internal/manifest: Relax SeqNum overlap invariant in L0

Changes Manifest.CheckOrdering to allow L0 SSTables to overlap
in sequence numbers.

Implemented for better compatibility with future versions of
Pebble which could support partitioned flushes and L0 sublevels.

Fixes #587.

internal/manifest/testdata/version_check_ordering

@@ -25,11 +25,58 @@ check-ordering
 L0
   c.SET.3-d.SET.4
   a.SET.1-b.SET.5
+  e.SET.2-f.SET.7
+  g.SET.6-h.SET.12
+  i.SET.8-j.SET.13
+  b.SET.15-d.SET.15
+  a.SET.14-j.SET.17
+  k.SET.16-n.SET.19
 ----
-L0 flushed file 000002 overlaps with the largest seqnum of a preceding flushed file: 1-5 vs 4
+OK
+
+# Add some ingested SSTables around the 14-19 seqnum cases.
+check-ordering
+L0
+  c.SET.3-d.SET.4
+  a.SET.1-b.SET.5
+  e.SET.2-f.SET.7
+  g.SET.6-h.SET.12
+  i.SET.8-j.SET.13
+  b.SET.15-d.SET.15
+  a.SET.14-j.SET.17
+  m.SET.18-n.SET.18
+  k.SET.16-n.SET.19
+  m.SET.20-n.SET.20
+----
+OK
+
+# Ensure that coincident sequence numbers are detected around sstables with
+# overlapping sequence numbers.
+check-ordering
+L0
+  c.SET.3-d.SET.4
+  a.SET.1-b.SET.5
+  e.SET.2-f.SET.7
+  g.SET.6-h.SET.12
+  i.SET.8-j.SET.13
+  b.SET.15-d.SET.15
+  a.SET.15-j.SET.17
+  m.SET.18-n.SET.18
+  k.SET.16-n.SET.19
+  m.SET.20-n.SET.20
+----
+L0 flushed file 000007 has smallest sequence number coincident with an ingested file : 15-17 vs 15
 0:
   000001:[c#3,SET-d#4,SET]
   000002:[a#1,SET-b#5,SET]
+  000003:[e#2,SET-f#7,SET]
+  000004:[g#6,SET-h#12,SET]
+  000005:[i#8,SET-j#13,SET]
+  000006:[b#15,SET-d#15,SET]
+  000007:[a#15,SET-j#17,SET]
+  000008:[m#18,SET-n#18,SET]
+  000009:[k#16,SET-n#19,SET]
+  000010:[m#20,SET-n#20,SET]
 
 check-ordering
 L0
@@ -66,7 +113,7 @@ L0
   a.SET.3-d.SET.3
   a.SET.3-d.SET.5
 ----
-L0 flushed file 000002 has an ingested file coincident with smallest seqnum: 3-5
+L0 flushed file 000002 has smallest sequence number coincident with an ingested file : 3-5 vs 3
 0:
   000001:[a#3,SET-d#3,SET]
   000002:[a#3,SET-d#5,SET]
@@ -94,7 +141,7 @@ L0
   a.SET.5-d.SET.5
   a.SET.4-d.SET.6
 ----
-L0 flushed file 000003 has an ingested file coincident with smallest seqnum: 4-6
+L0 flushed file 000003 has smallest sequence number coincident with an ingested file : 4-6 vs 4
 0:
   000001:[a#4,SET-d#4,SET]
   000002:[a#5,SET-d#5,SET]

internal/manifest/testdata/version_edit_apply

@@ -53,7 +53,11 @@ edit
   L0
    4:b.SET.3-d.SET.5
 ----
-pebble: internal error: L0 flushed file 000004 overlaps with the largest seqnum of a preceding flushed file: 3-5 vs 4
+0:
+  000001:[a#1,SET-c#2,SET]
+  000002:[c#3,SET-d#4,SET]
+  000004:[b#3,SET-d#5,SET]
+zombies []
 
 apply
  L0

internal/manifest/version.go

@@ -447,7 +447,7 @@ func CheckOrdering(cmp Compare, format base.Formatter, level int, files []*FileM
 		//   file (except for the 0 sequence number case below).
 		// - Files with multiple sequence numbers: these are necessarily flushed files.
 		//
-		// Two cases of overlapping sequence numbers:
+		// Three cases of overlapping sequence numbers:
 		// Case 1:
 		// An ingested file contained in the sequence numbers of the flushed file -- it must be
 		// fully contained (not coincident with either end of the flushed file) since the memtable
@@ -469,9 +469,17 @@ func CheckOrdering(cmp Compare, format base.Formatter, level int, files []*FileM
 		// in the file key intervals. This file is placed in L0 since it overlaps in the file
 		// key intervals but since it has no overlapping data, it is assigned a sequence number
 		// of 0 in RocksDB. We handle this case for compatibility with RocksDB.
-
-		// The largest sequence number of a flushed file. Increasing.
-		var largestFlushedSeqNum uint64
+		//
+		// Case 3:
+		// A sequence of flushed files that overlap in sequence numbers with one another,
+		// but do not overlap in keys inside the sstables. These files correspond to
+		// partitioned flushes or the results of intra-L0 compactions of partitioned
+		// flushes.
+		//
+		// Since these types of SSTables violate most other sequence number
+		// overlap invariants, and handling this case is important for compatibility
+		// with future versions of pebble, this method relaxes most L0 invariant
+		// checks except for those concerning ingested SSTables.
 
 		// The largest sequence number of any file. Increasing.
 		var largestSeqNum uint64
@@ -510,20 +518,13 @@ func CheckOrdering(cmp Compare, format base.Formatter, level int, files []*FileM
 				uncheckedIngestedSeqNums = append(uncheckedIngestedSeqNums, f.LargestSeqNum)
 			} else {
 				// Flushed file.
-				// Two flushed files cannot overlap.
-				if largestFlushedSeqNum > 0 && f.SmallestSeqNum <= largestFlushedSeqNum {
-					return fmt.Errorf("L0 flushed file %06d overlaps with the largest seqnum of a "+
-						"preceding flushed file: %d-%d vs %d", f.FileNum, f.SmallestSeqNum, f.LargestSeqNum,
-						largestFlushedSeqNum)
-				}
-				largestFlushedSeqNum = f.LargestSeqNum
 				// Check that unchecked ingested sequence numbers are not coincident with f.SmallestSeqNum.
 				// We do not need to check that they are not coincident with f.LargestSeqNum because we
 				// have already confirmed that LargestSeqNums were increasing.
 				for _, seq := range uncheckedIngestedSeqNums {
 					if seq == f.SmallestSeqNum {
-						return fmt.Errorf("L0 flushed file %06d has an ingested file coincident with "+
-							"smallest seqnum: %d-%d", f.FileNum, f.SmallestSeqNum, f.LargestSeqNum)
+						return fmt.Errorf("L0 flushed file %06d has smallest sequence number coincident with an ingested file "+
+							": %d-%d vs %d", f.FileNum, f.SmallestSeqNum, f.LargestSeqNum, seq)
 					}
 				}
 				uncheckedIngestedSeqNums = uncheckedIngestedSeqNums[:0]

testdata/compaction_check_ordering

@@ -18,12 +18,13 @@ L0
 ----
 fatal: L0 files 000001 and 000002 are not properly ordered: 3-4 vs 1-2
 
+# Seqnum overlaps are allowed in L0 as long as no key ranges overlap.
 check-ordering
 L0
   c.SET.3-d.SET.4
   a.SET.1-b.SET.5
 ----
-fatal: L0 flushed file 000002 overlaps with the largest seqnum of a preceding flushed file: 1-5 vs 4
+OK
 
 check-ordering
 L0

tool/make_incorrect_manifests.go

@@ -39,13 +39,13 @@ func makeManifest1() {
 	ve.NextFileNum = 5
 	ve.LastSeqNum = 20
 	ve.NewFiles = []manifest.NewFileEntry{
-		{0, &manifest.FileMetadata{
+		{6, &manifest.FileMetadata{
 			FileNum: 1, SmallestSeqNum: 2, LargestSeqNum: 5}}}
 	writeVE(writer, &ve)
 
 	ve.MinUnflushedLogNum = 3
 	ve.NewFiles = []manifest.NewFileEntry{
-		{0, &manifest.FileMetadata{
+		{6, &manifest.FileMetadata{
 			FileNum: 2, SmallestSeqNum: 1, LargestSeqNum: 4}}}
 	writeVE(writer, &ve)
 

tool/testdata/manifest_dump

@@ -162,31 +162,31 @@ MANIFEST-invalid
   log-num:       2
   next-file-num: 5
   last-seq-num:  20
-  added:         L0 000001:0<#2-#5>[#0,DEL-#0,DEL]
+  added:         L6 000001:0<#2-#5>[#0,DEL-#0,DEL]
 65
   comparer:     leveldb.BytewiseComparator
   log-num:       3
   next-file-num: 5
   last-seq-num:  20
-  added:         L0 000002:0<#1-#4>[#0,DEL-#0,DEL]
+  added:         L6 000002:0<#1-#4>[#0,DEL-#0,DEL]
 EOF
-pebble: internal error: L0 flushed file 000001 overlaps with the largest seqnum of a preceding flushed file: 2-5 vs 4
+pebble: internal error: L6 files 000001 and 000002 have overlapping ranges: #0,DEL-#0,DEL vs #0,DEL-#0,DEL
 
 manifest check
 ./testdata/MANIFEST-invalid
 ----
-MANIFEST-invalid: offset: 65 err: pebble: internal error: L0 flushed file 000001 overlaps with the largest seqnum of a preceding flushed file: 2-5 vs 4
+MANIFEST-invalid: offset: 65 err: pebble: internal error: L6 files 000001 and 000002 have overlapping ranges: #0,DEL-#0,DEL vs #0,DEL-#0,DEL
 Version state before failed Apply
 --- L0 ---
-  000001:0<#2-#5>[#0,DEL-#0,DEL]
 --- L1 ---
 --- L2 ---
 --- L3 ---
 --- L4 ---
 --- L5 ---
 --- L6 ---
+  000001:0<#2-#5>[#0,DEL-#0,DEL]
 Version edit that failed
-  added: L0 000002:0<#1-#4>[#0,DEL-#0,DEL]
+  added: L6 000002:0<#1-#4>[#0,DEL-#0,DEL]
 
 manifest dump
 ./testdata/find-db/MANIFEST-000001