | | /// but also delete the version in rowset meta vector. |
571 | | void _delete_stale_rowset_by_version(const Version& version); |
572 | | |
573 | | uint32_t _calc_cumulative_compaction_score( |
574 | | std::shared_ptr<CumulativeCompactionPolicy> cumulative_compaction_policy); |
575 | | uint32_t _calc_base_compaction_score() const; |
576 | | uint32_t _calc_binlog_compaction_score(int8_t* prefer_compaction_level = nullptr); |
577 | | |
578 | | std::vector<RowsetSharedPtr> _pick_visible_rowsets_to_compaction(int64_t min_start_version, |
579 | | int64_t max_start_version); |
580 | | |
581 | | void _init_context_common_fields(RowsetWriterContext& context); |
582 | | |
583 | | //////////////////////////////////////////////////////////////////////////// |
584 | | // begin cooldown functions |
585 | | //////////////////////////////////////////////////////////////////////////// |
586 | | Status _cooldown_data(RowsetSharedPtr rowset); |
587 | | Status _follow_cooldowned_data(); |
588 | | Status _read_cooldown_meta(const StorageResource& storage_resource, |
589 | | TabletMetaPB* tablet_meta_pb); |
590 | | bool _has_data_to_cooldown(); |
591 | | int64_t _get_newest_cooldown_time(const RowsetSharedPtr& rowset); |
592 | | //////////////////////////////////////////////////////////////////////////// |
593 | | // end cooldown functions |
594 | | //////////////////////////////////////////////////////////////////////////// |
595 | | |
596 | | void _clear_cache_by_rowset(const BetaRowsetSharedPtr& rowset); |
597 | | void check_table_size_correctness(); |
598 | | std::string get_segment_path(const RowsetMetaSharedPtr& rs_meta, int64_t seg_id); |
599 | | int64_t get_segment_file_size(const RowsetMetaSharedPtr& rs_meta); |
600 | | int64_t get_inverted_index_file_size(const RowsetMetaSharedPtr& rs_meta); |
601 | | |
602 | | public: |
603 | | static const int64_t K_INVALID_CUMULATIVE_POINT = -1; |
604 | | |
605 | | private: |
606 | | StorageEngine& _engine; |
607 | | DataDir* _data_dir = nullptr; |
608 | | |
609 | | std::string _tablet_path; |
610 | | |
611 | | DorisCallOnce<Status> _init_once; |
612 | | // meta store lock is used for prevent 2 threads do checkpoint concurrently |
613 | | // it will be used in econ-mode in the future |
614 | | std::shared_mutex _meta_store_lock; |
615 | | std::mutex _ingest_lock; |
616 | | std::mutex _base_compaction_lock; |
617 | | std::mutex _cumulative_compaction_lock; |
618 | | std::mutex _binlog_compaction_lock; |
619 | | std::shared_timed_mutex _migration_lock; |
620 | | std::mutex _build_inverted_index_lock; |
621 | | |
622 | | // In unique key table with MoW, we should guarantee that only one |
623 | | // writer can update rowset and delete bitmap at the same time. |
624 | | // We use a separate lock rather than _meta_lock, to avoid blocking read queries |
625 | | // during publish_txn, which might take hundreds of milliseconds |
626 | | mutable std::mutex _rowset_update_lock; |
627 | | |
628 | | // if this tablet is broken, set to true. default is false |
629 | | std::atomic<bool> _is_bad; |
630 | | // timestamp of last cumu compaction failure |
631 | | std::atomic<int64_t> _last_cumu_compaction_failure_millis; |
632 | | // timestamp of last base compaction failure |
633 | | std::atomic<int64_t> _last_base_compaction_failure_millis; |
634 | | // timestamp of last full compaction failure |
635 | | std::atomic<int64_t> _last_full_compaction_failure_millis; |
636 | | // timestamp of last cumu compaction success |
637 | | std::atomic<int64_t> _last_cumu_compaction_success_millis; |
638 | | // timestamp of last base compaction success |
639 | | std::atomic<int64_t> _last_base_compaction_success_millis; |
640 | | // timestamp of last full compaction success |
641 | | std::atomic<int64_t> _last_full_compaction_success_millis; |
642 | | // timestamp of last binlog compaction success for each compaction level |
643 | | std::array<std::atomic<int64_t>, BinlogCompactionPolicy::kBinlogCompactionMaxLevel> |
644 | | _last_binlog_compaction_success_millis; |
645 | | // timestamp of last binlog compaction failure |
646 | | std::atomic<int64_t> _last_binlog_compaction_failure_millis; |
647 | | // timestamp of last cumu compaction schedule time |
648 | | std::atomic<int64_t> _last_cumu_compaction_schedule_millis; |
649 | | // timestamp of last base compaction schedule time |
650 | | std::atomic<int64_t> _last_base_compaction_schedule_millis; |
651 | | // timestamp of last full compaction schedule time |
652 | | std::atomic<int64_t> _last_full_compaction_schedule_millis; |
653 | | std::atomic<int64_t> _cumulative_point; |
654 | | std::atomic<int64_t> _cumulative_promotion_size; |
655 | | std::atomic<int32_t> _newly_created_rowset_num; |
656 | | std::atomic<int64_t> _last_checkpoint_time; |
657 | | std::string _last_cumu_compaction_status; |
658 | | std::string _last_base_compaction_status; |
659 | | std::string _last_full_compaction_status; |
660 | | std::string _last_binlog_compaction_status; |
661 | | |
662 | | // cumulative compaction policy |
663 | | std::shared_ptr<CumulativeCompactionPolicy> _cumulative_compaction_policy; |
664 | | std::string_view _cumulative_compaction_type; |
665 | | std::shared_ptr<BinlogCompactionPolicy> _binlog_compaction_policy; |
666 | | |
667 | | // use a separate thread to check all tablets paths existence |
668 | | std::atomic<bool> _is_tablet_path_exists; |
669 | | |
670 | | int64_t _last_missed_version; |
671 | | int64_t _last_missed_time_s; |
672 | | |
673 | | bool _skip_cumu_compaction = false; |
674 | | int64_t _skip_cumu_compaction_ts; |
675 | | |
676 | | bool _skip_base_compaction = false; |
677 | | int64_t _skip_base_compaction_ts; |
678 | | |
679 | | bool _skip_binlog_compaction = false; |
680 | | int64_t _skip_binlog_compaction_ts; |
681 | | |
682 | | // cooldown related |
683 | | CooldownConf _cooldown_conf; |
684 | | // `_cooldown_conf_lock` is used to serialize update cooldown conf and all operations that: |
685 | | // 1. read cooldown conf |
686 | | // 2. upload rowsets to remote storage |
687 | | // 3. update cooldown meta id |
688 | | mutable std::shared_mutex _cooldown_conf_lock; |
689 | | // `_cold_compaction_lock` is used to serialize cold data compaction and all operations that |
690 | | // may delete compaction input rowsets. |
691 | | std::mutex _cold_compaction_lock; |
692 | | int64_t _last_failed_follow_cooldown_time = 0; |
693 | | |
694 | | int64_t _io_error_times = 0; |
695 | | |
696 | | // partition's visible version. it sync from fe, but not real-time. |
697 | | atomic_shared_ptr<const VersionWithTime> _visible_version; |
698 | | |
699 | | std::atomic_bool _is_full_compaction_running = false; |
700 | | |
701 | | int32_t _compaction_score = -1; |
702 | | int32_t _score_check_cnt = 0; |
703 | | }; |
704 | | |
705 | 28.7k | inline CumulativeCompactionPolicy* Tablet::cumulative_compaction_policy() { |
706 | 28.7k | return _cumulative_compaction_policy.get(); |
707 | 28.7k | } |
708 | | |
709 | 146 | inline BinlogCompactionPolicy* Tablet::binlog_compaction_policy() { |
710 | 146 | return _binlog_compaction_policy.get(); |
711 | 146 | } |
712 | | |
713 | 6.65M | inline bool Tablet::init_succeeded() { |
714 | 6.66M | return _init_once.has_called() && _init_once.stored_result().ok(); |
715 | 6.65M | } |
716 | | |
717 | 7.05M | inline bool Tablet::is_used() { |
718 | 7.05M | return !_is_bad && _data_dir->is_used(); |
719 | 7.05M | } |
720 | | |
721 | 140k | inline void Tablet::register_tablet_into_dir() { |
722 | 140k | _data_dir->register_tablet(this); |
723 | 140k | } |
724 | | |
725 | 1.08k | inline void Tablet::deregister_tablet_from_dir() { |
726 | 1.08k | _data_dir->deregister_tablet(this); |
727 | 1.08k | } |
728 | | |
729 | 385k | inline int64_t Tablet::cumulative_layer_point() const { |
730 | 385k | return _cumulative_point; |
731 | 385k | } |
732 | | |
733 | 1.26k | inline void Tablet::set_cumulative_layer_point(int64_t new_point) { |
734 | | // cumulative point should only be reset to -1, or be increased |
735 | 1.26k | CHECK(new_point == Tablet::K_INVALID_CUMULATIVE_POINT || new_point >= _cumulative_point) |
736 | 0 | << "Unexpected cumulative point: " << new_point |
737 | 0 | << ", origin: " << _cumulative_point.load(); |
738 | 1.26k | _cumulative_point = new_point; |
739 | 1.26k | } |
740 | | |
741 | 28.7k | inline int64_t Tablet::cumulative_promotion_size() const { |
742 | 28.7k | return _cumulative_promotion_size; |
743 | 28.7k | } |
744 | | |
745 | 43.5k | inline void Tablet::set_cumulative_promotion_size(int64_t new_size) { |
746 | 43.5k | _cumulative_promotion_size = new_size; |
747 | 43.5k | } |
748 | | |
749 | | // TODO(lingbin): Why other methods that need to get information from _tablet_meta |
750 | | // are not locked, here needs a comment to explain. |
751 | 739 | inline size_t Tablet::tablet_footprint() { |
752 | 739 | std::shared_lock rdlock(_meta_lock); |
753 | 739 | return _tablet_meta->tablet_footprint(); |
754 | 739 | } |
755 | | |
756 | 1.39M | inline size_t Tablet::tablet_local_size() { |
757 | 1.39M | std::shared_lock rdlock(_meta_lock); |
758 | 1.39M | return _tablet_meta->tablet_local_size(); |
759 | 1.39M | } |
760 | | |
761 | 1.39M | inline size_t Tablet::tablet_remote_size() { |
762 | 1.39M | std::shared_lock rdlock(_meta_lock); |
763 | 1.39M | return _tablet_meta->tablet_remote_size(); |
764 | 1.39M | } |
765 | | |
766 | | // TODO(lingbin): Why other methods which need to get information from _tablet_meta |
767 | | // are not locked, here needs a comment to explain. |
768 | | inline size_t Tablet::num_rows() { |
769 | | std::shared_lock rdlock(_meta_lock); |
770 | | return _tablet_meta->num_rows(); |
771 | | } |
772 | | |
773 | 1.96M | inline size_t Tablet::version_count() const { |
774 | 1.96M | std::shared_lock rdlock(_meta_lock); |
775 | 1.96M | return _tablet_meta->version_count(); |
776 | 1.96M | } |
777 | | |
778 | 136 | inline size_t Tablet::stale_version_count() const { |
779 | 136 | std::shared_lock rdlock(_meta_lock); |
780 | 136 | return _tablet_meta->stale_version_count(); |
781 | 136 | } |
782 | | |
783 | 168k | inline Version Tablet::max_version() const { |
784 | 168k | std::shared_lock rdlock(_meta_lock); |
785 | 168k | return _tablet_meta->max_version(); |
786 | 168k | } |
787 | | |
788 | 1.96M | inline uint64_t Tablet::segment_count() const { |
789 | 1.96M | std::shared_lock rdlock(_meta_lock); |
790 | 1.96M | uint64_t segment_nums = 0; |
791 | 3.36M | for (const auto& [_, rs_meta] : _tablet_meta->all_rs_metas()) { |
792 | 3.36M | segment_nums += rs_meta->num_segments(); |
793 | 3.36M | } |
794 | 1.96M | return segment_nums; |
795 | 1.96M | } |
796 | | |
797 | 0 | inline SortType Tablet::sort_type() const { |
798 | 0 | return _tablet_meta->tablet_schema()->sort_type(); |
799 | 0 | } |
800 | | |
801 | 0 | inline size_t Tablet::sort_col_num() const { |
802 | 0 | return _tablet_meta->tablet_schema()->sort_col_num(); |
803 | 0 | } |
804 | | |
805 | 0 | inline size_t Tablet::num_columns() const { |
806 | 0 | return _tablet_meta->tablet_schema()->num_columns(); |
807 | 0 | } |
808 | | |
809 | 0 | inline size_t Tablet::num_null_columns() const { |
810 | 0 | return _tablet_meta->tablet_schema()->num_null_columns(); |
811 | 0 | } |
812 | | |
813 | 0 | inline size_t Tablet::num_short_key_columns() const { |
814 | 0 | return _tablet_meta->tablet_schema()->num_short_key_columns(); |
815 | 0 | } |
816 | | |
817 | 0 | inline size_t Tablet::num_rows_per_row_block() const { |
818 | 0 | return _tablet_meta->tablet_schema()->num_rows_per_row_block(); |
819 | 0 | } |
820 | | |
821 | 0 | inline double Tablet::bloom_filter_fpp() const { |
822 | 0 | return _tablet_meta->tablet_schema()->bloom_filter_fpp(); |
823 | 0 | } |
824 | | |
825 | 0 | inline size_t Tablet::next_unique_id() const { |
826 | 0 | return _tablet_meta->tablet_schema()->next_column_unique_id(); |
827 | 0 | } |
828 | | |
829 | 136 | inline int64_t Tablet::avg_rs_meta_serialize_size() const { |
830 | 136 | return _tablet_meta->avg_rs_meta_serialize_size(); |
831 | 136 | } |
832 | | |
833 | | class TabletCopyType { |
834 | | public: |
835 | | static constexpr int32_t DEFAULT = TTabletCopyType::DATA | TTabletCopyType::CCR_BINLOG; |
836 | | |
837 | 4 | static bool has(int32_t copy_type, TTabletCopyType::type flag) { |
838 | 4 | return (copy_type & static_cast<int32_t>(flag)) != 0; |
839 | 4 | } |
840 | | |
841 | 2 | static Status validate(int32_t copy_type) { |
842 | 2 | if (copy_type <= 0 || (copy_type & ~all_types()) != 0) { |
843 | 0 | return Status::Error<ErrorCode::INVALID_ARGUMENT>( |
844 | 0 | "invalid copy_type bitmask: {}, valid bits: {}", copy_type, all_types()); |
845 | 0 | } |
846 | 2 | return Status::OK(); |
847 | 2 | } |
848 | | |
849 | | private: |
850 | 2 | static constexpr int32_t all_types() { |
851 | 2 | return TTabletCopyType::DATA | TTabletCopyType::ROW_BINLOG | TTabletCopyType::CCR_BINLOG; |
852 | 2 | } |
853 | | }; |
854 | | |
855 | | } // namespace doris |