leveldb 學習記錄(八) compact

隨着運行時間的增長，memtable會慢慢 轉化成 sstable。

sstable會愈來愈多 咱們就須要進行整合 compact

代碼會在寫入查詢key值 db寫入時等多出位置調用MaybeScheduleCompaction () 

檢測是否須要進行compact

 1 void DBImpl::MaybeScheduleCompaction() {
 2   mutex_.AssertHeld();
 3   if (bg_compaction_scheduled_) {
 4     // Already scheduled
 5   } else if (shutting_down_.Acquire_Load()) {
 6     // DB is being deleted; no more background compactions
 7   } else if (imm_ == NULL &&
 8              manual_compaction_ == NULL &&
 9              !versions_->NeedsCompaction()) {
10     // No work to be done
11   } else {
12     bg_compaction_scheduled_ = true;
13     env_->Schedule(&DBImpl::BGWork, this);
14   }
15 }
16 
17 void DBImpl::BGWork(void* db) {
18   reinterpret_cast<DBImpl*>(db)->BackgroundCall();
19 }
20 
21 void DBImpl::BackgroundCall() {
22   MutexLock l(&mutex_);
23   assert(bg_compaction_scheduled_);
24   if (!shutting_down_.Acquire_Load()) {
25     BackgroundCompaction();
26   }
27   bg_compaction_scheduled_ = false;
28 
29   // Previous compaction may have produced too many files in a level,
30   // so reschedule another compaction if needed.
31   MaybeScheduleCompaction();
32   bg_cv_.SignalAll();
33 }

實際進行compact的函數是 void DBImpl::BackgroundCompaction() 

1 手動觸發狀況下 會填寫class DBImpl下的一個變量 ManualCompaction manual_compaction_

1 struct ManualCompaction {
2     int level;
3     bool done;
4     const InternalKey* begin;   // NULL means beginning of key range
5     const InternalKey* end;     // NULL means end of key range
6     InternalKey tmp_storage;    // Used to keep track of compaction progress
7   };

 

 

爲了不外部指定的 key-range 過大，一次 compact 過多的 sstable 文件， manual_compaction 可能不會一次作完，因此有 done 來標

識是否已經所有完成， tmp_storage 保存上一次 compact 到的 end-key，即下一次的 startkey。
指定的beg end KEY會賦值到 versions_中，以便後面進行compact。  versions_->CompactRange(m->level, m->begin, m->end);

2 經過 versions_->PickCompaction() 選擇須要compact的level 和 key range

 1 Compaction* VersionSet::PickCompaction() {
 2   Compaction* c;
 3   int level;
 4 
 5   // We prefer compactions triggered by too much data in a level over
 6   // the compactions triggered by seeks.
 7   const bool size_compaction = (current_->compaction_score_ >= 1);
 8   const bool seek_compaction = (current_->file_to_compact_ != NULL);
 9   if (size_compaction) {
10     level = current_->compaction_level_;
11     assert(level >= 0);
12     assert(level+1 < config::kNumLevels);
13     c = new Compaction(level);
14 
15     // Pick the first file that comes after compact_pointer_[level]
16     for (size_t i = 0; i < current_->files_[level].size(); i++) {
17       FileMetaData* f = current_->files_[level][i];
18       if (compact_pointer_[level].empty() ||
19           icmp_.Compare(f->largest.Encode(), compact_pointer_[level]) > 0) {
20         c->inputs_[0].push_back(f);
21         break;
22       }
23     }
24     if (c->inputs_[0].empty()) {
25       // Wrap-around to the beginning of the key space
26       c->inputs_[0].push_back(current_->files_[level][0]);
27     }
28   } else if (seek_compaction) {
29     level = current_->file_to_compact_level_;
30     c = new Compaction(level);
31     c->inputs_[0].push_back(current_->file_to_compact_);
32   } else {
33     return NULL;
34   }
35 
36   c->input_version_ = current_;
37   c->input_version_->Ref();
38 
39   // Files in level 0 may overlap each other, so pick up all overlapping ones
40   if (level == 0) {
41     InternalKey smallest, largest;
42     GetRange(c->inputs_[0], &smallest, &largest);
43     // Note that the next call will discard the file we placed in
44     // c->inputs_[0] earlier and replace it with an overlapping set
45     // which will include the picked file.
46     current_->GetOverlappingInputs(0, &smallest, &largest, &c->inputs_[0]);
47     assert(!c->inputs_[0].empty());
48   }
49 
50   SetupOtherInputs(c);
51 
52   return c;
53 }

PickCompaction函數中 根據 文件尺寸和被seek屢次 來確認compact的文件

使用 c = new Compaction(level)   記錄要compact的level和文件指針

 

 todo  實際的compact操做  CompactMemTable()  DoCompactionWork()

 

 

參考

《leveldb實現解析》 淘寶 那巖