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 <limits> |
29 | | #include <memory> |
30 | | #include <mutex> |
31 | | #include <optional> |
32 | | #include <thread> |
33 | | #include <unordered_map> |
34 | | #include <vector> |
35 | | |
36 | | #include "io/cache/block_file_cache_ttl_mgr.h" |
37 | | #include "io/cache/cache_lru_dumper.h" |
38 | | #include "io/cache/file_block.h" |
39 | | #include "io/cache/file_cache_common.h" |
40 | | #include "io/cache/file_cache_storage.h" |
41 | | #include "io/cache/lru_queue_recorder.h" |
42 | | #include "runtime/runtime_profile.h" |
43 | | #include "util/threadpool.h" |
44 | | |
45 | | namespace doris::io { |
46 | | using RecycleFileCacheKeys = moodycamel::ConcurrentQueue<FileCacheKey>; |
47 | | |
48 | | class LockScopedTimer { |
49 | | public: |
50 | 7.54k | LockScopedTimer() : start_(std::chrono::steady_clock::now()) {} |
51 | 7.54k | ~LockScopedTimer() { |
52 | 7.54k | auto end = std::chrono::steady_clock::now(); |
53 | 7.54k | auto duration_us = |
54 | 7.54k | std::chrono::duration_cast<std::chrono::microseconds>(end - start_).count(); |
55 | 7.54k | if (duration_us > config::cache_lock_held_long_tail_threshold_us) { |
56 | 0 | LOG(WARNING) << "Lock held time " << std::to_string(duration_us) << "us. " |
57 | 0 | << get_stack_trace(); |
58 | 0 | } |
59 | 7.54k | } |
60 | | |
61 | | private: |
62 | | std::chrono::time_point<std::chrono::steady_clock> start_; |
63 | | }; |
64 | | |
65 | | // Note: the cache_lock is scoped, so do not add do...while(0) here. |
66 | | #define SCOPED_CACHE_LOCK(MUTEX, cache) \ |
67 | 7.54k | std::chrono::time_point<std::chrono::steady_clock> start_time = \ |
68 | 7.54k | std::chrono::steady_clock::now(); \ |
69 | 7.54k | std::lock_guard cache_lock(MUTEX); \ |
70 | 7.54k | std::chrono::time_point<std::chrono::steady_clock> acq_time = \ |
71 | 7.54k | std::chrono::steady_clock::now(); \ |
72 | 7.54k | auto duration_us = \ |
73 | 7.54k | std::chrono::duration_cast<std::chrono::microseconds>(acq_time - start_time).count(); \ |
74 | 7.54k | *(cache->_cache_lock_wait_time_us) << duration_us; \ |
75 | 7.54k | if (duration_us > config::cache_lock_wait_long_tail_threshold_us) { \ |
76 | 0 | LOG(WARNING) << "Lock wait time " << std::to_string(duration_us) << "us. " \ |
77 | 0 | << get_stack_trace() << std::endl; \ |
78 | 0 | } \ |
79 | 7.54k | LockScopedTimer cache_lock_timer; |
80 | | |
81 | | class FSFileCacheStorage; |
82 | | |
83 | | // NeedUpdateLRUBlocks keeps FileBlockSPtr entries that require LRU updates in a |
84 | | // deduplicated, sharded container. Entries are keyed by the raw FileBlock |
85 | | // pointer so that multiple shared_ptr copies of the same block are treated as a |
86 | | // single pending update. The structure is thread-safe and optimized for high |
87 | | // contention insert/drain workloads in the background update thread. |
88 | | // Note that Blocks are updated in batch, internal order is not important. |
89 | | class NeedUpdateLRUBlocks { |
90 | | public: |
91 | 185 | NeedUpdateLRUBlocks() = default; |
92 | | |
93 | | // Insert a block into the pending set. Returns true only when the block |
94 | | // was not already queued. Null inputs are ignored. |
95 | | bool insert(FileBlockSPtr block, size_t max_queue_size = std::numeric_limits<size_t>::max()); |
96 | | |
97 | | // Drain up to `limit` unique blocks into `output`. The method returns how |
98 | | // many blocks were actually drained and shrinks the internal size |
99 | | // accordingly. |
100 | | size_t drain(size_t limit, std::vector<FileBlockSPtr>* output); |
101 | | |
102 | | // Remove every pending block from the structure and reset the size. |
103 | | void clear(); |
104 | | |
105 | | // Thread-safe approximate size of queued unique blocks. |
106 | 1.67k | size_t size() const { return _size.load(std::memory_order_relaxed); } |
107 | | |
108 | | private: |
109 | | static constexpr size_t kShardCount = 64; |
110 | | static constexpr size_t kShardMask = kShardCount - 1; |
111 | | |
112 | | struct Shard { |
113 | | std::mutex mutex; |
114 | | std::unordered_map<FileBlock*, FileBlockSPtr> entries; |
115 | | }; |
116 | | |
117 | | size_t shard_index(FileBlock* ptr) const; |
118 | | void decrease_size(size_t delta); |
119 | | |
120 | | std::array<Shard, kShardCount> _shards; |
121 | | std::atomic<size_t> _size {0}; |
122 | | }; |
123 | | |
124 | | // The BlockFileCache is responsible for the management of the blocks |
125 | | // The current strategies are lru and ttl. |
126 | | |
127 | | struct FileBlockCell { |
128 | | friend class FileBlock; |
129 | | |
130 | | FileBlockSPtr file_block; |
131 | | /// Iterator is put here on first reservation attempt, if successful. |
132 | | std::optional<LRUQueue::Iterator> queue_iterator; |
133 | | |
134 | | mutable int64_t atime {0}; |
135 | 730k | void update_atime() const { |
136 | 730k | atime = std::chrono::duration_cast<std::chrono::seconds>( |
137 | 730k | std::chrono::steady_clock::now().time_since_epoch()) |
138 | 730k | .count(); |
139 | 730k | } |
140 | | |
141 | | /// Pointer to file block is always hold by the cache itself. |
142 | | /// Apart from pointer in cache, it can be hold by cache users, when they call |
143 | | /// getorSet(), but cache users always hold it via FileBlocksHolder. |
144 | 1.90k | bool releasable() const { |
145 | 1.90k | return (file_block.use_count() == 1 || |
146 | 1.90k | (file_block.use_count() == 2 && file_block->_owned_by_cached_reader)); |
147 | 1.90k | } |
148 | | |
149 | 544k | size_t size() const { return file_block->_block_range.size(); } |
150 | | |
151 | 70 | FileBlockCell() = default; |
152 | | FileBlockCell(FileBlockSPtr file_block, std::lock_guard<std::mutex>& cache_lock); |
153 | | FileBlockCell(FileBlockCell&& other) noexcept |
154 | 27.2k | : file_block(std::move(other.file_block)), |
155 | 27.2k | queue_iterator(other.queue_iterator), |
156 | 27.2k | atime(other.atime) { |
157 | 27.2k | file_block->cell = this; |
158 | 27.2k | } |
159 | | |
160 | | FileBlockCell& operator=(const FileBlockCell&) = delete; |
161 | | FileBlockCell(const FileBlockCell&) = delete; |
162 | | |
163 | 0 | size_t dowloading_size() const { return file_block->_downloaded_size; } |
164 | | }; |
165 | | |
166 | | class BlockFileCache { |
167 | | friend class FSFileCacheStorage; |
168 | | friend class MemFileCacheStorage; |
169 | | friend class FileBlock; |
170 | | friend struct FileBlocksHolder; |
171 | | friend class CacheLRUDumper; |
172 | | friend class LRUQueueRecorder; |
173 | | friend struct FileBlockCell; |
174 | | friend class BlockFileCacheTest; |
175 | | |
176 | | public: |
177 | | // hash the file_name to uint128 |
178 | | static UInt128Wrapper hash(const std::string& path); |
179 | | |
180 | | BlockFileCache(const std::string& cache_base_path, const FileCacheSettings& cache_settings); |
181 | | |
182 | 180 | virtual ~BlockFileCache() { |
183 | 180 | { |
184 | 180 | std::lock_guard lock(_close_mtx); |
185 | 180 | _close = true; |
186 | 180 | } |
187 | 180 | _close_cv.notify_all(); |
188 | 180 | if (_cache_background_monitor_thread.joinable()) { |
189 | 155 | _cache_background_monitor_thread.join(); |
190 | 155 | } |
191 | 180 | if (_cache_background_gc_thread.joinable()) { |
192 | 155 | _cache_background_gc_thread.join(); |
193 | 155 | } |
194 | 180 | if (_cache_background_evict_in_advance_thread.joinable()) { |
195 | 155 | _cache_background_evict_in_advance_thread.join(); |
196 | 155 | } |
197 | 180 | if (_cache_background_lru_dump_thread.joinable()) { |
198 | 155 | _cache_background_lru_dump_thread.join(); |
199 | 155 | } |
200 | 180 | if (_cache_background_lru_log_replay_thread.joinable()) { |
201 | 155 | _cache_background_lru_log_replay_thread.join(); |
202 | 155 | } |
203 | 180 | if (_cache_background_block_lru_update_thread.joinable()) { |
204 | 155 | _cache_background_block_lru_update_thread.join(); |
205 | 155 | } |
206 | 180 | if (_ttl_mgr) { |
207 | 133 | _ttl_mgr.reset(); |
208 | 133 | } |
209 | 180 | } |
210 | | |
211 | | /// Restore cache from local filesystem. |
212 | | Status initialize(); |
213 | | |
214 | | /// Cache capacity in bytes. |
215 | 14 | [[nodiscard]] size_t capacity() const { return _capacity; } |
216 | | |
217 | | // try to release all releasable block |
218 | | // it maybe hang the io/system |
219 | | size_t try_release(); |
220 | | |
221 | 147 | [[nodiscard]] const std::string& get_base_path() const { return _cache_base_path; } |
222 | | |
223 | | // Get storage for inspection |
224 | 0 | FileCacheStorage* get_storage() const { return _storage.get(); } |
225 | | |
226 | | /** |
227 | | * Given an `offset` and `size` representing [offset, offset + size) bytes interval, |
228 | | * return list of cached non-overlapping non-empty |
229 | | * file blocks `[block1, ..., blockN]` which intersect with given interval. |
230 | | * |
231 | | * blocks in returned list are ordered in ascending order and represent a full contiguous |
232 | | * interval (no holes). Each block in returned list has state: DOWNLOADED, DOWNLOADING or EMPTY. |
233 | | * |
234 | | * As long as pointers to returned file blocks are hold |
235 | | * it is guaranteed that these file blocks are not removed from cache. |
236 | | */ |
237 | | FileBlocksHolder get_or_set(const UInt128Wrapper& hash, size_t offset, size_t size, |
238 | | CacheContext& context); |
239 | | |
240 | | /** |
241 | | * record blocks read directly by CachedRemoteFileReader |
242 | | */ |
243 | | void add_need_update_lru_block(FileBlockSPtr block); |
244 | | |
245 | | /** |
246 | | * Clear all cached data for this cache instance async |
247 | | * |
248 | | * @returns summary message |
249 | | */ |
250 | | std::string clear_file_cache_async(); |
251 | | std::string clear_file_cache_directly(); |
252 | | |
253 | | /** |
254 | | * Reset the cache capacity. If the new_capacity is smaller than _capacity, the redundant data will be remove async. |
255 | | * |
256 | | * @returns summary message |
257 | | */ |
258 | | std::string reset_capacity(size_t new_capacity); |
259 | | |
260 | | std::map<size_t, FileBlockSPtr> get_blocks_by_key(const UInt128Wrapper& hash); |
261 | | |
262 | | /// For debug and UT |
263 | | std::string dump_structure(const UInt128Wrapper& hash); |
264 | | std::string dump_single_cache_type(const UInt128Wrapper& hash, size_t offset); |
265 | | |
266 | | void dump_lru_queues(bool force); |
267 | | |
268 | | [[nodiscard]] size_t get_used_cache_size(FileCacheType type) const; |
269 | | |
270 | | [[nodiscard]] size_t get_file_blocks_num(FileCacheType type) const; |
271 | | |
272 | | // change the block cache type |
273 | | void change_cache_type(const UInt128Wrapper& hash, size_t offset, FileCacheType new_type, |
274 | | std::lock_guard<std::mutex>& cache_lock); |
275 | | |
276 | | // remove all blocks that belong to the key |
277 | | void remove_if_cached(const UInt128Wrapper& key); |
278 | | void remove_if_cached_async(const UInt128Wrapper& key); |
279 | | |
280 | | // Reset the block size and keep FileBlock, LRU queue, and cache counters consistent. |
281 | | void reset_range(const UInt128Wrapper&, size_t offset, size_t old_size, size_t new_size, |
282 | | std::lock_guard<std::mutex>& cache_lock); |
283 | | |
284 | | // get the hotest blocks message by key |
285 | | // The tuple is composed of <offset, size, cache_type, expiration_time> |
286 | | [[nodiscard]] std::vector<std::tuple<size_t, size_t, FileCacheType, uint64_t>> |
287 | | get_hot_blocks_meta(const UInt128Wrapper& hash) const; |
288 | | |
289 | 539 | [[nodiscard]] bool get_async_open_success() const { return _async_open_done; } |
290 | | |
291 | | BlockFileCache& operator=(const BlockFileCache&) = delete; |
292 | | BlockFileCache(const BlockFileCache&) = delete; |
293 | | |
294 | | // try to reserve the new space for the new block if the cache is full |
295 | | bool try_reserve(const UInt128Wrapper& hash, const CacheContext& context, size_t offset, |
296 | | size_t size, std::lock_guard<std::mutex>& cache_lock); |
297 | | |
298 | | /** |
299 | | * Proactively evict cache blocks to free up space before cache is full. |
300 | | * |
301 | | * This function attempts to evict blocks from both NORMAL and TTL queues to maintain |
302 | | * cache size below high watermark. Unlike try_reserve() which blocks until space is freed, |
303 | | * this function initiates asynchronous eviction in background. |
304 | | * |
305 | | * @param size Number of bytes to try to evict |
306 | | * @param cache_lock Lock that must be held while accessing cache data structures |
307 | | * |
308 | | * @pre Caller must hold cache_lock |
309 | | * @pre _need_evict_cache_in_advance must be true |
310 | | * @pre _recycle_keys queue must have capacity for evicted blocks |
311 | | */ |
312 | | void try_evict_in_advance(size_t size, std::lock_guard<std::mutex>& cache_lock); |
313 | | |
314 | | void update_ttl_atime(const UInt128Wrapper& hash); |
315 | | |
316 | | void pause_ttl_manager(); |
317 | | void resume_ttl_manager(); |
318 | | |
319 | | std::map<std::string, double> get_stats(); |
320 | | |
321 | | // for be UTs |
322 | | std::map<std::string, double> get_stats_unsafe(); |
323 | 5 | [[nodiscard]] size_t need_update_lru_blocks_size_unsafe() const { |
324 | 5 | return _need_update_lru_blocks.size(); |
325 | 5 | } |
326 | | |
327 | | using AccessRecord = |
328 | | std::unordered_map<AccessKeyAndOffset, LRUQueue::Iterator, KeyAndOffsetHash>; |
329 | | |
330 | | /// Used to track and control the cache access of each query. |
331 | | /// Through it, we can realize the processing of different queries by the cache layer. |
332 | | struct QueryFileCacheContext { |
333 | | LRUQueue lru_queue; |
334 | | AccessRecord records; |
335 | | |
336 | 5 | QueryFileCacheContext(size_t max_cache_size) : lru_queue(max_cache_size, 0, 0) {} |
337 | | |
338 | | void remove(const UInt128Wrapper& hash, size_t offset, |
339 | | std::lock_guard<std::mutex>& cache_lock); |
340 | | |
341 | | void reserve(const UInt128Wrapper& hash, size_t offset, size_t size, |
342 | | std::lock_guard<std::mutex>& cache_lock); |
343 | | |
344 | 69 | size_t get_max_cache_size() const { return lru_queue.get_max_size(); } |
345 | | |
346 | 48 | size_t get_cache_size(std::lock_guard<std::mutex>& cache_lock) const { |
347 | 48 | return lru_queue.get_capacity(cache_lock); |
348 | 48 | } |
349 | | |
350 | 50 | LRUQueue& queue() { return lru_queue; } |
351 | | }; |
352 | | |
353 | | using QueryFileCacheContextPtr = std::shared_ptr<QueryFileCacheContext>; |
354 | | using QueryFileCacheContextMap = std::unordered_map<TUniqueId, QueryFileCacheContextPtr>; |
355 | | |
356 | | QueryFileCacheContextPtr get_query_context(const TUniqueId& query_id, |
357 | | std::lock_guard<std::mutex>&); |
358 | | |
359 | | void remove_query_context(const TUniqueId& query_id); |
360 | | |
361 | | QueryFileCacheContextPtr get_or_set_query_context(const TUniqueId& query_id, |
362 | | std::lock_guard<std::mutex>& cache_lock, |
363 | | int file_cache_query_limit_percent); |
364 | | |
365 | | /// Save a query context information, and adopt different cache policies |
366 | | /// for different queries through the context cache layer. |
367 | | struct QueryFileCacheContextHolder { |
368 | | QueryFileCacheContextHolder(const TUniqueId& query_id, BlockFileCache* mgr, |
369 | | QueryFileCacheContextPtr context) |
370 | 7 | : query_id(query_id), mgr(mgr), context(context) {} |
371 | | |
372 | | QueryFileCacheContextHolder& operator=(const QueryFileCacheContextHolder&) = delete; |
373 | | QueryFileCacheContextHolder(const QueryFileCacheContextHolder&) = delete; |
374 | | |
375 | 7 | ~QueryFileCacheContextHolder() { |
376 | | /// If only the query_map and the current holder hold the context_query, |
377 | | /// the query has been completed and the query_context is released. |
378 | 7 | if (context) { |
379 | 6 | context.reset(); |
380 | 6 | mgr->remove_query_context(query_id); |
381 | 6 | } |
382 | 7 | } |
383 | | |
384 | | const TUniqueId& query_id; |
385 | | BlockFileCache* mgr = nullptr; |
386 | | QueryFileCacheContextPtr context; |
387 | | }; |
388 | | using QueryFileCacheContextHolderPtr = std::unique_ptr<QueryFileCacheContextHolder>; |
389 | | QueryFileCacheContextHolderPtr get_query_context_holder(const TUniqueId& query_id, |
390 | | int file_cache_query_limit_percent); |
391 | | |
392 | 0 | int64_t approximate_available_cache_size() const { |
393 | 0 | return std::max<int64_t>( |
394 | 0 | _cache_capacity_metrics->get_value() - _cur_cache_size_metrics->get_value(), 0); |
395 | 0 | } |
396 | | |
397 | | Status report_file_cache_inconsistency(std::vector<std::string>& results); |
398 | | Status check_file_cache_consistency(InconsistencyContext& inconsistency_context); |
399 | | |
400 | | private: |
401 | | LRUQueue& get_queue(FileCacheType type); |
402 | | const LRUQueue& get_queue(FileCacheType type) const; |
403 | | |
404 | | template <class T, class U> |
405 | | requires IsXLock<T> && IsXLock<U> |
406 | | void remove(FileBlockSPtr file_block, T& cache_lock, U& segment_lock, bool sync = true); |
407 | | |
408 | | FileBlocks get_impl(const UInt128Wrapper& hash, const CacheContext& context, |
409 | | const FileBlock::Range& range, std::lock_guard<std::mutex>& cache_lock); |
410 | | |
411 | | template <class T> |
412 | | requires IsXLock<T> |
413 | | FileBlockCell* get_cell(const UInt128Wrapper& hash, size_t offset, T& cache_lock); |
414 | | |
415 | | virtual FileBlockCell* add_cell(const UInt128Wrapper& hash, const CacheContext& context, |
416 | | size_t offset, size_t size, FileBlock::State state, |
417 | | std::lock_guard<std::mutex>& cache_lock); |
418 | | |
419 | | Status initialize_unlocked(std::lock_guard<std::mutex>& cache_lock); |
420 | | |
421 | | void update_block_lru(FileBlockSPtr block, std::lock_guard<std::mutex>& cache_lock); |
422 | | |
423 | | void use_cell(const FileBlockCell& cell, FileBlocks* result, bool not_need_move, |
424 | | std::lock_guard<std::mutex>& cache_lock); |
425 | | |
426 | | bool try_reserve_for_lru(const UInt128Wrapper& hash, QueryFileCacheContextPtr query_context, |
427 | | const CacheContext& context, size_t offset, size_t size, |
428 | | std::lock_guard<std::mutex>& cache_lock, |
429 | | bool evict_in_advance = false); |
430 | | |
431 | | bool try_reserve_during_async_load(size_t size, std::lock_guard<std::mutex>& cache_lock); |
432 | | |
433 | | std::vector<FileCacheType> get_other_cache_type(FileCacheType cur_cache_type); |
434 | | std::vector<FileCacheType> get_other_cache_type_without_ttl(FileCacheType cur_cache_type); |
435 | | |
436 | | bool try_reserve_from_other_queue(FileCacheType cur_cache_type, size_t offset, int64_t cur_time, |
437 | | std::lock_guard<std::mutex>& cache_lock, |
438 | | bool evict_in_advance = false); |
439 | | |
440 | | size_t get_available_cache_size(FileCacheType cache_type) const; |
441 | | |
442 | | FileBlocks split_range_into_cells(const UInt128Wrapper& hash, const CacheContext& context, |
443 | | size_t offset, size_t size, FileBlock::State state, |
444 | | std::lock_guard<std::mutex>& cache_lock); |
445 | | |
446 | | std::string dump_structure_unlocked(const UInt128Wrapper& hash, |
447 | | std::lock_guard<std::mutex>& cache_lock); |
448 | | |
449 | | std::string dump_single_cache_type_unlocked(const UInt128Wrapper& hash, size_t offset, |
450 | | std::lock_guard<std::mutex>& cache_lock); |
451 | | |
452 | | void fill_holes_with_empty_file_blocks(FileBlocks& file_blocks, const UInt128Wrapper& hash, |
453 | | const CacheContext& context, |
454 | | const FileBlock::Range& range, |
455 | | std::lock_guard<std::mutex>& cache_lock); |
456 | | |
457 | | size_t get_used_cache_size_unlocked(FileCacheType type, |
458 | | std::lock_guard<std::mutex>& cache_lock) const; |
459 | | |
460 | | void check_disk_resource_limit(); |
461 | | void check_need_evict_cache_in_advance(); |
462 | | |
463 | | size_t get_available_cache_size_unlocked(FileCacheType type, |
464 | | std::lock_guard<std::mutex>& cache_lock) const; |
465 | | |
466 | | size_t get_file_blocks_num_unlocked(FileCacheType type, |
467 | | std::lock_guard<std::mutex>& cache_lock) const; |
468 | | |
469 | | bool need_to_move(FileCacheType cell_type, FileCacheType query_type) const; |
470 | | |
471 | | void run_background_monitor(); |
472 | | void run_background_gc(); |
473 | | void run_background_lru_log_replay(); |
474 | | size_t replay_lru_logs_once(); |
475 | | void run_background_lru_dump(); |
476 | | void restore_lru_queues_from_disk(std::lock_guard<std::mutex>& cache_lock); |
477 | | void run_background_evict_in_advance(); |
478 | | void run_background_block_lru_update(); |
479 | | |
480 | | bool try_reserve_from_other_queue_by_time_interval(FileCacheType cur_type, |
481 | | std::vector<FileCacheType> other_cache_types, |
482 | | size_t size, int64_t cur_time, |
483 | | std::lock_guard<std::mutex>& cache_lock, |
484 | | bool evict_in_advance); |
485 | | |
486 | | bool try_reserve_from_other_queue_by_size(FileCacheType cur_type, |
487 | | std::vector<FileCacheType> other_cache_types, |
488 | | size_t size, std::lock_guard<std::mutex>& cache_lock, |
489 | | bool evict_in_advance); |
490 | | |
491 | | bool is_overflow(size_t removed_size, size_t need_size, size_t cur_cache_size, |
492 | | bool evict_in_advance) const; |
493 | | |
494 | | void remove_file_blocks(std::vector<FileBlockCell*>&, std::lock_guard<std::mutex>&, bool sync, |
495 | | std::string& reason); |
496 | | |
497 | | void find_evict_candidates(LRUQueue& queue, size_t size, size_t cur_cache_size, |
498 | | size_t& removed_size, std::vector<FileBlockCell*>& to_evict, |
499 | | std::lock_guard<std::mutex>& cache_lock, size_t& cur_removed_size, |
500 | | bool evict_in_advance); |
501 | | |
502 | | Status check_ofstream_status(std::ofstream& out, std::string& filename); |
503 | | Status dump_one_lru_entry(std::ofstream& out, std::string& filename, const UInt128Wrapper& hash, |
504 | | size_t offset, size_t size); |
505 | | Status finalize_dump(std::ofstream& out, size_t entry_num, std::string& tmp_filename, |
506 | | std::string& final_filename, size_t& file_size); |
507 | | Status check_ifstream_status(std::ifstream& in, std::string& filename); |
508 | | Status parse_dump_footer(std::ifstream& in, std::string& filename, size_t& entry_num); |
509 | | Status parse_one_lru_entry(std::ifstream& in, std::string& filename, UInt128Wrapper& hash, |
510 | | size_t& offset, size_t& size); |
511 | | void remove_lru_dump_files(); |
512 | | |
513 | | void clear_need_update_lru_blocks(); |
514 | | |
515 | | // info |
516 | | std::string _cache_base_path; |
517 | | size_t _capacity = 0; |
518 | | size_t _max_file_block_size = 0; |
519 | | |
520 | | mutable std::mutex _mutex; |
521 | | bool _close {false}; |
522 | | std::mutex _close_mtx; |
523 | | std::condition_variable _close_cv; |
524 | | std::thread _cache_background_monitor_thread; |
525 | | std::thread _cache_background_gc_thread; |
526 | | std::thread _cache_background_evict_in_advance_thread; |
527 | | std::thread _cache_background_lru_dump_thread; |
528 | | std::thread _cache_background_lru_log_replay_thread; |
529 | | std::thread _cache_background_block_lru_update_thread; |
530 | | std::atomic_bool _async_open_done {false}; |
531 | | // disk space or inode is less than the specified value |
532 | | bool _disk_resource_limit_mode {false}; |
533 | | bool _need_evict_cache_in_advance {false}; |
534 | | bool _is_initialized {false}; |
535 | | |
536 | | // strategy |
537 | | using FileBlocksByOffset = std::map<size_t, FileBlockCell>; |
538 | | using CachedFiles = std::unordered_map<UInt128Wrapper, FileBlocksByOffset, KeyHash>; |
539 | | CachedFiles _files; |
540 | | QueryFileCacheContextMap _query_map; |
541 | | size_t _cur_cache_size = 0; |
542 | | size_t _cur_ttl_size = 0; |
543 | | std::multimap<uint64_t, UInt128Wrapper> _time_to_key; |
544 | | std::unordered_map<UInt128Wrapper, uint64_t, KeyHash> _key_to_time; |
545 | | // The three queues are level queue. |
546 | | // It means as level1/level2/level3 queue. |
547 | | // but the level2 is maximum. |
548 | | // If some datas are importance, we can cache it into index queue |
549 | | // If some datas are just use once, we can cache it into disposable queue |
550 | | // The size proportion is [1:17:2]. |
551 | | LRUQueue _index_queue; |
552 | | LRUQueue _normal_queue; |
553 | | LRUQueue _disposable_queue; |
554 | | LRUQueue _ttl_queue; |
555 | | |
556 | | // keys for async remove |
557 | | RecycleFileCacheKeys _recycle_keys; |
558 | | |
559 | | std::unique_ptr<LRUQueueRecorder> _lru_recorder; |
560 | | std::unique_ptr<CacheLRUDumper> _lru_dumper; |
561 | | std::unique_ptr<BlockFileCacheTtlMgr> _ttl_mgr; |
562 | | |
563 | | // metrics |
564 | | std::shared_ptr<bvar::Status<size_t>> _cache_capacity_metrics; |
565 | | std::shared_ptr<bvar::Status<size_t>> _cur_cache_size_metrics; |
566 | | std::shared_ptr<bvar::Status<size_t>> _cur_ttl_cache_size_metrics; |
567 | | std::shared_ptr<bvar::Status<size_t>> _cur_ttl_cache_lru_queue_cache_size_metrics; |
568 | | std::shared_ptr<bvar::Status<size_t>> _cur_ttl_cache_lru_queue_element_count_metrics; |
569 | | std::shared_ptr<bvar::Status<size_t>> _cur_normal_queue_element_count_metrics; |
570 | | std::shared_ptr<bvar::Status<size_t>> _cur_normal_queue_cache_size_metrics; |
571 | | std::shared_ptr<bvar::Status<size_t>> _cur_index_queue_element_count_metrics; |
572 | | std::shared_ptr<bvar::Status<size_t>> _cur_index_queue_cache_size_metrics; |
573 | | std::shared_ptr<bvar::Status<size_t>> _cur_disposable_queue_element_count_metrics; |
574 | | std::shared_ptr<bvar::Status<size_t>> _cur_disposable_queue_cache_size_metrics; |
575 | | std::array<std::shared_ptr<bvar::Adder<size_t>>, 4> _queue_evict_size_metrics; |
576 | | std::shared_ptr<bvar::Adder<size_t>> _total_read_size_metrics; |
577 | | std::shared_ptr<bvar::Adder<size_t>> _total_hit_size_metrics; |
578 | | std::shared_ptr<bvar::Adder<size_t>> _total_evict_size_metrics; |
579 | | std::shared_ptr<bvar::Adder<size_t>> _gc_evict_bytes_metrics; |
580 | | std::shared_ptr<bvar::Adder<size_t>> _gc_evict_count_metrics; |
581 | | std::shared_ptr<bvar::Adder<size_t>> _evict_by_time_metrics_matrix[4][4]; |
582 | | std::shared_ptr<bvar::Adder<size_t>> _evict_by_size_metrics_matrix[4][4]; |
583 | | std::shared_ptr<bvar::Adder<size_t>> _evict_by_self_lru_metrics_matrix[4]; |
584 | | std::shared_ptr<bvar::Adder<size_t>> _evict_by_try_release; |
585 | | |
586 | | std::shared_ptr<bvar::Window<bvar::Adder<size_t>>> _num_hit_blocks_5m; |
587 | | std::shared_ptr<bvar::Window<bvar::Adder<size_t>>> _num_read_blocks_5m; |
588 | | std::shared_ptr<bvar::Window<bvar::Adder<size_t>>> _num_hit_blocks_1h; |
589 | | std::shared_ptr<bvar::Window<bvar::Adder<size_t>>> _num_read_blocks_1h; |
590 | | |
591 | | std::shared_ptr<bvar::Adder<size_t>> _num_read_blocks; |
592 | | std::shared_ptr<bvar::Adder<size_t>> _num_hit_blocks; |
593 | | std::shared_ptr<bvar::Adder<size_t>> _num_removed_blocks; |
594 | | |
595 | | std::shared_ptr<bvar::Adder<size_t>> _no_warmup_num_read_blocks; |
596 | | std::shared_ptr<bvar::Adder<size_t>> _no_warmup_num_hit_blocks; |
597 | | |
598 | | std::shared_ptr<bvar::Window<bvar::Adder<size_t>>> _no_warmup_num_hit_blocks_5m; |
599 | | std::shared_ptr<bvar::Window<bvar::Adder<size_t>>> _no_warmup_num_read_blocks_5m; |
600 | | std::shared_ptr<bvar::Window<bvar::Adder<size_t>>> _no_warmup_num_hit_blocks_1h; |
601 | | std::shared_ptr<bvar::Window<bvar::Adder<size_t>>> _no_warmup_num_read_blocks_1h; |
602 | | |
603 | | std::shared_ptr<bvar::Status<double>> _hit_ratio; |
604 | | std::shared_ptr<bvar::Status<double>> _hit_ratio_5m; |
605 | | std::shared_ptr<bvar::Status<double>> _hit_ratio_1h; |
606 | | std::shared_ptr<bvar::Status<double>> _no_warmup_hit_ratio; |
607 | | std::shared_ptr<bvar::Status<double>> _no_warmup_hit_ratio_5m; |
608 | | std::shared_ptr<bvar::Status<double>> _no_warmup_hit_ratio_1h; |
609 | | std::shared_ptr<bvar::Status<size_t>> _disk_limit_mode_metrics; |
610 | | std::shared_ptr<bvar::Status<size_t>> _need_evict_cache_in_advance_metrics; |
611 | | std::shared_ptr<bvar::Status<size_t>> _meta_store_write_queue_size_metrics; |
612 | | |
613 | | std::shared_ptr<bvar::LatencyRecorder> _cache_lock_wait_time_us; |
614 | | std::shared_ptr<bvar::LatencyRecorder> _get_or_set_latency_us; |
615 | | std::shared_ptr<bvar::LatencyRecorder> _storage_sync_remove_latency_us; |
616 | | std::shared_ptr<bvar::LatencyRecorder> _storage_retry_sync_remove_latency_us; |
617 | | std::shared_ptr<bvar::LatencyRecorder> _storage_async_remove_latency_us; |
618 | | std::shared_ptr<bvar::LatencyRecorder> _evict_in_advance_latency_us; |
619 | | std::shared_ptr<bvar::LatencyRecorder> _recycle_keys_length_recorder; |
620 | | std::shared_ptr<bvar::LatencyRecorder> _update_lru_blocks_latency_us; |
621 | | std::shared_ptr<bvar::LatencyRecorder> _need_update_lru_blocks_length_recorder; |
622 | | std::shared_ptr<bvar::Adder<size_t>> _need_update_lru_blocks_produce_metrics; |
623 | | std::shared_ptr<bvar::Adder<size_t>> _need_update_lru_blocks_consume_metrics; |
624 | | std::shared_ptr<bvar::LatencyRecorder> _ttl_gc_latency_us; |
625 | | |
626 | | std::shared_ptr<bvar::LatencyRecorder> _shadow_queue_levenshtein_distance; |
627 | | std::array<std::shared_ptr<bvar::LatencyRecorder>, 4> _lru_recorder_queue_length_recorder; |
628 | | std::array<std::shared_ptr<bvar::Adder<size_t>>, 4> _lru_recorder_queue_produce_metrics; |
629 | | std::array<std::shared_ptr<bvar::Adder<size_t>>, 4> _lru_recorder_queue_consume_metrics; |
630 | | std::shared_ptr<bvar::Adder<size_t>> _lru_recorder_log_replay_idle_metrics; |
631 | | // keep _storage last so it will deconstruct first |
632 | | // otherwise, load_cache_info_into_memory might crash |
633 | | // coz it will use other members of BlockFileCache |
634 | | // so join this async load thread first |
635 | | std::unique_ptr<FileCacheStorage> _storage; |
636 | | std::shared_ptr<bvar::LatencyRecorder> _lru_dump_latency_us; |
637 | | std::mutex _dump_lru_queues_mtx; |
638 | | NeedUpdateLRUBlocks _need_update_lru_blocks; |
639 | | }; |
640 | | |
641 | | } // namespace doris::io |