@@ -269,11 +269,17 @@ type IterConfig struct {
269269 TombstoneElision TombstoneElision
270270 RangeKeyElision TombstoneElision
271271
272- // AllowZeroSeqNum allows the sequence number of KVs in the bottom snapshot
273- // stripe to be simplified to 0 (which improves compression and enables an
274- // optimization during forward iteration). This can be enabled if there are no
275- // tables overlapping the output at lower levels (than the output) in the LSM.
276- AllowZeroSeqNum bool
272+ // IsBottommostDataLayer indicates that the compaction inputs form the
273+ // bottommost layer of data for the compaction's key range. This allows the
274+ // sequence number of KVs in the bottom snapshot stripe to be simplified to
275+ // 0 (which improves compression and enables an optimization during forward
276+ // iteration). This can be enabled if there are no tables overlapping the
277+ // output at lower levels (than the output) in the LSM.
278+ //
279+ // This field may be false even when nothing is overlapping in lower levels.
280+ // At the time of writing, flushes always set this to false (because flushes
281+ // almost never form the bottommost layer of data).
282+ IsBottommostDataLayer bool
277283
278284 // IneffectualPointDeleteCallback is called if a SINGLEDEL is being elided
279285 // without deleting a point set/merge. False positives are rare but possible
@@ -613,8 +619,24 @@ func (i *Iter) Next() *base.InternalKV {
613619 }
614620 var needDelete bool
615621 if i .err == nil {
616- // includesBase is true whenever we've transformed the MERGE record
617- // into a SET.
622+ // If this is the oldest version of this key (the bottommost
623+ // snapshot stripe), we can transform the sequence number to
624+ // zero. This can improve compression and enables an
625+ // optimization during forward iteration to skip some key
626+ // comparisons. Additionally, we can transform the key kind to
627+ // SET so that iteration and future compactions do not need to
628+ // invoke the user's Merge operator.
629+ if i .isBottommostSnapshotStripe (origSnapshotIdx ) {
630+ i .kv .K .SetSeqNum (base .SeqNumZero )
631+ // During the merge (see mergeNext), we may have already
632+ // transformed the key kind to SET or SETWITHDEL, in which case we want to preserve the existing key kind.
633+ if i .kv .K .Kind () == base .InternalKeyKindMerge {
634+ i .kv .K .SetKind (base .InternalKeyKindSet )
635+ }
636+ }
637+
638+ // includesBase is true when we've merged the oldest operand in
639+ // the LSM.
618640 var includesBase bool
619641 switch i .kv .K .Kind () {
620642 case base .InternalKeyKindSet , base .InternalKeyKindSetWithDelete :
@@ -633,8 +655,6 @@ func (i *Iter) Next() *base.InternalKV {
633655 }
634656 continue
635657 }
636-
637- i .maybeZeroSeqnum (origSnapshotIdx )
638658 return & i .kv
639659 }
640660 if i .err != nil {
@@ -799,7 +819,14 @@ func (i *Iter) setNext() {
799819 // Save the current key.
800820 i .saveKey ()
801821 i .kv .V = i .iterKV .V
802- i .maybeZeroSeqnum (i .curSnapshotIdx )
822+
823+ // If this is the oldest version of this key (the bottommost snapshot
824+ // stripe), we can transform the sequence number to zero. This can improve
825+ // compression and enables an optimization during forward iteration to skip
826+ // some key comparisons.
827+ if i .isBottommostSnapshotStripe (i .curSnapshotIdx ) {
828+ i .kv .K .SetSeqNum (base .SeqNumZero )
829+ }
803830
804831 // If this key is already a SETWITHDEL we can early return and skip the remaining
805832 // records in the stripe:
@@ -1415,23 +1442,21 @@ func (i *Iter) lastRangeDelSpanFrontierReached(key []byte) []byte {
14151442 return nil
14161443}
14171444
1418- // maybeZeroSeqnum attempts to set the seqnum for the current key to 0. Doing
1419- // so improves compression and enables an optimization during forward iteration
1420- // to skip some key comparisons. The seqnum for an entry can be zeroed if the
1421- // entry is on the bottom snapshot stripe and on the bottom level of the LSM.
1422- func (i * Iter ) maybeZeroSeqnum (snapshotIdx int ) {
1423- if ! i .cfg .AllowZeroSeqNum {
1424- // TODO(peter): allowZeroSeqNum applies to the entire compaction. We could
1425- // make the determination on a key by key basis, similar to what is done
1426- // for elideTombstone. Need to add a benchmark for Iter to verify
1427- // that isn't too expensive.
1428- return
1429- }
1430- if snapshotIdx > 0 {
1431- // This is not the last snapshot
1432- return
1433- }
1434- i .kv .K .SetSeqNum (base .SeqNumZero )
1445+ // isBottommostSnapshotStripe returns true if the compaction's inputs form the
1446+ // bottommost layer of the LSM for the compaction's key range and the provided
1447+ // snapshot stripe is the last stripe.
1448+ //
1449+ // When isBottommostSnapshotStripe returns true, it is guaranteed there does not
1450+ // exist any overlapping keys with lower sequence numbers than the keys in the
1451+ // provided snapshot stripe. However isBottommostSnapshotStripe is permitted to
1452+ // return false even when there is no overlapping data in lower levels (eg,
1453+ // flushes).
1454+ func (i * Iter ) isBottommostSnapshotStripe (snapshotIdx int ) bool {
1455+ // TODO(peter): This determination applies to the entire compaction. We
1456+ // could make the determination on a key by key basis, similar to what is
1457+ // done for elideTombstone. Need to add a benchmark for Iter to verify that
1458+ // isn't too expensive.
1459+ return i .cfg .IsBottommostDataLayer && snapshotIdx == 0
14351460}
14361461
14371462func finishValueMerger (
0 commit comments