When selecting an index, detect when one produces strictly fewer rows than another.

Author: nicktobey

sql/analyzer/costed_index_scan.go

@@ -203,6 +203,9 @@ func getCostedIndexScan(ctx *sql.Context, statsProv sql.StatsProvider, rt sql.Ta
 		if !ok {
 			stat, err = uniformDistStatisticsForIndex(ctx, statsProv, iat, idx)
 		}
+		if err != nil {
+			return nil, nil, nil, err
+		}
 		err := c.cost(root, stat, idx)
 		if err != nil {
 			return nil, nil, nil, err
@@ -534,6 +537,16 @@ func (c *indexCoster) updateBest(s sql.Statistic, hist []sql.HistogramBucket, fd
 			return
 		}
 
+		// If one index uses a superset of the filters of the other, we should always pick the superset.
+		if c.bestFilters.SubsetOf(filters) {
+			update = true
+			return
+		}
+
+		if filters.SubsetOf(c.bestFilters) {
+			return
+		}
+
 		bestKey, bok := best.StrictKey()
 		cmpKey, cok := cmp.StrictKey()
 		if cok && !bok {
@@ -575,6 +588,10 @@ func (c *indexCoster) updateBest(s sql.Statistic, hist []sql.HistogramBucket, fd
 			return
 		}
 
+		if filters.Len() < c.bestFilters.Len() {
+			return
+		}
+
 		if s.ColSet().Len()-filters.Len() < c.bestStat.ColSet().Len()-c.bestFilters.Len() {
 			// prefer 1 range filter over 1 column index (1 - 1 = 0)
 			// vs. 1 range filter over 2 column index (2 - 1 = 1)
@@ -1237,11 +1254,6 @@ func (c *indexCoster) costIndexScanAnd(ctx *sql.Context, filter *iScanAnd, s sql
 		conjFDs = conj.getFds()
 	}
 
-	if exact.Len()+conj.applied.Len() == filter.childCnt() {
-		// matched all filters
-		return conj.hist, conjFDs, sql.NewFastIntSet(int(filter.id)), conj.missingPrefix, nil
-	}
-
 	return conj.hist, conjFDs, exact.Union(conj.applied), conj.missingPrefix, nil
 }