be/src/io/cache/block_file_cache.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 <bvar/bvar.h> |
21 | | #include <concurrentqueue.h> |
22 | | |
23 | | #include <algorithm> |
24 | | #include <array> |
25 | | #include <atomic> |
26 | | #include <boost/lockfree/spsc_queue.hpp> |
27 | | #include <functional> |
28 | | #include <memory> |
29 | | #include <mutex> |
30 | | #include <optional> |
31 | | #include <thread> |
32 | | #include <unordered_map> |
33 | | #include <vector> |
34 | | |
35 | | #include "io/cache/block_file_cache_ttl_mgr.h" |
36 | | #include "io/cache/cache_lru_dumper.h" |
37 | | #include "io/cache/file_block.h" |
38 | | #include "io/cache/file_cache_common.h" |
39 | | #include "io/cache/file_cache_storage.h" |
40 | | #include "io/cache/lru_queue_recorder.h" |
41 | | #include "runtime/runtime_profile.h" |
42 | | #include "util/threadpool.h" |
43 | | |
44 | | namespace doris::io { |
45 | | using RecycleFileCacheKeys = moodycamel::ConcurrentQueue<FileCacheKey>; |
46 | | |
47 | | class LockScopedTimer { |
48 | | public: |
49 | 191k | LockScopedTimer() : start_(std::chrono::steady_clock::now()) {} |
50 | 191k | ~LockScopedTimer() { |
51 | 191k | auto end = std::chrono::steady_clock::now(); |
52 | 191k | auto duration_us = |
53 | 191k | std::chrono::duration_cast<std::chrono::microseconds>(end - start_).count(); |
54 | 191k | if (duration_us > config::cache_lock_held_long_tail_threshold_us) { |
55 | 0 | LOG(WARNING) << "Lock held time " << std::to_string(duration_us) << "us. " |
56 | 0 | << get_stack_trace(); |
57 | 0 | } |
58 | 191k | } |
59 | | |
60 | | private: |
61 | | std::chrono::time_point<std::chrono::steady_clock> start_; |
62 | | }; |
63 | | |
64 | | // Note: the cache_lock is scoped, so do not add do...while(0) here. |
65 | | #define SCOPED_CACHE_LOCK(MUTEX, cache) \ |
66 | 191k | std::chrono::time_point<std::chrono::steady_clock> start_time = \ |
67 | 191k | std::chrono::steady_clock::now(); \ |
68 | 191k | std::lock_guard cache_lock(MUTEX); \ |
69 | 191k | std::chrono::time_point<std::chrono::steady_clock> acq_time = \ |
70 | 191k | std::chrono::steady_clock::now(); \ |
71 | 191k | auto duration_us = \ |
72 | 191k | std::chrono::duration_cast<std::chrono::microseconds>(acq_time - start_time).count(); \ |
73 | 191k | *(cache->_cache_lock_wait_time_us) << duration_us; \ |
74 | 191k | if (duration_us > config::cache_lock_wait_long_tail_threshold_us) { \ |
75 | 0 | LOG(WARNING) << "Lock wait time " << std::to_string(duration_us) << "us. " \ |
76 | 0 | << get_stack_trace() << std::endl; \ |
77 | 0 | } \ |
78 | 191k | LockScopedTimer cache_lock_timer; |
79 | | |
80 | | class FSFileCacheStorage; |
81 | | |
82 | | // NeedUpdateLRUBlocks keeps FileBlockSPtr entries that require LRU updates in a |
83 | | // deduplicated, sharded container. Entries are keyed by the raw FileBlock |
84 | | // pointer so that multiple shared_ptr copies of the same block are treated as a |
85 | | // single pending update. The structure is thread-safe and optimized for high |
86 | | // contention insert/drain workloads in the background update thread. |
87 | | // Note that Blocks are updated in batch, internal order is not important. |
88 | | class NeedUpdateLRUBlocks { |
89 | | public: |
90 | 169 | NeedUpdateLRUBlocks() = default; |
91 | | |
92 | | // Insert a block into the pending set. Returns true only when the block |
93 | | // was not already queued. Null inputs are ignored. |
94 | | bool insert(FileBlockSPtr block); |
95 | | |
96 | | // Drain up to `limit` unique blocks into `output`. The method returns how |
97 | | // many blocks were actually drained and shrinks the internal size |
98 | | // accordingly. |
99 | | size_t drain(size_t limit, std::vector<FileBlockSPtr>* output); |
100 | | |
101 | | // Remove every pending block from the structure and reset the size. |
102 | | void clear(); |
103 | | |
104 | | // Thread-safe approximate size of queued unique blocks. |
105 | 12.4k | size_t size() const { return _size.load(std::memory_order_relaxed); } |
106 | | |
107 | | private: |
108 | | static constexpr size_t kShardCount = 64; |
109 | | static constexpr size_t kShardMask = kShardCount - 1; |
110 | | |
111 | | struct Shard { |
112 | | std::mutex mutex; |
113 | | std::unordered_map<FileBlock*, FileBlockSPtr> entries; |
114 | | }; |
115 | | |
116 | | size_t shard_index(FileBlock* ptr) const; |
117 | | |
118 | | std::array<Shard, kShardCount> _shards; |
119 | | std::atomic<size_t> _size {0}; |
120 | | }; |
121 | | |
122 | | // The BlockFileCache is responsible for the management of the blocks |
123 | | // The current strategies are lru and ttl. |
124 | | |
125 | | struct FileBlockCell { |
126 | | friend class FileBlock; |
127 | | |
128 | | FileBlockSPtr file_block; |
129 | | /// Iterator is put here on first reservation attempt, if successful. |
130 | | std::optional<LRUQueue::Iterator> queue_iterator; |
131 | | |
132 | | mutable int64_t atime {0}; |
133 | 3.73M | void update_atime() const { |
134 | 3.73M | atime = std::chrono::duration_cast<std::chrono::seconds>( |
135 | 3.73M | std::chrono::steady_clock::now().time_since_epoch()) |
136 | 3.73M | .count(); |
137 | 3.73M | } |
138 | | |
139 | | /// Pointer to file block is always hold by the cache itself. |
140 | | /// Apart from pointer in cache, it can be hold by cache users, when they call |
141 | | /// getorSet(), but cache users always hold it via FileBlocksHolder. |
142 | 112k | bool releasable() const { |
143 | 112k | return (file_block.use_count() == 1 || |
144 | 112k | (file_block.use_count() == 2 && file_block->_owned_by_cached_reader)); |
145 | 112k | } |
146 | | |
147 | 2.67M | size_t size() const { return file_block->_block_range.size(); } |
148 | | |
149 | | FileBlockCell() = default; |
150 | | FileBlockCell(FileBlockSPtr file_block, std::lock_guard<std::mutex>& cache_lock); |
151 | | FileBlockCell(FileBlockCell&& other) noexcept |
152 | 294k | : file_block(std::move(other.file_block)), |
153 | 294k | queue_iterator(other.queue_iterator), |
154 | 294k | atime(other.atime) { |
155 | 294k | file_block->cell = this; |
156 | 294k | } |
157 | | |
158 | | FileBlockCell& operator=(const FileBlockCell&) = delete; |
159 | | FileBlockCell(const FileBlockCell&) = delete; |
160 | | |
161 | 0 | size_t dowloading_size() const { return file_block->_downloaded_size; } |
162 | | }; |
163 | | |
164 | | class BlockFileCache { |
165 | | friend class FSFileCacheStorage; |
166 | | friend class MemFileCacheStorage; |
167 | | friend class FileBlock; |
168 | | friend struct FileBlocksHolder; |
169 | | friend class CacheLRUDumper; |
170 | | friend class LRUQueueRecorder; |
171 | | friend struct FileBlockCell; |
172 | | |
173 | | public: |
174 | | // hash the file_name to uint128 |
175 | | static UInt128Wrapper hash(const std::string& path); |
176 | | |
177 | | BlockFileCache(const std::string& cache_base_path, const FileCacheSettings& cache_settings); |
178 | | |
179 | 161 | virtual ~BlockFileCache() { |
180 | 161 | { |
181 | 161 | std::lock_guard lock(_close_mtx); |
182 | 161 | _close = true; |
183 | 161 | } |
184 | 161 | _close_cv.notify_all(); |
185 | 161 | if (_cache_background_monitor_thread.joinable()) { |
186 | 141 | _cache_background_monitor_thread.join(); |
187 | 141 | } |
188 | 161 | if (_cache_background_gc_thread.joinable()) { |
189 | 141 | _cache_background_gc_thread.join(); |
190 | 141 | } |
191 | 161 | if (_cache_background_evict_in_advance_thread.joinable()) { |
192 | 141 | _cache_background_evict_in_advance_thread.join(); |
193 | 141 | } |
194 | 161 | if (_cache_background_lru_dump_thread.joinable()) { |
195 | 141 | _cache_background_lru_dump_thread.join(); |
196 | 141 | } |
197 | 161 | if (_cache_background_lru_log_replay_thread.joinable()) { |
198 | 141 | _cache_background_lru_log_replay_thread.join(); |
199 | 141 | } |
200 | 161 | if (_cache_background_block_lru_update_thread.joinable()) { |
201 | 141 | _cache_background_block_lru_update_thread.join(); |
202 | 141 | } |
203 | 161 | } |
204 | | |
205 | | /// Restore cache from local filesystem. |
206 | | Status initialize(); |
207 | | |
208 | | /// Cache capacity in bytes. |
209 | 14 | [[nodiscard]] size_t capacity() const { return _capacity; } |
210 | | |
211 | | // try to release all releasable block |
212 | | // it maybe hang the io/system |
213 | | size_t try_release(); |
214 | | |
215 | 1.71k | [[nodiscard]] const std::string& get_base_path() const { return _cache_base_path; } |
216 | | |
217 | | // Get storage for inspection |
218 | 1 | FileCacheStorage* get_storage() const { return _storage.get(); } |
219 | | |
220 | | /** |
221 | | * Given an `offset` and `size` representing [offset, offset + size) bytes interval, |
222 | | * return list of cached non-overlapping non-empty |
223 | | * file blocks `[block1, ..., blockN]` which intersect with given interval. |
224 | | * |
225 | | * blocks in returned list are ordered in ascending order and represent a full contiguous |
226 | | * interval (no holes). Each block in returned list has state: DOWNLOADED, DOWNLOADING or EMPTY. |
227 | | * |
228 | | * As long as pointers to returned file blocks are hold |
229 | | * it is guaranteed that these file blocks are not removed from cache. |
230 | | */ |
231 | | FileBlocksHolder get_or_set(const UInt128Wrapper& hash, size_t offset, size_t size, |
232 | | CacheContext& context); |
233 | | |
234 | | /** |
235 | | * record blocks read directly by CachedRemoteFileReader |
236 | | */ |
237 | | void add_need_update_lru_block(FileBlockSPtr block); |
238 | | |
239 | | /** |
240 | | * Clear all cached data for this cache instance async |
241 | | * |
242 | | * @returns summary message |
243 | | */ |
244 | | std::string clear_file_cache_async(); |
245 | | std::string clear_file_cache_directly(); |
246 | | |
247 | | /** |
248 | | * Reset the cache capacity. If the new_capacity is smaller than _capacity, the redundant data will be remove async. |
249 | | * |
250 | | * @returns summary message |
251 | | */ |
252 | | std::string reset_capacity(size_t new_capacity); |
253 | | |
254 | | std::map<size_t, FileBlockSPtr> get_blocks_by_key(const UInt128Wrapper& hash); |
255 | | |
256 | | /// For debug and UT |
257 | | std::string dump_structure(const UInt128Wrapper& hash); |
258 | | std::string dump_single_cache_type(const UInt128Wrapper& hash, size_t offset); |
259 | | |
260 | | void dump_lru_queues(bool force); |
261 | | |
262 | | [[nodiscard]] size_t get_used_cache_size(FileCacheType type) const; |
263 | | |
264 | | [[nodiscard]] size_t get_file_blocks_num(FileCacheType type) const; |
265 | | |
266 | | // change the block cache type |
267 | | void change_cache_type(const UInt128Wrapper& hash, size_t offset, FileCacheType new_type, |
268 | | std::lock_guard<std::mutex>& cache_lock); |
269 | | |
270 | | // remove all blocks that belong to the key |
271 | | void remove_if_cached(const UInt128Wrapper& key); |
272 | | void remove_if_cached_async(const UInt128Wrapper& key); |
273 | | |
274 | | // Shrink the block size. old_size is always larged than new_size. |
275 | | void reset_range(const UInt128Wrapper&, size_t offset, size_t old_size, size_t new_size, |
276 | | std::lock_guard<std::mutex>& cache_lock); |
277 | | |
278 | | // get the hotest blocks message by key |
279 | | // The tuple is composed of <offset, size, cache_type, expiration_time> |
280 | | [[nodiscard]] std::vector<std::tuple<size_t, size_t, FileCacheType, uint64_t>> |
281 | | get_hot_blocks_meta(const UInt128Wrapper& hash) const; |
282 | | |
283 | | [[nodiscard]] bool get_async_open_success() const { return _async_open_done; } |
284 | | |
285 | | BlockFileCache& operator=(const BlockFileCache&) = delete; |
286 | | BlockFileCache(const BlockFileCache&) = delete; |
287 | | |
288 | | // try to reserve the new space for the new block if the cache is full |
289 | | bool try_reserve(const UInt128Wrapper& hash, const CacheContext& context, size_t offset, |
290 | | size_t size, std::lock_guard<std::mutex>& cache_lock); |
291 | | |
292 | | /** |
293 | | * Proactively evict cache blocks to free up space before cache is full. |
294 | | * |
295 | | * This function attempts to evict blocks from both NORMAL and TTL queues to maintain |
296 | | * cache size below high watermark. Unlike try_reserve() which blocks until space is freed, |
297 | | * this function initiates asynchronous eviction in background. |
298 | | * |
299 | | * @param size Number of bytes to try to evict |
300 | | * @param cache_lock Lock that must be held while accessing cache data structures |
301 | | * |
302 | | * @pre Caller must hold cache_lock |
303 | | * @pre _need_evict_cache_in_advance must be true |
304 | | * @pre _recycle_keys queue must have capacity for evicted blocks |
305 | | */ |
306 | | void try_evict_in_advance(size_t size, std::lock_guard<std::mutex>& cache_lock); |
307 | | |
308 | | void update_ttl_atime(const UInt128Wrapper& hash); |
309 | | |
310 | | std::map<std::string, double> get_stats(); |
311 | | |
312 | | // for be UTs |
313 | | std::map<std::string, double> get_stats_unsafe(); |
314 | | |
315 | | using AccessRecord = |
316 | | std::unordered_map<AccessKeyAndOffset, LRUQueue::Iterator, KeyAndOffsetHash>; |
317 | | |
318 | | /// Used to track and control the cache access of each query. |
319 | | /// Through it, we can realize the processing of different queries by the cache layer. |
320 | | struct QueryFileCacheContext { |
321 | | LRUQueue lru_queue; |
322 | | AccessRecord records; |
323 | | |
324 | 5 | QueryFileCacheContext(size_t max_cache_size) : lru_queue(max_cache_size, 0, 0) {} |
325 | | |
326 | | void remove(const UInt128Wrapper& hash, size_t offset, |
327 | | std::lock_guard<std::mutex>& cache_lock); |
328 | | |
329 | | void reserve(const UInt128Wrapper& hash, size_t offset, size_t size, |
330 | | std::lock_guard<std::mutex>& cache_lock); |
331 | | |
332 | 69 | size_t get_max_cache_size() const { return lru_queue.get_max_size(); } |
333 | | |
334 | 48 | size_t get_cache_size(std::lock_guard<std::mutex>& cache_lock) const { |
335 | 48 | return lru_queue.get_capacity(cache_lock); |
336 | 48 | } |
337 | | |
338 | 50 | LRUQueue& queue() { return lru_queue; } |
339 | | }; |
340 | | |
341 | | using QueryFileCacheContextPtr = std::shared_ptr<QueryFileCacheContext>; |
342 | | using QueryFileCacheContextMap = std::unordered_map<TUniqueId, QueryFileCacheContextPtr>; |
343 | | |
344 | | QueryFileCacheContextPtr get_query_context(const TUniqueId& query_id, |
345 | | std::lock_guard<std::mutex>&); |
346 | | |
347 | | void remove_query_context(const TUniqueId& query_id); |
348 | | |
349 | | QueryFileCacheContextPtr get_or_set_query_context(const TUniqueId& query_id, |
350 | | std::lock_guard<std::mutex>& cache_lock, |
351 | | int file_cache_query_limit_percent); |
352 | | |
353 | | /// Save a query context information, and adopt different cache policies |
354 | | /// for different queries through the context cache layer. |
355 | | struct QueryFileCacheContextHolder { |
356 | | QueryFileCacheContextHolder(const TUniqueId& query_id, BlockFileCache* mgr, |
357 | | QueryFileCacheContextPtr context) |
358 | 7 | : query_id(query_id), mgr(mgr), context(context) {} |
359 | | |
360 | | QueryFileCacheContextHolder& operator=(const QueryFileCacheContextHolder&) = delete; |
361 | | QueryFileCacheContextHolder(const QueryFileCacheContextHolder&) = delete; |
362 | | |
363 | 7 | ~QueryFileCacheContextHolder() { |
364 | | /// If only the query_map and the current holder hold the context_query, |
365 | | /// the query has been completed and the query_context is released. |
366 | 7 | if (context) { |
367 | 6 | context.reset(); |
368 | 6 | mgr->remove_query_context(query_id); |
369 | 6 | } |
370 | 7 | } |
371 | | |
372 | | const TUniqueId& query_id; |
373 | | BlockFileCache* mgr = nullptr; |
374 | | QueryFileCacheContextPtr context; |
375 | | }; |
376 | | using QueryFileCacheContextHolderPtr = std::unique_ptr<QueryFileCacheContextHolder>; |
377 | | QueryFileCacheContextHolderPtr get_query_context_holder(const TUniqueId& query_id, |
378 | | int file_cache_query_limit_percent); |
379 | | |
380 | 6.24k | int64_t approximate_available_cache_size() const { |
381 | 6.24k | return std::max<int64_t>( |
382 | 6.24k | _cache_capacity_metrics->get_value() - _cur_cache_size_metrics->get_value(), 0); |
383 | 6.24k | } |
384 | | |
385 | | Status report_file_cache_inconsistency(std::vector<std::string>& results); |
386 | | Status check_file_cache_consistency(InconsistencyContext& inconsistency_context); |
387 | | |
388 | | private: |
389 | | LRUQueue& get_queue(FileCacheType type); |
390 | | const LRUQueue& get_queue(FileCacheType type) const; |
391 | | |
392 | | template <class T, class U> |
393 | | requires IsXLock<T> && IsXLock<U> |
394 | | void remove(FileBlockSPtr file_block, T& cache_lock, U& segment_lock, bool sync = true); |
395 | | |
396 | | FileBlocks get_impl(const UInt128Wrapper& hash, const CacheContext& context, |
397 | | const FileBlock::Range& range, std::lock_guard<std::mutex>& cache_lock); |
398 | | |
399 | | template <class T> |
400 | | requires IsXLock<T> |
401 | | FileBlockCell* get_cell(const UInt128Wrapper& hash, size_t offset, T& cache_lock); |
402 | | |
403 | | virtual FileBlockCell* add_cell(const UInt128Wrapper& hash, const CacheContext& context, |
404 | | size_t offset, size_t size, FileBlock::State state, |
405 | | std::lock_guard<std::mutex>& cache_lock); |
406 | | |
407 | | Status initialize_unlocked(std::lock_guard<std::mutex>& cache_lock); |
408 | | |
409 | | void update_block_lru(FileBlockSPtr block, std::lock_guard<std::mutex>& cache_lock); |
410 | | |
411 | | void use_cell(const FileBlockCell& cell, FileBlocks* result, bool not_need_move, |
412 | | std::lock_guard<std::mutex>& cache_lock); |
413 | | |
414 | | bool try_reserve_for_lru(const UInt128Wrapper& hash, QueryFileCacheContextPtr query_context, |
415 | | const CacheContext& context, size_t offset, size_t size, |
416 | | std::lock_guard<std::mutex>& cache_lock, |
417 | | bool evict_in_advance = false); |
418 | | |
419 | | bool try_reserve_during_async_load(size_t size, std::lock_guard<std::mutex>& cache_lock); |
420 | | |
421 | | std::vector<FileCacheType> get_other_cache_type(FileCacheType cur_cache_type); |
422 | | std::vector<FileCacheType> get_other_cache_type_without_ttl(FileCacheType cur_cache_type); |
423 | | |
424 | | bool try_reserve_from_other_queue(FileCacheType cur_cache_type, size_t offset, int64_t cur_time, |
425 | | std::lock_guard<std::mutex>& cache_lock, |
426 | | bool evict_in_advance = false); |
427 | | |
428 | | size_t get_available_cache_size(FileCacheType cache_type) const; |
429 | | |
430 | | FileBlocks split_range_into_cells(const UInt128Wrapper& hash, const CacheContext& context, |
431 | | size_t offset, size_t size, FileBlock::State state, |
432 | | std::lock_guard<std::mutex>& cache_lock); |
433 | | |
434 | | std::string dump_structure_unlocked(const UInt128Wrapper& hash, |
435 | | std::lock_guard<std::mutex>& cache_lock); |
436 | | |
437 | | std::string dump_single_cache_type_unlocked(const UInt128Wrapper& hash, size_t offset, |
438 | | std::lock_guard<std::mutex>& cache_lock); |
439 | | |
440 | | void fill_holes_with_empty_file_blocks(FileBlocks& file_blocks, const UInt128Wrapper& hash, |
441 | | const CacheContext& context, |
442 | | const FileBlock::Range& range, |
443 | | std::lock_guard<std::mutex>& cache_lock); |
444 | | |
445 | | size_t get_used_cache_size_unlocked(FileCacheType type, |
446 | | std::lock_guard<std::mutex>& cache_lock) const; |
447 | | |
448 | | void check_disk_resource_limit(); |
449 | | void check_need_evict_cache_in_advance(); |
450 | | |
451 | | size_t get_available_cache_size_unlocked(FileCacheType type, |
452 | | std::lock_guard<std::mutex>& cache_lock) const; |
453 | | |
454 | | size_t get_file_blocks_num_unlocked(FileCacheType type, |
455 | | std::lock_guard<std::mutex>& cache_lock) const; |
456 | | |
457 | | bool need_to_move(FileCacheType cell_type, FileCacheType query_type) const; |
458 | | |
459 | | void run_background_monitor(); |
460 | | void run_background_gc(); |
461 | | void run_background_lru_log_replay(); |
462 | | void run_background_lru_dump(); |
463 | | void restore_lru_queues_from_disk(std::lock_guard<std::mutex>& cache_lock); |
464 | | void run_background_evict_in_advance(); |
465 | | void run_background_block_lru_update(); |
466 | | |
467 | | bool try_reserve_from_other_queue_by_time_interval(FileCacheType cur_type, |
468 | | std::vector<FileCacheType> other_cache_types, |
469 | | size_t size, int64_t cur_time, |
470 | | std::lock_guard<std::mutex>& cache_lock, |
471 | | bool evict_in_advance); |
472 | | |
473 | | bool try_reserve_from_other_queue_by_size(FileCacheType cur_type, |
474 | | std::vector<FileCacheType> other_cache_types, |
475 | | size_t size, std::lock_guard<std::mutex>& cache_lock, |
476 | | bool evict_in_advance); |
477 | | |
478 | | bool is_overflow(size_t removed_size, size_t need_size, size_t cur_cache_size, |
479 | | bool evict_in_advance) const; |
480 | | |
481 | | void remove_file_blocks(std::vector<FileBlockCell*>&, std::lock_guard<std::mutex>&, bool sync, |
482 | | std::string& reason); |
483 | | |
484 | | void find_evict_candidates(LRUQueue& queue, size_t size, size_t cur_cache_size, |
485 | | size_t& removed_size, std::vector<FileBlockCell*>& to_evict, |
486 | | std::lock_guard<std::mutex>& cache_lock, size_t& cur_removed_size, |
487 | | bool evict_in_advance); |
488 | | |
489 | | Status check_ofstream_status(std::ofstream& out, std::string& filename); |
490 | | Status dump_one_lru_entry(std::ofstream& out, std::string& filename, const UInt128Wrapper& hash, |
491 | | size_t offset, size_t size); |
492 | | Status finalize_dump(std::ofstream& out, size_t entry_num, std::string& tmp_filename, |
493 | | std::string& final_filename, size_t& file_size); |
494 | | Status check_ifstream_status(std::ifstream& in, std::string& filename); |
495 | | Status parse_dump_footer(std::ifstream& in, std::string& filename, size_t& entry_num); |
496 | | Status parse_one_lru_entry(std::ifstream& in, std::string& filename, UInt128Wrapper& hash, |
497 | | size_t& offset, size_t& size); |
498 | | void remove_lru_dump_files(); |
499 | | |
500 | | void clear_need_update_lru_blocks(); |
501 | | |
502 | | // info |
503 | | std::string _cache_base_path; |
504 | | size_t _capacity = 0; |
505 | | size_t _max_file_block_size = 0; |
506 | | |
507 | | mutable std::mutex _mutex; |
508 | | bool _close {false}; |
509 | | std::mutex _close_mtx; |
510 | | std::condition_variable _close_cv; |
511 | | std::thread _cache_background_monitor_thread; |
512 | | std::thread _cache_background_gc_thread; |
513 | | std::thread _cache_background_evict_in_advance_thread; |
514 | | std::thread _cache_background_lru_dump_thread; |
515 | | std::thread _cache_background_lru_log_replay_thread; |
516 | | std::thread _cache_background_block_lru_update_thread; |
517 | | std::atomic_bool _async_open_done {false}; |
518 | | // disk space or inode is less than the specified value |
519 | | bool _disk_resource_limit_mode {false}; |
520 | | bool _need_evict_cache_in_advance {false}; |
521 | | bool _is_initialized {false}; |
522 | | |
523 | | // strategy |
524 | | using FileBlocksByOffset = std::map<size_t, FileBlockCell>; |
525 | | using CachedFiles = std::unordered_map<UInt128Wrapper, FileBlocksByOffset, KeyHash>; |
526 | | CachedFiles _files; |
527 | | QueryFileCacheContextMap _query_map; |
528 | | size_t _cur_cache_size = 0; |
529 | | size_t _cur_ttl_size = 0; |
530 | | std::multimap<uint64_t, UInt128Wrapper> _time_to_key; |
531 | | std::unordered_map<UInt128Wrapper, uint64_t, KeyHash> _key_to_time; |
532 | | // The three queues are level queue. |
533 | | // It means as level1/level2/level3 queue. |
534 | | // but the level2 is maximum. |
535 | | // If some datas are importance, we can cache it into index queue |
536 | | // If some datas are just use once, we can cache it into disposable queue |
537 | | // The size proportion is [1:17:2]. |
538 | | LRUQueue _index_queue; |
539 | | LRUQueue _normal_queue; |
540 | | LRUQueue _disposable_queue; |
541 | | LRUQueue _ttl_queue; |
542 | | |
543 | | // keys for async remove |
544 | | RecycleFileCacheKeys _recycle_keys; |
545 | | |
546 | | std::unique_ptr<LRUQueueRecorder> _lru_recorder; |
547 | | std::unique_ptr<CacheLRUDumper> _lru_dumper; |
548 | | std::unique_ptr<BlockFileCacheTtlMgr> _ttl_mgr; |
549 | | |
550 | | // metrics |
551 | | std::shared_ptr<bvar::Status<size_t>> _cache_capacity_metrics; |
552 | | std::shared_ptr<bvar::Status<size_t>> _cur_cache_size_metrics; |
553 | | std::shared_ptr<bvar::Status<size_t>> _cur_ttl_cache_size_metrics; |
554 | | std::shared_ptr<bvar::Status<size_t>> _cur_ttl_cache_lru_queue_cache_size_metrics; |
555 | | std::shared_ptr<bvar::Status<size_t>> _cur_ttl_cache_lru_queue_element_count_metrics; |
556 | | std::shared_ptr<bvar::Status<size_t>> _cur_normal_queue_element_count_metrics; |
557 | | std::shared_ptr<bvar::Status<size_t>> _cur_normal_queue_cache_size_metrics; |
558 | | std::shared_ptr<bvar::Status<size_t>> _cur_index_queue_element_count_metrics; |
559 | | std::shared_ptr<bvar::Status<size_t>> _cur_index_queue_cache_size_metrics; |
560 | | std::shared_ptr<bvar::Status<size_t>> _cur_disposable_queue_element_count_metrics; |
561 | | std::shared_ptr<bvar::Status<size_t>> _cur_disposable_queue_cache_size_metrics; |
562 | | std::array<std::shared_ptr<bvar::Adder<size_t>>, 4> _queue_evict_size_metrics; |
563 | | std::shared_ptr<bvar::Adder<size_t>> _total_read_size_metrics; |
564 | | std::shared_ptr<bvar::Adder<size_t>> _total_hit_size_metrics; |
565 | | std::shared_ptr<bvar::Adder<size_t>> _total_evict_size_metrics; |
566 | | std::shared_ptr<bvar::Adder<size_t>> _gc_evict_bytes_metrics; |
567 | | std::shared_ptr<bvar::Adder<size_t>> _gc_evict_count_metrics; |
568 | | std::shared_ptr<bvar::Adder<size_t>> _evict_by_time_metrics_matrix[4][4]; |
569 | | std::shared_ptr<bvar::Adder<size_t>> _evict_by_size_metrics_matrix[4][4]; |
570 | | std::shared_ptr<bvar::Adder<size_t>> _evict_by_self_lru_metrics_matrix[4]; |
571 | | std::shared_ptr<bvar::Adder<size_t>> _evict_by_try_release; |
572 | | |
573 | | std::shared_ptr<bvar::Window<bvar::Adder<size_t>>> _num_hit_blocks_5m; |
574 | | std::shared_ptr<bvar::Window<bvar::Adder<size_t>>> _num_read_blocks_5m; |
575 | | std::shared_ptr<bvar::Window<bvar::Adder<size_t>>> _num_hit_blocks_1h; |
576 | | std::shared_ptr<bvar::Window<bvar::Adder<size_t>>> _num_read_blocks_1h; |
577 | | |
578 | | std::shared_ptr<bvar::Adder<size_t>> _num_read_blocks; |
579 | | std::shared_ptr<bvar::Adder<size_t>> _num_hit_blocks; |
580 | | std::shared_ptr<bvar::Adder<size_t>> _num_removed_blocks; |
581 | | |
582 | | std::shared_ptr<bvar::Adder<size_t>> _no_warmup_num_read_blocks; |
583 | | std::shared_ptr<bvar::Adder<size_t>> _no_warmup_num_hit_blocks; |
584 | | |
585 | | std::shared_ptr<bvar::Window<bvar::Adder<size_t>>> _no_warmup_num_hit_blocks_5m; |
586 | | std::shared_ptr<bvar::Window<bvar::Adder<size_t>>> _no_warmup_num_read_blocks_5m; |
587 | | std::shared_ptr<bvar::Window<bvar::Adder<size_t>>> _no_warmup_num_hit_blocks_1h; |
588 | | std::shared_ptr<bvar::Window<bvar::Adder<size_t>>> _no_warmup_num_read_blocks_1h; |
589 | | |
590 | | std::shared_ptr<bvar::Status<double>> _hit_ratio; |
591 | | std::shared_ptr<bvar::Status<double>> _hit_ratio_5m; |
592 | | std::shared_ptr<bvar::Status<double>> _hit_ratio_1h; |
593 | | std::shared_ptr<bvar::Status<double>> _no_warmup_hit_ratio; |
594 | | std::shared_ptr<bvar::Status<double>> _no_warmup_hit_ratio_5m; |
595 | | std::shared_ptr<bvar::Status<double>> _no_warmup_hit_ratio_1h; |
596 | | std::shared_ptr<bvar::Status<size_t>> _disk_limit_mode_metrics; |
597 | | std::shared_ptr<bvar::Status<size_t>> _need_evict_cache_in_advance_metrics; |
598 | | std::shared_ptr<bvar::Status<size_t>> _meta_store_write_queue_size_metrics; |
599 | | |
600 | | std::shared_ptr<bvar::LatencyRecorder> _cache_lock_wait_time_us; |
601 | | std::shared_ptr<bvar::LatencyRecorder> _get_or_set_latency_us; |
602 | | std::shared_ptr<bvar::LatencyRecorder> _storage_sync_remove_latency_us; |
603 | | std::shared_ptr<bvar::LatencyRecorder> _storage_retry_sync_remove_latency_us; |
604 | | std::shared_ptr<bvar::LatencyRecorder> _storage_async_remove_latency_us; |
605 | | std::shared_ptr<bvar::LatencyRecorder> _evict_in_advance_latency_us; |
606 | | std::shared_ptr<bvar::LatencyRecorder> _recycle_keys_length_recorder; |
607 | | std::shared_ptr<bvar::LatencyRecorder> _update_lru_blocks_latency_us; |
608 | | std::shared_ptr<bvar::LatencyRecorder> _need_update_lru_blocks_length_recorder; |
609 | | std::shared_ptr<bvar::LatencyRecorder> _ttl_gc_latency_us; |
610 | | |
611 | | std::shared_ptr<bvar::LatencyRecorder> _shadow_queue_levenshtein_distance; |
612 | | // keep _storage last so it will deconstruct first |
613 | | // otherwise, load_cache_info_into_memory might crash |
614 | | // coz it will use other members of BlockFileCache |
615 | | // so join this async load thread first |
616 | | std::unique_ptr<FileCacheStorage> _storage; |
617 | | std::shared_ptr<bvar::LatencyRecorder> _lru_dump_latency_us; |
618 | | std::mutex _dump_lru_queues_mtx; |
619 | | NeedUpdateLRUBlocks _need_update_lru_blocks; |
620 | | }; |
621 | | |
622 | | } // namespace doris::io |