db: add range key table stats

data_test.go

@@ -796,6 +796,7 @@ func runTableStatsCmd(td *datadriven.TestData, d *DB) string {
 			var b bytes.Buffer
 			fmt.Fprintf(&b, "num-entries: %d\n", f.Stats.NumEntries)
 			fmt.Fprintf(&b, "num-deletions: %d\n", f.Stats.NumDeletions)
+			fmt.Fprintf(&b, "num-range-keys: %d\n", f.Stats.NumRangeKeys)
 			fmt.Fprintf(&b, "point-deletions-bytes-estimate: %d\n", f.Stats.PointDeletionsBytesEstimate)
 			fmt.Fprintf(&b, "range-deletions-bytes-estimate: %d\n", f.Stats.RangeDeletionsBytesEstimate)
 			return b.String()

internal/manifest/version.go

@@ -50,6 +50,8 @@ type TableStats struct {
 	NumEntries uint64
 	// The number of point and range deletion entries in the table.
 	NumDeletions uint64
+	// NumRangeKeys is the total number of range keys in the table.
+	NumRangeKeys uint64
 	// Estimate of the total disk space that may be dropped by this table's
 	// point deletions by compacting them.
 	PointDeletionsBytesEstimate uint64

table_stats.go

@@ -253,79 +253,20 @@ func (d *DB) loadTableStats(
 		stats.NumEntries = r.Properties.NumEntries
 		stats.NumDeletions = r.Properties.NumDeletions
 		if r.Properties.NumPointDeletions() > 0 {
-			// TODO(jackson): If the file has a wide keyspace, the average
-			// value size beneath the entire file might not be representative
-			// of the size of the keys beneath the point tombstones.
-			// We could write the ranges of 'clusters' of point tombstones to
-			// a sstable property and call averageValueSizeBeneath for each of
-			// these narrower ranges to improve the estimate.
-			avgKeySize, avgValSize, err := d.averageEntrySizeBeneath(v, level, meta)
-			if err != nil {
-				return err
+			if err = d.loadTablePointKeyStats(r, v, level, meta, &stats); err != nil {
+				return
 			}
-			stats.PointDeletionsBytesEstimate = pointDeletionsBytesEstimate(&r.Properties, avgKeySize, avgValSize)
-		}
-
-		if r.Properties.NumRangeDeletions == 0 {
-			return nil
 		}
-		// We iterate over the defragmented range tombstones, which ensures
-		// we don't double count ranges deleted at different sequence numbers.
-		// Also, merging abutting tombstones reduces the number of calls to
-		// estimateSizeBeneath which is costly, and improves the accuracy of
-		// our overall estimate.
-		rangeDelIter, err := r.NewRawRangeDelIter()
-		if err != nil {
-			return err
+		if r.Properties.NumRangeDeletions > 0 {
+			if compactionHints, err = d.loadTableRangeDelStats(r, v, level, meta, &stats); err != nil {
+				return
+			}
 		}
-		defer rangeDelIter.Close()
-		// Truncate tombstones to the containing file's bounds if necessary.
-		// See docs/range_deletions.md for why this is necessary.
-		rangeDelIter = keyspan.Truncate(
-			d.cmp, rangeDelIter, meta.Smallest.UserKey, meta.Largest.UserKey, nil, nil)
-		err = foreachDefragmentedTombstone(rangeDelIter, d.cmp,
-			func(startUserKey, endUserKey []byte, smallestSeqNum, largestSeqNum uint64) error {
-				// If the file is in the last level of the LSM, there is no
-				// data beneath it. The fact that there is still a range
-				// tombstone in a bottommost file suggests that an open
-				// snapshot kept the tombstone around. Estimate disk usage
-				// within the file itself.
-				if level == numLevels-1 {
-					size, err := r.EstimateDiskUsage(startUserKey, endUserKey)
-					if err != nil {
-						return err
-					}
-					stats.RangeDeletionsBytesEstimate += size
-					return nil
-				}
-
-				estimate, hintSeqNum, err := d.estimateSizeBeneath(v, level, meta, startUserKey, endUserKey)
-				if err != nil {
-					return err
-				}
-				stats.RangeDeletionsBytesEstimate += estimate
-
-				// If any files were completely contained with the range,
-				// hintSeqNum is the smallest sequence number contained in any
-				// such file.
-				if hintSeqNum == math.MaxUint64 {
-					return nil
-				}
-				hint := deleteCompactionHint{
-					start:                   make([]byte, len(startUserKey)),
-					end:                     make([]byte, len(endUserKey)),
-					tombstoneFile:           meta,
-					tombstoneLevel:          level,
-					tombstoneLargestSeqNum:  largestSeqNum,
-					tombstoneSmallestSeqNum: smallestSeqNum,
-					fileSmallestSeqNum:      hintSeqNum,
-				}
-				copy(hint.start, startUserKey)
-				copy(hint.end, endUserKey)
-				compactionHints = append(compactionHints, hint)
-				return nil
-			})
-		return err
+		// TODO(travers): Once we have real-world data, consider collecting
+		// additional stats that may provide improved heuristics for compaction
+		// picking.
+		stats.NumRangeKeys = r.Properties.NumRangeKeys()
+		return
 	})
 	if err != nil {
 		return stats, nil, err
@@ -334,6 +275,91 @@ func (d *DB) loadTableStats(
 	return stats, compactionHints, nil
 }
 
+// loadTablePointKeyStats calculates the point key statistics for the given
+// table. The provided manifest.TableStats are updated.
+func (d *DB) loadTablePointKeyStats(
+	r *sstable.Reader, v *version, level int, meta *fileMetadata, stats *manifest.TableStats,
+) error {
+	// TODO(jackson): If the file has a wide keyspace, the average
+	// value size beneath the entire file might not be representative
+	// of the size of the keys beneath the point tombstones.
+	// We could write the ranges of 'clusters' of point tombstones to
+	// a sstable property and call averageValueSizeBeneath for each of
+	// these narrower ranges to improve the estimate.
+	avgKeySize, avgValSize, err := d.averageEntrySizeBeneath(v, level, meta)
+	if err != nil {
+		return err
+	}
+	stats.PointDeletionsBytesEstimate =
+		pointDeletionsBytesEstimate(&r.Properties, avgKeySize, avgValSize)
+	return nil
+}
+
+// loadTableRangeDelStats calculates the range deletion statistics for the given
+// table.
+func (d *DB) loadTableRangeDelStats(
+	r *sstable.Reader, v *version, level int, meta *fileMetadata, stats *manifest.TableStats,
+) ([]deleteCompactionHint, error) {
+	var compactionHints []deleteCompactionHint
+	// We iterate over the defragmented range tombstones, which ensures
+	// we don't double count ranges deleted at different sequence numbers.
+	// Also, merging abutting tombstones reduces the number of calls to
+	// estimateSizeBeneath which is costly, and improves the accuracy of
+	// our overall estimate.
+	rangeDelIter, err := r.NewRawRangeDelIter()
+	if err != nil {
+		return nil, err
+	}
+	defer rangeDelIter.Close()
+	// Truncate tombstones to the containing file's bounds if necessary.
+	// See docs/range_deletions.md for why this is necessary.
+	rangeDelIter = keyspan.Truncate(
+		d.cmp, rangeDelIter, meta.Smallest.UserKey, meta.Largest.UserKey, nil, nil)
+	err = foreachDefragmentedTombstone(rangeDelIter, d.cmp,
+		func(startUserKey, endUserKey []byte, smallestSeqNum, largestSeqNum uint64) error {
+			// If the file is in the last level of the LSM, there is no
+			// data beneath it. The fact that there is still a range
+			// tombstone in a bottommost file suggests that an open
+			// snapshot kept the tombstone around. Estimate disk usage
+			// within the file itself.
+			if level == numLevels-1 {
+				size, err := r.EstimateDiskUsage(startUserKey, endUserKey)
+				if err != nil {
+					return err
+				}
+				stats.RangeDeletionsBytesEstimate += size
+				return nil
+			}
+
+			estimate, hintSeqNum, err := d.estimateSizeBeneath(v, level, meta, startUserKey, endUserKey)
+			if err != nil {
+				return err
+			}
+			stats.RangeDeletionsBytesEstimate += estimate
+
+			// If any files were completely contained with the range,
+			// hintSeqNum is the smallest sequence number contained in any
+			// such file.
+			if hintSeqNum == math.MaxUint64 {
+				return nil
+			}
+			hint := deleteCompactionHint{
+				start:                   make([]byte, len(startUserKey)),
+				end:                     make([]byte, len(endUserKey)),
+				tombstoneFile:           meta,
+				tombstoneLevel:          level,
+				tombstoneLargestSeqNum:  largestSeqNum,
+				tombstoneSmallestSeqNum: smallestSeqNum,
+				fileSmallestSeqNum:      hintSeqNum,
+			}
+			copy(hint.start, startUserKey)
+			copy(hint.end, endUserKey)
+			compactionHints = append(compactionHints, hint)
+			return nil
+		})
+	return compactionHints, err
+}
+
 func (d *DB) averageEntrySizeBeneath(
 	v *version, level int, meta *fileMetadata,
 ) (avgKeySize, avgValueSize uint64, err error) {
@@ -510,6 +536,7 @@ func maybeSetStatsFromProperties(meta *fileMetadata, props *sstable.Properties)
 		Valid:                       true,
 		NumEntries:                  props.NumEntries,
 		NumDeletions:                props.NumDeletions,
+		NumRangeKeys:                props.NumRangeKeys(),
 		PointDeletionsBytesEstimate: pointEstimate,
 		RangeDeletionsBytesEstimate: 0,
 	}

table_stats_test.go

@@ -11,6 +11,7 @@ import (
 	"github.com/cockroachdb/pebble/internal/base"
 	"github.com/cockroachdb/pebble/internal/datadriven"
 	"github.com/cockroachdb/pebble/internal/keyspan"
+	"github.com/cockroachdb/pebble/internal/testkeys"
 	"github.com/cockroachdb/pebble/vfs"
 	"github.com/stretchr/testify/require"
 )
@@ -28,6 +29,9 @@ func TestTableStats(t *testing.T) {
 		},
 	}
 	opts.DisableAutomaticCompactions = true
+	opts.Comparer = testkeys.Comparer
+	opts.Experimental.RangeKeys = new(RangeKeysArena)
+	opts.FormatMajorVersion = FormatRangeKeys
 
 	d, err := Open("", opts)
 	require.NoError(t, err)
@@ -92,6 +96,18 @@ func TestTableStats(t *testing.T) {
 			d.mu.Unlock()
 			return s
 
+		case "ingest":
+			if err = runBuildCmd(td, d, d.opts.FS); err != nil {
+				return err.Error()
+			}
+			if err = runIngestCmd(td, d, d.opts.FS); err != nil {
+				return err.Error()
+			}
+			d.mu.Lock()
+			s := d.mu.versions.currentVersion().String()
+			d.mu.Unlock()
+			return s
+
 		case "wait-pending-table-stats":
 			return runTableStatsCmd(td, d)
 

testdata/compaction_delete_only_hints

@@ -206,6 +206,7 @@ wait-pending-table-stats
 ----
 num-entries: 2
 num-deletions: 1
+num-range-keys: 0
 point-deletions-bytes-estimate: 0
 range-deletions-bytes-estimate: 26
 

testdata/compaction_tombstones

@@ -13,6 +13,7 @@ wait-pending-table-stats
 ----
 num-entries: 2
 num-deletions: 2
+num-range-keys: 0
 point-deletions-bytes-estimate: 0
 range-deletions-bytes-estimate: 0
 
@@ -34,6 +35,7 @@ wait-pending-table-stats
 ----
 num-entries: 2
 num-deletions: 2
+num-range-keys: 0
 point-deletions-bytes-estimate: 0
 range-deletions-bytes-estimate: 0
 
@@ -54,6 +56,7 @@ wait-pending-table-stats
 ----
 num-entries: 2
 num-deletions: 1
+num-range-keys: 0
 point-deletions-bytes-estimate: 0
 range-deletions-bytes-estimate: 26
 
@@ -76,6 +79,7 @@ wait-pending-table-stats
 ----
 num-entries: 2
 num-deletions: 1
+num-range-keys: 0
 point-deletions-bytes-estimate: 0
 range-deletions-bytes-estimate: 0
 
@@ -114,6 +118,7 @@ wait-pending-table-stats
 ----
 num-entries: 6
 num-deletions: 2
+num-range-keys: 0
 point-deletions-bytes-estimate: 0
 range-deletions-bytes-estimate: 76
 
@@ -146,6 +151,7 @@ wait-pending-table-stats
 ----
 num-entries: 11
 num-deletions: 1
+num-range-keys: 0
 point-deletions-bytes-estimate: 149
 range-deletions-bytes-estimate: 0
 
@@ -189,6 +195,7 @@ wait-pending-table-stats
 ----
 num-entries: 5
 num-deletions: 1
+num-range-keys: 0
 point-deletions-bytes-estimate: 0
 range-deletions-bytes-estimate: 16488
 
@@ -225,6 +232,7 @@ wait-pending-table-stats
 ----
 num-entries: 3
 num-deletions: 3
+num-range-keys: 0
 point-deletions-bytes-estimate: 13167
 range-deletions-bytes-estimate: 0
 

testdata/ingest

@@ -74,6 +74,7 @@ wait-pending-table-stats
 ----
 num-entries: 2
 num-deletions: 0
+num-range-keys: 0
 point-deletions-bytes-estimate: 0
 range-deletions-bytes-estimate: 0
 
@@ -346,6 +347,7 @@ wait-pending-table-stats
 ----
 num-entries: 2
 num-deletions: 2
+num-range-keys: 0
 point-deletions-bytes-estimate: 0
 range-deletions-bytes-estimate: 1666
 

testdata/manual_compaction

@@ -86,6 +86,7 @@ wait-pending-table-stats
 ----
 num-entries: 1
 num-deletions: 1
+num-range-keys: 0
 point-deletions-bytes-estimate: 0
 range-deletions-bytes-estimate: 1552
 
@@ -103,6 +104,7 @@ wait-pending-table-stats
 ----
 num-entries: 2
 num-deletions: 1
+num-range-keys: 0
 point-deletions-bytes-estimate: 0
 range-deletions-bytes-estimate: 776
 

testdata/manual_compaction_set_with_del

@@ -86,6 +86,7 @@ wait-pending-table-stats
 ----
 num-entries: 1
 num-deletions: 1
+num-range-keys: 0
 point-deletions-bytes-estimate: 0
 range-deletions-bytes-estimate: 1552
 
@@ -103,6 +104,7 @@ wait-pending-table-stats
 ----
 num-entries: 2
 num-deletions: 1
+num-range-keys: 0
 point-deletions-bytes-estimate: 0
 range-deletions-bytes-estimate: 776