1 // Copyright (c) 2011 The LevelDB Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style license that can be
3 // found in the LICENSE file. See the AUTHORS file for names of contributors.
5 #include "db/version_set.h"
9 #include "db/filename.h"
10 #include "db/log_reader.h"
11 #include "db/log_writer.h"
12 #include "db/memtable.h"
13 #include "db/table_cache.h"
14 #include "leveldb/env.h"
15 #include "leveldb/table_builder.h"
16 #include "table/merger.h"
17 #include "table/two_level_iterator.h"
18 #include "util/coding.h"
19 #include "util/logging.h"
23 static const int kTargetFileSize = 2 * 1048576;
25 // Maximum bytes of overlaps in grandparent (i.e., level+2) before we
26 // stop building a single file in a level->level+1 compaction.
27 static const int64_t kMaxGrandParentOverlapBytes = 10 * kTargetFileSize;
29 // Maximum number of bytes in all compacted files. We avoid expanding
30 // the lower level file set of a compaction if it would make the
31 // total compaction cover more than this many bytes.
32 static const int64_t kExpandedCompactionByteSizeLimit = 25 * kTargetFileSize;
34 static double MaxBytesForLevel(int level) {
35 // Note: the result for level zero is not really used since we set
36 // the level-0 compaction threshold based on number of files.
37 double result = 10 * 1048576.0; // Result for both level-0 and level-1
45 static uint64_t MaxFileSizeForLevel(int level) {
46 return kTargetFileSize; // We could vary per level to reduce number of files?
49 static int64_t TotalFileSize(const std::vector<FileMetaData*>& files) {
51 for (size_t i = 0; i < files.size(); i++) {
52 sum += files[i]->file_size;
58 std::string IntSetToString(const std::set<uint64_t>& s) {
59 std::string result = "{";
60 for (std::set<uint64_t>::const_iterator it = s.begin();
63 result += (result.size() > 1) ? "," : "";
64 result += NumberToString(*it);
74 // Remove from linked list
78 // Drop references to files
79 for (int level = 0; level < config::kNumLevels; level++) {
80 for (size_t i = 0; i < files_[level].size(); i++) {
81 FileMetaData* f = files_[level][i];
91 int FindFile(const InternalKeyComparator& icmp,
92 const std::vector<FileMetaData*>& files,
95 uint32_t right = files.size();
96 while (left < right) {
97 uint32_t mid = (left + right) / 2;
98 const FileMetaData* f = files[mid];
99 if (icmp.InternalKeyComparator::Compare(f->largest.Encode(), key) < 0) {
100 // Key at "mid.largest" is < "target". Therefore all
101 // files at or before "mid" are uninteresting.
104 // Key at "mid.largest" is >= "target". Therefore all files
105 // after "mid" are uninteresting.
112 static bool AfterFile(const Comparator* ucmp,
113 const Slice* user_key, const FileMetaData* f) {
114 // NULL user_key occurs before all keys and is therefore never after *f
115 return (user_key != NULL &&
116 ucmp->Compare(*user_key, f->largest.user_key()) > 0);
119 static bool BeforeFile(const Comparator* ucmp,
120 const Slice* user_key, const FileMetaData* f) {
121 // NULL user_key occurs after all keys and is therefore never before *f
122 return (user_key != NULL &&
123 ucmp->Compare(*user_key, f->smallest.user_key()) < 0);
126 bool SomeFileOverlapsRange(
127 const InternalKeyComparator& icmp,
128 bool disjoint_sorted_files,
129 const std::vector<FileMetaData*>& files,
130 const Slice* smallest_user_key,
131 const Slice* largest_user_key) {
132 const Comparator* ucmp = icmp.user_comparator();
133 if (!disjoint_sorted_files) {
134 // Need to check against all files
135 for (int i = 0; i < files.size(); i++) {
136 const FileMetaData* f = files[i];
137 if (AfterFile(ucmp, smallest_user_key, f) ||
138 BeforeFile(ucmp, largest_user_key, f)) {
141 return true; // Overlap
147 // Binary search over file list
149 if (smallest_user_key != NULL) {
150 // Find the earliest possible internal key for smallest_user_key
151 InternalKey small(*smallest_user_key, kMaxSequenceNumber,kValueTypeForSeek);
152 index = FindFile(icmp, files, small.Encode());
155 if (index >= files.size()) {
156 // beginning of range is after all files, so no overlap.
160 return !BeforeFile(ucmp, largest_user_key, files[index]);
163 // An internal iterator. For a given version/level pair, yields
164 // information about the files in the level. For a given entry, key()
165 // is the largest key that occurs in the file, and value() is an
166 // 16-byte value containing the file number and file size, both
167 // encoded using EncodeFixed64.
168 class Version::LevelFileNumIterator : public Iterator {
170 LevelFileNumIterator(const InternalKeyComparator& icmp,
171 const std::vector<FileMetaData*>* flist)
174 index_(flist->size()) { // Marks as invalid
176 virtual bool Valid() const {
177 return index_ < flist_->size();
179 virtual void Seek(const Slice& target) {
180 index_ = FindFile(icmp_, *flist_, target);
182 virtual void SeekToFirst() { index_ = 0; }
183 virtual void SeekToLast() {
184 index_ = flist_->empty() ? 0 : flist_->size() - 1;
186 virtual void Next() {
190 virtual void Prev() {
193 index_ = flist_->size(); // Marks as invalid
200 return (*flist_)[index_]->largest.Encode();
202 Slice value() const {
204 EncodeFixed64(value_buf_, (*flist_)[index_]->number);
205 EncodeFixed64(value_buf_+8, (*flist_)[index_]->file_size);
206 return Slice(value_buf_, sizeof(value_buf_));
208 virtual Status status() const { return Status::OK(); }
210 const InternalKeyComparator icmp_;
211 const std::vector<FileMetaData*>* const flist_;
214 // Backing store for value(). Holds the file number and size.
215 mutable char value_buf_[16];
218 static Iterator* GetFileIterator(void* arg,
219 const ReadOptions& options,
220 const Slice& file_value) {
221 TableCache* cache = reinterpret_cast<TableCache*>(arg);
222 if (file_value.size() != 16) {
223 return NewErrorIterator(
224 Status::Corruption("FileReader invoked with unexpected value"));
226 return cache->NewIterator(options,
227 DecodeFixed64(file_value.data()),
228 DecodeFixed64(file_value.data() + 8));
232 Iterator* Version::NewConcatenatingIterator(const ReadOptions& options,
234 return NewTwoLevelIterator(
235 new LevelFileNumIterator(vset_->icmp_, &files_[level]),
236 &GetFileIterator, vset_->table_cache_, options);
239 void Version::AddIterators(const ReadOptions& options,
240 std::vector<Iterator*>* iters) {
241 // Merge all level zero files together since they may overlap
242 for (size_t i = 0; i < files_[0].size(); i++) {
244 vset_->table_cache_->NewIterator(
245 options, files_[0][i]->number, files_[0][i]->file_size));
248 // For levels > 0, we can use a concatenating iterator that sequentially
249 // walks through the non-overlapping files in the level, opening them
251 for (int level = 1; level < config::kNumLevels; level++) {
252 if (!files_[level].empty()) {
253 iters->push_back(NewConcatenatingIterator(options, level));
258 // If "*iter" points at a value or deletion for user_key, store
259 // either the value, or a NotFound error and return true.
260 // Else return false.
261 static bool GetValue(const Comparator* cmp,
262 Iterator* iter, const Slice& user_key,
265 if (!iter->Valid()) {
268 ParsedInternalKey parsed_key;
269 if (!ParseInternalKey(iter->key(), &parsed_key)) {
270 *s = Status::Corruption("corrupted key for ", user_key);
273 if (cmp->Compare(parsed_key.user_key, user_key) != 0) {
276 switch (parsed_key.type) {
278 *s = Status::NotFound(Slice()); // Use an empty error message for speed
281 Slice v = iter->value();
282 value->assign(v.data(), v.size());
289 static bool NewestFirst(FileMetaData* a, FileMetaData* b) {
290 return a->number > b->number;
293 Status Version::Get(const ReadOptions& options,
297 Slice ikey = k.internal_key();
298 Slice user_key = k.user_key();
299 const Comparator* ucmp = vset_->icmp_.user_comparator();
302 stats->seek_file = NULL;
303 stats->seek_file_level = -1;
304 FileMetaData* last_file_read = NULL;
305 int last_file_read_level = -1;
307 // We can search level-by-level since entries never hop across
308 // levels. Therefore we are guaranteed that if we find data
309 // in an smaller level, later levels are irrelevant.
310 std::vector<FileMetaData*> tmp;
312 for (int level = 0; level < config::kNumLevels; level++) {
313 size_t num_files = files_[level].size();
314 if (num_files == 0) continue;
316 // Get the list of files to search in this level
317 FileMetaData* const* files = &files_[level][0];
319 // Level-0 files may overlap each other. Find all files that
320 // overlap user_key and process them in order from newest to oldest.
321 tmp.reserve(num_files);
322 for (uint32_t i = 0; i < num_files; i++) {
323 FileMetaData* f = files[i];
324 if (ucmp->Compare(user_key, f->smallest.user_key()) >= 0 &&
325 ucmp->Compare(user_key, f->largest.user_key()) <= 0) {
329 if (tmp.empty()) continue;
331 std::sort(tmp.begin(), tmp.end(), NewestFirst);
333 num_files = tmp.size();
335 // Binary search to find earliest index whose largest key >= ikey.
336 uint32_t index = FindFile(vset_->icmp_, files_[level], ikey);
337 if (index >= num_files) {
342 if (ucmp->Compare(user_key, tmp2->smallest.user_key()) < 0) {
343 // All of "tmp2" is past any data for user_key
353 for (uint32_t i = 0; i < num_files; ++i) {
354 if (last_file_read != NULL && stats->seek_file == NULL) {
355 // We have had more than one seek for this read. Charge the 1st file.
356 stats->seek_file = last_file_read;
357 stats->seek_file_level = last_file_read_level;
360 FileMetaData* f = files[i];
362 last_file_read_level = level;
364 Iterator* iter = vset_->table_cache_->NewIterator(
369 const bool done = GetValue(ucmp, iter, user_key, value, &s);
370 if (!iter->status().ok()) {
383 return Status::NotFound(Slice()); // Use an empty error message for speed
386 bool Version::UpdateStats(const GetStats& stats) {
387 FileMetaData* f = stats.seek_file;
390 if (f->allowed_seeks <= 0 && file_to_compact_ == NULL) {
391 file_to_compact_ = f;
392 file_to_compact_level_ = stats.seek_file_level;
399 void Version::Ref() {
403 void Version::Unref() {
404 assert(this != &vset_->dummy_versions_);
412 bool Version::OverlapInLevel(int level,
413 const Slice* smallest_user_key,
414 const Slice* largest_user_key) {
415 return SomeFileOverlapsRange(vset_->icmp_, (level > 0), files_[level],
416 smallest_user_key, largest_user_key);
419 int Version::PickLevelForMemTableOutput(
420 const Slice& smallest_user_key,
421 const Slice& largest_user_key) {
423 if (!OverlapInLevel(0, &smallest_user_key, &largest_user_key)) {
424 // Push to next level if there is no overlap in next level,
425 // and the #bytes overlapping in the level after that are limited.
426 InternalKey start(smallest_user_key, kMaxSequenceNumber, kValueTypeForSeek);
427 InternalKey limit(largest_user_key, 0, static_cast<ValueType>(0));
428 std::vector<FileMetaData*> overlaps;
429 while (level < config::kMaxMemCompactLevel) {
430 if (OverlapInLevel(level + 1, &smallest_user_key, &largest_user_key)) {
433 GetOverlappingInputs(level + 2, &start, &limit, &overlaps);
434 const int64_t sum = TotalFileSize(overlaps);
435 if (sum > kMaxGrandParentOverlapBytes) {
444 // Store in "*inputs" all files in "level" that overlap [begin,end]
445 void Version::GetOverlappingInputs(
447 const InternalKey* begin,
448 const InternalKey* end,
449 std::vector<FileMetaData*>* inputs) {
451 Slice user_begin, user_end;
453 user_begin = begin->user_key();
456 user_end = end->user_key();
458 const Comparator* user_cmp = vset_->icmp_.user_comparator();
459 for (size_t i = 0; i < files_[level].size(); ) {
460 FileMetaData* f = files_[level][i++];
461 const Slice file_start = f->smallest.user_key();
462 const Slice file_limit = f->largest.user_key();
463 if (begin != NULL && user_cmp->Compare(file_limit, user_begin) < 0) {
464 // "f" is completely before specified range; skip it
465 } else if (end != NULL && user_cmp->Compare(file_start, user_end) > 0) {
466 // "f" is completely after specified range; skip it
468 inputs->push_back(f);
470 // Level-0 files may overlap each other. So check if the newly
471 // added file has expanded the range. If so, restart search.
472 if (begin != NULL && user_cmp->Compare(file_start, user_begin) < 0) {
473 user_begin = file_start;
476 } else if (end != NULL && user_cmp->Compare(file_limit, user_end) > 0) {
477 user_end = file_limit;
486 std::string Version::DebugString() const {
488 for (int level = 0; level < config::kNumLevels; level++) {
491 // 17:123['a' .. 'd']
493 r.append("--- level ");
494 AppendNumberTo(&r, level);
496 const std::vector<FileMetaData*>& files = files_[level];
497 for (size_t i = 0; i < files.size(); i++) {
499 AppendNumberTo(&r, files[i]->number);
501 AppendNumberTo(&r, files[i]->file_size);
503 r.append(files[i]->smallest.DebugString());
505 r.append(files[i]->largest.DebugString());
512 // A helper class so we can efficiently apply a whole sequence
513 // of edits to a particular state without creating intermediate
514 // Versions that contain full copies of the intermediate state.
515 class VersionSet::Builder {
517 // Helper to sort by v->files_[file_number].smallest
518 struct BySmallestKey {
519 const InternalKeyComparator* internal_comparator;
521 bool operator()(FileMetaData* f1, FileMetaData* f2) const {
522 int r = internal_comparator->Compare(f1->smallest, f2->smallest);
526 // Break ties by file number
527 return (f1->number < f2->number);
532 typedef std::set<FileMetaData*, BySmallestKey> FileSet;
534 std::set<uint64_t> deleted_files;
535 FileSet* added_files;
540 LevelState levels_[config::kNumLevels];
543 // Initialize a builder with the files from *base and other info from *vset
544 Builder(VersionSet* vset, Version* base)
549 cmp.internal_comparator = &vset_->icmp_;
550 for (int level = 0; level < config::kNumLevels; level++) {
551 levels_[level].added_files = new FileSet(cmp);
556 for (int level = 0; level < config::kNumLevels; level++) {
557 const FileSet* added = levels_[level].added_files;
558 std::vector<FileMetaData*> to_unref;
559 to_unref.reserve(added->size());
560 for (FileSet::const_iterator it = added->begin();
561 it != added->end(); ++it) {
562 to_unref.push_back(*it);
565 for (uint32_t i = 0; i < to_unref.size(); i++) {
566 FileMetaData* f = to_unref[i];
576 // Apply all of the edits in *edit to the current state.
577 void Apply(VersionEdit* edit) {
578 // Update compaction pointers
579 for (size_t i = 0; i < edit->compact_pointers_.size(); i++) {
580 const int level = edit->compact_pointers_[i].first;
581 vset_->compact_pointer_[level] =
582 edit->compact_pointers_[i].second.Encode().ToString();
586 const VersionEdit::DeletedFileSet& del = edit->deleted_files_;
587 for (VersionEdit::DeletedFileSet::const_iterator iter = del.begin();
590 const int level = iter->first;
591 const uint64_t number = iter->second;
592 levels_[level].deleted_files.insert(number);
596 for (size_t i = 0; i < edit->new_files_.size(); i++) {
597 const int level = edit->new_files_[i].first;
598 FileMetaData* f = new FileMetaData(edit->new_files_[i].second);
601 // We arrange to automatically compact this file after
602 // a certain number of seeks. Let's assume:
603 // (1) One seek costs 10ms
604 // (2) Writing or reading 1MB costs 10ms (100MB/s)
605 // (3) A compaction of 1MB does 25MB of IO:
606 // 1MB read from this level
607 // 10-12MB read from next level (boundaries may be misaligned)
608 // 10-12MB written to next level
609 // This implies that 25 seeks cost the same as the compaction
610 // of 1MB of data. I.e., one seek costs approximately the
611 // same as the compaction of 40KB of data. We are a little
612 // conservative and allow approximately one seek for every 16KB
613 // of data before triggering a compaction.
614 f->allowed_seeks = (f->file_size / 16384);
615 if (f->allowed_seeks < 100) f->allowed_seeks = 100;
617 levels_[level].deleted_files.erase(f->number);
618 levels_[level].added_files->insert(f);
622 // Save the current state in *v.
623 void SaveTo(Version* v) {
625 cmp.internal_comparator = &vset_->icmp_;
626 for (int level = 0; level < config::kNumLevels; level++) {
627 // Merge the set of added files with the set of pre-existing files.
628 // Drop any deleted files. Store the result in *v.
629 const std::vector<FileMetaData*>& base_files = base_->files_[level];
630 std::vector<FileMetaData*>::const_iterator base_iter = base_files.begin();
631 std::vector<FileMetaData*>::const_iterator base_end = base_files.end();
632 const FileSet* added = levels_[level].added_files;
633 v->files_[level].reserve(base_files.size() + added->size());
634 for (FileSet::const_iterator added_iter = added->begin();
635 added_iter != added->end();
637 // Add all smaller files listed in base_
638 for (std::vector<FileMetaData*>::const_iterator bpos
639 = std::upper_bound(base_iter, base_end, *added_iter, cmp);
642 MaybeAddFile(v, level, *base_iter);
645 MaybeAddFile(v, level, *added_iter);
648 // Add remaining base files
649 for (; base_iter != base_end; ++base_iter) {
650 MaybeAddFile(v, level, *base_iter);
654 // Make sure there is no overlap in levels > 0
656 for (uint32_t i = 1; i < v->files_[level].size(); i++) {
657 const InternalKey& prev_end = v->files_[level][i-1]->largest;
658 const InternalKey& this_begin = v->files_[level][i]->smallest;
659 if (vset_->icmp_.Compare(prev_end, this_begin) >= 0) {
660 fprintf(stderr, "overlapping ranges in same level %s vs. %s\n",
661 prev_end.DebugString().c_str(),
662 this_begin.DebugString().c_str());
671 void MaybeAddFile(Version* v, int level, FileMetaData* f) {
672 if (levels_[level].deleted_files.count(f->number) > 0) {
673 // File is deleted: do nothing
675 std::vector<FileMetaData*>* files = &v->files_[level];
676 if (level > 0 && !files->empty()) {
678 assert(vset_->icmp_.Compare((*files)[files->size()-1]->largest,
687 VersionSet::VersionSet(const std::string& dbname,
688 const Options* options,
689 TableCache* table_cache,
690 const InternalKeyComparator* cmp)
691 : env_(options->env),
694 table_cache_(table_cache),
696 next_file_number_(2),
697 manifest_file_number_(0), // Filled by Recover()
701 descriptor_file_(NULL),
702 descriptor_log_(NULL),
703 dummy_versions_(this),
705 AppendVersion(new Version(this));
708 VersionSet::~VersionSet() {
710 assert(dummy_versions_.next_ == &dummy_versions_); // List must be empty
711 delete descriptor_log_;
712 delete descriptor_file_;
715 void VersionSet::AppendVersion(Version* v) {
717 assert(v->refs_ == 0);
718 assert(v != current_);
719 if (current_ != NULL) {
725 // Append to linked list
726 v->prev_ = dummy_versions_.prev_;
727 v->next_ = &dummy_versions_;
732 Status VersionSet::LogAndApply(VersionEdit* edit, port::Mutex* mu) {
733 if (edit->has_log_number_) {
734 assert(edit->log_number_ >= log_number_);
735 assert(edit->log_number_ < next_file_number_);
737 edit->SetLogNumber(log_number_);
740 if (!edit->has_prev_log_number_) {
741 edit->SetPrevLogNumber(prev_log_number_);
744 edit->SetNextFile(next_file_number_);
745 edit->SetLastSequence(last_sequence_);
747 Version* v = new Version(this);
749 Builder builder(this, current_);
755 // Initialize new descriptor log file if necessary by creating
756 // a temporary file that contains a snapshot of the current version.
757 std::string new_manifest_file;
759 if (descriptor_log_ == NULL) {
760 // No reason to unlock *mu here since we only hit this path in the
761 // first call to LogAndApply (when opening the database).
762 assert(descriptor_file_ == NULL);
763 new_manifest_file = DescriptorFileName(dbname_, manifest_file_number_);
764 edit->SetNextFile(next_file_number_);
765 s = env_->NewWritableFile(new_manifest_file, &descriptor_file_);
767 descriptor_log_ = new log::Writer(descriptor_file_);
768 s = WriteSnapshot(descriptor_log_);
772 // Unlock during expensive MANIFEST log write
776 // Write new record to MANIFEST log
779 edit->EncodeTo(&record);
780 s = descriptor_log_->AddRecord(record);
782 s = descriptor_file_->Sync();
786 // If we just created a new descriptor file, install it by writing a
787 // new CURRENT file that points to it.
788 if (s.ok() && !new_manifest_file.empty()) {
789 s = SetCurrentFile(env_, dbname_, manifest_file_number_);
795 // Install the new version
798 log_number_ = edit->log_number_;
799 prev_log_number_ = edit->prev_log_number_;
802 if (!new_manifest_file.empty()) {
803 delete descriptor_log_;
804 delete descriptor_file_;
805 descriptor_log_ = NULL;
806 descriptor_file_ = NULL;
807 env_->DeleteFile(new_manifest_file);
814 Status VersionSet::Recover() {
815 struct LogReporter : public log::Reader::Reporter {
817 virtual void Corruption(size_t bytes, const Status& s) {
818 if (this->status->ok()) *this->status = s;
822 // Read "CURRENT" file, which contains a pointer to the current manifest file
824 Status s = ReadFileToString(env_, CurrentFileName(dbname_), ¤t);
828 if (current.empty() || current[current.size()-1] != '\n') {
829 return Status::Corruption("CURRENT file does not end with newline");
831 current.resize(current.size() - 1);
833 std::string dscname = dbname_ + "/" + current;
834 SequentialFile* file;
835 s = env_->NewSequentialFile(dscname, &file);
840 bool have_log_number = false;
841 bool have_prev_log_number = false;
842 bool have_next_file = false;
843 bool have_last_sequence = false;
844 uint64_t next_file = 0;
845 uint64_t last_sequence = 0;
846 uint64_t log_number = 0;
847 uint64_t prev_log_number = 0;
848 Builder builder(this, current_);
851 LogReporter reporter;
852 reporter.status = &s;
853 log::Reader reader(file, &reporter, true/*checksum*/, 0/*initial_offset*/);
856 while (reader.ReadRecord(&record, &scratch) && s.ok()) {
858 s = edit.DecodeFrom(record);
860 if (edit.has_comparator_ &&
861 edit.comparator_ != icmp_.user_comparator()->Name()) {
862 s = Status::InvalidArgument(
863 edit.comparator_ + "does not match existing comparator ",
864 icmp_.user_comparator()->Name());
869 builder.Apply(&edit);
872 if (edit.has_log_number_) {
873 log_number = edit.log_number_;
874 have_log_number = true;
877 if (edit.has_prev_log_number_) {
878 prev_log_number = edit.prev_log_number_;
879 have_prev_log_number = true;
882 if (edit.has_next_file_number_) {
883 next_file = edit.next_file_number_;
884 have_next_file = true;
887 if (edit.has_last_sequence_) {
888 last_sequence = edit.last_sequence_;
889 have_last_sequence = true;
897 if (!have_next_file) {
898 s = Status::Corruption("no meta-nextfile entry in descriptor");
899 } else if (!have_log_number) {
900 s = Status::Corruption("no meta-lognumber entry in descriptor");
901 } else if (!have_last_sequence) {
902 s = Status::Corruption("no last-sequence-number entry in descriptor");
905 if (!have_prev_log_number) {
909 MarkFileNumberUsed(prev_log_number);
910 MarkFileNumberUsed(log_number);
914 Version* v = new Version(this);
916 // Install recovered version
919 manifest_file_number_ = next_file;
920 next_file_number_ = next_file + 1;
921 last_sequence_ = last_sequence;
922 log_number_ = log_number;
923 prev_log_number_ = prev_log_number;
929 void VersionSet::MarkFileNumberUsed(uint64_t number) {
930 if (next_file_number_ <= number) {
931 next_file_number_ = number + 1;
935 void VersionSet::Finalize(Version* v) {
936 // Precomputed best level for next compaction
938 double best_score = -1;
940 for (int level = 0; level < config::kNumLevels-1; level++) {
943 // We treat level-0 specially by bounding the number of files
944 // instead of number of bytes for two reasons:
946 // (1) With larger write-buffer sizes, it is nice not to do too
947 // many level-0 compactions.
949 // (2) The files in level-0 are merged on every read and
950 // therefore we wish to avoid too many files when the individual
951 // file size is small (perhaps because of a small write-buffer
952 // setting, or very high compression ratios, or lots of
953 // overwrites/deletions).
954 score = v->files_[level].size() /
955 static_cast<double>(config::kL0_CompactionTrigger);
957 // Compute the ratio of current size to size limit.
958 const uint64_t level_bytes = TotalFileSize(v->files_[level]);
959 score = static_cast<double>(level_bytes) / MaxBytesForLevel(level);
962 if (score > best_score) {
968 v->compaction_level_ = best_level;
969 v->compaction_score_ = best_score;
972 Status VersionSet::WriteSnapshot(log::Writer* log) {
973 // TODO: Break up into multiple records to reduce memory usage on recovery?
977 edit.SetComparatorName(icmp_.user_comparator()->Name());
979 // Save compaction pointers
980 for (int level = 0; level < config::kNumLevels; level++) {
981 if (!compact_pointer_[level].empty()) {
983 key.DecodeFrom(compact_pointer_[level]);
984 edit.SetCompactPointer(level, key);
989 for (int level = 0; level < config::kNumLevels; level++) {
990 const std::vector<FileMetaData*>& files = current_->files_[level];
991 for (size_t i = 0; i < files.size(); i++) {
992 const FileMetaData* f = files[i];
993 edit.AddFile(level, f->number, f->file_size, f->smallest, f->largest);
998 edit.EncodeTo(&record);
999 return log->AddRecord(record);
1002 int VersionSet::NumLevelFiles(int level) const {
1004 assert(level < config::kNumLevels);
1005 return current_->files_[level].size();
1008 const char* VersionSet::LevelSummary(LevelSummaryStorage* scratch) const {
1009 // Update code if kNumLevels changes
1010 assert(config::kNumLevels == 7);
1011 snprintf(scratch->buffer, sizeof(scratch->buffer),
1012 "files[ %d %d %d %d %d %d %d ]",
1013 int(current_->files_[0].size()),
1014 int(current_->files_[1].size()),
1015 int(current_->files_[2].size()),
1016 int(current_->files_[3].size()),
1017 int(current_->files_[4].size()),
1018 int(current_->files_[5].size()),
1019 int(current_->files_[6].size()));
1020 return scratch->buffer;
1023 uint64_t VersionSet::ApproximateOffsetOf(Version* v, const InternalKey& ikey) {
1024 uint64_t result = 0;
1025 for (int level = 0; level < config::kNumLevels; level++) {
1026 const std::vector<FileMetaData*>& files = v->files_[level];
1027 for (size_t i = 0; i < files.size(); i++) {
1028 if (icmp_.Compare(files[i]->largest, ikey) <= 0) {
1029 // Entire file is before "ikey", so just add the file size
1030 result += files[i]->file_size;
1031 } else if (icmp_.Compare(files[i]->smallest, ikey) > 0) {
1032 // Entire file is after "ikey", so ignore
1034 // Files other than level 0 are sorted by meta->smallest, so
1035 // no further files in this level will contain data for
1040 // "ikey" falls in the range for this table. Add the
1041 // approximate offset of "ikey" within the table.
1043 Iterator* iter = table_cache_->NewIterator(
1044 ReadOptions(), files[i]->number, files[i]->file_size, &tableptr);
1045 if (tableptr != NULL) {
1046 result += tableptr->ApproximateOffsetOf(ikey.Encode());
1055 void VersionSet::AddLiveFiles(std::set<uint64_t>* live) {
1056 for (Version* v = dummy_versions_.next_;
1057 v != &dummy_versions_;
1059 for (int level = 0; level < config::kNumLevels; level++) {
1060 const std::vector<FileMetaData*>& files = v->files_[level];
1061 for (size_t i = 0; i < files.size(); i++) {
1062 live->insert(files[i]->number);
1068 int64_t VersionSet::NumLevelBytes(int level) const {
1070 assert(level < config::kNumLevels);
1071 return TotalFileSize(current_->files_[level]);
1074 int64_t VersionSet::MaxNextLevelOverlappingBytes() {
1076 std::vector<FileMetaData*> overlaps;
1077 for (int level = 1; level < config::kNumLevels - 1; level++) {
1078 for (size_t i = 0; i < current_->files_[level].size(); i++) {
1079 const FileMetaData* f = current_->files_[level][i];
1080 current_->GetOverlappingInputs(level+1, &f->smallest, &f->largest,
1082 const int64_t sum = TotalFileSize(overlaps);
1091 // Stores the minimal range that covers all entries in inputs in
1092 // *smallest, *largest.
1093 // REQUIRES: inputs is not empty
1094 void VersionSet::GetRange(const std::vector<FileMetaData*>& inputs,
1095 InternalKey* smallest,
1096 InternalKey* largest) {
1097 assert(!inputs.empty());
1100 for (size_t i = 0; i < inputs.size(); i++) {
1101 FileMetaData* f = inputs[i];
1103 *smallest = f->smallest;
1104 *largest = f->largest;
1106 if (icmp_.Compare(f->smallest, *smallest) < 0) {
1107 *smallest = f->smallest;
1109 if (icmp_.Compare(f->largest, *largest) > 0) {
1110 *largest = f->largest;
1116 // Stores the minimal range that covers all entries in inputs1 and inputs2
1117 // in *smallest, *largest.
1118 // REQUIRES: inputs is not empty
1119 void VersionSet::GetRange2(const std::vector<FileMetaData*>& inputs1,
1120 const std::vector<FileMetaData*>& inputs2,
1121 InternalKey* smallest,
1122 InternalKey* largest) {
1123 std::vector<FileMetaData*> all = inputs1;
1124 all.insert(all.end(), inputs2.begin(), inputs2.end());
1125 GetRange(all, smallest, largest);
1128 Iterator* VersionSet::MakeInputIterator(Compaction* c) {
1129 ReadOptions options;
1130 options.verify_checksums = options_->paranoid_checks;
1131 options.fill_cache = false;
1133 // Level-0 files have to be merged together. For other levels,
1134 // we will make a concatenating iterator per level.
1135 // TODO(opt): use concatenating iterator for level-0 if there is no overlap
1136 const int space = (c->level() == 0 ? c->inputs_[0].size() + 1 : 2);
1137 Iterator** list = new Iterator*[space];
1139 for (int which = 0; which < 2; which++) {
1140 if (!c->inputs_[which].empty()) {
1141 if (c->level() + which == 0) {
1142 const std::vector<FileMetaData*>& files = c->inputs_[which];
1143 for (size_t i = 0; i < files.size(); i++) {
1144 list[num++] = table_cache_->NewIterator(
1145 options, files[i]->number, files[i]->file_size);
1148 // Create concatenating iterator for the files from this level
1149 list[num++] = NewTwoLevelIterator(
1150 new Version::LevelFileNumIterator(icmp_, &c->inputs_[which]),
1151 &GetFileIterator, table_cache_, options);
1155 assert(num <= space);
1156 Iterator* result = NewMergingIterator(&icmp_, list, num);
1161 Compaction* VersionSet::PickCompaction() {
1165 // We prefer compactions triggered by too much data in a level over
1166 // the compactions triggered by seeks.
1167 const bool size_compaction = (current_->compaction_score_ >= 1);
1168 const bool seek_compaction = (current_->file_to_compact_ != NULL);
1169 if (size_compaction) {
1170 level = current_->compaction_level_;
1172 assert(level+1 < config::kNumLevels);
1173 c = new Compaction(level);
1175 // Pick the first file that comes after compact_pointer_[level]
1176 for (size_t i = 0; i < current_->files_[level].size(); i++) {
1177 FileMetaData* f = current_->files_[level][i];
1178 if (compact_pointer_[level].empty() ||
1179 icmp_.Compare(f->largest.Encode(), compact_pointer_[level]) > 0) {
1180 c->inputs_[0].push_back(f);
1184 if (c->inputs_[0].empty()) {
1185 // Wrap-around to the beginning of the key space
1186 c->inputs_[0].push_back(current_->files_[level][0]);
1188 } else if (seek_compaction) {
1189 level = current_->file_to_compact_level_;
1190 c = new Compaction(level);
1191 c->inputs_[0].push_back(current_->file_to_compact_);
1196 c->input_version_ = current_;
1197 c->input_version_->Ref();
1199 // Files in level 0 may overlap each other, so pick up all overlapping ones
1201 InternalKey smallest, largest;
1202 GetRange(c->inputs_[0], &smallest, &largest);
1203 // Note that the next call will discard the file we placed in
1204 // c->inputs_[0] earlier and replace it with an overlapping set
1205 // which will include the picked file.
1206 current_->GetOverlappingInputs(0, &smallest, &largest, &c->inputs_[0]);
1207 assert(!c->inputs_[0].empty());
1210 SetupOtherInputs(c);
1215 void VersionSet::SetupOtherInputs(Compaction* c) {
1216 const int level = c->level();
1217 InternalKey smallest, largest;
1218 GetRange(c->inputs_[0], &smallest, &largest);
1220 current_->GetOverlappingInputs(level+1, &smallest, &largest, &c->inputs_[1]);
1222 // Get entire range covered by compaction
1223 InternalKey all_start, all_limit;
1224 GetRange2(c->inputs_[0], c->inputs_[1], &all_start, &all_limit);
1226 // See if we can grow the number of inputs in "level" without
1227 // changing the number of "level+1" files we pick up.
1228 if (!c->inputs_[1].empty()) {
1229 std::vector<FileMetaData*> expanded0;
1230 current_->GetOverlappingInputs(level, &all_start, &all_limit, &expanded0);
1231 const int64_t inputs0_size = TotalFileSize(c->inputs_[0]);
1232 const int64_t inputs1_size = TotalFileSize(c->inputs_[1]);
1233 const int64_t expanded0_size = TotalFileSize(expanded0);
1234 if (expanded0.size() > c->inputs_[0].size() &&
1235 inputs1_size + expanded0_size < kExpandedCompactionByteSizeLimit) {
1236 InternalKey new_start, new_limit;
1237 GetRange(expanded0, &new_start, &new_limit);
1238 std::vector<FileMetaData*> expanded1;
1239 current_->GetOverlappingInputs(level+1, &new_start, &new_limit,
1241 if (expanded1.size() == c->inputs_[1].size()) {
1242 Log(options_->info_log,
1243 "Expanding@%d %d+%d (%ld+%ld bytes) to %d+%d (%ld+%ld bytes)\n",
1245 int(c->inputs_[0].size()),
1246 int(c->inputs_[1].size()),
1247 long(inputs0_size), long(inputs1_size),
1248 int(expanded0.size()),
1249 int(expanded1.size()),
1250 long(expanded0_size), long(inputs1_size));
1251 smallest = new_start;
1252 largest = new_limit;
1253 c->inputs_[0] = expanded0;
1254 c->inputs_[1] = expanded1;
1255 GetRange2(c->inputs_[0], c->inputs_[1], &all_start, &all_limit);
1260 // Compute the set of grandparent files that overlap this compaction
1261 // (parent == level+1; grandparent == level+2)
1262 if (level + 2 < config::kNumLevels) {
1263 current_->GetOverlappingInputs(level + 2, &all_start, &all_limit,
1268 Log(options_->info_log, "Compacting %d '%s' .. '%s'",
1270 smallest.DebugString().c_str(),
1271 largest.DebugString().c_str());
1274 // Update the place where we will do the next compaction for this level.
1275 // We update this immediately instead of waiting for the VersionEdit
1276 // to be applied so that if the compaction fails, we will try a different
1277 // key range next time.
1278 compact_pointer_[level] = largest.Encode().ToString();
1279 c->edit_.SetCompactPointer(level, largest);
1282 Compaction* VersionSet::CompactRange(
1284 const InternalKey* begin,
1285 const InternalKey* end) {
1286 std::vector<FileMetaData*> inputs;
1287 current_->GetOverlappingInputs(level, begin, end, &inputs);
1288 if (inputs.empty()) {
1292 // Avoid compacting too much in one shot in case the range is large.
1293 const uint64_t limit = MaxFileSizeForLevel(level);
1295 for (int i = 0; i < inputs.size(); i++) {
1296 uint64_t s = inputs[i]->file_size;
1298 if (total >= limit) {
1299 inputs.resize(i + 1);
1304 Compaction* c = new Compaction(level);
1305 c->input_version_ = current_;
1306 c->input_version_->Ref();
1307 c->inputs_[0] = inputs;
1308 SetupOtherInputs(c);
1312 Compaction::Compaction(int level)
1314 max_output_file_size_(MaxFileSizeForLevel(level)),
1315 input_version_(NULL),
1316 grandparent_index_(0),
1318 overlapped_bytes_(0) {
1319 for (int i = 0; i < config::kNumLevels; i++) {
1324 Compaction::~Compaction() {
1325 if (input_version_ != NULL) {
1326 input_version_->Unref();
1330 bool Compaction::IsTrivialMove() const {
1331 // Avoid a move if there is lots of overlapping grandparent data.
1332 // Otherwise, the move could create a parent file that will require
1333 // a very expensive merge later on.
1334 return (num_input_files(0) == 1 &&
1335 num_input_files(1) == 0 &&
1336 TotalFileSize(grandparents_) <= kMaxGrandParentOverlapBytes);
1339 void Compaction::AddInputDeletions(VersionEdit* edit) {
1340 for (int which = 0; which < 2; which++) {
1341 for (size_t i = 0; i < inputs_[which].size(); i++) {
1342 edit->DeleteFile(level_ + which, inputs_[which][i]->number);
1347 bool Compaction::IsBaseLevelForKey(const Slice& user_key) {
1348 // Maybe use binary search to find right entry instead of linear search?
1349 const Comparator* user_cmp = input_version_->vset_->icmp_.user_comparator();
1350 for (int lvl = level_ + 2; lvl < config::kNumLevels; lvl++) {
1351 const std::vector<FileMetaData*>& files = input_version_->files_[lvl];
1352 for (; level_ptrs_[lvl] < files.size(); ) {
1353 FileMetaData* f = files[level_ptrs_[lvl]];
1354 if (user_cmp->Compare(user_key, f->largest.user_key()) <= 0) {
1355 // We've advanced far enough
1356 if (user_cmp->Compare(user_key, f->smallest.user_key()) >= 0) {
1357 // Key falls in this file's range, so definitely not base level
1368 bool Compaction::ShouldStopBefore(const Slice& internal_key) {
1369 // Scan to find earliest grandparent file that contains key.
1370 const InternalKeyComparator* icmp = &input_version_->vset_->icmp_;
1371 while (grandparent_index_ < grandparents_.size() &&
1372 icmp->Compare(internal_key,
1373 grandparents_[grandparent_index_]->largest.Encode()) > 0) {
1375 overlapped_bytes_ += grandparents_[grandparent_index_]->file_size;
1377 grandparent_index_++;
1381 if (overlapped_bytes_ > kMaxGrandParentOverlapBytes) {
1382 // Too much overlap for current output; start new output
1383 overlapped_bytes_ = 0;
1390 void Compaction::ReleaseInputs() {
1391 if (input_version_ != NULL) {
1392 input_version_->Unref();
1393 input_version_ = NULL;
1397 } // namespace leveldb