Merge branch 'master' into columnfamilies

Conflicts:
	db/db_impl.cc
	db/db_test.cc
	include/rocksdb/db.h

.arcconfig

@@ -6,6 +6,5 @@
     "linters"
   ],
   "lint.engine" : "FacebookFbcodeLintEngine",
-  "lint.engine.single.linter" : "FbcodeCppLinter",
+  "lint.engine.single.linter" : "FbcodeCppLinter"
-  "lint.cpplint.prefix" : "linters"
 }

.gitignore

@@ -12,6 +12,7 @@ build_config.mk
 *_test
 *_bench
 *_stress
+*.out
 
 ldb
 manifest_dump

Makefile

@@ -97,7 +97,7 @@ TOOLS = \
 	blob_store_bench
 
 
-PROGRAMS = db_bench signal_test $(TESTS) $(TOOLS)
+PROGRAMS = db_bench signal_test table_reader_bench $(TESTS) $(TOOLS)
 BENCHMARKS = db_bench_sqlite3 db_bench_tree_db table_reader_bench
 
 # The library name is configurable since we are maintaining libraries of both

db/db_impl.cc

@@ -3604,6 +3604,23 @@ Status DBImpl::MakeRoomForWrite(ColumnFamilyData* cfd, bool force) {
   return s;
 }
 
+Status DBImpl::GetPropertiesOfAllTables(TablePropertiesCollection* props) {
+  // Increment the ref count
+  mutex_.Lock();
+  auto version = versions_->current();
+  version->Ref();
+  mutex_.Unlock();
+
+  auto s = version->GetPropertiesOfAllTables(props);
+
+  // Decrement the ref count
+  mutex_.Lock();
+  version->Unref();
+  mutex_.Unlock();
+
+  return s;
+}
+
 const std::string& DBImpl::GetName() const {
   return dbname_;
 }

db/db_impl.h

@@ -494,6 +494,9 @@ class DBImpl : public DB {
   void InstallSuperVersion(ColumnFamilyData* cfd,
                            DeletionState& deletion_state);
 
+  virtual Status GetPropertiesOfAllTables(TablePropertiesCollection* props)
+      override;
+
   // Function that Get and KeyMayExist call with no_io true or false
   // Note: 'value_found' from KeyMayExist propagates here
   Status GetImpl(const ReadOptions& options, ColumnFamilyHandle* column_family,

db/db_test.cc

@@ -10,6 +10,7 @@
 #include <algorithm>
 #include <set>
 #include <unistd.h>
+#include <unordered_set>
 
 #include "db/dbformat.h"
 #include "db/db_impl.h"
@@ -26,6 +27,7 @@
 #include "rocksdb/slice.h"
 #include "rocksdb/slice_transform.h"
 #include "rocksdb/table.h"
+#include "rocksdb/table_properties.h"
 #include "table/block_based_table_factory.h"
 #include "util/hash.h"
 #include "util/hash_linklist_rep.h"
@@ -953,6 +955,28 @@ static long TestGetTickerCount(const Options& options, Tickers ticker_type) {
   return options.statistics->getTickerCount(ticker_type);
 }
 
+// A helper function that ensures the table properties returned in
+// `GetPropertiesOfAllTablesTest` is correct.
+// This test assumes entries size is differnt for each of the tables.
+void VerifyTableProperties(DB* db, uint64_t expected_entries_size) {
+  TablePropertiesCollection props;
+  ASSERT_OK(db->GetPropertiesOfAllTables(&props));
+
+  assert(props.size() == 4);
+  ASSERT_EQ(4U, props.size());
+  std::unordered_set<uint64_t> unique_entries;
+
+  // Indirect test
+  uint64_t sum = 0;
+  for (const auto& item : props) {
+    unique_entries.insert(item.second->num_entries);
+    sum += item.second->num_entries;
+  }
+
+  ASSERT_EQ(props.size(), unique_entries.size());
+  ASSERT_EQ(expected_entries_size, sum);
+}
+
 TEST(DBTest, Empty) {
   do {
     Options options = CurrentOptions();
@@ -1041,6 +1065,41 @@ TEST(DBTest, IndexAndFilterBlocksOfNewTableAddedToCache) {
             TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
 }
 
+TEST(DBTest, GetPropertiesOfAllTablesTest) {
+  Options options = CurrentOptions();
+  Reopen(&options);
+  // Create 4 tables
+  for (int table = 0; table < 4; ++table) {
+    for (int i = 0; i < 10 + table; ++i) {
+      db_->Put(WriteOptions(), std::to_string(table * 100 + i), "val");
+    }
+    db_->Flush(FlushOptions());
+  }
+
+  // 1. Read table properties directly from file
+  Reopen(&options);
+  VerifyTableProperties(db_, 10 + 11 + 12 + 13);
+
+  // 2. Put two tables to table cache and
+  Reopen(&options);
+  // fetch key from 1st and 2nd table, which will internally place that table to
+  // the table cache.
+  for (int i = 0; i < 2; ++i) {
+    Get(std::to_string(i * 100 + 0));
+  }
+
+  VerifyTableProperties(db_, 10 + 11 + 12 + 13);
+
+  // 3. Put all tables to table cache
+  Reopen(&options);
+  // fetch key from 1st and 2nd table, which will internally place that table to
+  // the table cache.
+  for (int i = 0; i < 4; ++i) {
+    Get(std::to_string(i * 100 + 0));
+  }
+  VerifyTableProperties(db_, 10 + 11 + 12 + 13);
+}
+
 TEST(DBTest, LevelLimitReopen) {
   Options options = CurrentOptions();
   CreateAndReopenWithCF({"pikachu"}, &options);
@@ -5073,6 +5132,11 @@ class ModelDB: public DB {
                           Status::NotSupported("Not implemented."));
     return s;
   }
+
+  virtual Status GetPropertiesOfAllTables(TablePropertiesCollection* props) {
+    return Status();
+  }
+
   using DB::KeyMayExist;
   virtual bool KeyMayExist(const ReadOptions& options,
                            ColumnFamilyHandle* column_family, const Slice& key,

db/plain_table_db_test.cc

@@ -22,7 +22,9 @@
 #include "rocksdb/filter_policy.h"
 #include "rocksdb/slice_transform.h"
 #include "rocksdb/table.h"
+#include "table/meta_blocks.h"
 #include "table/plain_table_factory.h"
+#include "table/plain_table_reader.h"
 #include "util/hash.h"
 #include "util/logging.h"
 #include "util/mutexlock.h"
@@ -46,6 +48,7 @@ class PlainTableDBTest {
 
  public:
   PlainTableDBTest() : env_(Env::Default()) {
+    ro_.prefix_seek = true;
     dbname_ = test::TmpDir() + "/plain_table_db_test";
     ASSERT_OK(DestroyDB(dbname_, Options()));
     db_ = nullptr;
@@ -57,10 +60,12 @@ class PlainTableDBTest {
     ASSERT_OK(DestroyDB(dbname_, Options()));
   }
 
+  ReadOptions ro_;
+
   // Return the current option configuration.
   Options CurrentOptions() {
     Options options;
-    options.table_factory.reset(new PlainTableFactory(16, 2, 0.8));
+    options.table_factory.reset(NewPlainTableFactory(16, 2, 0.8, 3));
     options.prefix_extractor = prefix_transform.get();
     options.allow_mmap_reads = true;
     return options;
@@ -119,7 +124,7 @@ class PlainTableDBTest {
   }
 
   std::string Get(const std::string& k, const Snapshot* snapshot = nullptr) {
-    ReadOptions options;
+    ReadOptions options = ro_;
     options.snapshot = snapshot;
     std::string result;
     Status s = db_->Get(options, k, &result);
@@ -176,25 +181,296 @@ TEST(PlainTableDBTest, Empty) {
   ASSERT_EQ("NOT_FOUND", Get("0000000000000foo"));
 }
 
-TEST(PlainTableDBTest, ReadWrite) {
+class TestPlainTableReader : public PlainTableReader {
-  ASSERT_OK(Put("1000000000000foo", "v1"));
+ public:
-  ASSERT_EQ("v1", Get("1000000000000foo"));
+  TestPlainTableReader(const EnvOptions& storage_options,
-  ASSERT_OK(Put("0000000000000bar", "v2"));
+                       const InternalKeyComparator& icomparator,
-  ASSERT_OK(Put("1000000000000foo", "v3"));
+                       uint64_t file_size, int bloom_bits_per_key,
-  ASSERT_EQ("v3", Get("1000000000000foo"));
+                       double hash_table_ratio, size_t index_sparseness,
-  ASSERT_EQ("v2", Get("0000000000000bar"));
+                       const TableProperties* table_properties,
-}
+                       unique_ptr<RandomAccessFile>&& file,
+                       const Options& options, bool* expect_bloom_not_match)
+      : PlainTableReader(options, std::move(file), storage_options, icomparator,
+                         file_size, bloom_bits_per_key, hash_table_ratio,
+                         index_sparseness, table_properties),
+        expect_bloom_not_match_(expect_bloom_not_match) {
+    Status s = PopulateIndex();
+    ASSERT_TRUE(s.ok());
+  }
+
+  virtual ~TestPlainTableReader() {}
+
+ private:
+  virtual bool MatchBloom(uint32_t hash) const override {
+    bool ret = PlainTableReader::MatchBloom(hash);
+    ASSERT_TRUE(!*expect_bloom_not_match_ || !ret);
+    return ret;
+  }
+  bool* expect_bloom_not_match_;
+};
+
+extern const uint64_t kPlainTableMagicNumber;
+class TestPlainTableFactory : public PlainTableFactory {
+ public:
+  explicit TestPlainTableFactory(bool* expect_bloom_not_match,
+                                 uint32_t user_key_len =
+                                     kPlainTableVariableLength,
+                                 int bloom_bits_per_key = 0,
+                                 double hash_table_ratio = 0.75,
+                                 size_t index_sparseness = 16)
+      : PlainTableFactory(user_key_len, user_key_len, hash_table_ratio,
+                          hash_table_ratio),
+        user_key_len_(user_key_len),
+        bloom_bits_per_key_(bloom_bits_per_key),
+        hash_table_ratio_(hash_table_ratio),
+        index_sparseness_(index_sparseness),
+        expect_bloom_not_match_(expect_bloom_not_match) {}
+
+  Status NewTableReader(const Options& options, const EnvOptions& soptions,
+                        const InternalKeyComparator& internal_comparator,
+                        unique_ptr<RandomAccessFile>&& file, uint64_t file_size,
+                        unique_ptr<TableReader>* table) const override {
+    TableProperties* props = nullptr;
+    auto s = ReadTableProperties(file.get(), file_size, kPlainTableMagicNumber,
+                                 options.env, options.info_log.get(), &props);
+    ASSERT_TRUE(s.ok());
+
+    std::unique_ptr<PlainTableReader> new_reader(new TestPlainTableReader(
+        soptions, internal_comparator, file_size, bloom_bits_per_key_,
+        hash_table_ratio_, index_sparseness_, props, std::move(file), options,
+        expect_bloom_not_match_));
+
+    *table = std::move(new_reader);
+    return s;
+  }
+
+ private:
+  uint32_t user_key_len_;
+  int bloom_bits_per_key_;
+  double hash_table_ratio_;
+  size_t index_sparseness_;
+  bool* expect_bloom_not_match_;
+};
 
 TEST(PlainTableDBTest, Flush) {
-  ASSERT_OK(Put("1000000000000foo", "v1"));
+  for (int bloom_bits = 0; bloom_bits <= 8; bloom_bits += 8) {
-  ASSERT_OK(Put("0000000000000bar", "v2"));
+    for (int total_order = 0; total_order <= 1; total_order++) {
-  ASSERT_OK(Put("1000000000000foo", "v3"));
+      Options options = CurrentOptions();
-  dbfull()->TEST_FlushMemTable();
+      options.create_if_missing = true;
-  ASSERT_EQ("v3", Get("1000000000000foo"));
+      // Set only one bucket to force bucket conflict.
-  ASSERT_EQ("v2", Get("0000000000000bar"));
+      // Test index interval for the same prefix to be 1, 2 and 4
+      if (total_order) {
+        options.table_factory.reset(
+            NewTotalOrderPlainTableFactory(16, bloom_bits, 2));
+      } else {
+        options.table_factory.reset(NewPlainTableFactory(16, bloom_bits));
+      }
+      DestroyAndReopen(&options);
+
+      ASSERT_OK(Put("1000000000000foo", "v1"));
+      ASSERT_OK(Put("0000000000000bar", "v2"));
+      ASSERT_OK(Put("1000000000000foo", "v3"));
+      dbfull()->TEST_FlushMemTable();
+      ASSERT_EQ("v3", Get("1000000000000foo"));
+      ASSERT_EQ("v2", Get("0000000000000bar"));
+    }
+  }
+}
+
+TEST(PlainTableDBTest, Flush2) {
+  for (int bloom_bits = 0; bloom_bits <= 10; bloom_bits += 10) {
+    for (int total_order = 0; total_order <= 1; total_order++) {
+      bool expect_bloom_not_match = false;
+      Options options = CurrentOptions();
+      options.create_if_missing = true;
+      // Set only one bucket to force bucket conflict.
+      // Test index interval for the same prefix to be 1, 2 and 4
+      if (total_order) {
+        options.prefix_extractor = nullptr;
+        options.table_factory.reset(new TestPlainTableFactory(
+            &expect_bloom_not_match, 16, bloom_bits, 0, 2));
+      } else {
+        options.table_factory.reset(
+            new TestPlainTableFactory(&expect_bloom_not_match, 16, bloom_bits));
+      }
+      DestroyAndReopen(&options);
+      ASSERT_OK(Put("0000000000000bar", "b"));
+      ASSERT_OK(Put("1000000000000foo", "v1"));
+      dbfull()->TEST_FlushMemTable();
+
+      ASSERT_OK(Put("1000000000000foo", "v2"));
+      dbfull()->TEST_FlushMemTable();
+      ASSERT_EQ("v2", Get("1000000000000foo"));
+
+      ASSERT_OK(Put("0000000000000eee", "v3"));
+      dbfull()->TEST_FlushMemTable();
+      ASSERT_EQ("v3", Get("0000000000000eee"));
+
+      ASSERT_OK(Delete("0000000000000bar"));
+      dbfull()->TEST_FlushMemTable();
+      ASSERT_EQ("NOT_FOUND", Get("0000000000000bar"));
+
+      ASSERT_OK(Put("0000000000000eee", "v5"));
+      ASSERT_OK(Put("9000000000000eee", "v5"));
+      dbfull()->TEST_FlushMemTable();
+      ASSERT_EQ("v5", Get("0000000000000eee"));
+
+      // Test Bloom Filter
+      if (bloom_bits > 0) {
+        // Neither key nor value should exist.
+        expect_bloom_not_match = true;
+        ASSERT_EQ("NOT_FOUND", Get("5_not00000000bar"));
+
+        // Key doesn't exist any more but prefix exists.
+        if (total_order) {
+          ASSERT_EQ("NOT_FOUND", Get("1000000000000not"));
+          ASSERT_EQ("NOT_FOUND", Get("0000000000000not"));
+        }
+        expect_bloom_not_match = false;
+      }
+    }
+  }
 }
 
 TEST(PlainTableDBTest, Iterator) {
+  for (int bloom_bits = 0; bloom_bits <= 8; bloom_bits += 8) {
+    for (int total_order = 0; total_order <= 1; total_order++) {
+      bool expect_bloom_not_match = false;
+      Options options = CurrentOptions();
+      options.create_if_missing = true;
+      // Set only one bucket to force bucket conflict.
+      // Test index interval for the same prefix to be 1, 2 and 4
+      if (total_order) {
+        options.prefix_extractor = nullptr;
+        options.table_factory.reset(new TestPlainTableFactory(
+            &expect_bloom_not_match, 16, bloom_bits, 0, 2));
+      } else {
+        options.table_factory.reset(
+            new TestPlainTableFactory(&expect_bloom_not_match, 16, bloom_bits));
+      }
+      DestroyAndReopen(&options);
+
+      ASSERT_OK(Put("1000000000foo002", "v_2"));
+      ASSERT_OK(Put("0000000000000bar", "random"));
+      ASSERT_OK(Put("1000000000foo001", "v1"));
+      ASSERT_OK(Put("3000000000000bar", "bar_v"));
+      ASSERT_OK(Put("1000000000foo003", "v__3"));
+      ASSERT_OK(Put("1000000000foo004", "v__4"));
+      ASSERT_OK(Put("1000000000foo005", "v__5"));
+      ASSERT_OK(Put("1000000000foo007", "v__7"));
+      ASSERT_OK(Put("1000000000foo008", "v__8"));
+      dbfull()->TEST_FlushMemTable();
+      ASSERT_EQ("v1", Get("1000000000foo001"));
+      ASSERT_EQ("v__3", Get("1000000000foo003"));
+      Iterator* iter = dbfull()->NewIterator(ro_);
+      iter->Seek("1000000000foo000");
+      ASSERT_TRUE(iter->Valid());
+      ASSERT_EQ("1000000000foo001", iter->key().ToString());
+      ASSERT_EQ("v1", iter->value().ToString());
+
+      iter->Next();
+      ASSERT_TRUE(iter->Valid());
+      ASSERT_EQ("1000000000foo002", iter->key().ToString());
+      ASSERT_EQ("v_2", iter->value().ToString());
+
+      iter->Next();
+      ASSERT_TRUE(iter->Valid());
+      ASSERT_EQ("1000000000foo003", iter->key().ToString());
+      ASSERT_EQ("v__3", iter->value().ToString());
+
+      iter->Next();
+      ASSERT_TRUE(iter->Valid());
+      ASSERT_EQ("1000000000foo004", iter->key().ToString());
+      ASSERT_EQ("v__4", iter->value().ToString());
+
+      iter->Seek("3000000000000bar");
+      ASSERT_TRUE(iter->Valid());
+      ASSERT_EQ("3000000000000bar", iter->key().ToString());
+      ASSERT_EQ("bar_v", iter->value().ToString());
+
+      iter->Seek("1000000000foo000");
+      ASSERT_TRUE(iter->Valid());
+      ASSERT_EQ("1000000000foo001", iter->key().ToString());
+      ASSERT_EQ("v1", iter->value().ToString());
+
+      iter->Seek("1000000000foo005");
+      ASSERT_TRUE(iter->Valid());
+      ASSERT_EQ("1000000000foo005", iter->key().ToString());
+      ASSERT_EQ("v__5", iter->value().ToString());
+
+      iter->Seek("1000000000foo006");
+      ASSERT_TRUE(iter->Valid());
+      ASSERT_EQ("1000000000foo007", iter->key().ToString());
+      ASSERT_EQ("v__7", iter->value().ToString());
+
+      iter->Seek("1000000000foo008");
+      ASSERT_TRUE(iter->Valid());
+      ASSERT_EQ("1000000000foo008", iter->key().ToString());
+      ASSERT_EQ("v__8", iter->value().ToString());
+
+      if (total_order == 0) {
+        iter->Seek("1000000000foo009");
+        ASSERT_TRUE(iter->Valid());
+        ASSERT_EQ("3000000000000bar", iter->key().ToString());
+      }
+
+      // Test Bloom Filter
+      if (bloom_bits > 0) {
+        // Neither key nor value should exist.
+        expect_bloom_not_match = true;
+        iter->Seek("2not000000000bar");
+        ASSERT_TRUE(!iter->Valid());
+
+        // Key doesn't exist any more but prefix exists.
+        if (total_order) {
+          iter->Seek("2not000000000bar");
+          ASSERT_TRUE(!iter->Valid());
+        }
+        expect_bloom_not_match = false;
+      }
+
+      delete iter;
+    }
+  }
+}
+
+// A test comparator which compare two strings in this way:
+// (1) first compare prefix of 8 bytes in alphabet order,
+// (2) if two strings share the same prefix, sort the other part of the string
+//     in the reverse alphabet order.
+class SimpleSuffixReverseComparator : public Comparator {
+ public:
+  SimpleSuffixReverseComparator() {}
+
+  virtual const char* Name() const { return "SimpleSuffixReverseComparator"; }
+
+  virtual int Compare(const Slice& a, const Slice& b) const {
+    Slice prefix_a = Slice(a.data(), 8);
+    Slice prefix_b = Slice(b.data(), 8);
+    int prefix_comp = prefix_a.compare(prefix_b);
+    if (prefix_comp != 0) {
+      return prefix_comp;
+    } else {
+      Slice suffix_a = Slice(a.data() + 8, a.size() - 8);
+      Slice suffix_b = Slice(b.data() + 8, b.size() - 8);
+      return -(suffix_a.compare(suffix_b));
+    }
+  }
+  virtual void FindShortestSeparator(std::string* start,
+                                     const Slice& limit) const {}
+
+  virtual void FindShortSuccessor(std::string* key) const {}
+};
+
+TEST(PlainTableDBTest, IteratorReverseSuffixComparator) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  // Set only one bucket to force bucket conflict.
+  // Test index interval for the same prefix to be 1, 2 and 4
+  SimpleSuffixReverseComparator comp;
+  options.comparator = &comp;
+  DestroyAndReopen(&options);
+
   ASSERT_OK(Put("1000000000foo002", "v_2"));
   ASSERT_OK(Put("0000000000000bar", "random"));
   ASSERT_OK(Put("1000000000foo001", "v1"));
@@ -207,22 +483,21 @@ TEST(PlainTableDBTest, Iterator) {
   dbfull()->TEST_FlushMemTable();
   ASSERT_EQ("v1", Get("1000000000foo001"));
   ASSERT_EQ("v__3", Get("1000000000foo003"));
-  ReadOptions ro;
+  Iterator* iter = dbfull()->NewIterator(ro_);
-  Iterator* iter = dbfull()->NewIterator(ro);
+  iter->Seek("1000000000foo009");
-  iter->Seek("1000000000foo001");
   ASSERT_TRUE(iter->Valid());
-  ASSERT_EQ("1000000000foo001", iter->key().ToString());
+  ASSERT_EQ("1000000000foo008", iter->key().ToString());
-  ASSERT_EQ("v1", iter->value().ToString());
+  ASSERT_EQ("v__8", iter->value().ToString());
 
   iter->Next();
   ASSERT_TRUE(iter->Valid());
-  ASSERT_EQ("1000000000foo002", iter->key().ToString());
+  ASSERT_EQ("1000000000foo007", iter->key().ToString());
-  ASSERT_EQ("v_2", iter->value().ToString());
+  ASSERT_EQ("v__7", iter->value().ToString());
 
   iter->Next();
   ASSERT_TRUE(iter->Valid());
-  ASSERT_EQ("1000000000foo003", iter->key().ToString());
+  ASSERT_EQ("1000000000foo005", iter->key().ToString());
-  ASSERT_EQ("v__3", iter->value().ToString());
+  ASSERT_EQ("v__5", iter->value().ToString());
 
   iter->Next();
   ASSERT_TRUE(iter->Valid());
@@ -234,11 +509,6 @@ TEST(PlainTableDBTest, Iterator) {
   ASSERT_EQ("3000000000000bar", iter->key().ToString());
   ASSERT_EQ("bar_v", iter->value().ToString());
 
-  iter->Seek("1000000000foo000");
-  ASSERT_TRUE(iter->Valid());
-  ASSERT_EQ("1000000000foo001", iter->key().ToString());
-  ASSERT_EQ("v1", iter->value().ToString());
-
   iter->Seek("1000000000foo005");
   ASSERT_TRUE(iter->Valid());
   ASSERT_EQ("1000000000foo005", iter->key().ToString());
@@ -246,42 +516,220 @@ TEST(PlainTableDBTest, Iterator) {
 
   iter->Seek("1000000000foo006");
   ASSERT_TRUE(iter->Valid());
-  ASSERT_EQ("1000000000foo007", iter->key().ToString());
+  ASSERT_EQ("1000000000foo005", iter->key().ToString());
-  ASSERT_EQ("v__7", iter->value().ToString());
+  ASSERT_EQ("v__5", iter->value().ToString());
 
   iter->Seek("1000000000foo008");
   ASSERT_TRUE(iter->Valid());
   ASSERT_EQ("1000000000foo008", iter->key().ToString());
   ASSERT_EQ("v__8", iter->value().ToString());
 
-  iter->Seek("1000000000foo009");
+  iter->Seek("1000000000foo000");
   ASSERT_TRUE(iter->Valid());
   ASSERT_EQ("3000000000000bar", iter->key().ToString());
 
-
   delete iter;
 }
 
-TEST(PlainTableDBTest, Flush2) {
+TEST(PlainTableDBTest, HashBucketConflict) {
-  ASSERT_OK(Put("0000000000000bar", "b"));
+  for (unsigned char i = 1; i <= 3; i++) {
-  ASSERT_OK(Put("1000000000000foo", "v1"));
+    Options options = CurrentOptions();
-  dbfull()->TEST_FlushMemTable();
+    options.create_if_missing = true;
+    // Set only one bucket to force bucket conflict.
+    // Test index interval for the same prefix to be 1, 2 and 4
+    options.table_factory.reset(NewTotalOrderPlainTableFactory(16, 0, 2 ^ i));
+    DestroyAndReopen(&options);
+    ASSERT_OK(Put("5000000000000fo0", "v1"));
+    ASSERT_OK(Put("5000000000000fo1", "v2"));
+    ASSERT_OK(Put("5000000000000fo2", "v"));
+    ASSERT_OK(Put("2000000000000fo0", "v3"));
+    ASSERT_OK(Put("2000000000000fo1", "v4"));
+    ASSERT_OK(Put("2000000000000fo2", "v"));
+    ASSERT_OK(Put("2000000000000fo3", "v"));
+
+    dbfull()->TEST_FlushMemTable();
+
+    ASSERT_EQ("v1", Get("5000000000000fo0"));
+    ASSERT_EQ("v2", Get("5000000000000fo1"));
+    ASSERT_EQ("v3", Get("2000000000000fo0"));
+    ASSERT_EQ("v4", Get("2000000000000fo1"));
+
+    ASSERT_EQ("NOT_FOUND", Get("5000000000000bar"));
+    ASSERT_EQ("NOT_FOUND", Get("2000000000000bar"));
+    ASSERT_EQ("NOT_FOUND", Get("5000000000000fo8"));
+    ASSERT_EQ("NOT_FOUND", Get("2000000000000fo8"));
+
+    ReadOptions ro;
+    Iterator* iter = dbfull()->NewIterator(ro);
+
+    iter->Seek("5000000000000fo0");
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("5000000000000fo0", iter->key().ToString());
+    iter->Next();
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("5000000000000fo1", iter->key().ToString());
+
+    iter->Seek("5000000000000fo1");
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("5000000000000fo1", iter->key().ToString());
+
+    iter->Seek("2000000000000fo0");
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("2000000000000fo0", iter->key().ToString());
+    iter->Next();
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("2000000000000fo1", iter->key().ToString());
+
+    iter->Seek("2000000000000fo1");
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("2000000000000fo1", iter->key().ToString());
+
+    iter->Seek("2000000000000bar");
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("2000000000000fo0", iter->key().ToString());
+
+    iter->Seek("5000000000000bar");
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("5000000000000fo0", iter->key().ToString());
+
+    iter->Seek("2000000000000fo8");
+    ASSERT_TRUE(!iter->Valid() ||
+                options.comparator->Compare(iter->key(), "20000001") > 0);
+
+    iter->Seek("5000000000000fo8");
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->Seek("1000000000000fo2");
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->Seek("3000000000000fo2");
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->Seek("8000000000000fo2");
+    ASSERT_TRUE(!iter->Valid());
+
+    delete iter;
+  }
+}
 
-  ASSERT_OK(Put("1000000000000foo", "v2"));
+TEST(PlainTableDBTest, HashBucketConflictReverseSuffixComparator) {
-  dbfull()->TEST_FlushMemTable();
+  for (unsigned char i = 1; i <= 3; i++) {
-  ASSERT_EQ("v2", Get("1000000000000foo"));
+    Options options = CurrentOptions();
+    options.create_if_missing = true;
+    SimpleSuffixReverseComparator comp;
+    options.comparator = &comp;
+    // Set only one bucket to force bucket conflict.
+    // Test index interval for the same prefix to be 1, 2 and 4
+    options.table_factory.reset(NewTotalOrderPlainTableFactory(16, 0, 2 ^ i));
+    DestroyAndReopen(&options);
+    ASSERT_OK(Put("5000000000000fo0", "v1"));
+    ASSERT_OK(Put("5000000000000fo1", "v2"));
+    ASSERT_OK(Put("5000000000000fo2", "v"));
+    ASSERT_OK(Put("2000000000000fo0", "v3"));
+    ASSERT_OK(Put("2000000000000fo1", "v4"));
+    ASSERT_OK(Put("2000000000000fo2", "v"));
+    ASSERT_OK(Put("2000000000000fo3", "v"));
+
+    dbfull()->TEST_FlushMemTable();
+
+    ASSERT_EQ("v1", Get("5000000000000fo0"));
+    ASSERT_EQ("v2", Get("5000000000000fo1"));
+    ASSERT_EQ("v3", Get("2000000000000fo0"));
+    ASSERT_EQ("v4", Get("2000000000000fo1"));
+
+    ASSERT_EQ("NOT_FOUND", Get("5000000000000bar"));
+    ASSERT_EQ("NOT_FOUND", Get("2000000000000bar"));
+    ASSERT_EQ("NOT_FOUND", Get("5000000000000fo8"));
+    ASSERT_EQ("NOT_FOUND", Get("2000000000000fo8"));
+
+    ReadOptions ro;
+    Iterator* iter = dbfull()->NewIterator(ro);
+
+    iter->Seek("5000000000000fo1");
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("5000000000000fo1", iter->key().ToString());
+    iter->Next();
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("5000000000000fo0", iter->key().ToString());
+
+    iter->Seek("5000000000000fo1");
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("5000000000000fo1", iter->key().ToString());
+
+    iter->Seek("2000000000000fo1");
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("2000000000000fo1", iter->key().ToString());
+    iter->Next();
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("2000000000000fo0", iter->key().ToString());
+
+    iter->Seek("2000000000000fo1");
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("2000000000000fo1", iter->key().ToString());
+
+    iter->Seek("2000000000000var");
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("2000000000000fo3", iter->key().ToString());
+
+    iter->Seek("5000000000000var");
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("5000000000000fo2", iter->key().ToString());
+
+    std::string seek_key = "2000000000000bar";
+    iter->Seek(seek_key);
+    ASSERT_TRUE(!iter->Valid() ||
+                options.prefix_extractor->Transform(iter->key()) !=
+                    options.prefix_extractor->Transform(seek_key));
+
+    iter->Seek("1000000000000fo2");
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->Seek("3000000000000fo2");
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->Seek("8000000000000fo2");
+    ASSERT_TRUE(!iter->Valid());
+
+    delete iter;
+  }
+}
 
-  ASSERT_OK(Put("0000000000000eee", "v3"));
+TEST(PlainTableDBTest, NonExistingKeyToNonEmptyBucket) {
-  dbfull()->TEST_FlushMemTable();
+  Options options = CurrentOptions();
-  ASSERT_EQ("v3", Get("0000000000000eee"));
+  options.create_if_missing = true;
+  // Set only one bucket to force bucket conflict.
+  // Test index interval for the same prefix to be 1, 2 and 4
+  options.table_factory.reset(NewTotalOrderPlainTableFactory(16, 0, 5));
+  DestroyAndReopen(&options);
+  ASSERT_OK(Put("5000000000000fo0", "v1"));
+  ASSERT_OK(Put("5000000000000fo1", "v2"));
+  ASSERT_OK(Put("5000000000000fo2", "v3"));
 
-  ASSERT_OK(Delete("0000000000000bar"));
   dbfull()->TEST_FlushMemTable();
-  ASSERT_EQ("NOT_FOUND", Get("0000000000000bar"));
 
-  ASSERT_OK(Put("0000000000000eee", "v5"));
+  ASSERT_EQ("v1", Get("5000000000000fo0"));
-  dbfull()->TEST_FlushMemTable();
+  ASSERT_EQ("v2", Get("5000000000000fo1"));
-  ASSERT_EQ("v5", Get("0000000000000eee"));
+  ASSERT_EQ("v3", Get("5000000000000fo2"));
+
+  ASSERT_EQ("NOT_FOUND", Get("8000000000000bar"));
+  ASSERT_EQ("NOT_FOUND", Get("1000000000000bar"));
+
+  Iterator* iter = dbfull()->NewIterator(ro_);
+
+  iter->Seek("5000000000000bar");
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("5000000000000fo0", iter->key().ToString());
+
+  iter->Seek("5000000000000fo8");
+  ASSERT_TRUE(!iter->Valid());
+
+  iter->Seek("1000000000000fo2");
+  ASSERT_TRUE(!iter->Valid());
+
+  iter->Seek("8000000000000fo2");
+  ASSERT_TRUE(!iter->Valid());
+
+  delete iter;
 }
 
 static std::string Key(int i) {

db/table_cache.cc

@@ -158,6 +158,32 @@ Status TableCache::Get(const ReadOptions& options,
   }
   return s;
 }
+Status TableCache::GetTableProperties(
+    const EnvOptions& toptions,
+    const InternalKeyComparator& internal_comparator,
+    const FileMetaData& file_meta,
+    std::shared_ptr<const TableProperties>* properties, bool no_io) {
+  Status s;
+  auto table_handle = file_meta.table_reader_handle;
+  // table already been pre-loaded?
+  if (table_handle) {
+    auto table = GetTableReaderFromHandle(table_handle);
+    *properties = table->GetTableProperties();
+    return s;
+  }
+
+  bool table_io;
+  s = FindTable(toptions, internal_comparator, file_meta.number,
+                file_meta.file_size, &table_handle, &table_io, no_io);
+  if (!s.ok()) {
+    return s;
+  }
+  assert(table_handle);
+  auto table = GetTableReaderFromHandle(table_handle);
+  *properties = table->GetTableProperties();
+  ReleaseHandle(table_handle);
+  return s;
+}
 
 bool TableCache::PrefixMayMatch(const ReadOptions& options,
                                 const InternalKeyComparator& icomparator,

db/table_cache.h

@@ -73,6 +73,18 @@ class TableCache {
   // Get TableReader from a cache handle.
   TableReader* GetTableReaderFromHandle(Cache::Handle* handle);
 
+  // Get the table properties of a given table.
+  // @no_io: indicates if we should load table to the cache if it is not present
+  //         in table cache yet.
+  // @returns: `properties` will be reset on success. Please note that we will
+  //            return Status::Incomplete() if table is not present in cache and
+  //            we set `no_io` to be true.
+  Status GetTableProperties(const EnvOptions& toptions,
+                            const InternalKeyComparator& internal_comparator,
+                            const FileMetaData& file_meta,
+                            std::shared_ptr<const TableProperties>* properties,
+                            bool no_io = false);
+
   // Release the handle from a cache
   void ReleaseHandle(Cache::Handle* handle);
 

db/version_set.cc

@@ -27,6 +27,8 @@
 #include "table/table_reader.h"
 #include "table/merger.h"
 #include "table/two_level_iterator.h"
+#include "table/format.h"
+#include "table/meta_blocks.h"
 #include "util/coding.h"
 #include "util/logging.h"
 #include "util/stop_watch.h"
@@ -241,6 +243,59 @@ bool Version::PrefixMayMatch(const ReadOptions& options,
   return may_match;
 }
 
+Status Version::GetPropertiesOfAllTables(TablePropertiesCollection* props) {
+  auto table_cache = vset_->table_cache_;
+  auto options = vset_->options_;
+  for (int level = 0; level < num_levels_; level++) {
+    for (const auto& file_meta : files_[level]) {
+      auto fname = TableFileName(vset_->dbname_, file_meta->number);
+      // 1. If the table is already present in table cache, load table
+      // properties from there.
+      std::shared_ptr<const TableProperties> table_properties;
+      Status s = table_cache->GetTableProperties(
+          vset_->storage_options_, vset_->icmp_, *file_meta, &table_properties,
+          true /* no io */);
+      if (s.ok()) {
+        props->insert({fname, table_properties});
+        continue;
+      }
+
+      // We only ignore error type `Incomplete` since it's by design that we
+      // disallow table when it's not in table cache.
+      if (!s.IsIncomplete()) {
+        return s;
+      }
+
+      // 2. Table is not present in table cache, we'll read the table properties
+      // directly from the properties block in the file.
+      std::unique_ptr<RandomAccessFile> file;
+      s = vset_->env_->NewRandomAccessFile(fname, &file,
+                                           vset_->storage_options_);
+      if (!s.ok()) {
+        return s;
+      }
+
+      TableProperties* raw_table_properties;
+      // By setting the magic number to kInvalidTableMagicNumber, we can by
+      // pass the magic number check in the footer.
+      s = ReadTableProperties(
+          file.get(), file_meta->file_size,
+          Footer::kInvalidTableMagicNumber /* table's magic number */,
+          vset_->env_, options->info_log.get(), &raw_table_properties);
+      if (!s.ok()) {
+        return s;
+      }
+      RecordTick(options->statistics.get(),
+                 NUMBER_DIRECT_LOAD_TABLE_PROPERTIES);
+
+      props->insert({fname, std::shared_ptr<const TableProperties>(
+                                raw_table_properties)});
+    }
+  }
+
+  return Status::OK();
+}
+
 Iterator* Version::NewConcatenatingIterator(const ReadOptions& options,
                                             const EnvOptions& soptions,
                                             int level) const {

db/version_set.h

@@ -193,6 +193,12 @@ class Version {
   // Returns the version nuber of this version
   uint64_t GetVersionNumber() const { return version_number_; }
 
+  // REQUIRES: lock is held
+  // On success, *props will be populated with all SSTables' table properties.
+  // The keys of `props` are the sst file name, the values of `props` are the
+  // tables' propertis, represented as shared_ptr.
+  Status GetPropertiesOfAllTables(TablePropertiesCollection* props);
+
   // used to sort files by size
   struct Fsize {
     int index;

include/rocksdb/db.h

@@ -14,6 +14,7 @@
 #include <memory>
 #include <vector>
 #include <string>
+#include <unordered_map>
 #include "rocksdb/iterator.h"
 #include "rocksdb/options.h"
 #include "rocksdb/types.h"
@@ -47,6 +48,7 @@ struct Options;
 struct ReadOptions;
 struct WriteOptions;
 struct FlushOptions;
+struct TableProperties;
 class WriteBatch;
 class Env;
 
@@ -78,6 +80,12 @@ struct Range {
   Range(const Slice& s, const Slice& l) : start(s), limit(l) { }
 };
 
+// A collections of table properties objects, where
+//  key: is the table's file name.
+//  value: the table properties object of the given table.
+typedef std::unordered_map<std::string, std::shared_ptr<const TableProperties>>
+    TablePropertiesCollection;
+
 // A DB is a persistent ordered map from keys to values.
 // A DB is safe for concurrent access from multiple threads without
 // any external synchronization.
@@ -427,6 +435,8 @@ class DB {
   // Returns default column family handle
   virtual ColumnFamilyHandle* DefaultColumnFamily() const = 0;
 
+  virtual Status GetPropertiesOfAllTables(TablePropertiesCollection* props) = 0;
+
  private:
   // No copying allowed
   DB(const DB&);

include/rocksdb/statistics.h

@@ -57,9 +57,9 @@ enum Tickers {
    * COMPACTION_KEY_DROP_* count the reasons for key drop during compaction
    * There are 3 reasons currently.
    */
-  COMPACTION_KEY_DROP_NEWER_ENTRY, // key was written with a newer value.
+  COMPACTION_KEY_DROP_NEWER_ENTRY,  // key was written with a newer value.
-  COMPACTION_KEY_DROP_OBSOLETE, // The key is obsolete.
+  COMPACTION_KEY_DROP_OBSOLETE,     // The key is obsolete.
-  COMPACTION_KEY_DROP_USER, // user compaction function has dropped the key.
+  COMPACTION_KEY_DROP_USER,  // user compaction function has dropped the key.
 
   // Number of keys written to the database via the Put and Write call's
   NUMBER_KEYS_WRITTEN,
@@ -80,8 +80,7 @@ enum Tickers {
   // write throttle because of too many files in L0
   STALL_L0_NUM_FILES_MICROS,
   RATE_LIMIT_DELAY_MILLIS,
-
+  NO_ITERATORS,  // number of iterators currently open
-  NO_ITERATORS, // number of iterators currently open
 
   // Number of MultiGet calls, keys read, and bytes read
   NUMBER_MULTIGET_CALLS,
@@ -107,77 +106,77 @@ enum Tickers {
   // Record the number of calls to GetUpadtesSince. Useful to keep track of
   // transaction log iterator refreshes
   GET_UPDATES_SINCE_CALLS,
-
+  BLOCK_CACHE_COMPRESSED_MISS,  // miss in the compressed block cache
-  BLOCK_CACHE_COMPRESSED_MISS, // miss in the compressed block cache
+  BLOCK_CACHE_COMPRESSED_HIT,   // hit in the compressed block cache
-  BLOCK_CACHE_COMPRESSED_HIT,  // hit in the compressed block cache
+  WAL_FILE_SYNCED,              // Number of times WAL sync is done
-
+  WAL_FILE_BYTES,               // Number of bytes written to WAL
-  WAL_FILE_SYNCED,  // Number of times WAL sync is done
-  WAL_FILE_BYTES,   // Number of bytes written to WAL
 
   // Writes can be processed by requesting thread or by the thread at the
   // head of the writers queue.
   WRITE_DONE_BY_SELF,
   WRITE_DONE_BY_OTHER,
+  WRITE_WITH_WAL,       // Number of Write calls that request WAL
+  COMPACT_READ_BYTES,   // Bytes read during compaction
+  COMPACT_WRITE_BYTES,  // Bytes written during compaction
 
-  WRITE_WITH_WAL,      // Number of Write calls that request WAL
+  // Number of table's properties loaded directly from file, without creating
-
+  // table reader object.
-  COMPACT_READ_BYTES,  // Bytes read during compaction
+  NUMBER_DIRECT_LOAD_TABLE_PROPERTIES,
-  COMPACT_WRITE_BYTES, // Bytes written during compaction
-
   TICKER_ENUM_MAX
 };
 
 // The order of items listed in  Tickers should be the same as
 // the order listed in TickersNameMap
 const std::vector<std::pair<Tickers, std::string>> TickersNameMap = {
-  { BLOCK_CACHE_MISS, "rocksdb.block.cache.miss" },
+    {BLOCK_CACHE_MISS, "rocksdb.block.cache.miss"},
-  { BLOCK_CACHE_HIT, "rocksdb.block.cache.hit" },
+    {BLOCK_CACHE_HIT, "rocksdb.block.cache.hit"},
-  { BLOCK_CACHE_ADD, "rocksdb.block.cache.add" },
+    {BLOCK_CACHE_ADD, "rocksdb.block.cache.add"},
-  { BLOCK_CACHE_INDEX_MISS, "rocksdb.block.cache.index.miss" },
+    {BLOCK_CACHE_INDEX_MISS, "rocksdb.block.cache.index.miss"},
-  { BLOCK_CACHE_INDEX_HIT, "rocksdb.block.cache.index.hit" },
+    {BLOCK_CACHE_INDEX_HIT, "rocksdb.block.cache.index.hit"},
-  { BLOCK_CACHE_FILTER_MISS, "rocksdb.block.cache.filter.miss" },
+    {BLOCK_CACHE_FILTER_MISS, "rocksdb.block.cache.filter.miss"},
-  { BLOCK_CACHE_FILTER_HIT, "rocksdb.block.cache.filter.hit" },
+    {BLOCK_CACHE_FILTER_HIT, "rocksdb.block.cache.filter.hit"},
-  { BLOCK_CACHE_DATA_MISS, "rocksdb.block.cache.data.miss" },
+    {BLOCK_CACHE_DATA_MISS, "rocksdb.block.cache.data.miss"},
-  { BLOCK_CACHE_DATA_HIT, "rocksdb.block.cache.data.hit" },
+    {BLOCK_CACHE_DATA_HIT, "rocksdb.block.cache.data.hit"},
-  { BLOOM_FILTER_USEFUL, "rocksdb.bloom.filter.useful" },
+    {BLOOM_FILTER_USEFUL, "rocksdb.bloom.filter.useful"},
-  { MEMTABLE_HIT, "rocksdb.memtable.hit" },
+    {MEMTABLE_HIT, "rocksdb.memtable.hit"},
-  { MEMTABLE_MISS, "rocksdb.memtable.miss" },
+    {MEMTABLE_MISS, "rocksdb.memtable.miss"},
-  { COMPACTION_KEY_DROP_NEWER_ENTRY, "rocksdb.compaction.key.drop.new" },
+    {COMPACTION_KEY_DROP_NEWER_ENTRY, "rocksdb.compaction.key.drop.new"},
-  { COMPACTION_KEY_DROP_OBSOLETE, "rocksdb.compaction.key.drop.obsolete" },
+    {COMPACTION_KEY_DROP_OBSOLETE, "rocksdb.compaction.key.drop.obsolete"},
-  { COMPACTION_KEY_DROP_USER, "rocksdb.compaction.key.drop.user" },
+    {COMPACTION_KEY_DROP_USER, "rocksdb.compaction.key.drop.user"},
-  { NUMBER_KEYS_WRITTEN, "rocksdb.number.keys.written" },
+    {NUMBER_KEYS_WRITTEN, "rocksdb.number.keys.written"},
-  { NUMBER_KEYS_READ, "rocksdb.number.keys.read" },
+    {NUMBER_KEYS_READ, "rocksdb.number.keys.read"},
-  { NUMBER_KEYS_UPDATED, "rocksdb.number.keys.updated" },
+    {NUMBER_KEYS_UPDATED, "rocksdb.number.keys.updated"},
-  { BYTES_WRITTEN, "rocksdb.bytes.written" },
+    {BYTES_WRITTEN, "rocksdb.bytes.written"},
-  { BYTES_READ, "rocksdb.bytes.read" },
+    {BYTES_READ, "rocksdb.bytes.read"},
-  { NO_FILE_CLOSES, "rocksdb.no.file.closes" },
+    {NO_FILE_CLOSES, "rocksdb.no.file.closes"},
-  { NO_FILE_OPENS, "rocksdb.no.file.opens" },
+    {NO_FILE_OPENS, "rocksdb.no.file.opens"},
-  { NO_FILE_ERRORS, "rocksdb.no.file.errors" },
+    {NO_FILE_ERRORS, "rocksdb.no.file.errors"},
-  { STALL_L0_SLOWDOWN_MICROS, "rocksdb.l0.slowdown.micros" },
+    {STALL_L0_SLOWDOWN_MICROS, "rocksdb.l0.slowdown.micros"},
-  { STALL_MEMTABLE_COMPACTION_MICROS, "rocksdb.memtable.compaction.micros" },
+    {STALL_MEMTABLE_COMPACTION_MICROS, "rocksdb.memtable.compaction.micros"},
-  { STALL_L0_NUM_FILES_MICROS, "rocksdb.l0.num.files.stall.micros" },
+    {STALL_L0_NUM_FILES_MICROS, "rocksdb.l0.num.files.stall.micros"},
-  { RATE_LIMIT_DELAY_MILLIS, "rocksdb.rate.limit.delay.millis" },
+    {RATE_LIMIT_DELAY_MILLIS, "rocksdb.rate.limit.delay.millis"},
-  { NO_ITERATORS, "rocksdb.num.iterators" },
+    {NO_ITERATORS, "rocksdb.num.iterators"},
-  { NUMBER_MULTIGET_CALLS, "rocksdb.number.multiget.get" },
+    {NUMBER_MULTIGET_CALLS, "rocksdb.number.multiget.get"},
-  { NUMBER_MULTIGET_KEYS_READ, "rocksdb.number.multiget.keys.read" },
+    {NUMBER_MULTIGET_KEYS_READ, "rocksdb.number.multiget.keys.read"},
-  { NUMBER_MULTIGET_BYTES_READ, "rocksdb.number.multiget.bytes.read" },
+    {NUMBER_MULTIGET_BYTES_READ, "rocksdb.number.multiget.bytes.read"},
-  { NUMBER_FILTERED_DELETES, "rocksdb.number.deletes.filtered" },
+    {NUMBER_FILTERED_DELETES, "rocksdb.number.deletes.filtered"},
-  { NUMBER_MERGE_FAILURES, "rocksdb.number.merge.failures" },
+    {NUMBER_MERGE_FAILURES, "rocksdb.number.merge.failures"},
-  { SEQUENCE_NUMBER, "rocksdb.sequence.number" },
+    {SEQUENCE_NUMBER, "rocksdb.sequence.number"},
-  { BLOOM_FILTER_PREFIX_CHECKED, "rocksdb.bloom.filter.prefix.checked" },
+    {BLOOM_FILTER_PREFIX_CHECKED, "rocksdb.bloom.filter.prefix.checked"},
-  { BLOOM_FILTER_PREFIX_USEFUL, "rocksdb.bloom.filter.prefix.useful" },
+    {BLOOM_FILTER_PREFIX_USEFUL, "rocksdb.bloom.filter.prefix.useful"},
-  { NUMBER_OF_RESEEKS_IN_ITERATION, "rocksdb.number.reseeks.iteration" },
+    {NUMBER_OF_RESEEKS_IN_ITERATION, "rocksdb.number.reseeks.iteration"},
-  { GET_UPDATES_SINCE_CALLS, "rocksdb.getupdatessince.calls" },
+    {GET_UPDATES_SINCE_CALLS, "rocksdb.getupdatessince.calls"},
-  { BLOCK_CACHE_COMPRESSED_MISS, "rocksdb.block.cachecompressed.miss" },
+    {BLOCK_CACHE_COMPRESSED_MISS, "rocksdb.block.cachecompressed.miss"},
-  { BLOCK_CACHE_COMPRESSED_HIT, "rocksdb.block.cachecompressed.hit" },
+    {BLOCK_CACHE_COMPRESSED_HIT, "rocksdb.block.cachecompressed.hit"},
-  { WAL_FILE_SYNCED, "rocksdb.wal.synced" },
+    {WAL_FILE_SYNCED, "rocksdb.wal.synced"},
-  { WAL_FILE_BYTES, "rocksdb.wal.bytes" },
+    {WAL_FILE_BYTES, "rocksdb.wal.bytes"},
-  { WRITE_DONE_BY_SELF, "rocksdb.write.self" },
+    {WRITE_DONE_BY_SELF, "rocksdb.write.self"},
-  { WRITE_DONE_BY_OTHER, "rocksdb.write.other" },
+    {WRITE_DONE_BY_OTHER, "rocksdb.write.other"},
-  { WRITE_WITH_WAL, "rocksdb.write.wal" },
+    {WRITE_WITH_WAL, "rocksdb.write.wal"},
-  { COMPACT_READ_BYTES, "rocksdb.compact.read.bytes" },
+    {COMPACT_READ_BYTES, "rocksdb.compact.read.bytes"},
-  { COMPACT_WRITE_BYTES, "rocksdb.compact.write.bytes" },
+    {COMPACT_WRITE_BYTES, "rocksdb.compact.write.bytes"},
-};
+    {NUMBER_DIRECT_LOAD_TABLE_PROPERTIES,
+     "rocksdb.number.direct.load.table.properties"}, };
 
 /**
  * Keep adding histogram's here.

include/rocksdb/table.h

@@ -60,20 +60,44 @@ struct BlockBasedTableOptions {
 extern TableFactory* NewBlockBasedTableFactory(
     const BlockBasedTableOptions& table_options = BlockBasedTableOptions());
 
-// -- Plain Table
+// -- Plain Table with prefix-only seek
+// For this factory, you need to set Options.prefix_extrator properly to make it
+// work. Look-up will starts with prefix hash lookup for key prefix. Inside the
+// hash bucket found, a binary search is executed for hash conflicts. Finally,
+// a linear search is used.
 // @user_key_len: plain table has optimization for fix-sized keys, which can be
 //                specified via user_key_len.  Alternatively, you can pass
 //                `kPlainTableVariableLength` if your keys have variable
 //                lengths.
-// @bloom_bits_per_key: the number of bits used for bloom filer per key. You may
+// @bloom_bits_per_key: the number of bits used for bloom filer per prefix. You
-//                  disable it by passing a zero.
+//                      may disable it by passing a zero.
 // @hash_table_ratio: the desired utilization of the hash table used for prefix
 //                    hashing. hash_table_ratio = number of prefixes / #buckets
 //                    in the hash table
+// @index_sparseness: inside each prefix, need to build one index record for how
+//                    many keys for binary search inside each hash bucket.
 const uint32_t kPlainTableVariableLength = 0;
-extern TableFactory* NewPlainTableFactory(
+extern TableFactory* NewPlainTableFactory(uint32_t user_key_len =
+                                              kPlainTableVariableLength,
+                                          int bloom_bits_per_prefix = 10,
+                                          double hash_table_ratio = 0.75,
+                                          size_t index_sparseness = 16);
+
+// -- Plain Table
+// This factory of plain table ignores Options.prefix_extractor and assumes no
+// hashable prefix available to the key structure. Lookup will be based on
+// binary search index only. Total order seek() can be issued.
+// @user_key_len: plain table has optimization for fix-sized keys, which can be
+//                specified via user_key_len.  Alternatively, you can pass
+//                `kPlainTableVariableLength` if your keys have variable
+//                lengths.
+// @bloom_bits_per_key: the number of bits used for bloom filer per key. You may
+//                  disable it by passing a zero.
+// @index_sparseness: need to build one index record for how many keys for
+//                    binary search.
+extern TableFactory* NewTotalOrderPlainTableFactory(
     uint32_t user_key_len = kPlainTableVariableLength,
-    int bloom_bits_per_key = 10, double hash_table_ratio = 0.75);
+    int bloom_bits_per_key = 0, size_t index_sparseness = 16);
 
 // A base class for table factories.
 class TableFactory {

include/utilities/stackable_db.h

@@ -182,6 +182,10 @@ class StackableDB : public DB {
     return db_->GetDbIdentity(identity);
   }
 
+  virtual Status GetPropertiesOfAllTables(TablePropertiesCollection* props) {
+    return db_->GetPropertiesOfAllTables(props);
+  }
+
   virtual Status GetUpdatesSince(SequenceNumber seq_number,
                                  unique_ptr<TransactionLogIterator>* iter)
     override {

linters/__phutil_library_map__.php

@@ -13,6 +13,7 @@ phutil_register_library_map(array(
     'FacebookFbcodeLintEngine' => 'lint_engine/FacebookFbcodeLintEngine.php',
     'FbcodeCppLinter' => 'cpp_linter/FbcodeCppLinter.php',
     'PfffCppLinter' => 'cpp_linter/PfffCppLinter.php',
+    'ArcanistCpplintLinter' => 'cpp_linter/ArcanistCpplintLinter.php',
   ),
   'function' =>
   array(

linters/cpp_linter/ArcanistCpplintLinter.php

@@ -0,0 +1,113 @@
+<?php
+
+/**
+ * Uses google's cpplint.py to check code. RocksDB team forked this file from
+ * phabricator's /src/lint/linter/ArcanistCpplintLinter.php, and customized it
+ * for its own use.
+ *
+ * You can get it here:
+ * http://google-styleguide.googlecode.com/svn/trunk/cpplint/cpplint.py
+ * @group linter
+ */
+final class ArcanistCpplintLinter extends ArcanistLinter {
+
+  public function willLintPaths(array $paths) {
+    return;
+  }
+
+  public function getLinterName() {
+    return 'cpplint.py';
+  }
+
+  public function getLintOptions() {
+    $config = $this->getEngine()->getConfigurationManager();
+    $options = $config->getConfigFromAnySource('lint.cpplint.options', '');
+
+    return $options;
+  }
+
+  public function getLintPath() {
+    $config = $this->getEngine()->getConfigurationManager();
+    $prefix = $config->getConfigFromAnySource('lint.cpplint.prefix');
+    $bin = $config->getConfigFromAnySource('lint.cpplint.bin', 'cpplint.py');
+
+    if ($prefix !== null) {
+      if (!Filesystem::pathExists($prefix.'/'.$bin)) {
+        throw new ArcanistUsageException(
+          "Unable to find cpplint.py binary in a specified directory. Make ".
+          "sure that 'lint.cpplint.prefix' and 'lint.cpplint.bin' keys are ".
+          "set correctly. If you'd rather use a copy of cpplint installed ".
+          "globally, you can just remove these keys from your .arcconfig.");
+      }
+
+      $bin = csprintf("%s/%s", $prefix, $bin);
+
+      return $bin;
+    }
+
+    // Search under current dir
+    list($err) = exec_manual('which %s/%s', $this->linterDir(), $bin);
+    if (!$err) {
+      return $this->linterDir().'/'.$bin;
+    }
+
+    // Look for globally installed cpplint.py
+    list($err) = exec_manual('which %s', $bin);
+    if ($err) {
+      throw new ArcanistUsageException(
+        "cpplint.py does not appear to be installed on this system. Install ".
+        "it (e.g., with 'wget \"http://google-styleguide.googlecode.com/".
+        "svn/trunk/cpplint/cpplint.py\"') or configure 'lint.cpplint.prefix' ".
+        "in your .arcconfig to point to the directory where it resides. ".
+        "Also don't forget to chmod a+x cpplint.py!");
+    }
+
+    return $bin;
+  }
+
+  public function lintPath($path) {
+    $bin = $this->getLintPath();
+    $options = $this->getLintOptions();
+    $path = $this->rocksdbDir().'/'.$path;
+
+    $f = new ExecFuture("%C %C $path", $bin, $options);
+
+    list($err, $stdout, $stderr) = $f->resolve();
+
+    if ($err === 2) {
+      throw new Exception("cpplint failed to run correctly:\n".$stderr);
+    }
+
+    $lines = explode("\n", $stderr);
+    $messages = array();
+    foreach ($lines as $line) {
+      $line = trim($line);
+      $matches = null;
+      $regex = '/^[^:]+:(\d+):\s*(.*)\s*\[(.*)\] \[(\d+)\]$/';
+      if (!preg_match($regex, $line, $matches)) {
+        continue;
+      }
+      foreach ($matches as $key => $match) {
+        $matches[$key] = trim($match);
+      }
+      $message = new ArcanistLintMessage();
+      $message->setPath($path);
+      $message->setLine($matches[1]);
+      $message->setCode($matches[3]);
+      $message->setName($matches[3]);
+      $message->setDescription($matches[2]);
+      $message->setSeverity(ArcanistLintSeverity::SEVERITY_WARNING);
+      $this->addLintMessage($message);
+    }
+  }
+
+  // The path of this linter
+  private function linterDir() {
+    return dirname(__FILE__);
+  }
+
+  // TODO(kaili) a quick and dirty way to figure out rocksdb's root dir.
+  private function rocksdbDir() {
+    return $this->linterDir()."/../..";
+  }
+}

table/meta_blocks.cc

@@ -270,12 +270,20 @@ Status ReadTableMagicNumber(const std::string& file_path,
 
   uint64_t file_size;
   options.env->GetFileSize(file_path, &file_size);
+  return ReadTableMagicNumber(file.get(), file_size, options, env_options,
+                              table_magic_number);
+}
+
+Status ReadTableMagicNumber(RandomAccessFile* file, uint64_t file_size,
+                            const Options& options,
+                            const EnvOptions& env_options,
+                            uint64_t* table_magic_number) {
   if (file_size < Footer::kEncodedLength) {
     return Status::InvalidArgument("file is too short to be an sstable");
   }
 
   Footer footer;
-  s = ReadFooterFromFile(file.get(), file_size, &footer);
+  auto s = ReadFooterFromFile(file, file_size, &footer);
   if (!s.ok()) {
     return s;
   }

table/meta_blocks.h

@@ -124,4 +124,9 @@ Status ReadTableMagicNumber(const std::string& file_path,
                             const Options& options,
                             const EnvOptions& env_options,
                             uint64_t* table_magic_number);
+
+Status ReadTableMagicNumber(RandomAccessFile* file, uint64_t file_size,
+                            const Options& options,
+                            const EnvOptions& env_options,
+                            uint64_t* table_magic_number);
 }  // namespace rocksdb

table/plain_table_factory.cc

@@ -21,7 +21,7 @@ Status PlainTableFactory::NewTableReader(const Options& options,
                                          unique_ptr<TableReader>* table) const {
   return PlainTableReader::Open(options, soptions, icomp, std::move(file),
                                 file_size, table, bloom_bits_per_key_,
-                                hash_table_ratio_);
+                                hash_table_ratio_, index_sparseness_);
 }
 
 TableBuilder* PlainTableFactory::NewTableBuilder(
@@ -32,9 +32,17 @@ TableBuilder* PlainTableFactory::NewTableBuilder(
 
 extern TableFactory* NewPlainTableFactory(uint32_t user_key_len,
                                           int bloom_bits_per_key,
-                                          double hash_table_ratio) {
+                                          double hash_table_ratio,
+                                          size_t index_sparseness) {
   return new PlainTableFactory(user_key_len, bloom_bits_per_key,
-                               hash_table_ratio);
+                               hash_table_ratio, index_sparseness);
+}
+
+extern TableFactory* NewTotalOrderPlainTableFactory(uint32_t user_key_len,
+                                                    int bloom_bits_per_key,
+                                                    size_t index_sparseness) {
+  return new PlainTableFactory(user_key_len, bloom_bits_per_key, 0,
+                               index_sparseness);
 }
 
 }  // namespace rocksdb

table/plain_table_factory.h

@@ -48,12 +48,20 @@ class PlainTableFactory : public TableFactory {
   // number of bits used for bloom filer per key. hash_table_ratio is
   // the desired utilization of the hash table used for prefix hashing.
   // hash_table_ratio = number of prefixes / #buckets in the hash table
+  // hash_table_ratio = 0 means skip hash table but only replying on binary
+  // search.
+  // index_sparseness determines index interval for keys
+  // inside the same prefix. It will be the maximum number of linear search
+  // required after hash and binary search.
+  // index_sparseness = 0 means index for every key.
   explicit PlainTableFactory(uint32_t user_key_len = kPlainTableVariableLength,
                              int bloom_bits_per_key = 0,
-                             double hash_table_ratio = 0.75)
+                             double hash_table_ratio = 0.75,
+                             size_t index_sparseness = 16)
       : user_key_len_(user_key_len),
         bloom_bits_per_key_(bloom_bits_per_key),
-        hash_table_ratio_(hash_table_ratio) {}
+        hash_table_ratio_(hash_table_ratio),
+        index_sparseness_(index_sparseness) {}
   const char* Name() const override { return "PlainTable"; }
   Status NewTableReader(const Options& options, const EnvOptions& soptions,
                         const InternalKeyComparator& internal_comparator,
@@ -71,6 +79,7 @@ class PlainTableFactory : public TableFactory {
   uint32_t user_key_len_;
   int bloom_bits_per_key_;
   double hash_table_ratio_;
+  size_t index_sparseness_;
 };
 
 }  // namespace rocksdb

table/plain_table_reader.cc

@@ -5,6 +5,7 @@
 #include "table/plain_table_reader.h"
 
 #include <string>
+#include <vector>
 
 #include "db/dbformat.h"
 
@@ -35,7 +36,7 @@ namespace rocksdb {
 
 namespace {
 
-inline uint32_t GetSliceHash(Slice const& s) {
+inline uint32_t GetSliceHash(const Slice& s) {
   return Hash(s.data(), s.size(), 397) ;
 }
 
@@ -43,12 +44,18 @@ inline uint32_t GetBucketIdFromHash(uint32_t hash, uint32_t num_buckets) {
   return hash % num_buckets;
 }
 
+// Safely getting a uint32_t element from a char array, where, starting from
+// `base`, every 4 bytes are considered as an fixed 32 bit integer.
+inline uint32_t GetFixed32Element(const char* base, size_t offset) {
+  return DecodeFixed32(base + offset * sizeof(uint32_t));
+}
+
 }  // namespace
 
 // Iterator to iterate IndexedTable
 class PlainTableIterator : public Iterator {
  public:
-  explicit PlainTableIterator(PlainTableReader* table);
+  explicit PlainTableIterator(PlainTableReader* table, bool use_prefix_seek);
   ~PlainTableIterator();
 
   bool Valid() const;
@@ -71,6 +78,7 @@ class PlainTableIterator : public Iterator {
 
  private:
   PlainTableReader* table_;
+  bool use_prefix_seek_;
   uint32_t offset_;
   uint32_t next_offset_;
   Slice key_;
@@ -83,34 +91,34 @@ class PlainTableIterator : public Iterator {
 };
 
 extern const uint64_t kPlainTableMagicNumber;
-PlainTableReader::PlainTableReader(const EnvOptions& storage_options,
+PlainTableReader::PlainTableReader(
-                                   const InternalKeyComparator& icomparator,
+    const Options& options, unique_ptr<RandomAccessFile>&& file,
-                                   uint64_t file_size, int bloom_bits_per_key,
+    const EnvOptions& storage_options, const InternalKeyComparator& icomparator,
-                                   double hash_table_ratio,
+    uint64_t file_size, int bloom_bits_per_key, double hash_table_ratio,
-                                   const TableProperties* table_properties)
+    size_t index_sparseness, const TableProperties* table_properties)
-    : soptions_(storage_options),
+    : options_(options),
+      soptions_(storage_options),
+      file_(std::move(file)),
       internal_comparator_(icomparator),
       file_size_(file_size),
       kHashTableRatio(hash_table_ratio),
       kBloomBitsPerKey(bloom_bits_per_key),
+      kIndexIntervalForSamePrefixKeys(index_sparseness),
       table_properties_(table_properties),
       data_end_offset_(table_properties_->data_size),
-      user_key_len_(table_properties->fixed_key_len) {}
+      user_key_len_(table_properties->fixed_key_len) {
+  assert(kHashTableRatio >= 0.0);
+}
 
 PlainTableReader::~PlainTableReader() {
-  delete[] hash_table_;
+}
-  delete[] sub_index_;
+
-  delete bloom_;
+Status PlainTableReader::Open(
-}
+    const Options& options, const EnvOptions& soptions,
-
+    const InternalKeyComparator& internal_comparator,
-Status PlainTableReader::Open(const Options& options,
+    unique_ptr<RandomAccessFile>&& file, uint64_t file_size,
-                              const EnvOptions& soptions,
+    unique_ptr<TableReader>* table_reader, const int bloom_bits_per_key,
-                              const InternalKeyComparator& internal_comparator,
+    double hash_table_ratio, size_t index_sparseness) {
-                              unique_ptr<RandomAccessFile>&& file,
-                              uint64_t file_size,
-                              unique_ptr<TableReader>* table_reader,
-                              const int bloom_bits_per_key,
-                              double hash_table_ratio) {
   assert(options.allow_mmap_reads);
 
   if (file_size > kMaxFileSize) {
@@ -124,11 +132,9 @@ Status PlainTableReader::Open(const Options& options,
     return s;
   }
 
-  std::unique_ptr<PlainTableReader> new_reader(
+  std::unique_ptr<PlainTableReader> new_reader(new PlainTableReader(
-      new PlainTableReader(soptions, internal_comparator, file_size,
+      options, std::move(file), soptions, internal_comparator, file_size,
-                           bloom_bits_per_key, hash_table_ratio, props));
+      bloom_bits_per_key, hash_table_ratio, index_sparseness, props));
-  new_reader->file_ = std::move(file);
-  new_reader->options_ = options;
 
   // -- Populate Index
   s = new_reader->PopulateIndex();
@@ -148,7 +154,7 @@ bool PlainTableReader::PrefixMayMatch(const Slice& internal_prefix) {
 }
 
 Iterator* PlainTableReader::NewIterator(const ReadOptions& options) {
-  return new PlainTableIterator(this);
+  return new PlainTableIterator(this, options.prefix_seek);
 }
 
 struct PlainTableReader::IndexRecord {
@@ -197,6 +203,9 @@ class PlainTableReader::IndexRecordList {
     return result;
   }
 
+  // Each group in `groups_` contains fix-sized records (determined by
+  // kNumRecordsPerGroup). Which can help us minimize the cost if resizing
+  // occurs.
   const size_t kNumRecordsPerGroup;
   IndexRecord* current_group_;
   // List of arrays allocated
@@ -204,79 +213,96 @@ class PlainTableReader::IndexRecordList {
   size_t num_records_in_current_group_;
 };
 
-int PlainTableReader::PopulateIndexRecordList(IndexRecordList* record_list) {
+Status PlainTableReader::PopulateIndexRecordList(IndexRecordList* record_list,
+                                                 int* num_prefixes) const {
   Slice prev_key_prefix_slice;
   uint32_t prev_key_prefix_hash = 0;
   uint32_t pos = data_start_offset_;
-  int key_index_within_prefix = 0;
+  int num_keys_per_prefix = 0;
   bool is_first_record = true;
   HistogramImpl keys_per_prefix_hist;
   // Need map to be ordered to make sure sub indexes generated
   // are in order.
 
-  int num_prefixes = 0;
+  *num_prefixes = 0;
   while (pos < data_end_offset_) {
     uint32_t key_offset = pos;
     ParsedInternalKey key;
     Slice value_slice;
-    status_ = Next(pos, &key, &value_slice, pos);
+    Status s = Next(&pos, &key, &value_slice);
+    if (!s.ok()) {
+      return s;
+    }
+    if (bloom_) {
+      // total order mode and bloom filter is enabled.
+      bloom_->AddHash(GetSliceHash(key.user_key));
+    }
     Slice key_prefix_slice = GetPrefix(key);
 
     if (is_first_record || prev_key_prefix_slice != key_prefix_slice) {
-      ++num_prefixes;
+      ++(*num_prefixes);
       if (!is_first_record) {
-        keys_per_prefix_hist.Add(key_index_within_prefix);
+        keys_per_prefix_hist.Add(num_keys_per_prefix);
       }
-      key_index_within_prefix = 0;
+      num_keys_per_prefix = 0;
       prev_key_prefix_slice = key_prefix_slice;
       prev_key_prefix_hash = GetSliceHash(key_prefix_slice);
     }
 
-    if (key_index_within_prefix++ % kIndexIntervalForSamePrefixKeys == 0) {
+    if (kIndexIntervalForSamePrefixKeys == 0 ||
+        num_keys_per_prefix++ % kIndexIntervalForSamePrefixKeys == 0) {
       // Add an index key for every kIndexIntervalForSamePrefixKeys keys
       record_list->AddRecord(prev_key_prefix_hash, key_offset);
     }
     is_first_record = false;
   }
 
-  keys_per_prefix_hist.Add(key_index_within_prefix);
+  keys_per_prefix_hist.Add(num_keys_per_prefix);
   Log(options_.info_log, "Number of Keys per prefix Histogram: %s",
       keys_per_prefix_hist.ToString().c_str());
 
-  return num_prefixes;
+  return Status::OK();
 }
 
 void PlainTableReader::AllocateIndexAndBloom(int num_prefixes) {
-  delete[] hash_table_;
+  index_.reset();
+
+  if (options_.prefix_extractor != nullptr) {
+    uint32_t bloom_total_bits = num_prefixes * kBloomBitsPerKey;
+    if (bloom_total_bits > 0) {
+      bloom_.reset(new DynamicBloom(bloom_total_bits));
+    }
+  }
 
-  if (kBloomBitsPerKey > 0) {
+  if (options_.prefix_extractor == nullptr || kHashTableRatio <= 0) {
-    bloom_ = new DynamicBloom(num_prefixes * kBloomBitsPerKey);
+    // Fall back to pure binary search if the user fails to specify a prefix
+    // extractor.
+    index_size_ = 1;
+  } else {
+    double hash_table_size_multipier = 1.0 / kHashTableRatio;
+    index_size_ = num_prefixes * hash_table_size_multipier + 1;
   }
-  double hash_table_size_multipier =
+  index_.reset(new uint32_t[index_size_]);
-      (kHashTableRatio > 1.0) ? 1.0 : 1.0 / kHashTableRatio;
-  hash_table_size_ = num_prefixes * hash_table_size_multipier + 1;
-  hash_table_ = new uint32_t[hash_table_size_];
 }
 
 size_t PlainTableReader::BucketizeIndexesAndFillBloom(
-    IndexRecordList& record_list, int num_prefixes,
+    IndexRecordList* record_list, std::vector<IndexRecord*>* hash_to_offsets,
-    std::vector<IndexRecord*>* hash_to_offsets,
     std::vector<uint32_t>* bucket_count) {
   size_t sub_index_size_needed = 0;
   bool first = true;
   uint32_t prev_hash = 0;
-  size_t num_records = record_list.GetNumRecords();
+  size_t num_records = record_list->GetNumRecords();
   for (size_t i = 0; i < num_records; i++) {
-    IndexRecord* index_record = record_list.At(i);
+    IndexRecord* index_record = record_list->At(i);
     uint32_t cur_hash = index_record->hash;
     if (first || prev_hash != cur_hash) {
       prev_hash = cur_hash;
       first = false;
-      if (bloom_) {
+      if (bloom_ && !IsTotalOrderMode()) {
         bloom_->AddHash(cur_hash);
       }
     }
-    uint32_t bucket = GetBucketIdFromHash(cur_hash, hash_table_size_);
+    uint32_t bucket = GetBucketIdFromHash(cur_hash, index_size_);
     IndexRecord* prev_bucket_head = (*hash_to_offsets)[bucket];
     index_record->next = prev_bucket_head;
     (*hash_to_offsets)[bucket] = index_record;
@@ -306,27 +332,24 @@ void PlainTableReader::FillIndexes(
   size_t buffer_size = 8 * 8;
   size_t buffer_used = 0;
   sub_index_size_needed += buffer_size;
-  sub_index_ = new char[sub_index_size_needed];
+  sub_index_.reset(new char[sub_index_size_needed]);
   size_t sub_index_offset = 0;
-  char* prev_ptr;
+  for (int i = 0; i < index_size_; i++) {
-  char* cur_ptr;
-  uint32_t* sub_index_ptr;
-  for (int i = 0; i < hash_table_size_; i++) {
     uint32_t num_keys_for_bucket = bucket_count[i];
     switch (num_keys_for_bucket) {
     case 0:
       // No key for bucket
-      hash_table_[i] = data_end_offset_;
+      index_[i] = data_end_offset_;
       break;
     case 1:
       // point directly to the file offset
-      hash_table_[i] = hash_to_offsets[i]->offset;
+      index_[i] = hash_to_offsets[i]->offset;
       break;
     default:
       // point to second level indexes.
-      hash_table_[i] = sub_index_offset | kSubIndexMask;
+      index_[i] = sub_index_offset | kSubIndexMask;
-      prev_ptr = sub_index_ + sub_index_offset;
+      char* prev_ptr = &sub_index_[sub_index_offset];
-      cur_ptr = EncodeVarint32(prev_ptr, num_keys_for_bucket);
+      char* cur_ptr = EncodeVarint32(prev_ptr, num_keys_for_bucket);
       sub_index_offset += (cur_ptr - prev_ptr);
       if (cur_ptr - prev_ptr > 2
           || (cur_ptr - prev_ptr == 2 && num_keys_for_bucket <= 127)) {
@@ -339,17 +362,16 @@ void PlainTableReader::FillIndexes(
           sub_index_size_needed += buffer_size;
           buffer_size *= 2;
           char* new_sub_index = new char[sub_index_size_needed];
-          memcpy(new_sub_index, sub_index_, sub_index_offset);
+          memcpy(new_sub_index, sub_index_.get(), sub_index_offset);
-          delete[] sub_index_;
+          sub_index_.reset(new_sub_index);
-          sub_index_ = new_sub_index;
         }
       }
-      sub_index_ptr = (uint32_t*) (sub_index_ + sub_index_offset);
+      char* sub_index_pos = &sub_index_[sub_index_offset];
       IndexRecord* record = hash_to_offsets[i];
       int j;
       for (j = num_keys_for_bucket - 1; j >= 0 && record;
            j--, record = record->next) {
-        sub_index_ptr[j] = record->offset;
+        EncodeFixed32(sub_index_pos + j * sizeof(uint32_t), record->offset);
       }
       assert(j == -1 && record == nullptr);
       sub_index_offset += kOffsetLen * num_keys_for_bucket;
@@ -358,10 +380,16 @@ void PlainTableReader::FillIndexes(
   }
 
   Log(options_.info_log, "hash table size: %d, suffix_map length %zu",
-      hash_table_size_, sub_index_size_needed);
+      index_size_, sub_index_size_needed);
 }
 
 Status PlainTableReader::PopulateIndex() {
+  // options.prefix_extractor is requried for a hash-based look-up.
+  if (options_.prefix_extractor == nullptr && kHashTableRatio != 0) {
+    return Status::NotSupported(
+        "PlainTable requires a prefix extractor enable prefix hash mode.");
+  }
+
   // Get mmapped memory to file_data_.
   Status s = file_->Read(0, file_size_, &file_data_, nullptr);
   if (!s.ok()) {
@@ -373,17 +401,30 @@ Status PlainTableReader::PopulateIndex() {
   // for a prefix (starting from the first one), generate a record of (hash,
   // offset) and append it to IndexRecordList, which is a data structure created
   // to store them.
-  int num_prefixes = PopulateIndexRecordList(&record_list);
+  int num_prefixes;
+
+  // Allocate bloom filter here for total order mode.
+  if (IsTotalOrderMode()) {
+    uint32_t num_bloom_bits = table_properties_->num_entries * kBloomBitsPerKey;
+    if (num_bloom_bits > 0) {
+      bloom_.reset(new DynamicBloom(num_bloom_bits));
+    }
+  }
+
+  s = PopulateIndexRecordList(&record_list, &num_prefixes);
+  if (!s.ok()) {
+    return s;
+  }
   // Calculated hash table and bloom filter size and allocate memory for indexes
   // and bloom filter based on the number of prefixes.
   AllocateIndexAndBloom(num_prefixes);
 
   // Bucketize all the index records to a temp data structure, in which for
   // each bucket, we generate a linked list of IndexRecord, in reversed order.
-  std::vector<IndexRecord*> hash_to_offsets(hash_table_size_, nullptr);
+  std::vector<IndexRecord*> hash_to_offsets(index_size_, nullptr);
-  std::vector<uint32_t> bucket_count(hash_table_size_, 0);
+  std::vector<uint32_t> bucket_count(index_size_, 0);
   size_t sub_index_size_needed = BucketizeIndexesAndFillBloom(
-      record_list, num_prefixes, &hash_to_offsets, &bucket_count);
+      &record_list, &hash_to_offsets, &bucket_count);
   // From the temp data structure, populate indexes.
   FillIndexes(sub_index_size_needed, hash_to_offsets, bucket_count);
 
@@ -392,16 +433,16 @@ Status PlainTableReader::PopulateIndex() {
 
 Status PlainTableReader::GetOffset(const Slice& target, const Slice& prefix,
                                    uint32_t prefix_hash, bool& prefix_matched,
-                                   uint32_t& ret_offset) {
+                                   uint32_t* offset) const {
   prefix_matched = false;
-  int bucket = GetBucketIdFromHash(prefix_hash, hash_table_size_);
+  int bucket = GetBucketIdFromHash(prefix_hash, index_size_);
-  uint32_t bucket_value = hash_table_[bucket];
+  uint32_t bucket_value = index_[bucket];
   if (bucket_value == data_end_offset_) {
-    ret_offset = data_end_offset_;
+    *offset = data_end_offset_;
     return Status::OK();
   } else if ((bucket_value & kSubIndexMask) == 0) {
     // point directly to the file
-    ret_offset = bucket_value;
+    *offset = bucket_value;
     return Status::OK();
   }
 
@@ -409,11 +450,9 @@ Status PlainTableReader::GetOffset(const Slice& target, const Slice& prefix,
   uint32_t low = 0;
   uint64_t prefix_index_offset = bucket_value ^ kSubIndexMask;
 
-  const char* index_ptr = sub_index_ + prefix_index_offset;
+  const char* index_ptr = &sub_index_[prefix_index_offset];
   uint32_t upper_bound = 0;
-  const uint32_t* base_ptr = (const uint32_t*) GetVarint32Ptr(index_ptr,
+  const char* base_ptr = GetVarint32Ptr(index_ptr, index_ptr + 4, &upper_bound);
-                                                              index_ptr + 4,
-                                                              &upper_bound);
   uint32_t high = upper_bound;
   ParsedInternalKey mid_key;
   ParsedInternalKey parsed_target;
@@ -424,9 +463,9 @@ Status PlainTableReader::GetOffset(const Slice& target, const Slice& prefix,
   // The key is between [low, high). Do a binary search between it.
   while (high - low > 1) {
     uint32_t mid = (high + low) / 2;
-    uint32_t file_offset = base_ptr[mid];
+    uint32_t file_offset = GetFixed32Element(base_ptr, mid);
     size_t tmp;
-    Status s = ReadKey(file_data_.data() + file_offset, &mid_key, tmp);
+    Status s = ReadKey(file_data_.data() + file_offset, &mid_key, &tmp);
     if (!s.ok()) {
       return s;
     }
@@ -438,7 +477,7 @@ Status PlainTableReader::GetOffset(const Slice& target, const Slice& prefix,
         // Happen to have found the exact key or target is smaller than the
         // first key after base_offset.
         prefix_matched = true;
-        ret_offset = file_offset;
+        *offset = file_offset;
         return Status::OK();
       } else {
         high = mid;
@@ -450,48 +489,48 @@ Status PlainTableReader::GetOffset(const Slice& target, const Slice& prefix,
   // to the wrong prefix.
   ParsedInternalKey low_key;
   size_t tmp;
-  uint32_t low_key_offset = base_ptr[low];
+  uint32_t low_key_offset = GetFixed32Element(base_ptr, low);
-  Status s = ReadKey(file_data_.data() + low_key_offset, &low_key, tmp);
+  Status s = ReadKey(file_data_.data() + low_key_offset, &low_key, &tmp);
   if (GetPrefix(low_key) == prefix) {
     prefix_matched = true;
-    ret_offset = low_key_offset;
+    *offset = low_key_offset;
   } else if (low + 1 < upper_bound) {
     // There is possible a next prefix, return it
     prefix_matched = false;
-    ret_offset = base_ptr[low + 1];
+    *offset = GetFixed32Element(base_ptr, low + 1);
   } else {
     // target is larger than a key of the last prefix in this bucket
     // but with a different prefix. Key does not exist.
-    ret_offset = data_end_offset_;
+    *offset = data_end_offset_;
   }
   return Status::OK();
 }
 
-bool PlainTableReader::MayHavePrefix(uint32_t hash) {
+bool PlainTableReader::MatchBloom(uint32_t hash) const {
   return bloom_ == nullptr || bloom_->MayContainHash(hash);
 }
 
-Slice PlainTableReader::GetPrefix(const ParsedInternalKey& target) {
+Slice PlainTableReader::GetPrefix(const ParsedInternalKey& target) const {
-  return options_.prefix_extractor->Transform(target.user_key);
+  return GetPrefixFromUserKey(target.user_key);
 }
 
-Status PlainTableReader::ReadKey(const char* row_ptr, ParsedInternalKey* key,
+Status PlainTableReader::ReadKey(const char* start, ParsedInternalKey* key,
-                                 size_t& bytes_read) {
+                                 size_t* bytes_read) const {
   const char* key_ptr = nullptr;
-  bytes_read = 0;
+  *bytes_read = 0;
   size_t user_key_size = 0;
   if (IsFixedLength()) {
     user_key_size = user_key_len_;
-    key_ptr = row_ptr;
+    key_ptr = start;
   } else {
     uint32_t tmp_size = 0;
-    key_ptr = GetVarint32Ptr(row_ptr, file_data_.data() + data_end_offset_,
+    key_ptr =
-                             &tmp_size);
+        GetVarint32Ptr(start, file_data_.data() + data_end_offset_, &tmp_size);
     if (key_ptr == nullptr) {
       return Status::Corruption("Unable to read the next key");
     }
     user_key_size = (size_t)tmp_size;
-    bytes_read = key_ptr - row_ptr;
+    *bytes_read = key_ptr - start;
   }
   if (key_ptr + user_key_size + 1 >= file_data_.data() + data_end_offset_) {
     return Status::Corruption("Unable to read the next key");
@@ -502,43 +541,42 @@ Status PlainTableReader::ReadKey(const char* row_ptr, ParsedInternalKey* key,
     key->user_key = Slice(key_ptr, user_key_size);
     key->sequence = 0;
     key->type = kTypeValue;
-    bytes_read += user_key_size + 1;
+    *bytes_read += user_key_size + 1;
   } else {
-    if (row_ptr + user_key_size + 8 >= file_data_.data() + data_end_offset_) {
+    if (start + user_key_size + 8 >= file_data_.data() + data_end_offset_) {
       return Status::Corruption("Unable to read the next key");
     }
     if (!ParseInternalKey(Slice(key_ptr, user_key_size + 8), key)) {
       return Status::Corruption(Slice());
     }
-    bytes_read += user_key_size + 8;
+    *bytes_read += user_key_size + 8;
   }
 
   return Status::OK();
 }
 
-Status PlainTableReader::Next(uint32_t offset, ParsedInternalKey* key,
+Status PlainTableReader::Next(uint32_t* offset, ParsedInternalKey* key,
-                              Slice* value, uint32_t& next_offset) {
+                              Slice* value) const {
-  if (offset == data_end_offset_) {
+  if (*offset == data_end_offset_) {
-    next_offset = data_end_offset_;
+    *offset = data_end_offset_;
     return Status::OK();
   }
 
-  if (offset > data_end_offset_) {
+  if (*offset > data_end_offset_) {
     return Status::Corruption("Offset is out of file size");
   }
 
-  const char* row_ptr = file_data_.data() + offset;
+  const char* start = file_data_.data() + *offset;
   size_t bytes_for_key;
-  Status s = ReadKey(row_ptr, key, bytes_for_key);
+  Status s = ReadKey(start, key, &bytes_for_key);
   uint32_t value_size;
-  const char* value_ptr = GetVarint32Ptr(row_ptr + bytes_for_key,
+  const char* value_ptr = GetVarint32Ptr(
-                                         file_data_.data() + data_end_offset_,
+      start + bytes_for_key, file_data_.data() + data_end_offset_, &value_size);
-                                         &value_size);
   if (value_ptr == nullptr) {
     return Status::Corruption("Error reading value length.");
   }
-  next_offset = offset + (value_ptr - row_ptr) + value_size;
+  *offset = *offset + (value_ptr - start) + value_size;
-  if (next_offset > data_end_offset_) {
+  if (*offset > data_end_offset_) {
     return Status::Corruption("Reach end of file when reading value");
   }
   *value = Slice(value_ptr, value_size);
@@ -552,14 +590,28 @@ Status PlainTableReader::Get(const ReadOptions& ro, const Slice& target,
                                            const Slice&, bool),
                              void (*mark_key_may_exist)(void*)) {
   // Check bloom filter first.
-  Slice prefix_slice = GetPrefix(target);
+  Slice prefix_slice;
-  uint32_t prefix_hash = GetSliceHash(prefix_slice);
+  uint32_t prefix_hash;
-  if (!MayHavePrefix(prefix_hash)) {
+  if (IsTotalOrderMode()) {
-    return Status::OK();
+    // Match whole user key for bloom filter check.
+    if (!MatchBloom(GetSliceHash(GetUserKey(target)))) {
+      return Status::OK();
+    }
+    // in total order mode, there is only one bucket 0, and we always use empty
+    // prefix.
+    prefix_slice = Slice();
+    prefix_hash = 0;
+  } else {
+    prefix_slice = GetPrefix(target);
+    prefix_hash = GetSliceHash(prefix_slice);
+    if (!MatchBloom(prefix_hash)) {
+      return Status::OK();
+    }
   }
   uint32_t offset;
   bool prefix_match;
-  Status s = GetOffset(target, prefix_slice, prefix_hash, prefix_match, offset);
+  Status s =
+      GetOffset(target, prefix_slice, prefix_hash, prefix_match, &offset);
   if (!s.ok()) {
     return s;
   }
@@ -571,7 +623,7 @@ Status PlainTableReader::Get(const ReadOptions& ro, const Slice& target,
 
   Slice found_value;
   while (offset < data_end_offset_) {
-    Status s = Next(offset, &found_key, &found_value, offset);
+    Status s = Next(&offset, &found_key, &found_value);
     if (!s.ok()) {
       return s;
     }
@@ -596,8 +648,9 @@ uint64_t PlainTableReader::ApproximateOffsetOf(const Slice& key) {
   return 0;
 }
 
-PlainTableIterator::PlainTableIterator(PlainTableReader* table) :
+PlainTableIterator::PlainTableIterator(PlainTableReader* table,
-    table_(table) {
+                                       bool use_prefix_seek)
+    : table_(table), use_prefix_seek_(use_prefix_seek) {
   next_offset_ = offset_ = table_->data_end_offset_;
 }
 
@@ -620,18 +673,39 @@ void PlainTableIterator::SeekToFirst() {
 
 void PlainTableIterator::SeekToLast() {
   assert(false);
+  status_ = Status::NotSupported("SeekToLast() is not supported in PlainTable");
 }
 
 void PlainTableIterator::Seek(const Slice& target) {
-  Slice prefix_slice =  table_->GetPrefix(target);
+  // If the user doesn't set prefix seek option and we are not able to do a
-  uint32_t prefix_hash = GetSliceHash(prefix_slice);
+  // total Seek(). assert failure.
-  if (!table_->MayHavePrefix(prefix_hash)) {
+  if (!use_prefix_seek_ && table_->index_size_ > 1) {
+    assert(false);
+    status_ = Status::NotSupported(
+        "PlainTable cannot issue non-prefix seek unless in total order mode.");
+    offset_ = next_offset_ = table_->data_end_offset_;
+    return;
+  }
+
+  Slice prefix_slice = table_->GetPrefix(target);
+  uint32_t prefix_hash;
+  uint32_t bloom_hash;
+  if (table_->IsTotalOrderMode()) {
+    // The total order mode, there is only one hash bucket 0. The bloom filter
+    // is checked against the whole user key.
+    prefix_hash = 0;
+    bloom_hash = GetSliceHash(table_->GetUserKey(target));
+  } else {
+    prefix_hash = GetSliceHash(prefix_slice);
+    bloom_hash = prefix_hash;
+  }
+  if (!table_->MatchBloom(bloom_hash)) {
     offset_ = next_offset_ = table_->data_end_offset_;
     return;
   }
   bool prefix_match;
   status_ = table_->GetOffset(target, prefix_slice, prefix_hash, prefix_match,
-                              next_offset_);
+                              &next_offset_);
   if (!status_.ok()) {
     offset_ = next_offset_ = table_->data_end_offset_;
     return;
@@ -661,7 +735,7 @@ void PlainTableIterator::Next() {
   if (offset_ < table_->data_end_offset_) {
     Slice tmp_slice;
     ParsedInternalKey parsed_key;
-    status_ = table_->Next(next_offset_, &parsed_key, &value_, next_offset_);
+    status_ = table_->Next(&next_offset_, &parsed_key, &value_);
     if (status_.ok()) {
       // Make a copy in this case. TODO optimize.
       tmp_str_.clear();

table/plain_table_reader.h

@@ -49,7 +49,8 @@ class PlainTableReader: public TableReader {
                      const InternalKeyComparator& internal_comparator,
                      unique_ptr<RandomAccessFile>&& file, uint64_t file_size,
                      unique_ptr<TableReader>* table,
-                     const int bloom_bits_per_key, double hash_table_ratio);
+                     const int bloom_bits_per_key, double hash_table_ratio,
+                     size_t index_sparseness);
 
   bool PrefixMayMatch(const Slice& internal_prefix);
 
@@ -68,36 +69,97 @@ class PlainTableReader: public TableReader {
     return table_properties_;
   }
 
-  PlainTableReader(const EnvOptions& storage_options,
+  PlainTableReader(const Options& options, unique_ptr<RandomAccessFile>&& file,
+                   const EnvOptions& storage_options,
                    const InternalKeyComparator& internal_comparator,
                    uint64_t file_size, int bloom_num_bits,
-                   double hash_table_ratio,
+                   double hash_table_ratio, size_t index_sparseness,
                    const TableProperties* table_properties);
-  ~PlainTableReader();
+  virtual ~PlainTableReader();
+
+ protected:
+  // Check bloom filter to see whether it might contain this prefix.
+  // The hash of the prefix is given, since it can be reused for index lookup
+  // too.
+  virtual bool MatchBloom(uint32_t hash) const;
+
+  // PopulateIndex() builds index of keys. It must be called before any query
+  // to the table.
+  //
+  // index_ contains buckets size of index_size_, each is a
+  // 32-bit integer. The lower 31 bits contain an offset value (explained below)
+  // and the first bit of the integer indicates type of the offset.
+  //
+  // +--------------+------------------------------------------------------+
+  // | Flag (1 bit) | Offset to binary search buffer or file (31 bits)     +
+  // +--------------+------------------------------------------------------+
+  //
+  // Explanation for the "flag bit":
+  //
+  // 0 indicates that the bucket contains only one prefix (no conflict when
+  //   hashing this prefix), whose first row starts from this offset of the
+  // file.
+  // 1 indicates that the bucket contains more than one prefixes, or there
+  //   are too many rows for one prefix so we need a binary search for it. In
+  //   this case, the offset indicates the offset of sub_index_ holding the
+  //   binary search indexes of keys for those rows. Those binary search indexes
+  //   are organized in this way:
+  //
+  // The first 4 bytes, indicate how many indexes (N) are stored after it. After
+  // it, there are N 32-bit integers, each points of an offset of the file,
+  // which
+  // points to starting of a row. Those offsets need to be guaranteed to be in
+  // ascending order so the keys they are pointing to are also in ascending
+  // order
+  // to make sure we can use them to do binary searches. Below is visual
+  // presentation of a bucket.
+  //
+  // <begin>
+  //   number_of_records:  varint32
+  //   record 1 file offset:  fixedint32
+  //   record 2 file offset:  fixedint32
+  //    ....
+  //   record N file offset:  fixedint32
+  // <end>
+  Status PopulateIndex();
 
  private:
   struct IndexRecord;
   class IndexRecordList;
 
-  uint32_t* hash_table_ = nullptr;
+  // Plain table maintains an index and a sub index.
-  int hash_table_size_ = 0;
+  // index is implemented by a hash table.
-  char* sub_index_ = nullptr;
+  // subindex is a big of memory array.
+  // For more details about the in-memory index, please refer to:
+  // https://github.com/facebook/rocksdb/wiki/PlainTable-Format
+  // #wiki-in-memory-index-format
+  std::unique_ptr<uint32_t[]> index_;
+  int index_size_ = 0;
+  std::unique_ptr<char[]> sub_index_;
 
   Options options_;
   const EnvOptions& soptions_;
+  unique_ptr<RandomAccessFile> file_;
+
   const InternalKeyComparator internal_comparator_;
+  // represents plain table's current status.
   Status status_;
-  unique_ptr<RandomAccessFile> file_;
 
   Slice file_data_;
-  uint32_t version_;
   uint32_t file_size_;
 
   const double kHashTableRatio;
   const int kBloomBitsPerKey;
-  DynamicBloom* bloom_ = nullptr;
+  // To speed up the search for keys with same prefix, we'll add index key for
+  // every N keys, where the "N" is determined by
+  // kIndexIntervalForSamePrefixKeys
+  const size_t kIndexIntervalForSamePrefixKeys = 16;
+  // Bloom filter is used to rule out non-existent key
+  unique_ptr<DynamicBloom> bloom_;
 
   std::shared_ptr<const TableProperties> table_properties_;
+  // data_start_offset_ and data_end_offset_ defines the range of the
+  // sst file that stores data.
   const uint32_t data_start_offset_ = 0;
   const uint32_t data_end_offset_;
   const size_t user_key_len_;
@@ -107,10 +169,6 @@ class PlainTableReader: public TableReader {
   static const size_t kOffsetLen = sizeof(uint32_t);
   static const uint64_t kMaxFileSize = 1u << 31;
   static const size_t kRecordsPerGroup = 256;
-  // To speed up the search for keys with same prefix, we'll add index key for
-  // every N keys, where the "N" is determined by
-  // kIndexIntervalForSamePrefixKeys
-  static const size_t kIndexIntervalForSamePrefixKeys = 16;
 
   bool IsFixedLength() const {
     return user_key_len_ != kPlainTableVariableLength;
@@ -125,95 +183,72 @@ class PlainTableReader: public TableReader {
 
   // Internal helper function to generate an IndexRecordList object from all
   // the rows, which contains index records as a list.
-  int PopulateIndexRecordList(IndexRecordList* record_list);
+  // If bloom_ is not null, all the keys' full-key hash will be added to the
+  // bloom filter.
+  Status PopulateIndexRecordList(IndexRecordList* record_list,
+                                 int* num_prefixes) const;
 
   // Internal helper function to allocate memory for indexes and bloom filters
   void AllocateIndexAndBloom(int num_prefixes);
 
   // Internal helper function to bucket index record list to hash buckets.
-  // hash_to_offsets is sized of of hash_table_size_, each contains a linked
+  // bucket_header is a vector of size hash_table_size_, with each entry
-  // list
+  // containing a linklist of IndexRecord hashed to the same bucket, in reverse
+  // order.
   // of offsets for the hash, in reversed order.
-  // bucket_count is sized of hash_table_size_. The value is how many index
+  // bucket_count is sized of index_size_. The value is how many index
-  // records are there in hash_to_offsets for the same bucket.
+  // records are there in bucket_headers for the same bucket.
-  size_t BucketizeIndexesAndFillBloom(
+  size_t BucketizeIndexesAndFillBloom(IndexRecordList* record_list,
-      IndexRecordList& record_list, int num_prefixes,
+                                      std::vector<IndexRecord*>* bucket_headers,
-      std::vector<IndexRecord*>* hash_to_offsets,
+                                      std::vector<uint32_t>* bucket_count);
-      std::vector<uint32_t>* bucket_count);
 
   // Internal helper class to fill the indexes and bloom filters to internal
-  // data structures. hash_to_offsets and bucket_count are bucketized indexes
+  // data structures. bucket_headers and bucket_count are bucketized indexes
   // and counts generated by BucketizeIndexesAndFillBloom().
   void FillIndexes(size_t sub_index_size_needed,
-                   const std::vector<IndexRecord*>& hash_to_offsets,
+                   const std::vector<IndexRecord*>& bucket_headers,
                    const std::vector<uint32_t>& bucket_count);
 
-  // PopulateIndex() builds index of keys. It must be called before any query
+  // Read a plain table key from the position `start`. The read content
-  // to the table.
+  // will be written to `key` and the size of read bytes will be populated
-  //
+  // in `bytes_read`.
-  // hash_table_ contains buckets size of hash_table_size_, each is a 32-bit
-  // integer. The lower 31 bits contain an offset value (explained below) and
-  // the first bit of the integer indicates type of the offset.
-  //
-  // +--------------+------------------------------------------------------+
-  // | Flag (1 bit) | Offset to binary search buffer or file (31 bits)     +
-  // +--------------+------------------------------------------------------+
-  //
-  // Explanation for the "flag bit":
-  //
-  // 0 indicates that the bucket contains only one prefix (no conflict when
-  //   hashing this prefix), whose first row starts from this offset of the
-  // file.
-  // 1 indicates that the bucket contains more than one prefixes, or there
-  //   are too many rows for one prefix so we need a binary search for it. In
-  //   this case, the offset indicates the offset of sub_index_ holding the
-  //   binary search indexes of keys for those rows. Those binary search indexes
-  //   are organized in this way:
-  //
-  // The first 4 bytes, indicate how many indexes (N) are stored after it. After
-  // it, there are N 32-bit integers, each points of an offset of the file,
-  // which
-  // points to starting of a row. Those offsets need to be guaranteed to be in
-  // ascending order so the keys they are pointing to are also in ascending
-  // order
-  // to make sure we can use them to do binary searches. Below is visual
-  // presentation of a bucket.
-  //
-  // <begin>
-  //   number_of_records:  varint32
-  //   record 1 file offset:  fixedint32
-  //   record 2 file offset:  fixedint32
-  //    ....
-  //   record N file offset:  fixedint32
-  // <end>
-  Status PopulateIndex();
-
-  // Check bloom filter to see whether it might contain this prefix.
-  // The hash of the prefix is given, since it can be reused for index lookup
-  // too.
-  bool MayHavePrefix(uint32_t hash);
-
   Status ReadKey(const char* row_ptr, ParsedInternalKey* key,
-                 size_t& bytes_read);
+                 size_t* bytes_read) const;
-  // Read the key and value at offset to key and value.
+  // Read the key and value at `offset` to parameters `key` and `value`.
-  // tmp_slice is a tmp slice.
+  // On success, `offset` will be updated as the offset for the next key.
-  // return next_offset as the offset for the next key.
+  Status Next(uint32_t* offset, ParsedInternalKey* key, Slice* value) const;
-  Status Next(uint32_t offset, ParsedInternalKey* key, Slice* value,
-              uint32_t& next_offset);
   // Get file offset for key target.
   // return value prefix_matched is set to true if the offset is confirmed
   // for a key with the same prefix as target.
   Status GetOffset(const Slice& target, const Slice& prefix,
                    uint32_t prefix_hash, bool& prefix_matched,
-                   uint32_t& ret_offset);
+                   uint32_t* offset) const;
+
+  Slice GetUserKey(const Slice& key) const {
+    return Slice(key.data(), key.size() - 8);
+  }
 
-  Slice GetPrefix(const Slice& target) {
+  Slice GetPrefix(const Slice& target) const {
     assert(target.size() >= 8);  // target is internal key
-    return options_.prefix_extractor->Transform(
+    return GetPrefixFromUserKey(GetUserKey(target));
-        Slice(target.data(), target.size() - 8));
   }
 
-  Slice GetPrefix(const ParsedInternalKey& target);
+  inline Slice GetPrefix(const ParsedInternalKey& target) const;
+
+  Slice GetPrefixFromUserKey(const Slice& user_key) const {
+    if (!IsTotalOrderMode()) {
+      return options_.prefix_extractor->Transform(user_key);
+    } else {
+      // Use empty slice as prefix if prefix_extractor is not set. In that case,
+      // it falls back to pure binary search and total iterator seek is
+      // supported.
+      return Slice();
+    }
+  }
+
+  bool IsTotalOrderMode() const {
+    return (options_.prefix_extractor == nullptr);
+  }
 
   // No copying allowed
   explicit PlainTableReader(const TableReader&) = delete;

table/table_reader_bench.cc

@@ -13,6 +13,7 @@
 #include "port/atomic_pointer.h"
 #include "table/block_based_table_factory.h"
 #include "table/plain_table_factory.h"
+#include "table/table_builder.h"
 #include "util/histogram.h"
 #include "util/testharness.h"
 #include "util/testutil.h"
@@ -39,6 +40,10 @@ static bool DummySaveValue(void* arg, const ParsedInternalKey& ikey,
   return false;
 }
 
+uint64_t Now(Env* env, bool measured_by_nanosecond) {
+  return measured_by_nanosecond ? env->NowNanos() : env->NowMicros();
+}
+
 // A very simple benchmark that.
 // Create a table with roughly numKey1 * numKey2 keys,
 // where there are numKey1 prefixes of the key, each has numKey2 number of
@@ -56,13 +61,14 @@ void TableReaderBenchmark(Options& opts, EnvOptions& env_options,
                           ReadOptions& read_options, int num_keys1,
                           int num_keys2, int num_iter, int prefix_len,
                           bool if_query_empty_keys, bool for_iterator,
-                          bool through_db) {
+                          bool through_db, bool measured_by_nanosecond) {
+  rocksdb::InternalKeyComparator ikc(opts.comparator);
+
   Slice prefix = Slice();
 
   std::string file_name = test::TmpDir()
       + "/rocksdb_table_reader_benchmark";
   std::string dbname = test::TmpDir() + "/rocksdb_table_reader_bench_db";
-  ReadOptions ro;
   WriteOptions wo;
   unique_ptr<WritableFile> file;
   Env* env = Env::Default();
@@ -71,7 +77,7 @@ void TableReaderBenchmark(Options& opts, EnvOptions& env_options,
   Status s;
   if (!through_db) {
     env->NewWritableFile(file_name, &file, env_options);
-    tb = opts.table_factory->NewTableBuilder(opts, file.get(),
+    tb = opts.table_factory->NewTableBuilder(opts, ikc, file.get(),
                                              CompressionType::kNoCompression);
   } else {
     s = DB::Open(opts, dbname, &db);
@@ -102,8 +108,8 @@ void TableReaderBenchmark(Options& opts, EnvOptions& env_options,
     Status s = env->NewRandomAccessFile(file_name, &raf, env_options);
     uint64_t file_size;
     env->GetFileSize(file_name, &file_size);
-    s = opts.table_factory->NewTableReader(opts, env_options, std::move(raf),
+    s = opts.table_factory->NewTableReader(
-                                           file_size, &table_reader);
+        opts, env_options, ikc, std::move(raf), file_size, &table_reader);
   }
 
   Random rnd(301);
@@ -124,15 +130,16 @@ void TableReaderBenchmark(Options& opts, EnvOptions& env_options,
         if (!for_iterator) {
           // Query one existing key;
           std::string key = MakeKey(r1, r2, through_db);
-          uint64_t start_micros = env->NowMicros();
+          uint64_t start_time = Now(env, measured_by_nanosecond);
           port::MemoryBarrier();
           if (!through_db) {
-            s = table_reader->Get(ro, key, arg, DummySaveValue, nullptr);
+            s = table_reader->Get(read_options, key, arg, DummySaveValue,
+                                  nullptr);
           } else {
-            s = db->Get(ro, key, &result);
+            s = db->Get(read_options, key, &result);
           }
           port::MemoryBarrier();
-          hist.Add(env->NowMicros() - start_micros);
+          hist.Add(Now(env, measured_by_nanosecond) - start_time);
         } else {
           int r2_len;
           if (if_query_empty_keys) {
@@ -150,7 +157,7 @@ void TableReaderBenchmark(Options& opts, EnvOptions& env_options,
             read_options.prefix = &prefix;
           }
           uint64_t total_time = 0;
-          uint64_t start_micros = env->NowMicros();
+          uint64_t start_time = Now(env, measured_by_nanosecond);
           port::MemoryBarrier();
           Iterator* iter;
           if (!through_db) {
@@ -165,9 +172,9 @@ void TableReaderBenchmark(Options& opts, EnvOptions& env_options,
             }
             // verify key;
             port::MemoryBarrier();
-            total_time += env->NowMicros() - start_micros;
+            total_time += Now(env, measured_by_nanosecond) - start_time;
             assert(Slice(MakeKey(r1, r2 + count, through_db)) == iter->key());
-            start_micros = env->NowMicros();
+            start_time = Now(env, measured_by_nanosecond);
             if (++count >= r2_len) {
               break;
             }
@@ -180,7 +187,7 @@ void TableReaderBenchmark(Options& opts, EnvOptions& env_options,
           }
           delete iter;
           port::MemoryBarrier();
-          total_time += env->NowMicros() - start_micros;
+          total_time += Now(env, measured_by_nanosecond) - start_time;
           hist.Add(total_time);
         }
       }
@@ -195,9 +202,10 @@ void TableReaderBenchmark(Options& opts, EnvOptions& env_options,
       "num_key2: %5d  %10s\n"
       "==================================================="
       "===================================================="
-      "\nHistogram (unit: microseconds): \n%s",
+      "\nHistogram (unit: %s): \n%s",
       opts.table_factory->Name(), num_keys1, num_keys2,
-      for_iterator? "iterator" : (if_query_empty_keys ? "empty" : "non_empty"),
+      for_iterator ? "iterator" : (if_query_empty_keys ? "empty" : "non_empty"),
+      measured_by_nanosecond ? "nanosecond" : "microsecond",
       hist.ToString().c_str());
   if (!through_db) {
     env->DeleteFile(file_name);
@@ -207,7 +215,7 @@ void TableReaderBenchmark(Options& opts, EnvOptions& env_options,
     DestroyDB(dbname, opts);
   }
 }
-} // namespace rocksdb
+}  // namespace rocksdb
 
 DEFINE_bool(query_empty, false, "query non-existing keys instead of existing "
             "ones.");
@@ -220,7 +228,9 @@ DEFINE_bool(through_db, false, "If enable, a DB instance will be created and "
             "the query will be against DB. Otherwise, will be directly against "
             "a table reader.");
 DEFINE_bool(plain_table, false, "Use PlainTable");
-
+DEFINE_string(time_unit, "microsecond",
+              "The time unit used for measuring performance. User can specify "
+              "`microsecond` (default) or `nanosecond`");
 
 int main(int argc, char** argv) {
   google::SetUsageMessage(std::string("\nUSAGE:\n") + std::string(argv[0]) +
@@ -237,10 +247,9 @@ int main(int argc, char** argv) {
   rocksdb::EnvOptions env_options;
   options.create_if_missing = true;
   options.compression = rocksdb::CompressionType::kNoCompression;
-  options.internal_comparator =
-      new rocksdb::InternalKeyComparator(options.comparator);
 
   if (FLAGS_plain_table) {
+    ro.prefix_seek = true;
     options.allow_mmap_reads = true;
     env_options.use_mmap_reads = true;
     tf = new rocksdb::PlainTableFactory(16, (FLAGS_prefix_len == 16) ? 0 : 8,
@@ -250,11 +259,15 @@ int main(int argc, char** argv) {
   } else {
     tf = new rocksdb::BlockBasedTableFactory();
   }
+  // if user provides invalid options, just fall back to microsecond.
+  bool measured_by_nanosecond = FLAGS_time_unit == "nanosecond";
+
   options.table_factory =
       std::shared_ptr<rocksdb::TableFactory>(tf);
   TableReaderBenchmark(options, env_options, ro, FLAGS_num_keys1,
                        FLAGS_num_keys2, FLAGS_iter, FLAGS_prefix_len,
-                       FLAGS_query_empty, FLAGS_iterator, FLAGS_through_db);
+                       FLAGS_query_empty, FLAGS_iterator, FLAGS_through_db,
+                       measured_by_nanosecond);
   delete tf;
   return 0;
 }

table/table_test.cc

@@ -306,8 +306,11 @@ class KeyConvertingIterator: public Iterator {
 class TableConstructor: public Constructor {
  public:
   explicit TableConstructor(const Comparator* cmp,
-                            bool convert_to_internal_key = false)
+                            bool convert_to_internal_key = false,
-      : Constructor(cmp), convert_to_internal_key_(convert_to_internal_key) {}
+                            bool prefix_seek = false)
+      : Constructor(cmp),
+        convert_to_internal_key_(convert_to_internal_key),
+        prefix_seek_(prefix_seek) {}
   ~TableConstructor() { Reset(); }
 
   virtual Status FinishImpl(const Options& options,
@@ -347,7 +350,11 @@ class TableConstructor: public Constructor {
   }
 
   virtual Iterator* NewIterator() const {
-    Iterator* iter = table_reader_->NewIterator(ReadOptions());
+    ReadOptions ro;
+    if (prefix_seek_) {
+      ro.prefix_seek = true;
+    }
+    Iterator* iter = table_reader_->NewIterator(ro);
     if (convert_to_internal_key_) {
       return new KeyConvertingIterator(iter);
     } else {
@@ -380,6 +387,7 @@ class TableConstructor: public Constructor {
     source_.reset();
   }
   bool convert_to_internal_key_;
+  bool prefix_seek_;
 
   uint64_t uniq_id_;
   unique_ptr<StringSink> sink_;
@@ -549,6 +557,7 @@ enum TestType {
   BLOCK_BASED_TABLE_TEST,
   PLAIN_TABLE_SEMI_FIXED_PREFIX,
   PLAIN_TABLE_FULL_STR_PREFIX,
+  PLAIN_TABLE_TOTAL_ORDER,
   BLOCK_TEST,
   MEMTABLE_TEST,
   DB_TEST
@@ -565,8 +574,9 @@ static std::vector<TestArgs> GenerateArgList() {
   std::vector<TestArgs> test_args;
   std::vector<TestType> test_types = {
       BLOCK_BASED_TABLE_TEST,      PLAIN_TABLE_SEMI_FIXED_PREFIX,
-      PLAIN_TABLE_FULL_STR_PREFIX, BLOCK_TEST,
+      PLAIN_TABLE_FULL_STR_PREFIX, PLAIN_TABLE_TOTAL_ORDER,
-      MEMTABLE_TEST,               DB_TEST};
+      BLOCK_TEST,                  MEMTABLE_TEST,
+      DB_TEST};
   std::vector<bool> reverse_compare_types = {false, true};
   std::vector<int> restart_intervals = {16, 1, 1024};
 
@@ -689,8 +699,8 @@ class Harness {
         only_support_prefix_seek_ = true;
         options_.prefix_extractor = prefix_transform.get();
         options_.allow_mmap_reads = true;
-        options_.table_factory.reset(new PlainTableFactory());
+        options_.table_factory.reset(NewPlainTableFactory());
-        constructor_ = new TableConstructor(options_.comparator, true);
+        constructor_ = new TableConstructor(options_.comparator, true, true);
         internal_comparator_.reset(
             new InternalKeyComparator(options_.comparator));
         break;
@@ -699,8 +709,18 @@ class Harness {
         only_support_prefix_seek_ = true;
         options_.prefix_extractor = noop_transform.get();
         options_.allow_mmap_reads = true;
-        options_.table_factory.reset(new PlainTableFactory());
+        options_.table_factory.reset(NewPlainTableFactory());
-        constructor_ = new TableConstructor(options_.comparator, true);
+        constructor_ = new TableConstructor(options_.comparator, true, true);
+        internal_comparator_.reset(
+            new InternalKeyComparator(options_.comparator));
+        break;
+      case PLAIN_TABLE_TOTAL_ORDER:
+        support_prev_ = false;
+        only_support_prefix_seek_ = false;
+        options_.prefix_extractor = nullptr;
+        options_.allow_mmap_reads = true;
+        options_.table_factory.reset(NewTotalOrderPlainTableFactory());
+        constructor_ = new TableConstructor(options_.comparator, true, false);
         internal_comparator_.reset(
             new InternalKeyComparator(options_.comparator));
         break;