Merge in LevelDB 1.19 changes

20 files changed, 1236 insertions, 114 deletions

diff --git a/src/leveldb/db/corruption_test.cc b/src/leveldb/db/corruption_test.cc
index 96afc68913..37a484d25f 100644
--- a/src/leveldb/db/corruption_test.cc
+++ b/src/leveldb/db/corruption_test.cc
@@ -36,7 +36,7 @@ class CorruptionTest {
     tiny_cache_ = NewLRUCache(100);
     options_.env = &env_;
     options_.block_cache = tiny_cache_;
-    dbname_ = test::TmpDir() + "/db_test";
+    dbname_ = test::TmpDir() + "/corruption_test";
     DestroyDB(dbname_, options_);
 
     db_ = NULL;
diff --git a/src/leveldb/db/db_bench.cc b/src/leveldb/db/db_bench.cc
index 705a170aae..7a0f5e08cd 100644
--- a/src/leveldb/db/db_bench.cc
+++ b/src/leveldb/db/db_bench.cc
@@ -33,6 +33,7 @@
 //      readmissing   -- read N missing keys in random order
 //      readhot       -- read N times in random order from 1% section of DB
 //      seekrandom    -- N random seeks
+//      open          -- cost of opening a DB
 //      crc32c        -- repeated crc32c of 4K of data
 //      acquireload   -- load N*1000 times
 //   Meta operations:
@@ -99,6 +100,9 @@ static int FLAGS_bloom_bits = -1;
 // benchmark will fail.
 static bool FLAGS_use_existing_db = false;
 
+// If true, reuse existing log/MANIFEST files when re-opening a database.
+static bool FLAGS_reuse_logs = false;
+
 // Use the db with the following name.
 static const char* FLAGS_db = NULL;
 
@@ -138,6 +142,7 @@ class RandomGenerator {
   }
 };
 
+#if defined(__linux)
 static Slice TrimSpace(Slice s) {
   size_t start = 0;
   while (start < s.size() && isspace(s[start])) {
@@ -149,6 +154,7 @@ static Slice TrimSpace(Slice s) {
   }
   return Slice(s.data() + start, limit - start);
 }
+#endif
 
 static void AppendWithSpace(std::string* str, Slice msg) {
   if (msg.empty()) return;
@@ -442,7 +448,11 @@ class Benchmark {
       bool fresh_db = false;
       int num_threads = FLAGS_threads;
 
-      if (name == Slice("fillseq")) {
+      if (name == Slice("open")) {
+        method = &Benchmark::OpenBench;
+        num_ /= 10000;
+        if (num_ < 1) num_ = 1;
+      } else if (name == Slice("fillseq")) {
         fresh_db = true;
         method = &Benchmark::WriteSeq;
       } else if (name == Slice("fillbatch")) {
@@ -695,6 +705,7 @@ class Benchmark {
     options.write_buffer_size = FLAGS_write_buffer_size;
     options.max_open_files = FLAGS_open_files;
     options.filter_policy = filter_policy_;
+    options.reuse_logs = FLAGS_reuse_logs;
     Status s = DB::Open(options, FLAGS_db, &db_);
     if (!s.ok()) {
       fprintf(stderr, "open error: %s\n", s.ToString().c_str());
@@ -702,6 +713,14 @@ class Benchmark {
     }
   }
 
+  void OpenBench(ThreadState* thread) {
+    for (int i = 0; i < num_; i++) {
+      delete db_;
+      Open();
+      thread->stats.FinishedSingleOp();
+    }
+  }
+
   void WriteSeq(ThreadState* thread) {
     DoWrite(thread, true);
   }
@@ -941,6 +960,9 @@ int main(int argc, char** argv) {
     } else if (sscanf(argv[i], "--use_existing_db=%d%c", &n, &junk) == 1 &&
                (n == 0 || n == 1)) {
       FLAGS_use_existing_db = n;
+    } else if (sscanf(argv[i], "--reuse_logs=%d%c", &n, &junk) == 1 &&
+               (n == 0 || n == 1)) {
+      FLAGS_reuse_logs = n;
     } else if (sscanf(argv[i], "--num=%d%c", &n, &junk) == 1) {
       FLAGS_num = n;
     } else if (sscanf(argv[i], "--reads=%d%c", &n, &junk) == 1) {
diff --git a/src/leveldb/db/db_impl.cc b/src/leveldb/db/db_impl.cc
index 49b95953b4..60f4e66e55 100644
--- a/src/leveldb/db/db_impl.cc
+++ b/src/leveldb/db/db_impl.cc
@@ -125,7 +125,7 @@ DBImpl::DBImpl(const Options& raw_options, const std::string& dbname)
       db_lock_(NULL),
       shutting_down_(NULL),
       bg_cv_(&mutex_),
-      mem_(new MemTable(internal_comparator_)),
+      mem_(NULL),
       imm_(NULL),
       logfile_(NULL),
       logfile_number_(0),
@@ -134,7 +134,6 @@ DBImpl::DBImpl(const Options& raw_options, const std::string& dbname)
       tmp_batch_(new WriteBatch),
       bg_compaction_scheduled_(false),
       manual_compaction_(NULL) {
-  mem_->Ref();
   has_imm_.Release_Store(NULL);
 
   // Reserve ten files or so for other uses and give the rest to TableCache.
@@ -271,7 +270,7 @@ void DBImpl::DeleteObsoleteFiles() {
   }
 }
 
-Status DBImpl::Recover(VersionEdit* edit) {
+Status DBImpl::Recover(VersionEdit* edit, bool *save_manifest) {
   mutex_.AssertHeld();
 
   // Ignore error from CreateDir since the creation of the DB is
@@ -301,66 +300,69 @@ Status DBImpl::Recover(VersionEdit* edit) {
     }
   }
 
-  s = versions_->Recover();
-  if (s.ok()) {
-    SequenceNumber max_sequence(0);
-
-    // Recover from all newer log files than the ones named in the
-    // descriptor (new log files may have been added by the previous
-    // incarnation without registering them in the descriptor).
-    //
-    // Note that PrevLogNumber() is no longer used, but we pay
-    // attention to it in case we are recovering a database
-    // produced by an older version of leveldb.
-    const uint64_t min_log = versions_->LogNumber();
-    const uint64_t prev_log = versions_->PrevLogNumber();
-    std::vector<std::string> filenames;
-    s = env_->GetChildren(dbname_, &filenames);
+  s = versions_->Recover(save_manifest);
+  if (!s.ok()) {
+    return s;
+  }
+  SequenceNumber max_sequence(0);
+
+  // Recover from all newer log files than the ones named in the
+  // descriptor (new log files may have been added by the previous
+  // incarnation without registering them in the descriptor).
+  //
+  // Note that PrevLogNumber() is no longer used, but we pay
+  // attention to it in case we are recovering a database
+  // produced by an older version of leveldb.
+  const uint64_t min_log = versions_->LogNumber();
+  const uint64_t prev_log = versions_->PrevLogNumber();
+  std::vector<std::string> filenames;
+  s = env_->GetChildren(dbname_, &filenames);
+  if (!s.ok()) {
+    return s;
+  }
+  std::set<uint64_t> expected;
+  versions_->AddLiveFiles(&expected);
+  uint64_t number;
+  FileType type;
+  std::vector<uint64_t> logs;
+  for (size_t i = 0; i < filenames.size(); i++) {
+    if (ParseFileName(filenames[i], &number, &type)) {
+      expected.erase(number);
+      if (type == kLogFile && ((number >= min_log) || (number == prev_log)))
+        logs.push_back(number);
+    }
+  }
+  if (!expected.empty()) {
+    char buf[50];
+    snprintf(buf, sizeof(buf), "%d missing files; e.g.",
+             static_cast<int>(expected.size()));
+    return Status::Corruption(buf, TableFileName(dbname_, *(expected.begin())));
+  }
+
+  // Recover in the order in which the logs were generated
+  std::sort(logs.begin(), logs.end());
+  for (size_t i = 0; i < logs.size(); i++) {
+    s = RecoverLogFile(logs[i], (i == logs.size() - 1), save_manifest, edit,
+                       &max_sequence);
     if (!s.ok()) {
       return s;
     }
-    std::set<uint64_t> expected;
-    versions_->AddLiveFiles(&expected);
-    uint64_t number;
-    FileType type;
-    std::vector<uint64_t> logs;
-    for (size_t i = 0; i < filenames.size(); i++) {
-      if (ParseFileName(filenames[i], &number, &type)) {
-        expected.erase(number);
-        if (type == kLogFile && ((number >= min_log) || (number == prev_log)))
-          logs.push_back(number);
-      }
-    }
-    if (!expected.empty()) {
-      char buf[50];
-      snprintf(buf, sizeof(buf), "%d missing files; e.g.",
-               static_cast<int>(expected.size()));
-      return Status::Corruption(buf, TableFileName(dbname_, *(expected.begin())));
-    }
-
-    // Recover in the order in which the logs were generated
-    std::sort(logs.begin(), logs.end());
-    for (size_t i = 0; i < logs.size(); i++) {
-      s = RecoverLogFile(logs[i], edit, &max_sequence);
 
-      // The previous incarnation may not have written any MANIFEST
-      // records after allocating this log number.  So we manually
-      // update the file number allocation counter in VersionSet.
-      versions_->MarkFileNumberUsed(logs[i]);
-    }
+    // The previous incarnation may not have written any MANIFEST
+    // records after allocating this log number.  So we manually
+    // update the file number allocation counter in VersionSet.
+    versions_->MarkFileNumberUsed(logs[i]);
+  }
 
-    if (s.ok()) {
-      if (versions_->LastSequence() < max_sequence) {
-        versions_->SetLastSequence(max_sequence);
-      }
-    }
+  if (versions_->LastSequence() < max_sequence) {
+    versions_->SetLastSequence(max_sequence);
   }
 
-  return s;
+  return Status::OK();
 }
 
-Status DBImpl::RecoverLogFile(uint64_t log_number,
-                              VersionEdit* edit,
+Status DBImpl::RecoverLogFile(uint64_t log_number, bool last_log,
+                              bool* save_manifest, VersionEdit* edit,
                               SequenceNumber* max_sequence) {
   struct LogReporter : public log::Reader::Reporter {
     Env* env;
@@ -405,6 +407,7 @@ Status DBImpl::RecoverLogFile(uint64_t log_number,
   std::string scratch;
   Slice record;
   WriteBatch batch;
+  int compactions = 0;
   MemTable* mem = NULL;
   while (reader.ReadRecord(&record, &scratch) &&
          status.ok()) {
@@ -432,25 +435,52 @@ Status DBImpl::RecoverLogFile(uint64_t log_number,
     }
 
     if (mem->ApproximateMemoryUsage() > options_.write_buffer_size) {
+      compactions++;
+      *save_manifest = true;
       status = WriteLevel0Table(mem, edit, NULL);
+      mem->Unref();
+      mem = NULL;
       if (!status.ok()) {
         // Reflect errors immediately so that conditions like full
         // file-systems cause the DB::Open() to fail.
         break;
       }
-      mem->Unref();
-      mem = NULL;
     }
   }
 
-  if (status.ok() && mem != NULL) {
-    status = WriteLevel0Table(mem, edit, NULL);
-    // Reflect errors immediately so that conditions like full
-    // file-systems cause the DB::Open() to fail.
+  delete file;
+
+  // See if we should keep reusing the last log file.
+  if (status.ok() && options_.reuse_logs && last_log && compactions == 0) {
+    assert(logfile_ == NULL);
+    assert(log_ == NULL);
+    assert(mem_ == NULL);
+    uint64_t lfile_size;
+    if (env_->GetFileSize(fname, &lfile_size).ok() &&
+        env_->NewAppendableFile(fname, &logfile_).ok()) {
+      Log(options_.info_log, "Reusing old log %s \n", fname.c_str());
+      log_ = new log::Writer(logfile_, lfile_size);
+      logfile_number_ = log_number;
+      if (mem != NULL) {
+        mem_ = mem;
+        mem = NULL;
+      } else {
+        // mem can be NULL if lognum exists but was empty.
+        mem_ = new MemTable(internal_comparator_);
+        mem_->Ref();
+      }
+    }
+  }
+
+  if (mem != NULL) {
+    // mem did not get reused; compact it.
+    if (status.ok()) {
+      *save_manifest = true;
+      status = WriteLevel0Table(mem, edit, NULL);
+    }
+    mem->Unref();
   }
 
-  if (mem != NULL) mem->Unref();
-  delete file;
   return status;
 }
 
@@ -821,8 +851,9 @@ Status DBImpl::FinishCompactionOutputFile(CompactionState* compact,
     delete iter;
     if (s.ok()) {
       Log(options_.info_log,
-          "Generated table #%llu: %lld keys, %lld bytes",
+          "Generated table #%llu@%d: %lld keys, %lld bytes",
           (unsigned long long) output_number,
+          compact->compaction->level(),
           (unsigned long long) current_entries,
           (unsigned long long) current_bytes);
     }
@@ -1395,6 +1426,19 @@ bool DBImpl::GetProperty(const Slice& property, std::string* value) {
   } else if (in == "sstables") {
     *value = versions_->current()->DebugString();
     return true;
+  } else if (in == "approximate-memory-usage") {
+    size_t total_usage = options_.block_cache->TotalCharge();
+    if (mem_) {
+      total_usage += mem_->ApproximateMemoryUsage();
+    }
+    if (imm_) {
+      total_usage += imm_->ApproximateMemoryUsage();
+    }
+    char buf[50];
+    snprintf(buf, sizeof(buf), "%llu",
+             static_cast<unsigned long long>(total_usage));
+    value->append(buf);
+    return true;
   }
 
   return false;
@@ -1449,8 +1493,11 @@ Status DB::Open(const Options& options, const std::string& dbname,
   DBImpl* impl = new DBImpl(options, dbname);
   impl->mutex_.Lock();
   VersionEdit edit;
-  Status s = impl->Recover(&edit); // Handles create_if_missing, error_if_exists
-  if (s.ok()) {
+  // Recover handles create_if_missing, error_if_exists
+  bool save_manifest = false;
+  Status s = impl->Recover(&edit, &save_manifest);
+  if (s.ok() && impl->mem_ == NULL) {
+    // Create new log and a corresponding memtable.
     uint64_t new_log_number = impl->versions_->NewFileNumber();
     WritableFile* lfile;
     s = options.env->NewWritableFile(LogFileName(dbname, new_log_number),
@@ -1460,15 +1507,22 @@ Status DB::Open(const Options& options, const std::string& dbname,
       impl->logfile_ = lfile;
       impl->logfile_number_ = new_log_number;
       impl->log_ = new log::Writer(lfile);
-      s = impl->versions_->LogAndApply(&edit, &impl->mutex_);
-    }
-    if (s.ok()) {
-      impl->DeleteObsoleteFiles();
-      impl->MaybeScheduleCompaction();
+      impl->mem_ = new MemTable(impl->internal_comparator_);
+      impl->mem_->Ref();
     }
   }
+  if (s.ok() && save_manifest) {
+    edit.SetPrevLogNumber(0);  // No older logs needed after recovery.
+    edit.SetLogNumber(impl->logfile_number_);
+    s = impl->versions_->LogAndApply(&edit, &impl->mutex_);
+  }
+  if (s.ok()) {
+    impl->DeleteObsoleteFiles();
+    impl->MaybeScheduleCompaction();
+  }
   impl->mutex_.Unlock();
   if (s.ok()) {
+    assert(impl->mem_ != NULL);
     *dbptr = impl;
   } else {
     delete impl;
diff --git a/src/leveldb/db/db_impl.h b/src/leveldb/db/db_impl.h
index cfc998164a..8ff323e728 100644
--- a/src/leveldb/db/db_impl.h
+++ b/src/leveldb/db/db_impl.h
@@ -78,7 +78,8 @@ class DBImpl : public DB {
   // Recover the descriptor from persistent storage.  May do a significant
   // amount of work to recover recently logged updates.  Any changes to
   // be made to the descriptor are added to *edit.
-  Status Recover(VersionEdit* edit) EXCLUSIVE_LOCKS_REQUIRED(mutex_);
+  Status Recover(VersionEdit* edit, bool* save_manifest)
+      EXCLUSIVE_LOCKS_REQUIRED(mutex_);
 
   void MaybeIgnoreError(Status* s) const;
 
@@ -90,9 +91,8 @@ class DBImpl : public DB {
   // Errors are recorded in bg_error_.
   void CompactMemTable() EXCLUSIVE_LOCKS_REQUIRED(mutex_);
 
-  Status RecoverLogFile(uint64_t log_number,
-                        VersionEdit* edit,
-                        SequenceNumber* max_sequence)
+  Status RecoverLogFile(uint64_t log_number, bool last_log, bool* save_manifest,
+                        VersionEdit* edit, SequenceNumber* max_sequence)
       EXCLUSIVE_LOCKS_REQUIRED(mutex_);
 
   Status WriteLevel0Table(MemTable* mem, VersionEdit* edit, Version* base)
diff --git a/src/leveldb/db/db_test.cc b/src/leveldb/db/db_test.cc
index 0fed9137d5..a0b08bc19c 100644
--- a/src/leveldb/db/db_test.cc
+++ b/src/leveldb/db/db_test.cc
@@ -193,6 +193,7 @@ class DBTest {
   // Sequence of option configurations to try
   enum OptionConfig {
     kDefault,
+    kReuse,
     kFilter,
     kUncompressed,
     kEnd
@@ -237,7 +238,11 @@ class DBTest {
   // Return the current option configuration.
   Options CurrentOptions() {
     Options options;
+    options.reuse_logs = false;
     switch (option_config_) {
+      case kReuse:
+        options.reuse_logs = true;
+        break;
       case kFilter:
         options.filter_policy = filter_policy_;
         break;
@@ -558,6 +563,17 @@ TEST(DBTest, GetFromVersions) {
   } while (ChangeOptions());
 }
 
+TEST(DBTest, GetMemUsage) {
+  do {
+    ASSERT_OK(Put("foo", "v1"));
+    std::string val;
+    ASSERT_TRUE(db_->GetProperty("leveldb.approximate-memory-usage", &val));
+    int mem_usage = atoi(val.c_str());
+    ASSERT_GT(mem_usage, 0);
+    ASSERT_LT(mem_usage, 5*1024*1024);
+  } while (ChangeOptions());
+}
+
 TEST(DBTest, GetSnapshot) {
   do {
     // Try with both a short key and a long key
@@ -1080,6 +1096,14 @@ TEST(DBTest, ApproximateSizes) {
     // 0 because GetApproximateSizes() does not account for memtable space
     ASSERT_TRUE(Between(Size("", Key(50)), 0, 0));
 
+    if (options.reuse_logs) {
+      // Recovery will reuse memtable, and GetApproximateSizes() does not
+      // account for memtable usage;
+      Reopen(&options);
+      ASSERT_TRUE(Between(Size("", Key(50)), 0, 0));
+      continue;
+    }
+
     // Check sizes across recovery by reopening a few times
     for (int run = 0; run < 3; run++) {
       Reopen(&options);
@@ -1123,6 +1147,11 @@ TEST(DBTest, ApproximateSizes_MixOfSmallAndLarge) {
     ASSERT_OK(Put(Key(6), RandomString(&rnd, 300000)));
     ASSERT_OK(Put(Key(7), RandomString(&rnd, 10000)));
 
+    if (options.reuse_logs) {
+      // Need to force a memtable compaction since recovery does not do so.
+      ASSERT_OK(dbfull()->TEST_CompactMemTable());
+    }
+
     // Check sizes across recovery by reopening a few times
     for (int run = 0; run < 3; run++) {
       Reopen(&options);
@@ -2084,7 +2113,8 @@ void BM_LogAndApply(int iters, int num_base_files) {
   InternalKeyComparator cmp(BytewiseComparator());
   Options options;
   VersionSet vset(dbname, &options, NULL, &cmp);
-  ASSERT_OK(vset.Recover());
+  bool save_manifest;
+  ASSERT_OK(vset.Recover(&save_manifest));
   VersionEdit vbase;
   uint64_t fnum = 1;
   for (int i = 0; i < num_base_files; i++) {
diff --git a/src/leveldb/db/fault_injection_test.cc b/src/leveldb/db/fault_injection_test.cc
new file mode 100644
index 0000000000..875dfe81ee
--- /dev/null
+++ b/src/leveldb/db/fault_injection_test.cc
@@ -0,0 +1,554 @@
+// Copyright 2014 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+// This test uses a custom Env to keep track of the state of a filesystem as of
+// the last "sync". It then checks for data loss errors by purposely dropping
+// file data (or entire files) not protected by a "sync".
+
+#include "leveldb/db.h"
+
+#include <map>
+#include <set>
+#include "db/db_impl.h"
+#include "db/filename.h"
+#include "db/log_format.h"
+#include "db/version_set.h"
+#include "leveldb/cache.h"
+#include "leveldb/env.h"
+#include "leveldb/table.h"
+#include "leveldb/write_batch.h"
+#include "util/logging.h"
+#include "util/mutexlock.h"
+#include "util/testharness.h"
+#include "util/testutil.h"
+
+namespace leveldb {
+
+static const int kValueSize = 1000;
+static const int kMaxNumValues = 2000;
+static const size_t kNumIterations = 3;
+
+class FaultInjectionTestEnv;
+
+namespace {
+
+// Assume a filename, and not a directory name like "/foo/bar/"
+static std::string GetDirName(const std::string filename) {
+  size_t found = filename.find_last_of("/\\");
+  if (found == std::string::npos) {
+    return "";
+  } else {
+    return filename.substr(0, found);
+  }
+}
+
+Status SyncDir(const std::string& dir) {
+  // As this is a test it isn't required to *actually* sync this directory.
+  return Status::OK();
+}
+
+// A basic file truncation function suitable for this test.
+Status Truncate(const std::string& filename, uint64_t length) {
+  leveldb::Env* env = leveldb::Env::Default();
+
+  SequentialFile* orig_file;
+  Status s = env->NewSequentialFile(filename, &orig_file);
+  if (!s.ok())
+    return s;
+
+  char* scratch = new char[length];
+  leveldb::Slice result;
+  s = orig_file->Read(length, &result, scratch);
+  delete orig_file;
+  if (s.ok()) {
+    std::string tmp_name = GetDirName(filename) + "/truncate.tmp";
+    WritableFile* tmp_file;
+    s = env->NewWritableFile(tmp_name, &tmp_file);
+    if (s.ok()) {
+      s = tmp_file->Append(result);
+      delete tmp_file;
+      if (s.ok()) {
+        s = env->RenameFile(tmp_name, filename);
+      } else {
+        env->DeleteFile(tmp_name);
+      }
+    }
+  }
+
+  delete[] scratch;
+
+  return s;
+}
+
+struct FileState {
+  std::string filename_;
+  ssize_t pos_;
+  ssize_t pos_at_last_sync_;
+  ssize_t pos_at_last_flush_;
+
+  FileState(const std::string& filename)
+      : filename_(filename),
+        pos_(-1),
+        pos_at_last_sync_(-1),
+        pos_at_last_flush_(-1) { }
+
+  FileState() : pos_(-1), pos_at_last_sync_(-1), pos_at_last_flush_(-1) {}
+
+  bool IsFullySynced() const { return pos_ <= 0 || pos_ == pos_at_last_sync_; }
+
+  Status DropUnsyncedData() const;
+};
+
+}  // anonymous namespace
+
+// A wrapper around WritableFile which informs another Env whenever this file
+// is written to or sync'ed.
+class TestWritableFile : public WritableFile {
+ public:
+  TestWritableFile(const FileState& state,
+                   WritableFile* f,
+                   FaultInjectionTestEnv* env);
+  virtual ~TestWritableFile();
+  virtual Status Append(const Slice& data);
+  virtual Status Close();
+  virtual Status Flush();
+  virtual Status Sync();
+
+ private:
+  FileState state_;
+  WritableFile* target_;
+  bool writable_file_opened_;
+  FaultInjectionTestEnv* env_;
+
+  Status SyncParent();
+};
+
+class FaultInjectionTestEnv : public EnvWrapper {
+ public:
+  FaultInjectionTestEnv() : EnvWrapper(Env::Default()), filesystem_active_(true) {}
+  virtual ~FaultInjectionTestEnv() { }
+  virtual Status NewWritableFile(const std::string& fname,
+                                 WritableFile** result);
+  virtual Status NewAppendableFile(const std::string& fname,
+                                   WritableFile** result);
+  virtual Status DeleteFile(const std::string& f);
+  virtual Status RenameFile(const std::string& s, const std::string& t);
+
+  void WritableFileClosed(const FileState& state);
+  Status DropUnsyncedFileData();
+  Status DeleteFilesCreatedAfterLastDirSync();
+  void DirWasSynced();
+  bool IsFileCreatedSinceLastDirSync(const std::string& filename);
+  void ResetState();
+  void UntrackFile(const std::string& f);
+  // Setting the filesystem to inactive is the test equivalent to simulating a
+  // system reset. Setting to inactive will freeze our saved filesystem state so
+  // that it will stop being recorded. It can then be reset back to the state at
+  // the time of the reset.
+  bool IsFilesystemActive() const { return filesystem_active_; }
+  void SetFilesystemActive(bool active) { filesystem_active_ = active; }
+
+ private:
+  port::Mutex mutex_;
+  std::map<std::string, FileState> db_file_state_;
+  std::set<std::string> new_files_since_last_dir_sync_;
+  bool filesystem_active_;  // Record flushes, syncs, writes
+};
+
+TestWritableFile::TestWritableFile(const FileState& state,
+                                   WritableFile* f,
+                                   FaultInjectionTestEnv* env)
+    : state_(state),
+      target_(f),
+      writable_file_opened_(true),
+      env_(env) {
+  assert(f != NULL);
+}
+
+TestWritableFile::~TestWritableFile() {
+  if (writable_file_opened_) {
+    Close();
+  }
+  delete target_;
+}
+
+Status TestWritableFile::Append(const Slice& data) {
+  Status s = target_->Append(data);
+  if (s.ok() && env_->IsFilesystemActive()) {
+    state_.pos_ += data.size();
+  }
+  return s;
+}
+
+Status TestWritableFile::Close() {
+  writable_file_opened_ = false;
+  Status s = target_->Close();
+  if (s.ok()) {
+    env_->WritableFileClosed(state_);
+  }
+  return s;
+}
+
+Status TestWritableFile::Flush() {
+  Status s = target_->Flush();
+  if (s.ok() && env_->IsFilesystemActive()) {
+    state_.pos_at_last_flush_ = state_.pos_;
+  }
+  return s;
+}
+
+Status TestWritableFile::SyncParent() {
+  Status s = SyncDir(GetDirName(state_.filename_));
+  if (s.ok()) {
+    env_->DirWasSynced();
+  }
+  return s;
+}
+
+Status TestWritableFile::Sync() {
+  if (!env_->IsFilesystemActive()) {
+    return Status::OK();
+  }
+  // Ensure new files referred to by the manifest are in the filesystem.
+  Status s = target_->Sync();
+  if (s.ok()) {
+    state_.pos_at_last_sync_ = state_.pos_;
+  }
+  if (env_->IsFileCreatedSinceLastDirSync(state_.filename_)) {
+    Status ps = SyncParent();
+    if (s.ok() && !ps.ok()) {
+      s = ps;
+    }
+  }
+  return s;
+}
+
+Status FaultInjectionTestEnv::NewWritableFile(const std::string& fname,
+                                              WritableFile** result) {
+  WritableFile* actual_writable_file;
+  Status s = target()->NewWritableFile(fname, &actual_writable_file);
+  if (s.ok()) {
+    FileState state(fname);
+    state.pos_ = 0;
+    *result = new TestWritableFile(state, actual_writable_file, this);
+    // NewWritableFile doesn't append to files, so if the same file is
+    // opened again then it will be truncated - so forget our saved
+    // state.
+    UntrackFile(fname);
+    MutexLock l(&mutex_);
+    new_files_since_last_dir_sync_.insert(fname);
+  }
+  return s;
+}
+
+Status FaultInjectionTestEnv::NewAppendableFile(const std::string& fname,
+                                                WritableFile** result) {
+  WritableFile* actual_writable_file;
+  Status s = target()->NewAppendableFile(fname, &actual_writable_file);
+  if (s.ok()) {
+    FileState state(fname);
+    state.pos_ = 0;
+    {
+      MutexLock l(&mutex_);
+      if (db_file_state_.count(fname) == 0) {
+        new_files_since_last_dir_sync_.insert(fname);
+      } else {
+        state = db_file_state_[fname];
+      }
+    }
+    *result = new TestWritableFile(state, actual_writable_file, this);
+  }
+  return s;
+}
+
+Status FaultInjectionTestEnv::DropUnsyncedFileData() {
+  Status s;
+  MutexLock l(&mutex_);
+  for (std::map<std::string, FileState>::const_iterator it =
+           db_file_state_.begin();
+       s.ok() && it != db_file_state_.end(); ++it) {
+    const FileState& state = it->second;
+    if (!state.IsFullySynced()) {
+      s = state.DropUnsyncedData();
+    }
+  }
+  return s;
+}
+
+void FaultInjectionTestEnv::DirWasSynced() {
+  MutexLock l(&mutex_);
+  new_files_since_last_dir_sync_.clear();
+}
+
+bool FaultInjectionTestEnv::IsFileCreatedSinceLastDirSync(
+    const std::string& filename) {
+  MutexLock l(&mutex_);
+  return new_files_since_last_dir_sync_.find(filename) !=
+         new_files_since_last_dir_sync_.end();
+}
+
+void FaultInjectionTestEnv::UntrackFile(const std::string& f) {
+  MutexLock l(&mutex_);
+  db_file_state_.erase(f);
+  new_files_since_last_dir_sync_.erase(f);
+}
+
+Status FaultInjectionTestEnv::DeleteFile(const std::string& f) {
+  Status s = EnvWrapper::DeleteFile(f);
+  ASSERT_OK(s);
+  if (s.ok()) {
+    UntrackFile(f);
+  }
+  return s;
+}
+
+Status FaultInjectionTestEnv::RenameFile(const std::string& s,
+                                         const std::string& t) {
+  Status ret = EnvWrapper::RenameFile(s, t);
+
+  if (ret.ok()) {
+    MutexLock l(&mutex_);
+    if (db_file_state_.find(s) != db_file_state_.end()) {
+      db_file_state_[t] = db_file_state_[s];
+      db_file_state_.erase(s);
+    }
+
+    if (new_files_since_last_dir_sync_.erase(s) != 0) {
+      assert(new_files_since_last_dir_sync_.find(t) ==
+             new_files_since_last_dir_sync_.end());
+      new_files_since_last_dir_sync_.insert(t);
+    }
+  }
+
+  return ret;
+}
+
+void FaultInjectionTestEnv::ResetState() {
+  // Since we are not destroying the database, the existing files
+  // should keep their recorded synced/flushed state. Therefore
+  // we do not reset db_file_state_ and new_files_since_last_dir_sync_.
+  MutexLock l(&mutex_);
+  SetFilesystemActive(true);
+}
+
+Status FaultInjectionTestEnv::DeleteFilesCreatedAfterLastDirSync() {
+  // Because DeleteFile access this container make a copy to avoid deadlock
+  mutex_.Lock();
+  std::set<std::string> new_files(new_files_since_last_dir_sync_.begin(),
+                                  new_files_since_last_dir_sync_.end());
+  mutex_.Unlock();
+  Status s;
+  std::set<std::string>::const_iterator it;
+  for (it = new_files.begin(); s.ok() && it != new_files.end(); ++it) {
+    s = DeleteFile(*it);
+  }
+  return s;
+}
+
+void FaultInjectionTestEnv::WritableFileClosed(const FileState& state) {
+  MutexLock l(&mutex_);
+  db_file_state_[state.filename_] = state;
+}
+
+Status FileState::DropUnsyncedData() const {
+  ssize_t sync_pos = pos_at_last_sync_ == -1 ? 0 : pos_at_last_sync_;
+  return Truncate(filename_, sync_pos);
+}
+
+class FaultInjectionTest {
+ public:
+  enum ExpectedVerifResult { VAL_EXPECT_NO_ERROR, VAL_EXPECT_ERROR };
+  enum ResetMethod { RESET_DROP_UNSYNCED_DATA, RESET_DELETE_UNSYNCED_FILES };
+
+  FaultInjectionTestEnv* env_;
+  std::string dbname_;
+  Cache* tiny_cache_;
+  Options options_;
+  DB* db_;
+
+  FaultInjectionTest()
+      : env_(new FaultInjectionTestEnv),
+        tiny_cache_(NewLRUCache(100)),
+        db_(NULL) {
+    dbname_ = test::TmpDir() + "/fault_test";
+    DestroyDB(dbname_, Options());  // Destroy any db from earlier run
+    options_.reuse_logs = true;
+    options_.env = env_;
+    options_.paranoid_checks = true;
+    options_.block_cache = tiny_cache_;
+    options_.create_if_missing = true;
+  }
+
+  ~FaultInjectionTest() {
+    CloseDB();
+    DestroyDB(dbname_, Options());
+    delete tiny_cache_;
+    delete env_;
+  }
+
+  void ReuseLogs(bool reuse) {
+    options_.reuse_logs = reuse;
+  }
+
+  void Build(int start_idx, int num_vals) {
+    std::string key_space, value_space;
+    WriteBatch batch;
+    for (int i = start_idx; i < start_idx + num_vals; i++) {
+      Slice key = Key(i, &key_space);
+      batch.Clear();
+      batch.Put(key, Value(i, &value_space));
+      WriteOptions options;
+      ASSERT_OK(db_->Write(options, &batch));
+    }
+  }
+
+  Status ReadValue(int i, std::string* val) const {
+    std::string key_space, value_space;
+    Slice key = Key(i, &key_space);
+    Value(i, &value_space);
+    ReadOptions options;
+    return db_->Get(options, key, val);
+  }
+
+  Status Verify(int start_idx, int num_vals,
+                ExpectedVerifResult expected) const {
+    std::string val;
+    std::string value_space;
+    Status s;
+    for (int i = start_idx; i < start_idx + num_vals && s.ok(); i++) {
+      Value(i, &value_space);
+      s = ReadValue(i, &val);
+      if (expected == VAL_EXPECT_NO_ERROR) {
+        if (s.ok()) {
+          ASSERT_EQ(value_space, val);
+        }
+      } else if (s.ok()) {
+        fprintf(stderr, "Expected an error at %d, but was OK\n", i);
+        s = Status::IOError(dbname_, "Expected value error:");
+      } else {
+        s = Status::OK();  // An expected error
+      }
+    }
+    return s;
+  }
+
+  // Return the ith key
+  Slice Key(int i, std::string* storage) const {
+    char buf[100];
+    snprintf(buf, sizeof(buf), "%016d", i);
+    storage->assign(buf, strlen(buf));
+    return Slice(*storage);
+  }
+
+  // Return the value to associate with the specified key
+  Slice Value(int k, std::string* storage) const {
+    Random r(k);
+    return test::RandomString(&r, kValueSize, storage);
+  }
+
+  Status OpenDB() {
+    delete db_;
+    db_ = NULL;
+    env_->ResetState();
+    return DB::Open(options_, dbname_, &db_);
+  }
+
+  void CloseDB() {
+    delete db_;
+    db_ = NULL;
+  }
+
+  void DeleteAllData() {
+    Iterator* iter = db_->NewIterator(ReadOptions());
+    WriteOptions options;
+    for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
+      ASSERT_OK(db_->Delete(WriteOptions(), iter->key()));
+    }
+
+    delete iter;
+  }
+
+  void ResetDBState(ResetMethod reset_method) {
+    switch (reset_method) {
+      case RESET_DROP_UNSYNCED_DATA:
+        ASSERT_OK(env_->DropUnsyncedFileData());
+        break;
+      case RESET_DELETE_UNSYNCED_FILES:
+        ASSERT_OK(env_->DeleteFilesCreatedAfterLastDirSync());
+        break;
+      default:
+        assert(false);
+    }
+  }
+
+  void PartialCompactTestPreFault(int num_pre_sync, int num_post_sync) {
+    DeleteAllData();
+    Build(0, num_pre_sync);
+    db_->CompactRange(NULL, NULL);
+    Build(num_pre_sync, num_post_sync);
+  }
+
+  void PartialCompactTestReopenWithFault(ResetMethod reset_method,
+                                         int num_pre_sync,
+                                         int num_post_sync) {
+    env_->SetFilesystemActive(false);
+    CloseDB();
+    ResetDBState(reset_method);
+    ASSERT_OK(OpenDB());
+    ASSERT_OK(Verify(0, num_pre_sync, FaultInjectionTest::VAL_EXPECT_NO_ERROR));
+    ASSERT_OK(Verify(num_pre_sync, num_post_sync, FaultInjectionTest::VAL_EXPECT_ERROR));
+  }
+
+  void NoWriteTestPreFault() {
+  }
+
+  void NoWriteTestReopenWithFault(ResetMethod reset_method) {
+    CloseDB();
+    ResetDBState(reset_method);
+    ASSERT_OK(OpenDB());
+  }
+
+  void DoTest() {
+    Random rnd(0);
+    ASSERT_OK(OpenDB());
+    for (size_t idx = 0; idx < kNumIterations; idx++) {
+      int num_pre_sync = rnd.Uniform(kMaxNumValues);
+      int num_post_sync = rnd.Uniform(kMaxNumValues);
+
+      PartialCompactTestPreFault(num_pre_sync, num_post_sync);
+      PartialCompactTestReopenWithFault(RESET_DROP_UNSYNCED_DATA,
+                                        num_pre_sync,
+                                        num_post_sync);
+
+      NoWriteTestPreFault();
+      NoWriteTestReopenWithFault(RESET_DROP_UNSYNCED_DATA);
+
+      PartialCompactTestPreFault(num_pre_sync, num_post_sync);
+      // No new files created so we expect all values since no files will be
+      // dropped.
+      PartialCompactTestReopenWithFault(RESET_DELETE_UNSYNCED_FILES,
+                                        num_pre_sync + num_post_sync,
+                                        0);
+
+      NoWriteTestPreFault();
+      NoWriteTestReopenWithFault(RESET_DELETE_UNSYNCED_FILES);
+    }
+  }
+};
+
+TEST(FaultInjectionTest, FaultTestNoLogReuse) {
+  ReuseLogs(false);
+  DoTest();
+}
+
+TEST(FaultInjectionTest, FaultTestWithLogReuse) {
+  ReuseLogs(true);
+  DoTest();
+}
+
+}  // namespace leveldb
+
+int main(int argc, char** argv) {
+  return leveldb::test::RunAllTests();
+}
diff --git a/src/leveldb/db/leveldb_main.cc b/src/leveldb/db/leveldbutil.cc
index 9f4b7dd70c..9f4b7dd70c 100644
--- a/src/leveldb/db/leveldb_main.cc
+++ b/src/leveldb/db/leveldbutil.cc
diff --git a/src/leveldb/db/log_reader.cc b/src/leveldb/db/log_reader.cc
index e44b66c85b..a6d304545d 100644
--- a/src/leveldb/db/log_reader.cc
+++ b/src/leveldb/db/log_reader.cc
@@ -25,7 +25,8 @@ Reader::Reader(SequentialFile* file, Reporter* reporter, bool checksum,
       eof_(false),
       last_record_offset_(0),
       end_of_buffer_offset_(0),
-      initial_offset_(initial_offset) {
+      initial_offset_(initial_offset),
+      resyncing_(initial_offset > 0) {
 }
 
 Reader::~Reader() {
@@ -72,8 +73,25 @@ bool Reader::ReadRecord(Slice* record, std::string* scratch) {
 
   Slice fragment;
   while (true) {
-    uint64_t physical_record_offset = end_of_buffer_offset_ - buffer_.size();
     const unsigned int record_type = ReadPhysicalRecord(&fragment);
+
+    // ReadPhysicalRecord may have only had an empty trailer remaining in its
+    // internal buffer. Calculate the offset of the next physical record now
+    // that it has returned, properly accounting for its header size.
+    uint64_t physical_record_offset =
+        end_of_buffer_offset_ - buffer_.size() - kHeaderSize - fragment.size();
+
+    if (resyncing_) {
+      if (record_type == kMiddleType) {
+        continue;
+      } else if (record_type == kLastType) {
+        resyncing_ = false;
+        continue;
+      } else {
+        resyncing_ = false;
+      }
+    }
+
     switch (record_type) {
       case kFullType:
         if (in_fragmented_record) {
diff --git a/src/leveldb/db/log_reader.h b/src/leveldb/db/log_reader.h
index 6aff791716..8389d61f8f 100644
--- a/src/leveldb/db/log_reader.h
+++ b/src/leveldb/db/log_reader.h
@@ -73,6 +73,11 @@ class Reader {
   // Offset at which to start looking for the first record to return
   uint64_t const initial_offset_;
 
+  // True if we are resynchronizing after a seek (initial_offset_ > 0). In
+  // particular, a run of kMiddleType and kLastType records can be silently
+  // skipped in this mode
+  bool resyncing_;
+
   // Extend record types with the following special values
   enum {
     kEof = kMaxRecordType + 1,
diff --git a/src/leveldb/db/log_test.cc b/src/leveldb/db/log_test.cc
index dcf0562652..48a5928657 100644
--- a/src/leveldb/db/log_test.cc
+++ b/src/leveldb/db/log_test.cc
@@ -79,7 +79,7 @@ class LogTest {
     virtual Status Skip(uint64_t n) {
       if (n > contents_.size()) {
         contents_.clear();
-        return Status::NotFound("in-memory file skipepd past end");
+        return Status::NotFound("in-memory file skipped past end");
       }
 
       contents_.remove_prefix(n);
@@ -104,23 +104,34 @@ class LogTest {
   StringSource source_;
   ReportCollector report_;
   bool reading_;
-  Writer writer_;
-  Reader reader_;
+  Writer* writer_;
+  Reader* reader_;
 
   // Record metadata for testing initial offset functionality
   static size_t initial_offset_record_sizes_[];
   static uint64_t initial_offset_last_record_offsets_[];
+  static int num_initial_offset_records_;
 
  public:
   LogTest() : reading_(false),
-              writer_(&dest_),
-              reader_(&source_, &report_, true/*checksum*/,
-                      0/*initial_offset*/) {
+              writer_(new Writer(&dest_)),
+              reader_(new Reader(&source_, &report_, true/*checksum*/,
+                      0/*initial_offset*/)) {
+  }
+
+  ~LogTest() {
+    delete writer_;
+    delete reader_;
+  }
+
+  void ReopenForAppend() {
+    delete writer_;
+    writer_ = new Writer(&dest_, dest_.contents_.size());
   }
 
   void Write(const std::string& msg) {
     ASSERT_TRUE(!reading_) << "Write() after starting to read";
-    writer_.AddRecord(Slice(msg));
+    writer_->AddRecord(Slice(msg));
   }
 
   size_t WrittenBytes() const {
@@ -134,7 +145,7 @@ class LogTest {
     }
     std::string scratch;
     Slice record;
-    if (reader_.ReadRecord(&record, &scratch)) {
+    if (reader_->ReadRecord(&record, &scratch)) {
       return record.ToString();
     } else {
       return "EOF";
@@ -182,13 +193,18 @@ class LogTest {
   }
 
   void WriteInitialOffsetLog() {
-    for (int i = 0; i < 4; i++) {
+    for (int i = 0; i < num_initial_offset_records_; i++) {
       std::string record(initial_offset_record_sizes_[i],
                          static_cast<char>('a' + i));
       Write(record);
     }
   }
 
+  void StartReadingAt(uint64_t initial_offset) {
+    delete reader_;
+    reader_ = new Reader(&source_, &report_, true/*checksum*/, initial_offset);
+  }
+
   void CheckOffsetPastEndReturnsNoRecords(uint64_t offset_past_end) {
     WriteInitialOffsetLog();
     reading_ = true;
@@ -208,32 +224,48 @@ class LogTest {
     source_.contents_ = Slice(dest_.contents_);
     Reader* offset_reader = new Reader(&source_, &report_, true/*checksum*/,
                                        initial_offset);
-    Slice record;
-    std::string scratch;
-    ASSERT_TRUE(offset_reader->ReadRecord(&record, &scratch));
-    ASSERT_EQ(initial_offset_record_sizes_[expected_record_offset],
-              record.size());
-    ASSERT_EQ(initial_offset_last_record_offsets_[expected_record_offset],
-              offset_reader->LastRecordOffset());
-    ASSERT_EQ((char)('a' + expected_record_offset), record.data()[0]);
+
+    // Read all records from expected_record_offset through the last one.
+    ASSERT_LT(expected_record_offset, num_initial_offset_records_);
+    for (; expected_record_offset < num_initial_offset_records_;
+         ++expected_record_offset) {
+      Slice record;
+      std::string scratch;
+      ASSERT_TRUE(offset_reader->ReadRecord(&record, &scratch));
+      ASSERT_EQ(initial_offset_record_sizes_[expected_record_offset],
+                record.size());
+      ASSERT_EQ(initial_offset_last_record_offsets_[expected_record_offset],
+                offset_reader->LastRecordOffset());
+      ASSERT_EQ((char)('a' + expected_record_offset), record.data()[0]);
+    }
     delete offset_reader;
   }
-
 };
 
 size_t LogTest::initial_offset_record_sizes_[] =
     {10000,  // Two sizable records in first block
      10000,
      2 * log::kBlockSize - 1000,  // Span three blocks
-     1};
+     1,
+     13716,  // Consume all but two bytes of block 3.
+     log::kBlockSize - kHeaderSize, // Consume the entirety of block 4.
+    };
 
 uint64_t LogTest::initial_offset_last_record_offsets_[] =
     {0,
      kHeaderSize + 10000,
      2 * (kHeaderSize + 10000),
      2 * (kHeaderSize + 10000) +
-         (2 * log::kBlockSize - 1000) + 3 * kHeaderSize};
+         (2 * log::kBlockSize - 1000) + 3 * kHeaderSize,
+     2 * (kHeaderSize + 10000) +
+         (2 * log::kBlockSize - 1000) + 3 * kHeaderSize
+         + kHeaderSize + 1,
+     3 * log::kBlockSize,
+    };
 
+// LogTest::initial_offset_last_record_offsets_ must be defined before this.
+int LogTest::num_initial_offset_records_ =
+    sizeof(LogTest::initial_offset_last_record_offsets_)/sizeof(uint64_t);
 
 TEST(LogTest, Empty) {
   ASSERT_EQ("EOF", Read());
@@ -318,6 +350,15 @@ TEST(LogTest, AlignedEof) {
   ASSERT_EQ("EOF", Read());
 }
 
+TEST(LogTest, OpenForAppend) {
+  Write("hello");
+  ReopenForAppend();
+  Write("world");
+  ASSERT_EQ("hello", Read());
+  ASSERT_EQ("world", Read());
+  ASSERT_EQ("EOF", Read());
+}
+
 TEST(LogTest, RandomRead) {
   const int N = 500;
   Random write_rnd(301);
@@ -445,6 +486,22 @@ TEST(LogTest, PartialLastIsIgnored) {
   ASSERT_EQ(0, DroppedBytes());
 }
 
+TEST(LogTest, SkipIntoMultiRecord) {
+  // Consider a fragmented record:
+  //    first(R1), middle(R1), last(R1), first(R2)
+  // If initial_offset points to a record after first(R1) but before first(R2)
+  // incomplete fragment errors are not actual errors, and must be suppressed
+  // until a new first or full record is encountered.
+  Write(BigString("foo", 3*kBlockSize));
+  Write("correct");
+  StartReadingAt(kBlockSize);
+
+  ASSERT_EQ("correct", Read());
+  ASSERT_EQ("", ReportMessage());
+  ASSERT_EQ(0, DroppedBytes());
+  ASSERT_EQ("EOF", Read());
+}
+
 TEST(LogTest, ErrorJoinsRecords) {
   // Consider two fragmented records:
   //    first(R1) last(R1) first(R2) last(R2)
@@ -514,6 +571,10 @@ TEST(LogTest, ReadFourthStart) {
       3);
 }
 
+TEST(LogTest, ReadInitialOffsetIntoBlockPadding) {
+  CheckInitialOffsetRecord(3 * log::kBlockSize - 3, 5);
+}
+
 TEST(LogTest, ReadEnd) {
   CheckOffsetPastEndReturnsNoRecords(0);
 }
diff --git a/src/leveldb/db/log_writer.cc b/src/leveldb/db/log_writer.cc
index 2da99ac088..74a03270da 100644
--- a/src/leveldb/db/log_writer.cc
+++ b/src/leveldb/db/log_writer.cc
@@ -12,15 +12,24 @@
 namespace leveldb {
 namespace log {
 
-Writer::Writer(WritableFile* dest)
-    : dest_(dest),
-      block_offset_(0) {
+static void InitTypeCrc(uint32_t* type_crc) {
   for (int i = 0; i <= kMaxRecordType; i++) {
     char t = static_cast<char>(i);
-    type_crc_[i] = crc32c::Value(&t, 1);
+    type_crc[i] = crc32c::Value(&t, 1);
   }
 }
 
+Writer::Writer(WritableFile* dest)
+    : dest_(dest),
+      block_offset_(0) {
+  InitTypeCrc(type_crc_);
+}
+
+Writer::Writer(WritableFile* dest, uint64_t dest_length)
+    : dest_(dest), block_offset_(dest_length % kBlockSize) {
+  InitTypeCrc(type_crc_);
+}
+
 Writer::~Writer() {
 }
 
diff --git a/src/leveldb/db/log_writer.h b/src/leveldb/db/log_writer.h
index a3a954d967..9e7cc4705b 100644
--- a/src/leveldb/db/log_writer.h
+++ b/src/leveldb/db/log_writer.h
@@ -22,6 +22,12 @@ class Writer {
   // "*dest" must be initially empty.
   // "*dest" must remain live while this Writer is in use.
   explicit Writer(WritableFile* dest);
+
+  // Create a writer that will append data to "*dest".
+  // "*dest" must have initial length "dest_length".
+  // "*dest" must remain live while this Writer is in use.
+  Writer(WritableFile* dest, uint64_t dest_length);
+
   ~Writer();
 
   Status AddRecord(const Slice& slice);
diff --git a/src/leveldb/db/memtable.h b/src/leveldb/db/memtable.h
index 92e90bb099..9f41567cde 100644
--- a/src/leveldb/db/memtable.h
+++ b/src/leveldb/db/memtable.h
@@ -36,10 +36,7 @@ class MemTable {
   }
 
   // Returns an estimate of the number of bytes of data in use by this
-  // data structure.
-  //
-  // REQUIRES: external synchronization to prevent simultaneous
-  // operations on the same MemTable.
+  // data structure. It is safe to call when MemTable is being modified.
   size_t ApproximateMemoryUsage();
 
   // Return an iterator that yields the contents of the memtable.
diff --git a/src/leveldb/db/recovery_test.cc b/src/leveldb/db/recovery_test.cc
new file mode 100644
index 0000000000..9596f4288a
--- /dev/null
+++ b/src/leveldb/db/recovery_test.cc
@@ -0,0 +1,324 @@
+// Copyright (c) 2014 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "db/db_impl.h"
+#include "db/filename.h"
+#include "db/version_set.h"
+#include "db/write_batch_internal.h"
+#include "leveldb/db.h"
+#include "leveldb/env.h"
+#include "leveldb/write_batch.h"
+#include "util/logging.h"
+#include "util/testharness.h"
+#include "util/testutil.h"
+
+namespace leveldb {
+
+class RecoveryTest {
+ public:
+  RecoveryTest() : env_(Env::Default()), db_(NULL) {
+    dbname_ = test::TmpDir() + "/recovery_test";
+    DestroyDB(dbname_, Options());
+    Open();
+  }
+
+  ~RecoveryTest() {
+    Close();
+    DestroyDB(dbname_, Options());
+  }
+
+  DBImpl* dbfull() const { return reinterpret_cast<DBImpl*>(db_); }
+  Env* env() const { return env_; }
+
+  bool CanAppend() {
+    WritableFile* tmp;
+    Status s = env_->NewAppendableFile(CurrentFileName(dbname_), &tmp);
+    delete tmp;
+    if (s.IsNotSupportedError()) {
+      return false;
+    } else {
+      return true;
+    }
+  }
+
+  void Close() {
+    delete db_;
+    db_ = NULL;
+  }
+
+  void Open(Options* options = NULL) {
+    Close();
+    Options opts;
+    if (options != NULL) {
+      opts = *options;
+    } else {
+      opts.reuse_logs = true;  // TODO(sanjay): test both ways
+      opts.create_if_missing = true;
+    }
+    if (opts.env == NULL) {
+      opts.env = env_;
+    }
+    ASSERT_OK(DB::Open(opts, dbname_, &db_));
+    ASSERT_EQ(1, NumLogs());
+  }
+
+  Status Put(const std::string& k, const std::string& v) {
+    return db_->Put(WriteOptions(), k, v);
+  }
+
+  std::string Get(const std::string& k, const Snapshot* snapshot = NULL) {
+    std::string result;
+    Status s = db_->Get(ReadOptions(), k, &result);
+    if (s.IsNotFound()) {
+      result = "NOT_FOUND";
+    } else if (!s.ok()) {
+      result = s.ToString();
+    }
+    return result;
+  }
+
+  std::string ManifestFileName() {
+    std::string current;
+    ASSERT_OK(ReadFileToString(env_, CurrentFileName(dbname_), &current));
+    size_t len = current.size();
+    if (len > 0 && current[len-1] == '\n') {
+      current.resize(len - 1);
+    }
+    return dbname_ + "/" + current;
+  }
+
+  std::string LogName(uint64_t number) {
+    return LogFileName(dbname_, number);
+  }
+
+  size_t DeleteLogFiles() {
+    std::vector<uint64_t> logs = GetFiles(kLogFile);
+    for (size_t i = 0; i < logs.size(); i++) {
+      ASSERT_OK(env_->DeleteFile(LogName(logs[i]))) << LogName(logs[i]);
+    }
+    return logs.size();
+  }
+
+  uint64_t FirstLogFile() {
+    return GetFiles(kLogFile)[0];
+  }
+
+  std::vector<uint64_t> GetFiles(FileType t) {
+    std::vector<std::string> filenames;
+    ASSERT_OK(env_->GetChildren(dbname_, &filenames));
+    std::vector<uint64_t> result;
+    for (size_t i = 0; i < filenames.size(); i++) {
+      uint64_t number;
+      FileType type;
+      if (ParseFileName(filenames[i], &number, &type) && type == t) {
+        result.push_back(number);
+      }
+    }
+    return result;
+  }
+
+  int NumLogs() {
+    return GetFiles(kLogFile).size();
+  }
+
+  int NumTables() {
+    return GetFiles(kTableFile).size();
+  }
+
+  uint64_t FileSize(const std::string& fname) {
+    uint64_t result;
+    ASSERT_OK(env_->GetFileSize(fname, &result)) << fname;
+    return result;
+  }
+
+  void CompactMemTable() {
+    dbfull()->TEST_CompactMemTable();
+  }
+
+  // Directly construct a log file that sets key to val.
+  void MakeLogFile(uint64_t lognum, SequenceNumber seq, Slice key, Slice val) {
+    std::string fname = LogFileName(dbname_, lognum);
+    WritableFile* file;
+    ASSERT_OK(env_->NewWritableFile(fname, &file));
+    log::Writer writer(file);
+    WriteBatch batch;
+    batch.Put(key, val);
+    WriteBatchInternal::SetSequence(&batch, seq);
+    ASSERT_OK(writer.AddRecord(WriteBatchInternal::Contents(&batch)));
+    ASSERT_OK(file->Flush());
+    delete file;
+  }
+
+ private:
+  std::string dbname_;
+  Env* env_;
+  DB* db_;
+};
+
+TEST(RecoveryTest, ManifestReused) {
+  if (!CanAppend()) {
+    fprintf(stderr, "skipping test because env does not support appending\n");
+    return;
+  }
+  ASSERT_OK(Put("foo", "bar"));
+  Close();
+  std::string old_manifest = ManifestFileName();
+  Open();
+  ASSERT_EQ(old_manifest, ManifestFileName());
+  ASSERT_EQ("bar", Get("foo"));
+  Open();
+  ASSERT_EQ(old_manifest, ManifestFileName());
+  ASSERT_EQ("bar", Get("foo"));
+}
+
+TEST(RecoveryTest, LargeManifestCompacted) {
+  if (!CanAppend()) {
+    fprintf(stderr, "skipping test because env does not support appending\n");
+    return;
+  }
+  ASSERT_OK(Put("foo", "bar"));
+  Close();
+  std::string old_manifest = ManifestFileName();
+
+  // Pad with zeroes to make manifest file very big.
+  {
+    uint64_t len = FileSize(old_manifest);
+    WritableFile* file;
+    ASSERT_OK(env()->NewAppendableFile(old_manifest, &file));
+    std::string zeroes(3*1048576 - static_cast<size_t>(len), 0);
+    ASSERT_OK(file->Append(zeroes));
+    ASSERT_OK(file->Flush());
+    delete file;
+  }
+
+  Open();
+  std::string new_manifest = ManifestFileName();
+  ASSERT_NE(old_manifest, new_manifest);
+  ASSERT_GT(10000, FileSize(new_manifest));
+  ASSERT_EQ("bar", Get("foo"));
+
+  Open();
+  ASSERT_EQ(new_manifest, ManifestFileName());
+  ASSERT_EQ("bar", Get("foo"));
+}
+
+TEST(RecoveryTest, NoLogFiles) {
+  ASSERT_OK(Put("foo", "bar"));
+  ASSERT_EQ(1, DeleteLogFiles());
+  Open();
+  ASSERT_EQ("NOT_FOUND", Get("foo"));
+  Open();
+  ASSERT_EQ("NOT_FOUND", Get("foo"));
+}
+
+TEST(RecoveryTest, LogFileReuse) {
+  if (!CanAppend()) {
+    fprintf(stderr, "skipping test because env does not support appending\n");
+    return;
+  }
+  for (int i = 0; i < 2; i++) {
+    ASSERT_OK(Put("foo", "bar"));
+    if (i == 0) {
+      // Compact to ensure current log is empty
+      CompactMemTable();
+    }
+    Close();
+    ASSERT_EQ(1, NumLogs());
+    uint64_t number = FirstLogFile();
+    if (i == 0) {
+      ASSERT_EQ(0, FileSize(LogName(number)));
+    } else {
+      ASSERT_LT(0, FileSize(LogName(number)));
+    }
+    Open();
+    ASSERT_EQ(1, NumLogs());
+    ASSERT_EQ(number, FirstLogFile()) << "did not reuse log file";
+    ASSERT_EQ("bar", Get("foo"));
+    Open();
+    ASSERT_EQ(1, NumLogs());
+    ASSERT_EQ(number, FirstLogFile()) << "did not reuse log file";
+    ASSERT_EQ("bar", Get("foo"));
+  }
+}
+
+TEST(RecoveryTest, MultipleMemTables) {
+  // Make a large log.
+  const int kNum = 1000;
+  for (int i = 0; i < kNum; i++) {
+    char buf[100];
+    snprintf(buf, sizeof(buf), "%050d", i);
+    ASSERT_OK(Put(buf, buf));
+  }
+  ASSERT_EQ(0, NumTables());
+  Close();
+  ASSERT_EQ(0, NumTables());
+  ASSERT_EQ(1, NumLogs());
+  uint64_t old_log_file = FirstLogFile();
+
+  // Force creation of multiple memtables by reducing the write buffer size.
+  Options opt;
+  opt.reuse_logs = true;
+  opt.write_buffer_size = (kNum*100) / 2;
+  Open(&opt);
+  ASSERT_LE(2, NumTables());
+  ASSERT_EQ(1, NumLogs());
+  ASSERT_NE(old_log_file, FirstLogFile()) << "must not reuse log";
+  for (int i = 0; i < kNum; i++) {
+    char buf[100];
+    snprintf(buf, sizeof(buf), "%050d", i);
+    ASSERT_EQ(buf, Get(buf));
+  }
+}
+
+TEST(RecoveryTest, MultipleLogFiles) {
+  ASSERT_OK(Put("foo", "bar"));
+  Close();
+  ASSERT_EQ(1, NumLogs());
+
+  // Make a bunch of uncompacted log files.
+  uint64_t old_log = FirstLogFile();
+  MakeLogFile(old_log+1, 1000, "hello", "world");
+  MakeLogFile(old_log+2, 1001, "hi", "there");
+  MakeLogFile(old_log+3, 1002, "foo", "bar2");
+
+  // Recover and check that all log files were processed.
+  Open();
+  ASSERT_LE(1, NumTables());
+  ASSERT_EQ(1, NumLogs());
+  uint64_t new_log = FirstLogFile();
+  ASSERT_LE(old_log+3, new_log);
+  ASSERT_EQ("bar2", Get("foo"));
+  ASSERT_EQ("world", Get("hello"));
+  ASSERT_EQ("there", Get("hi"));
+
+  // Test that previous recovery produced recoverable state.
+  Open();
+  ASSERT_LE(1, NumTables());
+  ASSERT_EQ(1, NumLogs());
+  if (CanAppend()) {
+    ASSERT_EQ(new_log, FirstLogFile());
+  }
+  ASSERT_EQ("bar2", Get("foo"));
+  ASSERT_EQ("world", Get("hello"));
+  ASSERT_EQ("there", Get("hi"));
+
+  // Check that introducing an older log file does not cause it to be re-read.
+  Close();
+  MakeLogFile(old_log+1, 2000, "hello", "stale write");
+  Open();
+  ASSERT_LE(1, NumTables());
+  ASSERT_EQ(1, NumLogs());
+  if (CanAppend()) {
+    ASSERT_EQ(new_log, FirstLogFile());
+  }
+  ASSERT_EQ("bar2", Get("foo"));
+  ASSERT_EQ("world", Get("hello"));
+  ASSERT_EQ("there", Get("hi"));
+}
+
+}  // namespace leveldb
+
+int main(int argc, char** argv) {
+  return leveldb::test::RunAllTests();
+}
diff --git a/src/leveldb/db/skiplist.h b/src/leveldb/db/skiplist.h
index ed8b092203..8bd77764d8 100644
--- a/src/leveldb/db/skiplist.h
+++ b/src/leveldb/db/skiplist.h
@@ -1,10 +1,10 @@
-#ifndef STORAGE_LEVELDB_DB_SKIPLIST_H_
-#define STORAGE_LEVELDB_DB_SKIPLIST_H_
-
 // Copyright (c) 2011 The LevelDB Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file. See the AUTHORS file for names of contributors.
-//
+
+#ifndef STORAGE_LEVELDB_DB_SKIPLIST_H_
+#define STORAGE_LEVELDB_DB_SKIPLIST_H_
+
 // Thread safety
 // -------------
 //
diff --git a/src/leveldb/db/skiplist_test.cc b/src/leveldb/db/skiplist_test.cc
index c78f4b4fb1..aee1461e1b 100644
--- a/src/leveldb/db/skiplist_test.cc
+++ b/src/leveldb/db/skiplist_test.cc
@@ -250,7 +250,7 @@ class ConcurrentTest {
         // Note that generation 0 is never inserted, so it is ok if
         // <*,0,*> is missing.
         ASSERT_TRUE((gen(pos) == 0) ||
-                    (gen(pos) > initial_state.Get(key(pos)))
+                    (gen(pos) > static_cast<Key>(initial_state.Get(key(pos))))
                     ) << "key: " << key(pos)
                       << "; gen: " << gen(pos)
                       << "; initgen: "
diff --git a/src/leveldb/db/snapshot.h b/src/leveldb/db/snapshot.h
index e7f8fd2c37..6ed413c42d 100644
--- a/src/leveldb/db/snapshot.h
+++ b/src/leveldb/db/snapshot.h
@@ -5,6 +5,7 @@
 #ifndef STORAGE_LEVELDB_DB_SNAPSHOT_H_
 #define STORAGE_LEVELDB_DB_SNAPSHOT_H_
 
+#include "db/dbformat.h"
 #include "leveldb/db.h"
 
 namespace leveldb {
diff --git a/src/leveldb/db/version_set.cc b/src/leveldb/db/version_set.cc
index aa83df55e4..a5e0f77a6a 100644
--- a/src/leveldb/db/version_set.cc
+++ b/src/leveldb/db/version_set.cc
@@ -893,7 +893,7 @@ Status VersionSet::LogAndApply(VersionEdit* edit, port::Mutex* mu) {
   return s;
 }
 
-Status VersionSet::Recover() {
+Status VersionSet::Recover(bool *save_manifest) {
   struct LogReporter : public log::Reader::Reporter {
     Status* status;
     virtual void Corruption(size_t bytes, const Status& s) {
@@ -1003,11 +1003,49 @@ Status VersionSet::Recover() {
     last_sequence_ = last_sequence;
     log_number_ = log_number;
     prev_log_number_ = prev_log_number;
+
+    // See if we can reuse the existing MANIFEST file.
+    if (ReuseManifest(dscname, current)) {
+      // No need to save new manifest
+    } else {
+      *save_manifest = true;
+    }
   }
 
   return s;
 }
 
+bool VersionSet::ReuseManifest(const std::string& dscname,
+                               const std::string& dscbase) {
+  if (!options_->reuse_logs) {
+    return false;
+  }
+  FileType manifest_type;
+  uint64_t manifest_number;
+  uint64_t manifest_size;
+  if (!ParseFileName(dscbase, &manifest_number, &manifest_type) ||
+      manifest_type != kDescriptorFile ||
+      !env_->GetFileSize(dscname, &manifest_size).ok() ||
+      // Make new compacted MANIFEST if old one is too big
+      manifest_size >= kTargetFileSize) {
+    return false;
+  }
+
+  assert(descriptor_file_ == NULL);
+  assert(descriptor_log_ == NULL);
+  Status r = env_->NewAppendableFile(dscname, &descriptor_file_);
+  if (!r.ok()) {
+    Log(options_->info_log, "Reuse MANIFEST: %s\n", r.ToString().c_str());
+    assert(descriptor_file_ == NULL);
+    return false;
+  }
+
+  Log(options_->info_log, "Reusing MANIFEST %s\n", dscname.c_str());
+  descriptor_log_ = new log::Writer(descriptor_file_, manifest_size);
+  manifest_file_number_ = manifest_number;
+  return true;
+}
+
 void VersionSet::MarkFileNumberUsed(uint64_t number) {
   if (next_file_number_ <= number) {
     next_file_number_ = number + 1;
diff --git a/src/leveldb/db/version_set.h b/src/leveldb/db/version_set.h
index 8dc14b8e01..1dec745673 100644
--- a/src/leveldb/db/version_set.h
+++ b/src/leveldb/db/version_set.h
@@ -179,7 +179,7 @@ class VersionSet {
       EXCLUSIVE_LOCKS_REQUIRED(mu);
 
   // Recover the last saved descriptor from persistent storage.
-  Status Recover();
+  Status Recover(bool *save_manifest);
 
   // Return the current version.
   Version* current() const { return current_; }
@@ -274,6 +274,8 @@ class VersionSet {
   friend class Compaction;
   friend class Version;
 
+  bool ReuseManifest(const std::string& dscname, const std::string& dscbase);
+
   void Finalize(Version* v);
 
   void GetRange(const std::vector<FileMetaData*>& inputs,
diff --git a/src/leveldb/db/write_batch_internal.h b/src/leveldb/db/write_batch_internal.h
index 310a3c8912..9448ef7b21 100644
--- a/src/leveldb/db/write_batch_internal.h
+++ b/src/leveldb/db/write_batch_internal.h
@@ -5,6 +5,7 @@
 #ifndef STORAGE_LEVELDB_DB_WRITE_BATCH_INTERNAL_H_
 #define STORAGE_LEVELDB_DB_WRITE_BATCH_INTERNAL_H_
 
+#include "db/dbformat.h"
 #include "leveldb/write_batch.h"
 
 namespace leveldb {