manifest: add BackingValueSize field for blob refs

Track BackingValueSize for blob references, which is the uncompressed value
size in blob files referenced by the backing sstable. For non-virtual sstables,
this equals the existing ValueSize field. The BackingValueSize should only be
popualted for FMV >= FormatBackingValueSize, falling back to reading ValueSize
when the BackingValueSize is 0.

Use BackingValueSize in the estimation of referencedValueSize for the
CurrentBlobFileSet. This prevents wasteful blob file rewrite compactions when
virtual sstables are involved, as blob compactions currentlyr rewrite based
on value references tracked by the original table.

Part of: #4915

Author: xinhaoz

blob_rewrite_test.go

@@ -142,7 +142,11 @@ func TestBlobRewrite(t *testing.T) {
 				} else {
 					fmt.Fprintln(&buf, "blobrefs:[")
 					for i, ref := range meta.BlobReferences {
-						fmt.Fprintf(&buf, " %d: %s %d\n", i, ref.FileID, ref.ValueSize)
+						fmt.Fprintf(&buf, " %d: %s %d", i, ref.FileID, ref.ValueSize)
+						if ref.ValueSize != ref.BackingValueSize && ref.BackingValueSize > 0 {
+							fmt.Fprintf(&buf, "/%d", ref.BackingValueSize)
+						}
+						fmt.Fprintf(&buf, "\n")
 					}
 					fmt.Fprintln(&buf, "]")
 				}
@@ -309,13 +313,14 @@ func TestBlobRewriteRandomized(t *testing.T) {
 		))
 		require.NoError(t, tw.Close())
 		originalValueIndices[i] = i
+		valueSize := uint64(len(values[i]))
 		originalTables[i] = &manifest.TableMetadata{
 			TableNum: base.TableNum(i),
 			TableBacking: &manifest.TableBacking{
 				DiskFileNum: base.DiskFileNum(i),
 			},
 			BlobReferences: []manifest.BlobReference{
-				{FileID: base.BlobFileID(blobFileID), ValueSize: uint64(len(values[i]))},
+				{FileID: base.BlobFileID(blobFileID), ValueSize: valueSize, BackingValueSize: valueSize},
 			},
 		}
 	}

compaction_picker.go

@@ -1690,7 +1690,8 @@ func (p *compactionPickerByScore) pickBlobFileRewriteCompaction(
 		// Blob file rewrite compactions are disabled.
 		return nil
 	}
-	garbagePct := float64(aggregateStats.ValueSize-aggregateStats.ReferencedValueSize) /
+
+	garbagePct := float64(aggregateStats.ValueSize-aggregateStats.ReferencedBackingValueSize) /
 		float64(aggregateStats.ValueSize)
 	if garbagePct <= policy.TargetGarbageRatio {
 		// Not enough garbage to warrant a rewrite compaction.

compaction_test.go

@@ -1508,6 +1508,11 @@ func TestCompaction(t *testing.T) {
 			minVersion: formatDeprecatedExperimentalValueSeparation,
 			maxVersion: formatDeprecatedExperimentalValueSeparation,
 		},
+		"backing_value_size": {
+			minVersion: FormatBackingValueSize,
+			maxVersion: FormatBackingValueSize,
+			verbose:    true,
+		},
 	}
 	datadriven.Walk(t, "testdata/compaction", func(t *testing.T, path string) {
 		filename := filepath.Base(path)

excise.go

@@ -163,7 +163,7 @@ func (d *DB) exciseTable(
 			} else if err := determineExcisedTableSize(d.fileCache, m, leftTable); err != nil {
 				return nil, nil, err
 			}
-			determineExcisedTableBlobReferences(m.BlobReferences, m.Size, leftTable)
+			determineExcisedTableBlobReferences(m.BlobReferences, m.Size, leftTable, d.FormatMajorVersion())
 			if err := leftTable.Validate(d.cmp, d.opts.Comparer.FormatKey); err != nil {
 				return nil, nil, err
 			}
@@ -203,7 +203,7 @@ func (d *DB) exciseTable(
 			} else if err := determineExcisedTableSize(d.fileCache, m, rightTable); err != nil {
 				return nil, nil, err
 			}
-			determineExcisedTableBlobReferences(m.BlobReferences, m.Size, rightTable)
+			determineExcisedTableBlobReferences(m.BlobReferences, m.Size, rightTable, d.FormatMajorVersion())
 			if err := rightTable.Validate(d.cmp, d.opts.Comparer.FormatKey); err != nil {
 				return nil, nil, err
 			}
@@ -455,13 +455,17 @@ func determineExcisedTableBlobReferences(
 	originalBlobReferences manifest.BlobReferences,
 	originalSize uint64,
 	excisedTable *manifest.TableMetadata,
+	fmv FormatMajorVersion,
 ) {
 	if len(originalBlobReferences) == 0 {
 		return
 	}
 	newBlobReferences := make(manifest.BlobReferences, len(originalBlobReferences))
 	for i, bf := range originalBlobReferences {
 		bf.ValueSize = max(bf.ValueSize*excisedTable.Size/originalSize, 1)
+		if fmv < FormatBackingValueSize {
+			bf.BackingValueSize = 0
+		}
 		newBlobReferences[i] = bf
 	}
 	excisedTable.BlobReferences = newBlobReferences

internal/manifest/blob_metadata.go

@@ -37,6 +37,10 @@ type BlobReference struct {
 	// any of the referencing tables are virtualized tables, the ValueSize may
 	// be approximate.
 	ValueSize uint64
+	// BackingValueSize is the sum of the lengths of the uncompressed values within the
+	// blob file for which there exists a reference in the backing sstable. For non-virtual
+	// sstables, this is the same as ValueSize.
+	BackingValueSize uint64
 	// EstimatedPhysicalSize is an estimate of the physical size of the blob
 	// reference, in bytes. It's calculated by scaling the blob file's physical
 	// size according to the ValueSize of the blob reference relative to the
@@ -47,7 +51,7 @@ type BlobReference struct {
 // MakeBlobReference creates a BlobReference from the given file ID, value size,
 // and physical blob file.
 func MakeBlobReference(
-	fileID base.BlobFileID, valueSize uint64, phys *PhysicalBlobFile,
+	fileID base.BlobFileID, valueSize uint64, backingValueSize uint64, phys *PhysicalBlobFile,
 ) BlobReference {
 	if invariants.Enabled {
 		switch {
@@ -61,8 +65,9 @@ func MakeBlobReference(
 		}
 	}
 	return BlobReference{
-		FileID:    fileID,
-		ValueSize: valueSize,
+		FileID:           fileID,
+		ValueSize:        valueSize,
+		BackingValueSize: backingValueSize,
 		//                        valueSize
 		//   Reference size =  -----------------  ×  phys.Size
 		//                      phys.ValueSize
@@ -467,6 +472,12 @@ type AggregateBlobFileStats struct {
 	// in all blob files in the set that are still referenced by live tables
 	// (i.e., in the latest version).
 	ReferencedValueSize uint64
+	// ReferencedBackingValueSize is the sum of the length of the uncompressed
+	// values in all blob files in the set that are still referenced by live
+	// backing tables (i.e., in the latest version). Note that this may
+	// overcount values when virtual sstables with the same backing are present,
+	// as each virtual sstable contributes the full backing table's value size.
+	ReferencedBackingValueSize uint64
 	// ReferencesCount is the total number of tracked references in live tables
 	// (i.e., in the latest version). When virtual sstables are present, this
 	// count is per-virtual sstable (not per backing physical sstable).
@@ -475,8 +486,8 @@ type AggregateBlobFileStats struct {
 
 // String implements fmt.Stringer.
 func (s AggregateBlobFileStats) String() string {
-	return fmt.Sprintf("Files:{Count: %d, Size: %d, ValueSize: %d}, References:{ValueSize: %d, Count: %d}",
-		s.Count, s.PhysicalSize, s.ValueSize, s.ReferencedValueSize, s.ReferencesCount)
+	return fmt.Sprintf("Files:{Count: %d, Size: %d, ValueSize: %d}, References:{ValueSize: %d, BackingValueSize: %d, Count: %d}",
+		s.Count, s.PhysicalSize, s.ValueSize, s.ReferencedValueSize, s.ReferencedBackingValueSize, s.ReferencesCount)
 }
 
 // CurrentBlobFileSet describes the set of blob files that are currently live in
@@ -528,7 +539,10 @@ type currentBlobFile struct {
 	// (and we can pool the B-Tree nodes to further reduce allocs)
 	references map[*TableMetadata]struct{}
 	// referencedValueSize is the sum of the length of uncompressed values in
-	// this blob file that are still live.
+	// this blob file that are still live (referenced by backing tables).
+	// This value can be greater than the physical ValueSize of the blob file,
+	// since virtual sstables contribute the referenced value size of the original
+	// backing table.
 	referencedValueSize uint64
 	// heapState holds a pointer to the heap that the blob file is in (if any)
 	// and its index in the heap's items slice. If referencedValueSize is less
@@ -628,7 +642,11 @@ func (s *CurrentBlobFileSet) Init(bve *BulkVersionEdit, h BlobRewriteHeuristic)
 					panic(errors.AssertionFailedf("pebble: referenced blob file %d not found", ref.FileID))
 				}
 				cbf.references[table] = struct{}{}
-				cbf.referencedValueSize += ref.ValueSize
+				refSize := ref.BackingValueSize
+				if refSize == 0 {
+					refSize = ref.ValueSize
+				}
+				cbf.referencedValueSize += refSize
 				s.stats.ReferencedValueSize += ref.ValueSize
 				s.stats.ReferencesCount++
 			}
@@ -779,8 +797,14 @@ func (s *CurrentBlobFileSet) ApplyAndUpdateVersionEdit(ve *VersionEdit) error {
 				return errors.AssertionFailedf("pebble: referenced blob file %d not found", ref.FileID)
 			}
 			cbf.references[e.Meta] = struct{}{}
-			cbf.referencedValueSize += ref.ValueSize
+			refSize := ref.BackingValueSize
+			if refSize == 0 {
+				// Fall back to using ValueSize to estimate reference size.
+				refSize = ref.ValueSize
+			}
+			cbf.referencedValueSize += refSize
 			s.stats.ReferencedValueSize += ref.ValueSize
+			s.stats.ReferencedBackingValueSize += refSize
 			s.stats.ReferencesCount++
 		}
 	}
@@ -797,10 +821,6 @@ func (s *CurrentBlobFileSet) ApplyAndUpdateVersionEdit(ve *VersionEdit) error {
 				return errors.AssertionFailedf("pebble: referenced blob file %d not found", ref.FileID)
 			}
 			if invariants.Enabled {
-				if ref.ValueSize > cbf.referencedValueSize {
-					return errors.AssertionFailedf("pebble: referenced value size %d for blob file %s is greater than the referenced value size %d",
-						ref.ValueSize, cbf.metadata.FileID, cbf.referencedValueSize)
-				}
 				if _, ok := cbf.references[meta]; !ok {
 					return errors.AssertionFailedf("pebble: deleted table %s's reference to blob file %s not known",
 						meta.TableNum, ref.FileID)
@@ -810,7 +830,13 @@ func (s *CurrentBlobFileSet) ApplyAndUpdateVersionEdit(ve *VersionEdit) error {
 			// Decrement the stats for this reference. We decrement even if the
 			// table is being moved, because we incremented the stats when we
 			// iterated over the version edit's new tables.
-			cbf.referencedValueSize -= ref.ValueSize
+			refSize := ref.BackingValueSize
+			if refSize == 0 {
+				// Fall back to using ValueSize to estimate reference size.
+				refSize = ref.ValueSize
+			}
+			cbf.referencedValueSize -= refSize
+			s.stats.ReferencedBackingValueSize -= refSize
 			s.stats.ReferencedValueSize -= ref.ValueSize
 			s.stats.ReferencesCount--
 			if _, ok := newTables[meta.TableNum]; ok {

internal/manifest/table_metadata.go

@@ -850,7 +850,12 @@ func (m *TableMetadata) DebugString(format base.FormatKey, verbose bool) string
 			if i > 0 {
 				fmt.Fprint(&b, ", ")
 			}
-			fmt.Fprintf(&b, "(%s: %d)", r.FileID, r.ValueSize)
+			fmt.Fprintf(&b, "(%s: %d", r.FileID, r.ValueSize)
+			// BackingValueSize is only set for on virtual ssts with FMV >= FormatBackingValueSize.
+			if r.BackingValueSize > 0 && r.BackingValueSize != r.ValueSize {
+				fmt.Fprintf(&b, "/%d", r.BackingValueSize)
+			}
+			fmt.Fprintf(&b, ")")
 		}
 		fmt.Fprintf(&b, "; depth:%d]", m.BlobReferenceDepth)
 	}
@@ -937,6 +942,10 @@ func ParseTableMetadataDebug(s string) (_ *TableMetadata, err error) {
 				ref.FileID = p.BlobFileID()
 				p.Expect(":")
 				ref.ValueSize = p.Uint64()
+				if p.Peek() == "/" {
+					p.Expect("/")
+					ref.BackingValueSize = p.Uint64()
+				}
 				m.BlobReferences = append(m.BlobReferences, ref)
 				p.Expect(")")
 			}

internal/manifest/testdata/current_blob_file_set

@@ -1,7 +1,7 @@
 init
 ----
 CurrentBlobFileSet:
-Files:{Count: 0, Size: 0, ValueSize: 0}, References:{ValueSize: 0, Count: 0}
+Files:{Count: 0, Size: 0, ValueSize: 0}, References:{ValueSize: 0, BackingValueSize: 0, Count: 0}
 Counts:{FullyReferenced: 0, Eligible: 0, TooRecent: 0}
 
 # A version edit that does not contain blob files leaves the set unchanged.
@@ -14,7 +14,7 @@ modified version edit:
   add-table:     L3 000010:[d#1,SET-e#1,SET] seqnums:[0-0] points:[d#1,SET-e#1,SET]
 current blob file set:
 CurrentBlobFileSet:
-Files:{Count: 0, Size: 0, ValueSize: 0}, References:{ValueSize: 0, Count: 0}
+Files:{Count: 0, Size: 0, ValueSize: 0}, References:{ValueSize: 0, BackingValueSize: 0, Count: 0}
 Counts:{FullyReferenced: 0, Eligible: 0, TooRecent: 0}
 
 # A version edit that adds a new blob file records the new file and reference.
@@ -29,7 +29,7 @@ modified version edit:
   add-blob-file: B000012 physical:{000012 size:[20535 (20KB)] vals:[25935 (25KB)]}
 current blob file set:
 CurrentBlobFileSet:
-Files:{Count: 1, Size: 20535, ValueSize: 25935}, References:{ValueSize: 25935, Count: 1}
+Files:{Count: 1, Size: 20535, ValueSize: 25935}, References:{ValueSize: 25935, BackingValueSize: 25935, Count: 1}
 Counts:{FullyReferenced: 1, Eligible: 0, TooRecent: 0}
 
 # A version edit that moves a referencing table from one level to another should
@@ -45,7 +45,7 @@ modified version edit:
   add-table:     L4 000013:[d#1,SET-e#1,SET] seqnums:[0-0] points:[d#1,SET-e#1,SET] blobrefs:[(B000012: 25935); depth:1]
 current blob file set:
 CurrentBlobFileSet:
-Files:{Count: 1, Size: 20535, ValueSize: 25935}, References:{ValueSize: 25935, Count: 1}
+Files:{Count: 1, Size: 20535, ValueSize: 25935}, References:{ValueSize: 25935, BackingValueSize: 25935, Count: 1}
 Counts:{FullyReferenced: 1, Eligible: 0, TooRecent: 0}
 
 # A version edit that moves references from deleted tables to created tables
@@ -63,7 +63,7 @@ modified version edit:
   add-table:     L5 000015:[f#1,SET-g#1,SET] seqnums:[0-0] points:[f#1,SET-g#1,SET] blobrefs:[(B000012: 15935); depth:2]
 current blob file set:
 CurrentBlobFileSet:
-Files:{Count: 1, Size: 20535, ValueSize: 25935}, References:{ValueSize: 15945, Count: 2}
+Files:{Count: 1, Size: 20535, ValueSize: 25935}, References:{ValueSize: 15945, BackingValueSize: 15945, Count: 2}
 Counts:{FullyReferenced: 0, Eligible: 1, TooRecent: 0}
 NextRewrite: B000012 physical:{000012 size:[20535 (20KB)] vals:[25935 (25KB)]} (61.5% live, created at 0)
 
@@ -77,7 +77,7 @@ modified version edit:
   del-table:     L5 000014
 current blob file set:
 CurrentBlobFileSet:
-Files:{Count: 1, Size: 20535, ValueSize: 25935}, References:{ValueSize: 15935, Count: 1}
+Files:{Count: 1, Size: 20535, ValueSize: 25935}, References:{ValueSize: 15935, BackingValueSize: 15935, Count: 1}
 Counts:{FullyReferenced: 0, Eligible: 1, TooRecent: 0}
 NextRewrite: B000012 physical:{000012 size:[20535 (20KB)] vals:[25935 (25KB)]} (61.4% live, created at 0)
 
@@ -93,7 +93,7 @@ modified version edit:
   del-blob-file: B000012 000012
 current blob file set:
 CurrentBlobFileSet:
-Files:{Count: 0, Size: 0, ValueSize: 0}, References:{ValueSize: 0, Count: 0}
+Files:{Count: 0, Size: 0, ValueSize: 0}, References:{ValueSize: 0, BackingValueSize: 0, Count: 0}
 Counts:{FullyReferenced: 0, Eligible: 0, TooRecent: 0}
 
 applyAndUpdateVersionEdit
@@ -106,7 +106,7 @@ modified version edit:
   add-blob-file: B000017 physical:{000017 size:[20535 (20KB)] vals:[25935 (25KB)]}
 current blob file set:
 CurrentBlobFileSet:
-Files:{Count: 1, Size: 20535, ValueSize: 25935}, References:{ValueSize: 25935, Count: 1}
+Files:{Count: 1, Size: 20535, ValueSize: 25935}, References:{ValueSize: 25935, BackingValueSize: 25935, Count: 1}
 Counts:{FullyReferenced: 1, Eligible: 0, TooRecent: 0}
 
 applyAndUpdateVersionEdit
@@ -119,7 +119,7 @@ modified version edit:
   add-table:     L4 000018:[d#1,SET-e#1,SET] seqnums:[0-0] points:[d#1,SET-e#1,SET] blobrefs:[(B000017: 15935); depth:1]
 current blob file set:
 CurrentBlobFileSet:
-Files:{Count: 1, Size: 20535, ValueSize: 25935}, References:{ValueSize: 15935, Count: 1}
+Files:{Count: 1, Size: 20535, ValueSize: 25935}, References:{ValueSize: 15935, BackingValueSize: 15935, Count: 1}
 Counts:{FullyReferenced: 0, Eligible: 1, TooRecent: 0}
 NextRewrite: B000017 physical:{000017 size:[20535 (20KB)] vals:[25935 (25KB)]} (61.4% live, created at 0)
 
@@ -139,15 +139,15 @@ modified version edit:
   del-blob-file: B000017 000017
 current blob file set:
 CurrentBlobFileSet:
-Files:{Count: 1, Size: 10535, ValueSize: 15935}, References:{ValueSize: 15935, Count: 1}
+Files:{Count: 1, Size: 10535, ValueSize: 15935}, References:{ValueSize: 15935, BackingValueSize: 15935, Count: 1}
 Counts:{FullyReferenced: 1, Eligible: 0, TooRecent: 0}
 
 # Initialize a blob file set with a minimum rewrite age of 5 seconds.
 
 init rw-minimum-age=5s
 ----
 CurrentBlobFileSet:
-Files:{Count: 0, Size: 0, ValueSize: 0}, References:{ValueSize: 0, Count: 0}
+Files:{Count: 0, Size: 0, ValueSize: 0}, References:{ValueSize: 0, BackingValueSize: 0, Count: 0}
 Counts:{FullyReferenced: 0, Eligible: 0, TooRecent: 0}
 
 applyAndUpdateVersionEdit
@@ -160,7 +160,7 @@ modified version edit:
   add-blob-file: B000002 physical:{000002 size:[1000 (1000B)] vals:[2000 (2.0KB)]}
 current blob file set:
 CurrentBlobFileSet:
-Files:{Count: 1, Size: 1000, ValueSize: 2000}, References:{ValueSize: 1000, Count: 1}
+Files:{Count: 1, Size: 1000, ValueSize: 2000}, References:{ValueSize: 1000, BackingValueSize: 1000, Count: 1}
 Counts:{FullyReferenced: 1, Eligible: 0, TooRecent: 0}
 
 # Rewrite the referencing sstable to reduce the blob file's referenced value
@@ -176,7 +176,7 @@ modified version edit:
   add-table:     L6 000004:[d#1,SET-e#1,SET] seqnums:[0-0] points:[d#1,SET-e#1,SET] blobrefs:[(B000002: 500); depth:1]
 current blob file set:
 CurrentBlobFileSet:
-Files:{Count: 1, Size: 1000, ValueSize: 2000}, References:{ValueSize: 500, Count: 1}
+Files:{Count: 1, Size: 1000, ValueSize: 2000}, References:{ValueSize: 500, BackingValueSize: 500, Count: 1}
 Counts:{FullyReferenced: 0, Eligible: 0, TooRecent: 1}
 NextEligible: B000002 physical:{000002 size:[1000 (1000B)] vals:[2000 (2.0KB)]} (25.0% live, created at 1)
 
@@ -190,7 +190,7 @@ modified version edit:
   add-table:     L6 000005:[f#1,SET-g#1,SET] seqnums:[0-0] points:[f#1,SET-g#1,SET]
 current blob file set:
 CurrentBlobFileSet:
-Files:{Count: 1, Size: 1000, ValueSize: 2000}, References:{ValueSize: 500, Count: 1}
+Files:{Count: 1, Size: 1000, ValueSize: 2000}, References:{ValueSize: 500, BackingValueSize: 500, Count: 1}
 Counts:{FullyReferenced: 0, Eligible: 0, TooRecent: 1}
 NextEligible: B000002 physical:{000002 size:[1000 (1000B)] vals:[2000 (2.0KB)]} (25.0% live, created at 1)
 
@@ -202,7 +202,7 @@ modified version edit:
   add-table:     L0 000006:[a#2,SET-b#3,SET] seqnums:[0-0] points:[a#2,SET-b#3,SET]
 current blob file set:
 CurrentBlobFileSet:
-Files:{Count: 1, Size: 1000, ValueSize: 2000}, References:{ValueSize: 500, Count: 1}
+Files:{Count: 1, Size: 1000, ValueSize: 2000}, References:{ValueSize: 500, BackingValueSize: 500, Count: 1}
 Counts:{FullyReferenced: 0, Eligible: 0, TooRecent: 1}
 NextEligible: B000002 physical:{000002 size:[1000 (1000B)] vals:[2000 (2.0KB)]} (25.0% live, created at 1)
 
@@ -217,6 +217,6 @@ modified version edit:
   add-table:     L6 000007:[x#9,SET-y#10,SET] seqnums:[0-0] points:[x#9,SET-y#10,SET]
 current blob file set:
 CurrentBlobFileSet:
-Files:{Count: 1, Size: 1000, ValueSize: 2000}, References:{ValueSize: 500, Count: 1}
+Files:{Count: 1, Size: 1000, ValueSize: 2000}, References:{ValueSize: 500, BackingValueSize: 500, Count: 1}
 Counts:{FullyReferenced: 0, Eligible: 1, TooRecent: 0}
 NextRewrite: B000002 physical:{000002 size:[1000 (1000B)] vals:[2000 (2.0KB)]} (25.0% live, created at 1)