Trigger compaction in CompactOnDeletionCollector based on deletion ratio (facebook#6806)

Summary:
In level compaction, if the total size (even if compensated after taking account of the deletions) of a level hasn't exceeded the limit, but there are lots of deletion entries in some SST files of the level, these files should also be good candidates for compaction. Otherwise, queries for the deleted keys might be slow because they need to go over all the tombstones.

This PR adds an option `deletion_ratio` to the factory of `CompactOnDeletionCollector` to configure it to trigger compaction when the ratio of tombstones >= `deletion_ratio`.
Pull Request resolved: facebook#6806

Test Plan:
Added new unit test in `compact_on_deletion_collector_test.cc`.
make compact_on_deletion_collector_test && ./compact_on_deletion_collector_test

Reviewed By: ajkr

Differential Revision: D21511981

Pulled By: cheng-chang

fbshipit-source-id: 65a9d0150e8c9c00337787686475252e4535a3e1

Author: Cheng Chang

include/rocksdb/utilities/table_properties_collectors.h

@@ -14,7 +14,8 @@ namespace ROCKSDB_NAMESPACE {
 
 // A factory of a table property collector that marks a SST
 // file as need-compaction when it observe at least "D" deletion
-// entries in any "N" consecutive entires.
+// entries in any "N" consecutive entires or the ratio of tombstone
+// entries in the whole file >= the specified deletion ratio.
 class CompactOnDeletionCollectorFactory
     : public TablePropertiesCollectorFactory {
  public:
@@ -34,6 +35,13 @@ class CompactOnDeletionCollectorFactory
     deletion_trigger_.store(deletion_trigger);
   }
 
+  // Change deletion ratio.
+  // @param deletion_ratio, if <= 0 or > 1, disable triggering compaction
+  //     based on deletion ratio.
+  void SetDeletionRatio(double deletion_ratio) {
+    deletion_ratio_.store(deletion_ratio);
+  }
+
   const char* Name() const override {
     return "CompactOnDeletionCollector";
   }
@@ -43,34 +51,45 @@ class CompactOnDeletionCollectorFactory
  private:
   friend std::shared_ptr<CompactOnDeletionCollectorFactory>
   NewCompactOnDeletionCollectorFactory(size_t sliding_window_size,
-                                       size_t deletion_trigger);
+                                       size_t deletion_trigger,
+                                       double deletion_ratio);
   // A factory of a table property collector that marks a SST
   // file as need-compaction when it observe at least "D" deletion
-  // entries in any "N" consecutive entires.
+  // entries in any "N" consecutive entires, or the ratio of tombstone
+  // entries >= deletion_ratio.
   //
   // @param sliding_window_size "N"
   // @param deletion_trigger "D"
+  // @param deletion_ratio, if <= 0 or > 1, disable triggering compaction
+  //     based on deletion ratio.
   CompactOnDeletionCollectorFactory(size_t sliding_window_size,
-                                    size_t deletion_trigger)
+                                    size_t deletion_trigger,
+                                    double deletion_ratio)
       : sliding_window_size_(sliding_window_size),
-        deletion_trigger_(deletion_trigger) {}
+        deletion_trigger_(deletion_trigger),
+        deletion_ratio_(deletion_ratio) {}
 
   std::atomic<size_t> sliding_window_size_;
   std::atomic<size_t> deletion_trigger_;
+  std::atomic<double> deletion_ratio_;
 };
 
 // Creates a factory of a table property collector that marks a SST
 // file as need-compaction when it observe at least "D" deletion
-// entries in any "N" consecutive entires.
+// entries in any "N" consecutive entires, or the ratio of tombstone
+// entries >= deletion_ratio.
 //
 // @param sliding_window_size "N". Note that this number will be
 //     round up to the smallest multiple of 128 that is no less
 //     than the specified size.
 // @param deletion_trigger "D".  Note that even when "N" is changed,
 //     the specified number for "D" will not be changed.
+// @param deletion_ratio, if <= 0 or > 1, disable triggering compaction
+//     based on deletion ratio. Disabled by default.
 extern std::shared_ptr<CompactOnDeletionCollectorFactory>
 NewCompactOnDeletionCollectorFactory(size_t sliding_window_size,
-                                     size_t deletion_trigger);
+                                     size_t deletion_trigger,
+                                     double deletion_ratio = 0);
 }  // namespace ROCKSDB_NAMESPACE
 
 #endif  // !ROCKSDB_LITE

utilities/table_properties_collectors/compact_on_deletion_collector.cc

@@ -13,12 +13,14 @@
 namespace ROCKSDB_NAMESPACE {
 
 CompactOnDeletionCollector::CompactOnDeletionCollector(
-    size_t sliding_window_size, size_t deletion_trigger)
+    size_t sliding_window_size, size_t deletion_trigger, double deletion_ratio)
     : bucket_size_((sliding_window_size + kNumBuckets - 1) / kNumBuckets),
       current_bucket_(0),
       num_keys_in_current_bucket_(0),
       num_deletions_in_observation_window_(0),
       deletion_trigger_(deletion_trigger),
+      deletion_ratio_(deletion_ratio),
+      deletion_ratio_enabled_(deletion_ratio > 0 && deletion_ratio <= 1),
       need_compaction_(false),
       finished_(false) {
   memset(num_deletions_in_buckets_, 0, sizeof(size_t) * kNumBuckets);
@@ -35,7 +37,7 @@ Status CompactOnDeletionCollector::AddUserKey(const Slice& /*key*/,
                                               SequenceNumber /*seq*/,
                                               uint64_t /*file_size*/) {
   assert(!finished_);
-  if (bucket_size_ == 0) {
+  if (!bucket_size_ && !deletion_ratio_enabled_) {
     // This collector is effectively disabled
     return Status::OK();
   }
@@ -45,54 +47,76 @@ Status CompactOnDeletionCollector::AddUserKey(const Slice& /*key*/,
     return Status::OK();
   }
 
-  if (num_keys_in_current_bucket_ == bucket_size_) {
-    // When the current bucket is full, advance the cursor of the
-    // ring buffer to the next bucket.
-    current_bucket_ = (current_bucket_ + 1) % kNumBuckets;
-
-    // Update the current count of observed deletion keys by excluding
-    // the number of deletion keys in the oldest bucket in the
-    // observation window.
-    assert(num_deletions_in_observation_window_ >=
-        num_deletions_in_buckets_[current_bucket_]);
-    num_deletions_in_observation_window_ -=
-        num_deletions_in_buckets_[current_bucket_];
-    num_deletions_in_buckets_[current_bucket_] = 0;
-    num_keys_in_current_bucket_ = 0;
+  if (deletion_ratio_enabled_) {
+    total_entries_++;
+    if (type == kEntryDelete) {
+      deletion_entries_++;
+    }
   }
 
-  num_keys_in_current_bucket_++;
-  if (type == kEntryDelete) {
-    num_deletions_in_observation_window_++;
-    num_deletions_in_buckets_[current_bucket_]++;
-    if (num_deletions_in_observation_window_ >= deletion_trigger_) {
-      need_compaction_ = true;
+  if (bucket_size_) {
+    if (num_keys_in_current_bucket_ == bucket_size_) {
+      // When the current bucket is full, advance the cursor of the
+      // ring buffer to the next bucket.
+      current_bucket_ = (current_bucket_ + 1) % kNumBuckets;
+
+      // Update the current count of observed deletion keys by excluding
+      // the number of deletion keys in the oldest bucket in the
+      // observation window.
+      assert(num_deletions_in_observation_window_ >=
+             num_deletions_in_buckets_[current_bucket_]);
+      num_deletions_in_observation_window_ -=
+          num_deletions_in_buckets_[current_bucket_];
+      num_deletions_in_buckets_[current_bucket_] = 0;
+      num_keys_in_current_bucket_ = 0;
+    }
+
+    num_keys_in_current_bucket_++;
+    if (type == kEntryDelete) {
+      num_deletions_in_observation_window_++;
+      num_deletions_in_buckets_[current_bucket_]++;
+      if (num_deletions_in_observation_window_ >= deletion_trigger_) {
+        need_compaction_ = true;
+      }
     }
   }
+
+  return Status::OK();
+}
+
+Status CompactOnDeletionCollector::Finish(
+    UserCollectedProperties* /*properties*/) {
+  if (!need_compaction_ && deletion_ratio_enabled_ && total_entries_ > 0) {
+    double ratio = static_cast<double>(deletion_entries_) / total_entries_;
+    need_compaction_ = ratio >= deletion_ratio_;
+  }
+  finished_ = true;
   return Status::OK();
 }
 
 TablePropertiesCollector*
 CompactOnDeletionCollectorFactory::CreateTablePropertiesCollector(
     TablePropertiesCollectorFactory::Context /*context*/) {
-  return new CompactOnDeletionCollector(
-      sliding_window_size_.load(), deletion_trigger_.load());
+  return new CompactOnDeletionCollector(sliding_window_size_.load(),
+                                        deletion_trigger_.load(),
+                                        deletion_ratio_.load());
 }
 
 std::string CompactOnDeletionCollectorFactory::ToString() const {
   std::ostringstream cfg;
   cfg << Name() << " (Sliding window size = " << sliding_window_size_.load()
-       << " Deletion trigger = " << deletion_trigger_.load() << ')';
+      << " Deletion trigger = " << deletion_trigger_.load()
+      << " Deletion ratio = " << deletion_ratio_.load() << ')';
   return cfg.str();
 }
 
 std::shared_ptr<CompactOnDeletionCollectorFactory>
-    NewCompactOnDeletionCollectorFactory(
-        size_t sliding_window_size,
-        size_t deletion_trigger) {
+NewCompactOnDeletionCollectorFactory(size_t sliding_window_size,
+                                     size_t deletion_trigger,
+                                     double deletion_ratio) {
   return std::shared_ptr<CompactOnDeletionCollectorFactory>(
-      new CompactOnDeletionCollectorFactory(
-          sliding_window_size, deletion_trigger));
+      new CompactOnDeletionCollectorFactory(sliding_window_size,
+                                            deletion_trigger, deletion_ratio));
 }
 }  // namespace ROCKSDB_NAMESPACE
 #endif  // !ROCKSDB_LITE

utilities/table_properties_collectors/compact_on_deletion_collector.h

@@ -11,9 +11,8 @@ namespace ROCKSDB_NAMESPACE {
 
 class CompactOnDeletionCollector : public TablePropertiesCollector {
  public:
-  CompactOnDeletionCollector(
-      size_t sliding_window_size,
-      size_t deletion_trigger);
+  CompactOnDeletionCollector(size_t sliding_window_size,
+                             size_t deletion_trigger, double deletion_raatio);
 
   // AddUserKey() will be called when a new key/value pair is inserted into the
   // table.
@@ -28,10 +27,7 @@ class CompactOnDeletionCollector : public TablePropertiesCollector {
   // for writing the properties block.
   // @params properties  User will add their collected statistics to
   // `properties`.
-  virtual Status Finish(UserCollectedProperties* /*properties*/) override {
-    finished_ = true;
-    return Status::OK();
-  }
+  virtual Status Finish(UserCollectedProperties* /*properties*/) override;
 
   // Return the human-readable properties, where the key is property name and
   // the value is the human-readable form of value.
@@ -64,6 +60,10 @@ class CompactOnDeletionCollector : public TablePropertiesCollector {
   size_t num_keys_in_current_bucket_;
   size_t num_deletions_in_observation_window_;
   size_t deletion_trigger_;
+  const double deletion_ratio_;
+  const bool deletion_ratio_enabled_;
+  size_t total_entries_ = 0;
+  size_t deletion_entries_ = 0;
   // true if the current SST file needs to be compacted.
   bool need_compaction_;
   bool finished_;