add counter for deletion dropping optimization

Summary:
add this counter stat to track usage of deletion-dropping optimization. if usage is low, we can delete it to prevent bugs like #2726.
Closes https://github.com/facebook/rocksdb/pull/2761

Differential Revision: D5665421

Pulled By: ajkr

fbshipit-source-id: 881befa2d199838dac88709e7b376a43d304e3d4

db/compaction_iteration_stats.h

@@ -16,6 +16,8 @@ struct CompactionIterationStats {
   int64_t num_record_drop_obsolete = 0;
   int64_t num_record_drop_range_del = 0;
   int64_t num_range_del_drop_obsolete = 0;
+  // Deletions obsoleted before bottom level due to file gap optimization.
+  int64_t num_optimized_del_drop_obsolete = 0;
   uint64_t total_filter_time = 0;
 
   // Input statistics

db/compaction_iterator.cc

@@ -111,6 +111,7 @@ void CompactionIterator::ResetRecordCounts() {
   iter_stats_.num_record_drop_obsolete = 0;
   iter_stats_.num_record_drop_range_del = 0;
   iter_stats_.num_range_del_drop_obsolete = 0;
+  iter_stats_.num_optimized_del_drop_obsolete = 0;
 }
 
 void CompactionIterator::SeekToFirst() {
@@ -426,6 +427,9 @@ void CompactionIterator::NextFromInput() {
           // Can compact out this SingleDelete.
           ++iter_stats_.num_record_drop_obsolete;
           ++iter_stats_.num_single_del_fallthru;
+          if (!bottommost_level_) {
+            ++iter_stats_.num_optimized_del_drop_obsolete;
+          }
         } else {
           // Output SingleDelete
           valid_ = true;
@@ -467,6 +471,9 @@ void CompactionIterator::NextFromInput() {
       // Note:  Dropping this Delete will not affect TransactionDB
       // write-conflict checking since it is earlier than any snapshot.
       ++iter_stats_.num_record_drop_obsolete;
+      if (!bottommost_level_) {
+        ++iter_stats_.num_optimized_del_drop_obsolete;
+      }
       input_->Next();
     } else if (ikey_.type == kTypeMerge) {
       if (!merge_helper_->HasOperator()) {

db/compaction_job.cc

@@ -1014,6 +1014,10 @@ void CompactionJob::RecordDroppedKeys(
     RecordTick(stats_, COMPACTION_RANGE_DEL_DROP_OBSOLETE,
                c_iter_stats.num_range_del_drop_obsolete);
   }
+  if (c_iter_stats.num_optimized_del_drop_obsolete > 0) {
+    RecordTick(stats_, COMPACTION_OPTIMIZED_DEL_DROP_OBSOLETE,
+               c_iter_stats.num_optimized_del_drop_obsolete);
+  }
 }
 
 Status CompactionJob::FinishCompactionOutputFile(

db/db_compaction_test.cc

@@ -2684,6 +2684,46 @@ TEST_P(DBCompactionTestWithParam, IntraL0CompactionDoesNotObsoleteDeletions) {
   ASSERT_TRUE(db_->Get(roptions, Key(0), &result).IsNotFound());
 }
 
+TEST_F(DBCompactionTest, OptimizedDeletionObsoleting) {
+  // Deletions can be dropped when compacted to non-last level if they fall
+  // outside the lower-level files' key-ranges.
+  const int kNumL0Files = 4;
+  Options options = CurrentOptions();
+  options.level0_file_num_compaction_trigger = kNumL0Files;
+  options.statistics = rocksdb::CreateDBStatistics();
+  DestroyAndReopen(options);
+
+  // put key 1 and 3 in separate L1, L2 files.
+  // So key 0, 2, and 4+ fall outside these levels' key-ranges.
+  for (int level = 2; level >= 1; --level) {
+    for (int i = 0; i < 2; ++i) {
+      Put(Key(2 * i + 1), "val");
+      Flush();
+    }
+    MoveFilesToLevel(level);
+    ASSERT_EQ(2, NumTableFilesAtLevel(level));
+  }
+
+  // Delete keys in range [1, 4]. These L0 files will be compacted with L1:
+  // - Tombstones for keys 2 and 4 can be dropped early.
+  // - Tombstones for keys 1 and 3 must be kept due to L2 files' key-ranges.
+  for (int i = 0; i < kNumL0Files; ++i) {
+    Put(Key(0), "val");  // sentinel to prevent trivial move
+    Delete(Key(i + 1));
+    Flush();
+  }
+  dbfull()->TEST_WaitForCompact();
+
+  for (int i = 0; i < kNumL0Files; ++i) {
+    std::string value;
+    ASSERT_TRUE(db_->Get(ReadOptions(), Key(i + 1), &value).IsNotFound());
+  }
+  ASSERT_EQ(2, options.statistics->getTickerCount(
+                   COMPACTION_OPTIMIZED_DEL_DROP_OBSOLETE));
+  ASSERT_EQ(2,
+            options.statistics->getTickerCount(COMPACTION_KEY_DROP_OBSOLETE));
+}
+
 INSTANTIATE_TEST_CASE_P(DBCompactionTestWithParam, DBCompactionTestWithParam,
                         ::testing::Values(std::make_tuple(1, true),
                                           std::make_tuple(1, false),

include/rocksdb/statistics.h

@@ -105,8 +105,9 @@ enum Tickers : uint32_t {
   COMPACTION_KEY_DROP_OBSOLETE,     // The key is obsolete.
   COMPACTION_KEY_DROP_RANGE_DEL,    // key was covered by a range tombstone.
   COMPACTION_KEY_DROP_USER,  // user compaction function has dropped the key.
-
   COMPACTION_RANGE_DEL_DROP_OBSOLETE,  // all keys in range were deleted.
+  // Deletions obsoleted before bottom level due to file gap optimization.
+  COMPACTION_OPTIMIZED_DEL_DROP_OBSOLETE,
 
   // Number of keys written to the database via the Put and Write call's
   NUMBER_KEYS_WRITTEN,
@@ -264,7 +265,9 @@ const std::vector<std::pair<Tickers, std::string>> TickersNameMap = {
     {COMPACTION_KEY_DROP_RANGE_DEL, "rocksdb.compaction.key.drop.range_del"},
     {COMPACTION_KEY_DROP_USER, "rocksdb.compaction.key.drop.user"},
     {COMPACTION_RANGE_DEL_DROP_OBSOLETE,
-     "rocksdb.compaction.range_del.drop.obsolete"},
+      "rocksdb.compaction.range_del.drop.obsolete"},
+    {COMPACTION_OPTIMIZED_DEL_DROP_OBSOLETE,
+      "rocksdb.compaction.optimized.del.drop.obsolete"},
     {NUMBER_KEYS_WRITTEN, "rocksdb.number.keys.written"},
     {NUMBER_KEYS_READ, "rocksdb.number.keys.read"},
     {NUMBER_KEYS_UPDATED, "rocksdb.number.keys.updated"},