Add more black-box tests for PlainTable and explicitly support total order mode

Summary:
1. Add some more implementation-aware tests for PlainTable
2. move from a hard-coded one index per 16 rows in one prefix to a configurable number. Also, make hash table ratio = 0  means binary search only. Also fixes some divide 0 risks.
3. Explicitly support total order (only use binary search)
4. some code cleaning up.

Test Plan: make all check

Reviewers: haobo, kailiu

Reviewed By: haobo

CC: leveldb

Differential Revision: https://reviews.facebook.net/D16023

db/db_impl.cc

@@ -2754,7 +2754,10 @@ Iterator* DBImpl::NewInternalIterator(const ReadOptions& options,
 
 Iterator* DBImpl::TEST_NewInternalIterator() {
   SequenceNumber ignored;
-  return NewInternalIterator(ReadOptions(), &ignored);
+  ReadOptions read_options;
+  // Use prefix_seek to make the test function more useful.
+  read_options.prefix_seek = true;
+  return NewInternalIterator(read_options, &ignored);
 }
 
 std::pair<Iterator*, Iterator*> DBImpl::GetTailingIteratorPair(

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 +181,298 @@ 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(storage_options, icomparator, file_size,
+                         bloom_bits_per_key, hash_table_ratio, index_sparseness,
+                         table_properties),
+        expect_bloom_not_match_(expect_bloom_not_match) {
+    file_ = std::move(file);
+    options_ = options;
+    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 +485,21 @@ 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 +511,6 @@ 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 +518,220 @@ 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();
+    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;
+  }
+}
+
+TEST(PlainTableDBTest, HashBucketConflictReverseSuffixComparator) {
+  for (unsigned char i = 1; i <= 3; i++) {
+    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;
+  }
+}
+
+TEST(PlainTableDBTest, NonExistingKeyToNonEmptyBucket) {
+  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
+  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"));
+
   dbfull()->TEST_FlushMemTable();
 
-  ASSERT_OK(Put("1000000000000foo", "v2"));
+  ASSERT_EQ("v1", Get("5000000000000fo0"));
-  dbfull()->TEST_FlushMemTable();
+  ASSERT_EQ("v2", Get("5000000000000fo1"));
-  ASSERT_EQ("v2", Get("1000000000000foo"));
+  ASSERT_EQ("v3", Get("5000000000000fo2"));
 
-  ASSERT_OK(Put("0000000000000eee", "v3"));
+  ASSERT_EQ("NOT_FOUND", Get("8000000000000bar"));
-  dbfull()->TEST_FlushMemTable();
+  ASSERT_EQ("NOT_FOUND", Get("1000000000000bar"));
-  ASSERT_EQ("v3", Get("0000000000000eee"));
 
-  ASSERT_OK(Delete("0000000000000bar"));
+  Iterator* iter = dbfull()->NewIterator(ro_);
-  dbfull()->TEST_FlushMemTable();
-  ASSERT_EQ("NOT_FOUND", Get("0000000000000bar"));
 
-  ASSERT_OK(Put("0000000000000eee", "v5"));
+  iter->Seek("5000000000000bar");
-  dbfull()->TEST_FlushMemTable();
+  ASSERT_TRUE(iter->Valid());
-  ASSERT_EQ("v5", Get("0000000000000eee"));
+  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) {

include/rocksdb/table.h

@@ -60,20 +60,44 @@ struct BlockBasedTableOptions {
 extern TableFactory* NewBlockBasedTableFactory(
     const BlockBasedTableOptions& table_options = BlockBasedTableOptions());
 
+// -- 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 prefix. You
+//                      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(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.
-// @hash_table_ratio: the desired utilization of the hash table used for prefix
+// @index_sparseness: need to build one index record for how many keys for
-//                    hashing. hash_table_ratio = number of prefixes / #buckets
+//                    binary search.
-//                    in the hash table
+extern TableFactory* NewTotalOrderPlainTableFactory(
-const uint32_t kPlainTableVariableLength = 0;
-extern TableFactory* NewPlainTableFactory(
     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 {

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

@@ -48,7 +48,7 @@ inline uint32_t GetBucketIdFromHash(uint32_t hash, uint32_t num_buckets) {
 // 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 +71,7 @@ class PlainTableIterator : public Iterator {
 
  private:
   PlainTableReader* table_;
+  bool use_prefix_seek_;
   uint32_t offset_;
   uint32_t next_offset_;
   Slice key_;
@@ -87,12 +88,14 @@ PlainTableReader::PlainTableReader(const EnvOptions& storage_options,
                                    const InternalKeyComparator& icomparator,
                                    uint64_t file_size, int bloom_bits_per_key,
                                    double hash_table_ratio,
+                                   size_t index_sparseness,
                                    const TableProperties* table_properties)
     : soptions_(storage_options),
       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) {}
@@ -103,14 +106,12 @@ PlainTableReader::~PlainTableReader() {
   delete bloom_;
 }
 
-Status PlainTableReader::Open(const Options& options,
+Status PlainTableReader::Open(
-                              const EnvOptions& soptions,
+    const Options& options, const EnvOptions& soptions,
-                              const InternalKeyComparator& internal_comparator,
+    const InternalKeyComparator& internal_comparator,
-                              unique_ptr<RandomAccessFile>&& file,
+    unique_ptr<RandomAccessFile>&& file, uint64_t file_size,
-                              uint64_t file_size,
+    unique_ptr<TableReader>* table_reader, const int bloom_bits_per_key,
-                              unique_ptr<TableReader>* table_reader,
+    double hash_table_ratio, size_t index_sparseness) {
-                              const int bloom_bits_per_key,
-                              double hash_table_ratio) {
   assert(options.allow_mmap_reads);
 
   if (file_size > kMaxFileSize) {
@@ -124,9 +125,9 @@ 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,
+      soptions, internal_comparator, file_size, bloom_bits_per_key,
-                           bloom_bits_per_key, hash_table_ratio, props));
+      hash_table_ratio, index_sparseness, props));
   new_reader->file_ = std::move(file);
   new_reader->options_ = options;
 
@@ -148,7 +149,7 @@ 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 {
@@ -204,7 +205,9 @@ 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,
+                                                 DynamicBloom* bloom_) const {
   Slice prev_key_prefix_slice;
   uint32_t prev_key_prefix_hash = 0;
   uint32_t pos = data_start_offset_;
@@ -214,16 +217,23 @@ int PlainTableReader::PopulateIndexRecordList(IndexRecordList* record_list) {
   // 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, &pos);
+    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);
       }
@@ -232,7 +242,8 @@ int PlainTableReader::PopulateIndexRecordList(IndexRecordList* record_list) {
       prev_key_prefix_hash = GetSliceHash(key_prefix_slice);
     }
 
-    if (key_index_within_prefix++ % kIndexIntervalForSamePrefixKeys == 0) {
+    if (kIndexIntervalForSamePrefixKeys == 0 ||
+        key_index_within_prefix++ % kIndexIntervalForSamePrefixKeys == 0) {
       // Add an index key for every kIndexIntervalForSamePrefixKeys keys
       record_list->AddRecord(prev_key_prefix_hash, key_offset);
     }
@@ -243,18 +254,27 @@ int PlainTableReader::PopulateIndexRecordList(IndexRecordList* record_list) {
   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_;
 
-  if (kBloomBitsPerKey > 0) {
+  if (options_.prefix_extractor != nullptr) {
-    bloom_ = new DynamicBloom(num_prefixes * kBloomBitsPerKey);
+    uint32_t bloom_total_bits = num_prefixes * kBloomBitsPerKey;
+    if (bloom_total_bits > 0) {
+      bloom_ = new DynamicBloom(bloom_total_bits);
+    }
+  }
+
+  if (options_.prefix_extractor == nullptr || kHashTableRatio <= 0) {
+    // Fall back to pure binary search if the user fails to specify a prefix
+    // extractor.
+    hash_table_size_ = 1;
+  } else {
+    double hash_table_size_multipier = 1.0 / kHashTableRatio;
+    hash_table_size_ = num_prefixes * hash_table_size_multipier + 1;
   }
-  double hash_table_size_multipier =
-      (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_];
 }
 
@@ -272,7 +292,7 @@ size_t PlainTableReader::BucketizeIndexesAndFillBloom(
     if (first || prev_hash != cur_hash) {
       prev_hash = cur_hash;
       first = false;
-      if (bloom_) {
+      if (bloom_ && !IsTotalOrderMode()) {
         bloom_->AddHash(cur_hash);
       }
     }
@@ -362,6 +382,12 @@ void PlainTableReader::FillIndexes(
 }
 
 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()) {
@@ -373,7 +399,20 @@ 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_ = new DynamicBloom(num_bloom_bits);
+    }
+  }
+
+  s = PopulateIndexRecordList(&record_list, &num_prefixes, bloom_);
+  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);
@@ -392,16 +431,16 @@ 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* ret_offset) const {
   prefix_matched = false;
   int bucket = GetBucketIdFromHash(prefix_hash, hash_table_size_);
   uint32_t bucket_value = hash_table_[bucket];
   if (bucket_value == data_end_offset_) {
-    ret_offset = data_end_offset_;
+    *ret_offset = data_end_offset_;
     return Status::OK();
   } else if ((bucket_value & kSubIndexMask) == 0) {
     // point directly to the file
-    ret_offset = bucket_value;
+    *ret_offset = bucket_value;
     return Status::OK();
   }
 
@@ -426,7 +465,7 @@ Status PlainTableReader::GetOffset(const Slice& target, const Slice& prefix,
     uint32_t mid = (high + low) / 2;
     uint32_t file_offset = 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;
+        *ret_offset = file_offset;
         return Status::OK();
       } else {
         high = mid;
@@ -451,34 +490,34 @@ Status PlainTableReader::GetOffset(const Slice& target, const Slice& prefix,
   ParsedInternalKey low_key;
   size_t tmp;
   uint32_t low_key_offset = 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;
+    *ret_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];
+    *ret_offset = 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_;
+    *ret_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,
-                                 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_;
@@ -491,7 +530,7 @@ Status PlainTableReader::ReadKey(const char* row_ptr, ParsedInternalKey* key,
       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 - row_ptr;
   }
   if (key_ptr + user_key_size + 1 >= file_data_.data() + data_end_offset_) {
     return Status::Corruption("Unable to read the next key");
@@ -502,7 +541,7 @@ 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_) {
       return Status::Corruption("Unable to read the next key");
@@ -510,16 +549,16 @@ Status PlainTableReader::ReadKey(const char* row_ptr, ParsedInternalKey* 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,
-                              Slice* value, uint32_t& next_offset) {
+                              Slice* value, uint32_t* next_offset) const {
   if (offset == data_end_offset_) {
-    next_offset = data_end_offset_;
+    *next_offset = data_end_offset_;
     return Status::OK();
   }
 
@@ -529,7 +568,7 @@ Status PlainTableReader::Next(uint32_t offset, ParsedInternalKey* key,
 
   const char* row_ptr = file_data_.data() + offset;
   size_t bytes_for_key;
-  Status s = ReadKey(row_ptr, key, bytes_for_key);
+  Status s = ReadKey(row_ptr, key, &bytes_for_key);
   uint32_t value_size;
   const char* value_ptr = GetVarint32Ptr(row_ptr + bytes_for_key,
                                          file_data_.data() + data_end_offset_,
@@ -537,8 +576,8 @@ Status PlainTableReader::Next(uint32_t offset, ParsedInternalKey* key,
   if (value_ptr == nullptr) {
     return Status::Corruption("Error reading value length.");
   }
-  next_offset = offset + (value_ptr - row_ptr) + value_size;
+  *next_offset = offset + (value_ptr - row_ptr) + value_size;
-  if (next_offset > data_end_offset_) {
+  if (*next_offset > data_end_offset_) {
     return Status::Corruption("Reach end of file when reading value");
   }
   *value = Slice(value_ptr, value_size);
@@ -552,14 +591,28 @@ 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 +624,7 @@ 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, &offset);
     if (!s.ok()) {
       return s;
     }
@@ -596,8 +649,9 @@ 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 +674,39 @@ 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_->hash_table_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 +736,7 @@ 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_, &next_offset_);
     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);
 
@@ -71,82 +72,15 @@ class PlainTableReader: public TableReader {
   PlainTableReader(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();
 
- private:
+ protected:
-  struct IndexRecord;
+  // Check bloom filter to see whether it might contain this prefix.
-  class IndexRecordList;
+  // The hash of the prefix is given, since it can be reused for index lookup
-
+  // too.
-  uint32_t* hash_table_ = nullptr;
+  virtual bool MatchBloom(uint32_t hash) const;
-  int hash_table_size_ = 0;
-  char* sub_index_ = nullptr;
-
-  Options options_;
-  const EnvOptions& soptions_;
-  const InternalKeyComparator internal_comparator_;
-  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;
-
-  std::shared_ptr<const TableProperties> table_properties_;
-  const uint32_t data_start_offset_ = 0;
-  const uint32_t data_end_offset_;
-  const size_t user_key_len_;
-
-  static const size_t kNumInternalBytes = 8;
-  static const uint32_t kSubIndexMask = 0x80000000;
-  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;
-  }
-
-  size_t GetFixedInternalKeyLength() const {
-    return user_key_len_ + kNumInternalBytes;
-  }
-
-  friend class TableCache;
-  friend class PlainTableIterator;
-
-  // Internal helper function to generate an IndexRecordList object from all
-  // the rows, which contains index records as a list.
-  int PopulateIndexRecordList(IndexRecordList* record_list);
-
-  // 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
-  // list
-  // of offsets for the hash, in reversed order.
-  // bucket_count is sized of hash_table_size_. The value is how many index
-  // records are there in hash_to_offsets for the same bucket.
-  size_t BucketizeIndexesAndFillBloom(
-      IndexRecordList& record_list, int num_prefixes,
-      std::vector<IndexRecord*>* hash_to_offsets,
-      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
-  // and counts generated by BucketizeIndexesAndFillBloom().
-  void FillIndexes(size_t sub_index_size_needed,
-                   const std::vector<IndexRecord*>& hash_to_offsets,
-                   const std::vector<uint32_t>& bucket_count);
 
   // PopulateIndex() builds index of keys. It must be called before any query
   // to the table.
@@ -188,32 +122,122 @@ class PlainTableReader: public TableReader {
   // <end>
   Status PopulateIndex();
 
-  // Check bloom filter to see whether it might contain this prefix.
+  Options options_;
-  // The hash of the prefix is given, since it can be reused for index lookup
+  unique_ptr<RandomAccessFile> file_;
-  // too.
+
-  bool MayHavePrefix(uint32_t hash);
+ private:
+  struct IndexRecord;
+  class IndexRecordList;
+
+  uint32_t* hash_table_ = nullptr;
+  int hash_table_size_ = 0;
+  char* sub_index_ = nullptr;
+
+  const EnvOptions& soptions_;
+  const InternalKeyComparator internal_comparator_;
+  Status status_;
+
+  Slice file_data_;
+  uint32_t version_;
+  uint32_t file_size_;
+
+  const double kHashTableRatio;
+  const int kBloomBitsPerKey;
+  // 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;
+  DynamicBloom* bloom_ = nullptr;
+
+  std::shared_ptr<const TableProperties> table_properties_;
+  const uint32_t data_start_offset_ = 0;
+  const uint32_t data_end_offset_;
+  const size_t user_key_len_;
+
+  static const size_t kNumInternalBytes = 8;
+  static const uint32_t kSubIndexMask = 0x80000000;
+  static const size_t kOffsetLen = sizeof(uint32_t);
+  static const uint64_t kMaxFileSize = 1u << 31;
+  static const size_t kRecordsPerGroup = 256;
+
+  bool IsFixedLength() const {
+    return user_key_len_ != kPlainTableVariableLength;
+  }
+
+  size_t GetFixedInternalKeyLength() const {
+    return user_key_len_ + kNumInternalBytes;
+  }
+
+  friend class TableCache;
+  friend class PlainTableIterator;
+
+  // Internal helper function to generate an IndexRecordList object from all
+  // the rows, which contains index records as a 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, DynamicBloom* bloom_) 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
+  // list
+  // of offsets for the hash, in reversed order.
+  // bucket_count is sized of hash_table_size_. The value is how many index
+  // records are there in hash_to_offsets for the same bucket.
+  size_t BucketizeIndexesAndFillBloom(
+      IndexRecordList& record_list, int num_prefixes,
+      std::vector<IndexRecord*>* hash_to_offsets,
+      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
+  // and counts generated by BucketizeIndexesAndFillBloom().
+  void FillIndexes(size_t sub_index_size_needed,
+                   const std::vector<IndexRecord*>& hash_to_offsets,
+                   const std::vector<uint32_t>& bucket_count);
 
   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.
   // tmp_slice is a tmp slice.
   // return next_offset as the offset for the next key.
   Status Next(uint32_t offset, ParsedInternalKey* key, Slice* value,
-              uint32_t& next_offset);
+              uint32_t* next_offset) const;
   // 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* ret_offset) const;
 
-  Slice GetPrefix(const Slice& target) {
+  Slice GetUserKey(const Slice& key) const {
-    assert(target.size() >= 8);  // target is internal key
+    return Slice(key.data(), key.size() - 8);
-    return options_.prefix_extractor->Transform(
-        Slice(target.data(), target.size() - 8));
   }
 
-  Slice GetPrefix(const ParsedInternalKey& target);
+  Slice GetPrefix(const Slice& target) const {
+    assert(target.size() >= 8);  // target is internal key
+    return GetPrefixFromUserKey(GetUserKey(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_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_;
@@ -548,6 +556,7 @@ 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
@@ -564,8 +573,9 @@ 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};
 
@@ -688,8 +698,8 @@ 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;
@@ -698,8 +708,18 @@ 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;