[memo] merge joins must use globally sorted indexes (#2803)

* [memo] merge joins must use globally sorted indexes * explicitly label default unordered * second merge join side idx order check
dolthub · Dec 26, 2024 · 5425a89 · 5425a89
1 parent 999a371
commit 5425a89
Show file tree

Hide file tree

Showing 4 changed files with 27 additions and 10 deletions.
diff --git a/sql/analyzer/indexed_joins.go b/sql/analyzer/indexed_joins.go
@@ -1070,9 +1070,18 @@ func addMergeJoins(ctx *sql.Context, m *memo.Memo) error {
 		//    Check to see if any rIndexes match that set of filters
 		//    Remove the last matched filter
 		for _, lIndex := range lIndexes {
+			if lIndex.Order() == sql.IndexOrderNone {
+				// lookups can be unordered, merge indexes need to
+				// be globally ordered
+				continue
+			}
+
 			matchedEqFilters := matchedFiltersForLeftIndex(lIndex, join.Left.RelProps.FuncDeps().Constants(), eqFilters)
 			for len(matchedEqFilters) > 0 {
 				for _, rIndex := range rIndexes {
+					if rIndex.Order() == sql.IndexOrderNone {
+						continue
+					}
 					if rightIndexMatchesFilters(rIndex, join.Left.RelProps.FuncDeps().Constants(), matchedEqFilters) {
 						jb := join.Copy()
 						if d, ok := jb.Left.First.(*memo.Distinct); ok && lIndex.SqlIdx().IsUnique() {

diff --git a/sql/memo/memo.go b/sql/memo/memo.go
@@ -743,14 +743,15 @@ type SourceRel interface {
 
 type Index struct {
 	// ordered list of index columns
-	order []sql.ColumnId
+	cols []sql.ColumnId
 	// unordered column set
-	set sql.ColSet
-	idx sql.Index
+	set   sql.ColSet
+	idx   sql.Index
+	order sql.IndexOrder
 }
 
 func (i *Index) Cols() []sql.ColumnId {
-	return i.order
+	return i.cols
 }
 
 func (i *Index) ColSet() sql.ColSet {
@@ -761,6 +762,10 @@ func (i *Index) SqlIdx() sql.Index {
 	return i.idx
 }
 
+func (i *Index) Order() sql.IndexOrder {
+	return i.order
+}
+
 type sourceBase struct {
 	*relBase
 	indexes []*Index

diff --git a/sql/memo/rel_props.go b/sql/memo/rel_props.go
@@ -226,7 +226,10 @@ func (p *relProps) populateFds() {
 			// strict if primary key or all nonNull and unique
 			columns := idxExprsColumns(idx)
 			strict := true
-			normIdx := &Index{idx: idx, order: make([]sql.ColumnId, len(columns))}
+			normIdx := &Index{idx: idx, cols: make([]sql.ColumnId, len(columns)), order: sql.IndexOrderNone}
+			if oidx, ok := idx.(sql.OrderedIndex); ok {
+				normIdx.order = oidx.Order()
+			}
 			for i, c := range columns {
 				ord := sch.IndexOfColName(strings.ToLower(c))
 				idOffset := firstCol + sql.ColumnId(ord)
@@ -236,7 +239,7 @@ func (p *relProps) populateFds() {
 					p.grp.m.HandleErr(err)
 				}
 				normIdx.set.Add(colId)
-				normIdx.order[i] = colId
+				normIdx.cols[i] = colId
 				if !notNull.Contains(colId) {
 					strict = false
 				}

diff --git a/sql/memo/rel_props_test.go b/sql/memo/rel_props_test.go
@@ -39,8 +39,8 @@ func TestPopulateFDs(t *testing.T) {
 			notNull: sql.NewColSet(1, 2, 3),
 			indexes: []*Index{
 				{
-					order: []sql.ColumnId{2, 1},
-					set:   sql.NewColSet(1, 2),
+					cols: []sql.ColumnId{2, 1},
+					set:  sql.NewColSet(1, 2),
 				},
 			},
 			key: sql.NewColSet(1, 2),
@@ -62,8 +62,8 @@ func TestPopulateFDs(t *testing.T) {
 			notNull: sql.NewColSet(1, 2, 3),
 			indexes: []*Index{
 				{
-					order: []sql.ColumnId{1, 2, 3},
-					set:   sql.NewColSet(1, 2, 3),
+					cols: []sql.ColumnId{1, 2, 3},
+					set:  sql.NewColSet(1, 2, 3),
 				},
 			},
 			key: sql.NewColSet(1, 2, 3),