compaction: elision-only compactions for tables with only range keys

Currently, a table is marked as eligible for elision-only compaction by
the `elisionOnlyAnnotator` under the following circumstances:

- the table's range deletion estimate is greater than or equal to 10% of
  the total table size, OR
- the number of deletions is greater than or equal to 10% of the table's
  total point key entries.

If a table contains only range keys, the second predicate is true (given
that `0 >= 0`), scheduling an elision-only compaction. Howeve, if the
table contains only range key-sets, such keys cannot be elided, and the
compaction picker will continue to schedule the table for elision,
without effect. This can tie up compaction slots.

While it is _technically_ possible that a table with containing
exclusively range keys, but no range key sets _could_ be eligible for an
elision-only compaction (i.e. if there are no spans underneath, or
snapshots preventing the elision, etc.), the utility of such a
compaction is minimal, given that a compaction into a table containing a
few range keys would be inexpensive.

Tweak the elision-only compaction heuristics to skip elision-only
compactions of tables that contain exclusively range keys.

compaction_picker.go

@@ -1239,8 +1239,13 @@ func (a elisionOnlyAnnotator) Accumulate(f *fileMetadata, dst interface{}) (inte
 	// which may be collapsed. Ideally, we would have 'obsolete keys'
 	// statistics that would include tombstones, the keys that are
 	// dropped by tombstones and duplicated user keys. See #847.
+	//
+	// Note that tables that contain exclusively range keys (i.e. no point keys,
+	// `NumEntries` and `RangeDeletionsBytesEstimate` are both zero) are excluded
+	// from elision-only compactions.
+	// TODO(travers): Consider an alternative heuristic for elision of range-keys.
 	if f.Stats.RangeDeletionsBytesEstimate*10 < f.Size &&
-		f.Stats.NumDeletions*10 < f.Stats.NumEntries {
+		f.Stats.NumDeletions*10 <= f.Stats.NumEntries {
 		return dst, true
 	}
 	if dst == nil {

compaction_picker_test.go

@@ -1176,6 +1176,7 @@ func TestCompactionOutputFileSize(t *testing.T) {
 				}
 				m.Stats.Valid = true
 				m.Stats.RangeDeletionsBytesEstimate = uint64(v)
+				m.Stats.NumDeletions = 1 // At least one range del responsible for the deletion bytes.
 			}
 		}
 		m.SmallestSeqNum = m.Smallest.SeqNum()

compaction_test.go

@@ -2097,6 +2097,7 @@ func TestCompactionTombstones(t *testing.T) {
 							compactInfo = &info
 						},
 					},
+					FormatMajorVersion: FormatNewest,
 				}
 				var err error
 				d, err = runDBDefineCmd(td, opts)

testdata/compaction_tombstones

@@ -244,3 +244,50 @@ range-deletions-bytes-estimate: 0
 maybe-compact
 ----
 [JOB 100] compacted(default) L5 [000004] (794 B) + L6 [000006] (13 K) -> L6 [] (0 B), in 1.0s (2.0s total), output rate 0 B/s
+
+# A table containing only range keys is not eligible for elision.
+# RANGEKEYDEL or RANGEKEYUNSET.
+
+define
+L6
+  rangekey:a-b:{(#1,RANGEKEYDEL)}
+L6
+  rangekey:b-c:{(#2,RANGEKEYUNSET,@1)}
+L6
+  rangekey:c-d:{(#3,RANGEKEYSET,@1)}
+----
+6:
+  000004:[a#1,RANGEKEYDEL-b#72057594037927935,RANGEKEYDEL]
+  000005:[b#2,RANGEKEYUNSET-c#72057594037927935,RANGEKEYUNSET]
+  000006:[c#3,RANGEKEYSET-d#72057594037927935,RANGEKEYSET]
+
+wait-pending-table-stats
+000004
+----
+num-entries: 0
+num-deletions: 0
+num-range-keys: 1
+point-deletions-bytes-estimate: 0
+range-deletions-bytes-estimate: 0
+
+wait-pending-table-stats
+000005
+----
+num-entries: 0
+num-deletions: 0
+num-range-keys: 1
+point-deletions-bytes-estimate: 0
+range-deletions-bytes-estimate: 0
+
+wait-pending-table-stats
+000006
+----
+num-entries: 0
+num-deletions: 0
+num-range-keys: 1
+point-deletions-bytes-estimate: 0
+range-deletions-bytes-estimate: 0
+
+maybe-compact
+----
+(none)