db: support range keys in delete-only compactions

compaction.go

@@ -1816,13 +1816,63 @@ func (d *DB) maybeScheduleCompactionPicker(
 	}
 }
 
+// deleteCompactionHintType indicates whether the deleteCompactionHint was
+// generated from a span containing a range del (point key only), a range key
+// delete (range key only), or both a point and range key.
+type deleteCompactionHintType uint8
+
+const (
+	// NOTE: While these are primarily used as enumeration types, they are also
+	// used for some bitwise operations. Care should be taken when updating.
+	deleteCompactionHintTypeUnknown deleteCompactionHintType = iota
+	deleteCompactionHintTypePointKeyOnly
+	deleteCompactionHintTypeRangeKeyOnly
+	deleteCompactionHintTypePointAndRangeKey
+)
+
+// String implements fmt.Stringer.
+func (h deleteCompactionHintType) String() string {
+	switch h {
+	case deleteCompactionHintTypeUnknown:
+		return "unknown"
+	case deleteCompactionHintTypePointKeyOnly:
+		return "point-key-only"
+	case deleteCompactionHintTypeRangeKeyOnly:
+		return "range-key-only"
+	case deleteCompactionHintTypePointAndRangeKey:
+		return "point-and-range-key"
+	default:
+		panic(fmt.Sprintf("unknown hint type: %d", h))
+	}
+}
+
+// compactionHintFromKeys returns a deleteCompactionHintType given a slice of
+// keyspan.Keys.
+func compactionHintFromKeys(keys []keyspan.Key) deleteCompactionHintType {
+	var hintType deleteCompactionHintType
+	for _, k := range keys {
+		switch k.Kind() {
+		case base.InternalKeyKindRangeDelete:
+			hintType |= deleteCompactionHintTypePointKeyOnly
+		case base.InternalKeyKindRangeKeyDelete:
+			hintType |= deleteCompactionHintTypeRangeKeyOnly
+		default:
+			panic(fmt.Sprintf("unsupported key kind: %s", k.Kind()))
+		}
+	}
+	return hintType
+}
+
 // A deleteCompactionHint records a user key and sequence number span that has been
 // deleted by a range tombstone. A hint is recorded if at least one sstable
 // falls completely within both the user key and sequence number spans.
 // Once the tombstones and the observed completely-contained sstables fall
 // into the same snapshot stripe, a delete-only compaction may delete any
 // sstables within the range.
 type deleteCompactionHint struct {
+	// The type of key span that generated this hint (point key, range key, or
+	// both).
+	hintType deleteCompactionHintType
 	// start and end are user keys specifying a key range [start, end) of
 	// deleted keys.
 	start []byte
@@ -1850,9 +1900,12 @@ type deleteCompactionHint struct {
 }
 
 func (h deleteCompactionHint) String() string {
-	return fmt.Sprintf("L%d.%s %s-%s seqnums(tombstone=%d-%d, file-smallest=%d)",
+	return fmt.Sprintf(
+		"L%d.%s %s-%s seqnums(tombstone=%d-%d, file-smallest=%d, type=%s)",
 		h.tombstoneLevel, h.tombstoneFile.FileNum, h.start, h.end,
-		h.tombstoneSmallestSeqNum, h.tombstoneLargestSeqNum, h.fileSmallestSeqNum)
+		h.tombstoneSmallestSeqNum, h.tombstoneLargestSeqNum, h.fileSmallestSeqNum,
+		h.hintType,
+	)
 }
 
 func (h *deleteCompactionHint) canDelete(cmp Compare, m *fileMetadata, snapshots []uint64) bool {
@@ -1873,6 +1926,27 @@ func (h *deleteCompactionHint) canDelete(cmp Compare, m *fileMetadata, snapshots
 		return false
 	}
 
+	switch h.hintType {
+	case deleteCompactionHintTypePointKeyOnly:
+		// A hint generated by a range del span cannot delete tables that contain
+		// range keys.
+		if m.HasRangeKeys {
+			return false
+		}
+	case deleteCompactionHintTypeRangeKeyOnly:
+		// A hint generated by a range key del span cannot delete tables that
+		// contain point keys.
+		if m.HasPointKeys {
+			return false
+		}
+	case deleteCompactionHintTypePointAndRangeKey:
+		// A hint from a span that contains both range dels *and* range keys can
+		// only be deleted if both bounds fall within the hint. The next check takes
+		// care of this.
+	default:
+		panic(fmt.Sprintf("pebble: unknown delete compaction hint type: %d", h.hintType))
+	}
+
 	// The file's keys must be completely contained within the hint range.
 	return cmp(h.start, m.Smallest.UserKey) <= 0 && cmp(m.Largest.UserKey, h.end) < 0
 }
@@ -1991,12 +2065,6 @@ func checkDeleteCompactionHints(
 				if m.Compacting || !h.canDelete(cmp, m, snapshots) || files[m] {
 					continue
 				}
-				if m.HasRangeKeys {
-					// TODO(bilal): Remove this conditional when deletion hints work well
-					// with sstables containing range keys.
-					continue
-				}
-
 				if files == nil {
 					// Construct files lazily, assuming most calls will not
 					// produce delete-only compactions.

compaction_test.go

@@ -1848,32 +1848,67 @@ func TestCompactionDeleteOnlyHints(t *testing.T) {
 	}()
 
 	var compactInfo *CompactionInfo // protected by d.mu
+	reset := func() (*Options, error) {
+		if d != nil {
+			compactInfo = nil
+			if err := d.Close(); err != nil {
+				return nil, err
+			}
+		}
+		opts := &Options{
+			FS:         vfs.NewMem(),
+			DebugCheck: DebugCheckLevels,
+			EventListener: EventListener{
+				CompactionEnd: func(info CompactionInfo) {
+					if compactInfo != nil {
+						return
+					}
+					compactInfo = &info
+				},
+			},
+			FormatMajorVersion: FormatNewest,
+		}
 
+		// Collection of table stats can trigger compactions. As we want full
+		// control over when compactions are run, disable stats by default.
+		opts.private.disableTableStats = true
+
+		// Disable automatic compactions to prevent the range key-only tables from
+		// being compacted away after they are created. Compactions do not yet
+		// understand that these tables need to remain in the LSM.
+		// TODO(travers): Revisit this once compactions support range keys.
+		opts.DisableAutomaticCompactions = true
+
+		return opts, nil
+	}
+
+	compactionString := func() string {
+		for d.mu.compact.compactingCount > 0 {
+			d.mu.compact.cond.Wait()
+		}
+
+		s := "(none)"
+		if compactInfo != nil {
+			// Fix the job ID and durations for determinism.
+			compactInfo.JobID = 100
+			compactInfo.Duration = time.Second
+			compactInfo.TotalDuration = 2 * time.Second
+			s = compactInfo.String()
+			compactInfo = nil
+		}
+		return s
+	}
+
+	var err error
+	var opts *Options
 	datadriven.RunTest(t, "testdata/compaction_delete_only_hints",
 		func(td *datadriven.TestData) string {
 			switch td.Cmd {
 			case "define":
-				if d != nil {
-					compactInfo = nil
-					if err := d.Close(); err != nil {
-						return err.Error()
-					}
-				}
-				opts := &Options{
-					FS:         vfs.NewMem(),
-					DebugCheck: DebugCheckLevels,
-					EventListener: EventListener{
-						CompactionEnd: func(info CompactionInfo) {
-							compactInfo = &info
-						},
-					},
+				opts, err = reset()
+				if err != nil {
+					return err.Error()
 				}
-
-				// Collection of table stats can trigger compactions. As we want full
-				// control over when compactions are run, disable stats by default.
-				opts.private.disableTableStats = true
-
-				var err error
 				d, err = runDBDefineCmd(td, opts)
 				if err != nil {
 					return err.Error()
@@ -1902,7 +1937,20 @@ func TestCompactionDeleteOnlyHints(t *testing.T) {
 						FileNum: base.FileNum(parseUint64(parts[1])),
 					}
 
+					var hintType deleteCompactionHintType
+					switch typ := parts[7]; typ {
+					case "point_key_only":
+						hintType = deleteCompactionHintTypePointKeyOnly
+					case "range_key_only":
+						hintType = deleteCompactionHintTypeRangeKeyOnly
+					case "point_and_range_key":
+						hintType = deleteCompactionHintTypePointAndRangeKey
+					default:
+						return fmt.Sprintf("unknown hint type: %s", typ)
+					}
+
 					h := deleteCompactionHint{
+						hintType:                hintType,
 						start:                   start,
 						end:                     end,
 						fileSmallestSeqNum:      parseUint64(parts[4]),
@@ -1954,9 +2002,6 @@ func TestCompactionDeleteOnlyHints(t *testing.T) {
 			case "maybe-compact":
 				d.mu.Lock()
 				d.maybeScheduleCompaction()
-				for d.mu.compact.compactingCount > 0 {
-					d.mu.compact.cond.Wait()
-				}
 
 				var buf bytes.Buffer
 				fmt.Fprintf(&buf, "Deletion hints:\n")
@@ -1967,15 +2012,7 @@ func TestCompactionDeleteOnlyHints(t *testing.T) {
 					fmt.Fprintf(&buf, "  (none)\n")
 				}
 				fmt.Fprintf(&buf, "Compactions:\n")
-				s := "(none)"
-				if compactInfo != nil {
-					// Fix the job ID and durations for determinism.
-					compactInfo.JobID = 100
-					compactInfo.Duration = time.Second
-					compactInfo.TotalDuration = 2 * time.Second
-					s = compactInfo.String()
-				}
-				fmt.Fprintf(&buf, "  %s", s)
+				fmt.Fprintf(&buf, "  %s", compactionString())
 				d.mu.Unlock()
 				return buf.String()
 
@@ -2009,32 +2046,12 @@ func TestCompactionDeleteOnlyHints(t *testing.T) {
 					return err.Error()
 				}
 
-				compactionString := func() string {
-					for d.mu.compact.compactingCount > 0 {
-						d.mu.compact.cond.Wait()
-					}
-
-					s := "(none)"
-					if compactInfo != nil {
-						// Fix the job ID and durations for determinism.
-						compactInfo.JobID = 100
-						compactInfo.Duration = time.Second
-						compactInfo.TotalDuration = 2 * time.Second
-						s = compactInfo.String()
-						compactInfo = nil
-					}
-					return s
-				}
-
 				d.mu.Lock()
 				// Closing the snapshot may have triggered a compaction.
 				str := compactionString()
 				d.mu.Unlock()
 				return str
 
-			case "wait-pending-table-stats":
-				return runTableStatsCmd(td, d)
-
 			case "iter":
 				snap := Snapshot{
 					db:     d,
@@ -2043,6 +2060,36 @@ func TestCompactionDeleteOnlyHints(t *testing.T) {
 				iter := snap.NewIter(nil)
 				return runIterCmd(td, iter, true)
 
+			case "reset":
+				opts, err = reset()
+				if err != nil {
+					return err.Error()
+				}
+				d, err = Open("", opts)
+				if err != nil {
+					return err.Error()
+				}
+				return ""
+
+			case "ingest":
+				// Compactions / flushes do not yet fully support range keys. The
+				// "ingest" operation exists to allow tables containing range keys to be
+				// added to the LSM via ingest, rather than a flush.
+				// TODO(travers): Revisit this once compactions support range keys.
+				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()
+				}
+				return "OK"
+
+			case "describe-lsm":
+				d.mu.Lock()
+				s := d.mu.versions.currentVersion().String()
+				d.mu.Unlock()
+				return s
+
 			default:
 				return fmt.Sprintf("unknown command: %s", td.Cmd)
 			}

data_test.go

@@ -649,6 +649,14 @@ func runDBDefineCmd(td *datadriven.TestData, opts *Options) (*DB, error) {
 				return nil, errors.Errorf("%s: could not parse %q as bool: %s", td.Cmd, arg.Vals[0], err)
 			}
 			opts.private.disableTableStats = !enable
+		case "block-size":
+			size, err := strconv.Atoi(arg.Vals[0])
+			if err != nil {
+				return nil, err
+			}
+			for _, levelOpts := range opts.Levels {
+				levelOpts.BlockSize = size
+			}
 		default:
 			return nil, errors.Errorf("%s: unknown arg: %s", td.Cmd, arg.Key)
 		}

internal/keyspan/span.go

@@ -5,6 +5,7 @@
 package keyspan // import "github.com/cockroachdb/pebble/internal/keyspan"
 
 import (
+	"bytes"
 	"fmt"
 	"sort"
 	"strconv"
@@ -65,6 +66,15 @@ func (k Key) Kind() base.InternalKeyKind {
 	return base.InternalKeyKind(k.Trailer & 0xff)
 }
 
+// Equal returns true if this Key is equal to the given key. Two keys are said
+// to be equal if the two Keys have equal trailers, suffix and value. Suffix
+// comparison uses the provided base.Compare func. Value comparison is bytewise.
+func (k Key) Equal(cmp base.Compare, b Key) bool {
+	return k.Trailer == b.Trailer &&
+		cmp(k.Suffix, b.Suffix) == 0 &&
+		bytes.Equal(k.Value, b.Value)
+}
+
 // Valid returns true if the span is defined.
 func (s *Span) Valid() bool {
 	return s.Start != nil && s.End != nil