facebook · maysamyabandeh · Jun 7, 2018 · Aug 8, 2018 · Aug 8, 2018 · Aug 8, 2018
diff --git a/HISTORY.md b/HISTORY.md
@@ -2,6 +2,7 @@
 ## Unreleased
 ### Public API Change
 ### New Features
+* Changes the format of index blocks by delta encoding the index values, which are the block handles. This saves the encoding of BlockHandle::offset of the non-head index entries in each restart interval. The feature is backward compatible but not forward compatible. It is disabled by default unless format_version 4 or above is used.
 ### Bug Fixes
 
 ## 5.15.0 (7/17/2018)
@@ -13,7 +14,7 @@
 * The "rocksdb.num.entries" table property no longer counts range deletion tombstones as entries.
 
 ### New Features
-* Changes the format of index blocks by storing the key in their raw form rather than converting them to InternalKey. This saves 8 bytes per index key. The feature is backward compatbile but not forward compatible. It is disabled by default unless format_version 3 or above is used.
+* Changes the format of index blocks by storing the key in their raw form rather than converting them to InternalKey. This saves 8 bytes per index key. The feature is backward compatible but not forward compatible. It is disabled by default unless format_version 3 or above is used.
 * Avoid memcpy when reading mmap files with OpenReadOnly and max_open_files==-1.
 * Support dynamically changing `ColumnFamilyOptions::ttl` via `SetOptions()`.
 * Add a new table property, "rocksdb.num.range-deletions", which counts the number of range deletion tombstones in the table.

diff --git a/db/builder.h b/db/builder.h
@@ -35,7 +35,6 @@ class VersionEdit;
 class TableBuilder;
 class WritableFileWriter;
 class InternalStats;
-class InternalIterator;
 
 // @param column_family_name Name of the column family that is also identified
 //    by column_family_id, or empty string if unknown. It must outlive the

diff --git a/db/db_impl.cc b/db/db_impl.cc
@@ -1047,7 +1047,7 @@ InternalIterator* DBImpl::NewInternalIterator(
   } else {
     CleanupSuperVersion(super_version);
   }
-  return NewErrorInternalIterator(s, arena);
+  return NewErrorInternalIterator<Slice>(s, arena);
 }
 
 ColumnFamilyHandle* DBImpl::DefaultColumnFamily() const {

diff --git a/db/db_iter.h b/db/db_iter.h
@@ -23,7 +23,6 @@ namespace rocksdb {
 
 class Arena;
 class DBIter;
-class InternalIterator;
 
 // Return a new iterator that converts internal keys (yielded by
 // "*internal_iter") that were live at the specified "sequence" number

diff --git a/db/db_properties_test.cc b/db/db_properties_test.cc
@@ -180,17 +180,16 @@ void ParseTablePropertiesString(std::string tp_string, TableProperties* tp) {
   ResetTableProperties(tp);
   sscanf(tp_string.c_str(),
          "# data blocks %" SCNu64 " # entries %" SCNu64
-         " # range deletions %" SCNu64
-         " raw key size %" SCNu64
+         " # range deletions %" SCNu64 " raw key size %" SCNu64
          " raw average key size %lf "
          " raw value size %" SCNu64
          " raw average value size %lf "
          " data block size %" SCNu64 " index block size (user-key? %" SCNu64
-         ") %" SCNu64 " filter block size %" SCNu64,
+         ", delta-value? %" SCNu64 ") %" SCNu64 " filter block size %" SCNu64,
          &tp->num_data_blocks, &tp->num_entries, &tp->num_range_deletions,
          &tp->raw_key_size, &dummy_double, &tp->raw_value_size, &dummy_double,
-         &tp->data_size, &tp->index_key_is_user_key, &tp->index_size,
-         &tp->filter_size);
+         &tp->data_size, &tp->index_key_is_user_key,
+         &tp->index_value_is_delta_encoded, &tp->index_size, &tp->filter_size);
 }
 
 void VerifySimilar(uint64_t a, uint64_t b, double bias) {
@@ -224,14 +223,11 @@ void VerifyTableProperties(const TableProperties& base_tp,
   ASSERT_EQ(base_tp.num_range_deletions, new_tp.num_range_deletions);
 }
 
-void GetExpectedTableProperties(TableProperties* expected_tp,
-                                const int kKeySize, const int kValueSize,
-                                const int kKeysPerTable,
-                                const int kRangeDeletionsPerTable,
-                                const int kTableCount,
-                                const int kBloomBitsPerKey,
-                                const size_t kBlockSize,
-                                const bool index_key_is_user_key) {
+void GetExpectedTableProperties(
+    TableProperties* expected_tp, const int kKeySize, const int kValueSize,
+    const int kKeysPerTable, const int kRangeDeletionsPerTable,
+    const int kTableCount, const int kBloomBitsPerKey, const size_t kBlockSize,
+    const bool index_key_is_user_key, const bool value_delta_encoding) {
   const int kKeyCount = kTableCount * kKeysPerTable;
   const int kRangeDeletionCount = kTableCount * kRangeDeletionsPerTable;
   const int kAvgSuccessorSize = kKeySize / 5;
@@ -248,7 +244,9 @@ void GetExpectedTableProperties(TableProperties* expected_tp,
       kTableCount * (kKeysPerTable * (kKeySize + 8 + kValueSize));
   expected_tp->index_size =
       expected_tp->num_data_blocks *
-      (kAvgSuccessorSize + (index_key_is_user_key ? 0 : 8));
+      (kAvgSuccessorSize + (index_key_is_user_key ? 0 : 8) -
+       // discount 1 byte as value size is not encoded in value delta encoding
+       (value_delta_encoding ? 1 : 0));
   expected_tp->filter_size =
       kTableCount * (kKeysPerTable * kBloomBitsPerKey / 8);
 }
@@ -342,12 +340,14 @@ TEST_F(DBPropertiesTest, AggregatedTableProperties) {
     TableProperties output_tp;
     ParseTablePropertiesString(property, &output_tp);
     bool index_key_is_user_key = output_tp.index_key_is_user_key > 0;
+    bool value_is_delta_encoded = output_tp.index_value_is_delta_encoded > 0;
 
     TableProperties expected_tp;
     GetExpectedTableProperties(&expected_tp, kKeySize, kValueSize,
                                kKeysPerTable, kRangeDeletionsPerTable,
                                kTableCount, kBloomBitsPerKey,
-                               table_options.block_size, index_key_is_user_key);
+                               table_options.block_size, index_key_is_user_key,
+                               value_is_delta_encoded);
 
     VerifyTableProperties(expected_tp, output_tp);
   }
@@ -533,6 +533,7 @@ TEST_F(DBPropertiesTest, AggregatedTablePropertiesAtLevel) {
     db_->GetProperty(DB::Properties::kAggregatedTableProperties, &tp_string);
     ParseTablePropertiesString(tp_string, &tp);
     bool index_key_is_user_key = tp.index_key_is_user_key > 0;
+    bool value_is_delta_encoded = tp.index_value_is_delta_encoded > 0;
     ASSERT_EQ(sum_tp.data_size, tp.data_size);
     ASSERT_EQ(sum_tp.index_size, tp.index_size);
     ASSERT_EQ(sum_tp.filter_size, tp.filter_size);
@@ -542,10 +543,10 @@ TEST_F(DBPropertiesTest, AggregatedTablePropertiesAtLevel) {
     ASSERT_EQ(sum_tp.num_entries, tp.num_entries);
     ASSERT_EQ(sum_tp.num_range_deletions, tp.num_range_deletions);
     if (table > 3) {
-      GetExpectedTableProperties(
-          &expected_tp, kKeySize, kValueSize, kKeysPerTable,
-          kRangeDeletionsPerTable, table, kBloomBitsPerKey,
-          table_options.block_size, index_key_is_user_key);
+      GetExpectedTableProperties(&expected_tp, kKeySize, kValueSize,
+                                 kKeysPerTable, kRangeDeletionsPerTable, table,
+                                 kBloomBitsPerKey, table_options.block_size,
+                                 index_key_is_user_key, value_is_delta_encoded);
       // Gives larger bias here as index block size, filter block size,
       // and data block size become much harder to estimate in this test.
       VerifyTableProperties(expected_tp, tp, 0.5, 0.4, 0.4, 0.25);

diff --git a/db/db_test_util.cc b/db/db_test_util.cc
@@ -449,8 +449,8 @@ Options DBTestBase::GetOptions(
       options.prefix_extractor.reset(NewNoopTransform());
       break;
     }
-    case kBlockBasedTableWithPartitionedIndexFormat3: {
-      table_options.format_version = 3;
+    case kBlockBasedTableWithPartitionedIndexFormat4: {
+      table_options.format_version = 4;
       // Format 3 changes the binary index format. Since partitioned index is a
       // super-set of simple indexes, we are also using kTwoLevelIndexSearch to
       // test this format.

diff --git a/db/db_test_util.h b/db/db_test_util.h
@@ -698,7 +698,7 @@ class DBTestBase : public testing::Test {
     kLevelSubcompactions,
     kBlockBasedTableWithIndexRestartInterval,
     kBlockBasedTableWithPartitionedIndex,
-    kBlockBasedTableWithPartitionedIndexFormat3,
+    kBlockBasedTableWithPartitionedIndexFormat4,
     kPartitionedFilterWithNewTableReaderForCompactions,
     kUniversalSubcompactions,
     // This must be the last line

diff --git a/db/memtable.h b/db/memtable.h
@@ -34,7 +34,6 @@ namespace rocksdb {
 class Mutex;
 class MemTableIterator;
 class MergeContext;
-class InternalIterator;
 
 struct ImmutableMemTableOptions {
   explicit ImmutableMemTableOptions(const ImmutableCFOptions& ioptions,

diff --git a/db/merge_helper.h b/db/merge_helper.h
@@ -26,7 +26,6 @@ class Iterator;
 class Logger;
 class MergeOperator;
 class Statistics;
-class InternalIterator;
 
 class MergeHelper {
  public:

diff --git a/db/table_cache.cc b/db/table_cache.cc
@@ -238,7 +238,7 @@ InternalIterator* TableCache::NewIterator(
   if (s.ok()) {
     if (options.table_filter &&
         !options.table_filter(*table_reader->GetTableProperties())) {
-      result = NewEmptyInternalIterator(arena);
+      result = NewEmptyInternalIterator<Slice>(arena);
     } else {
       result = table_reader->NewIterator(options, prefix_extractor, arena,
                                          skip_filters, for_compaction);
@@ -279,7 +279,7 @@ InternalIterator* TableCache::NewIterator(
   }
   if (!s.ok()) {
     assert(result == nullptr);
-    result = NewErrorInternalIterator(s, arena);
+    result = NewErrorInternalIterator<Slice>(s, arena);
   }
   return result;
 }

diff --git a/db/table_cache.h b/db/table_cache.h
@@ -31,7 +31,6 @@ class Arena;
 struct FileDescriptor;
 class GetContext;
 class HistogramImpl;
-class InternalIterator;
 
 class TableCache {
  public:

diff --git a/db/version_edit.cc b/db/version_edit.cc
@@ -46,7 +46,7 @@ enum CustomTag : uint32_t {
   kTerminate = 1,  // The end of customized fields
   kNeedCompaction = 2,
   // Since Manifest is not entirely currently forward-compatible, and the only
-  // forward-compatbile part is the CutsomtTag of kNewFile, we currently encode
+  // forward-compatible part is the CutsomtTag of kNewFile, we currently encode
   // kMinLogNumberToKeep as part of a CustomTag as a hack. This should be
   // removed when manifest becomes forward-comptabile.
   kMinLogNumberToKeepHack = 3,
@@ -274,7 +274,7 @@ const char* VersionEdit::DecodeNewFile4From(Slice* input) {
           break;
         case kMinLogNumberToKeepHack:
           // This is a hack to encode kMinLogNumberToKeep in a
-          // forward-compatbile fashion.
+          // forward-compatible fashion.
           if (!GetFixed64(&field, &min_log_number_to_keep_)) {
             return "deleted log number malformatted";
           }

diff --git a/db/version_set.h b/db/version_set.h
@@ -52,7 +52,6 @@ class Writer;
 }
 
 class Compaction;
-class InternalIterator;
 class LogBuffer;
 class LookupKey;
 class MemTable;

diff --git a/include/rocksdb/table_properties.h b/include/rocksdb/table_properties.h
@@ -34,6 +34,7 @@ struct TablePropertiesNames {
   static const std::string kIndexPartitions;
   static const std::string kTopLevelIndexSize;
   static const std::string kIndexKeyIsUserKey;
+  static const std::string kIndexValueIsDeltaEncoded;
   static const std::string kFilterSize;
   static const std::string kRawKeySize;
   static const std::string kRawValueSize;
@@ -139,6 +140,8 @@ struct TableProperties {
   // Whether the index key is user key. Otherwise it includes 8 byte of sequence
   // number added by internal key format.
   uint64_t index_key_is_user_key = 0;
+  // Whether delta encoding is used to encode the index values.
+  uint64_t index_value_is_delta_encoded = 0;
   // the size of filter block.
   uint64_t filter_size = 0;
   // total raw key size