Change The Way Level Target And Compaction Score Are Calculated (#10057)

Summary:
The current level targets for dynamical leveling has a problem: the target level size will dramatically change after a L0->L1 compaction. When there are many L0 bytes, lower level compactions are delayed, but they will be resumed after the L0->L1 compaction finishes, so the expected write amplification benefits might not be realized. The proposal here is to revert the level targetting size, but instead relying on adjusting score for each level to prioritize levels that need to compact most.
Basic idea:
(1) target level size isn't adjusted, but score is adjusted. The reasoning is that with parallel compactions, holding compactions from happening might not be desirable, but we would like the compactions are scheduled from the level we feel most needed. For example, if we have a extra-large L2, we would like all compactions are scheduled for L2->L3 compactions, rather than L4->L5. This gets complicated when a large L0->L1 compaction is going on. Should we compact L2->L3 or L4->L5. So the proposal for that is:
(2) the score is calculated by actual level size / (target size + estimated upper bytes coming down). The reasoning is that if we have a large amount of pending L0/L1 bytes coming down, compacting L2->L3 might be more expensive, as when the L0 bytes are compacted down to L2, the actual L2->L3 fanout would change dramatically. On the other hand, when the amount of bytes coming down to L5, the impacts to L5->L6 fanout are much less. So when calculating target score, we can adjust it by adding estimated downward bytes to the target level size.

Pull Request resolved: https://github.com/facebook/rocksdb/pull/10057

Test Plan: Repurpose tests VersionStorageInfoTest.MaxBytesForLevelDynamicWithLargeL0_* tests to cover this scenario.

Reviewed By: ajkr

Differential Revision: D37539742

fbshipit-source-id: 9c154cbfe92023f918cf5d80875d8776ad4831a4

HISTORY.md

@@ -16,6 +16,9 @@
 * Fix a bug in which backup/checkpoint can include a WAL deleted by RocksDB.
 * Fix a bug where concurrent compactions might cause unnecessary further write stalling. In some cases, this might cause write rate to drop to minimum.
 
+## Behavior Change
+* In leveled compaction with dynamic levelling, level multiplier is not anymore adjusted due to oversized L0. Instead, compaction score is adjusted by increasing size level target by adding incoming bytes from upper levels. This would deprioritize compactions from upper levels if more data from L0 is coming. This is to fix some unnecessary full stalling due to drastic change of level targets, while not wasting write bandwidth for compaction while writes are overloaded.
+
 ## 7.4.0 (06/19/2022)
 ### Bug Fixes
 * Fixed a bug in calculating key-value integrity protection for users of in-place memtable updates. In particular, the affected users would be those who configure `protection_bytes_per_key > 0` on `WriteBatch` or `WriteOptions`, and configure `inplace_callback != nullptr`.

db/version_set.cc

@@ -2657,6 +2657,16 @@ uint32_t GetExpiredTtlFilesCount(const ImmutableOptions& ioptions,
 void VersionStorageInfo::ComputeCompactionScore(
     const ImmutableOptions& immutable_options,
     const MutableCFOptions& mutable_cf_options) {
+  double total_downcompact_bytes = 0.0;
+  // Historically, score is defined as actual bytes in a level divided by
+  // the level's target size, and 1.0 is the threshold for triggering
+  // compaction. Higher score means higher prioritization.
+  // Now we keep the compaction triggering condition, but consider more
+  // factors for priorization, while still keeping the 1.0 threshold.
+  // In order to provide flexibility for reducing score while still
+  // maintaining it to be over 1.0, we scale the original score by 10x
+  // if it is larger than 1.0.
+  const double kScoreScale = 10.0;
   for (int level = 0; level <= MaxInputLevel(); level++) {
     double score;
     if (level == 0) {
@@ -2674,6 +2684,7 @@ void VersionStorageInfo::ComputeCompactionScore(
       int num_sorted_runs = 0;
       uint64_t total_size = 0;
       for (auto* f : files_[level]) {
+        total_downcompact_bytes += static_cast<double>(f->fd.GetFileSize());
         if (!f->being_compacted) {
           total_size += f->compensated_file_size;
           num_sorted_runs++;
@@ -2737,18 +2748,40 @@ void VersionStorageInfo::ComputeCompactionScore(
           }
           score =
               std::max(score, static_cast<double>(total_size) / l0_target_size);
+          if (immutable_options.level_compaction_dynamic_level_bytes &&
+              score > 1.0) {
+            score *= kScoreScale;
+          }
         }
       }
     } else {
       // Compute the ratio of current size to size limit.
       uint64_t level_bytes_no_compacting = 0;
+      uint64_t level_total_bytes = 0;
       for (auto f : files_[level]) {
+        level_total_bytes += f->fd.GetFileSize();
         if (!f->being_compacted) {
           level_bytes_no_compacting += f->compensated_file_size;
         }
       }
+      if (!immutable_options.level_compaction_dynamic_level_bytes ||
+          level_bytes_no_compacting < MaxBytesForLevel(level)) {
         score = static_cast<double>(level_bytes_no_compacting) /
                 MaxBytesForLevel(level);
+      } else {
+        // If there are a large mount of data being compacted down to the
+        // current level soon, we would de-prioritize compaction from
+        // a level where the incoming data would be a large ratio. We do
+        // it by dividing level size not by target level size, but
+        // the target size and the incoming compaction bytes.
+        score = static_cast<double>(level_bytes_no_compacting) /
+                (MaxBytesForLevel(level) + total_downcompact_bytes) *
+                kScoreScale;
+      }
+      if (level_total_bytes > MaxBytesForLevel(level)) {
+        total_downcompact_bytes +=
+            static_cast<double>(level_total_bytes - MaxBytesForLevel(level));
+      }
     }
     compaction_level_[level] = level;
     compaction_score_[level] = score;
@@ -3775,13 +3808,7 @@ void VersionStorageInfo::CalculateBaseBytes(const ImmutableOptions& ioptions,
       // No compaction from L1+ needs to be scheduled.
       base_level_ = num_levels_ - 1;
     } else {
-      uint64_t l0_size = 0;
-      for (const auto& f : files_[0]) {
-        l0_size += f->fd.GetFileSize();
-      }
-
-      uint64_t base_bytes_max =
-          std::max(options.max_bytes_for_level_base, l0_size);
+      uint64_t base_bytes_max = options.max_bytes_for_level_base;
       uint64_t base_bytes_min = static_cast<uint64_t>(
           base_bytes_max / options.max_bytes_for_level_multiplier);
 
@@ -3823,26 +3850,6 @@ void VersionStorageInfo::CalculateBaseBytes(const ImmutableOptions& ioptions,
 
       level_multiplier_ = options.max_bytes_for_level_multiplier;
       assert(base_level_size > 0);
-      if (l0_size > base_level_size &&
-          (l0_size > options.max_bytes_for_level_base ||
-           static_cast<int>(files_[0].size() / 2) >=
-               options.level0_file_num_compaction_trigger)) {
-        // We adjust the base level according to actual L0 size, and adjust
-        // the level multiplier accordingly, when:
-        //   1. the L0 size is larger than level size base, or
-        //   2. number of L0 files reaches twice the L0->L1 compaction trigger
-        // We don't do this otherwise to keep the LSM-tree structure stable
-        // unless the L0 compaction is backlogged.
-        base_level_size = l0_size;
-        if (base_level_ == num_levels_ - 1) {
-          level_multiplier_ = 1.0;
-        } else {
-          level_multiplier_ = std::pow(
-              static_cast<double>(max_level_size) /
-                  static_cast<double>(base_level_size),
-              1.0 / static_cast<double>(num_levels_ - base_level_ - 1));
-        }
-      }
 
       uint64_t level_size = base_level_size;
       for (int i = base_level_; i < num_levels_; i++) {

db/version_set_test.cc

@@ -376,73 +376,80 @@ TEST_F(VersionStorageInfoTest, MaxBytesForLevelDynamicWithLargeL0_1) {
   ASSERT_EQ(2, vstorage_.base_level());
   // level multiplier should be 3.5
   ASSERT_EQ(vstorage_.level_multiplier(), 5.0);
-  // Level size should be around 30,000, 105,000, 367,500
   ASSERT_EQ(40000U, vstorage_.MaxBytesForLevel(2));
   ASSERT_EQ(51450U, vstorage_.MaxBytesForLevel(3));
   ASSERT_EQ(257250U, vstorage_.MaxBytesForLevel(4));
+
+  vstorage_.ComputeCompactionScore(ioptions_, mutable_cf_options_);
+  // Only L0 hits compaction.
+  ASSERT_EQ(vstorage_.CompactionScoreLevel(0), 0);
 }
 
 TEST_F(VersionStorageInfoTest, MaxBytesForLevelDynamicWithLargeL0_2) {
   ioptions_.level_compaction_dynamic_level_bytes = true;
   mutable_cf_options_.max_bytes_for_level_base = 10000;
   mutable_cf_options_.max_bytes_for_level_multiplier = 5;
-  mutable_cf_options_.level0_file_num_compaction_trigger = 2;
+  mutable_cf_options_.level0_file_num_compaction_trigger = 4;
 
   Add(0, 11U, "1", "2", 10000U);
   Add(0, 12U, "1", "2", 10000U);
   Add(0, 13U, "1", "2", 10000U);
 
+  // Level size should be around 10,000, 10,290, 51,450, 257,250
   Add(5, 4U, "1", "2", 1286250U);
-  Add(4, 5U, "1", "2", 200000U);
-  Add(3, 6U, "1", "2", 40000U);
-  Add(2, 7U, "1", "2", 8000U);
+  Add(4, 5U, "1", "2", 258000U);  // unadjusted score 1.003
+  Add(3, 6U, "1", "2", 53000U);   // unadjusted score 1.03
+  Add(2, 7U, "1", "2", 20000U);   // unadjusted score 1.94
 
   UpdateVersionStorageInfo();
 
   ASSERT_EQ(0, logger_->log_count);
-  ASSERT_EQ(2, vstorage_.base_level());
-  // level multiplier should be 3.5
-  ASSERT_LT(vstorage_.level_multiplier(), 3.6);
-  ASSERT_GT(vstorage_.level_multiplier(), 3.4);
-  // Level size should be around 30,000, 105,000, 367,500
-  ASSERT_EQ(30000U, vstorage_.MaxBytesForLevel(2));
-  ASSERT_LT(vstorage_.MaxBytesForLevel(3), 110000U);
-  ASSERT_GT(vstorage_.MaxBytesForLevel(3), 100000U);
-  ASSERT_LT(vstorage_.MaxBytesForLevel(4), 370000U);
-  ASSERT_GT(vstorage_.MaxBytesForLevel(4), 360000U);
+  ASSERT_EQ(1, vstorage_.base_level());
+  ASSERT_EQ(10000U, vstorage_.MaxBytesForLevel(1));
+  ASSERT_EQ(10290U, vstorage_.MaxBytesForLevel(2));
+  ASSERT_EQ(51450U, vstorage_.MaxBytesForLevel(3));
+  ASSERT_EQ(257250U, vstorage_.MaxBytesForLevel(4));
+
+  vstorage_.ComputeCompactionScore(ioptions_, mutable_cf_options_);
+  // Although L2 and l3 have higher unadjusted compaction score, considering
+  // a relatively large L0 being compacted down soon, L4 is picked up for
+  // compaction.
+  // L0 is still picked up for oversizing.
+  ASSERT_EQ(0, vstorage_.CompactionScoreLevel(0));
+  ASSERT_EQ(4, vstorage_.CompactionScoreLevel(1));
 }
 
 TEST_F(VersionStorageInfoTest, MaxBytesForLevelDynamicWithLargeL0_3) {
   ioptions_.level_compaction_dynamic_level_bytes = true;
-  mutable_cf_options_.max_bytes_for_level_base = 10000;
+  mutable_cf_options_.max_bytes_for_level_base = 20000;
   mutable_cf_options_.max_bytes_for_level_multiplier = 5;
-  mutable_cf_options_.level0_file_num_compaction_trigger = 2;
+  mutable_cf_options_.level0_file_num_compaction_trigger = 5;
 
-  Add(0, 11U, "1", "2", 5000U);
-  Add(0, 12U, "1", "2", 5000U);
-  Add(0, 13U, "1", "2", 5000U);
-  Add(0, 14U, "1", "2", 5000U);
-  Add(0, 15U, "1", "2", 5000U);
-  Add(0, 16U, "1", "2", 5000U);
+  Add(0, 11U, "1", "2", 2500U);
+  Add(0, 12U, "1", "2", 2500U);
+  Add(0, 13U, "1", "2", 2500U);
+  Add(0, 14U, "1", "2", 2500U);
 
+  // Level size should be around 20,000, 53000, 258000
   Add(5, 4U, "1", "2", 1286250U);
-  Add(4, 5U, "1", "2", 200000U);
-  Add(3, 6U, "1", "2", 40000U);
-  Add(2, 7U, "1", "2", 8000U);
+  Add(4, 5U, "1", "2", 260000U);  // Unadjusted score 1.01, adjusted about 4.3
+  Add(3, 6U, "1", "2", 85000U);   // Unadjusted score 1.42, adjusted about 11.6
+  Add(2, 7U, "1", "2", 30000);    // Unadjusted score 1.5, adjusted about 10.0
 
   UpdateVersionStorageInfo();
 
   ASSERT_EQ(0, logger_->log_count);
   ASSERT_EQ(2, vstorage_.base_level());
-  // level multiplier should be 3.5
-  ASSERT_LT(vstorage_.level_multiplier(), 3.6);
-  ASSERT_GT(vstorage_.level_multiplier(), 3.4);
-  // Level size should be around 30,000, 105,000, 367,500
-  ASSERT_EQ(30000U, vstorage_.MaxBytesForLevel(2));
-  ASSERT_LT(vstorage_.MaxBytesForLevel(3), 110000U);
-  ASSERT_GT(vstorage_.MaxBytesForLevel(3), 100000U);
-  ASSERT_LT(vstorage_.MaxBytesForLevel(4), 370000U);
-  ASSERT_GT(vstorage_.MaxBytesForLevel(4), 360000U);
+  ASSERT_EQ(20000U, vstorage_.MaxBytesForLevel(2));
+
+  vstorage_.ComputeCompactionScore(ioptions_, mutable_cf_options_);
+  // Although L2 has higher unadjusted compaction score, considering
+  // a relatively large L0 being compacted down soon, L3 is picked up for
+  // compaction.
+
+  ASSERT_EQ(3, vstorage_.CompactionScoreLevel(0));
+  ASSERT_EQ(2, vstorage_.CompactionScoreLevel(1));
+  ASSERT_EQ(4, vstorage_.CompactionScoreLevel(2));
 }
 
 TEST_F(VersionStorageInfoTest, EstimateLiveDataSize) {