plan: make agg's cost formula more reasonable. (#6307)

plan/cbo_test.go

@@ -248,7 +248,7 @@ func (s *testAnalyzeSuite) TestIndexRead(c *C) {
 		},
 		{
 			sql:  "select count(*) from t where e > 1 group by b",
-			best: "TableReader(Table(t)->Sel([gt(test.t.e, 1)])->HashAgg)->HashAgg",
+			best: "IndexLookUp(Index(t.b)[[<nil>,+inf]], Table(t)->Sel([gt(test.t.e, 1)]))->StreamAgg",
 		},
 		{
 			sql:  "select count(e) from t where t.b <= 20",

plan/physical_plan_builder.go

@@ -33,10 +33,13 @@ const (
 	scanFactor         = 2.0
 	descScanFactor     = 5 * scanFactor
 	memoryFactor       = 5.0
-	hashAggMemFactor   = 2.0
+	// 0.5 is the looking up agg context factor.
+	hashAggFactor      = 1.2 + 0.5
 	selectionFactor    = 0.8
 	distinctFactor     = 0.8
 	cpuFactor          = 0.9
+	distinctAggFactor  = 1.6
+	createAggCtxFactor = 6
 )
 
 // wholeTaskTypes records all possible kinds of task that a plan can return. For Agg, TopN and Limit, we will try to get

plan/physical_plans.go

@@ -304,6 +304,15 @@ type basePhysicalAgg struct {
 	GroupByItems []expression.Expression
 }
 
+func (p *basePhysicalAgg) hasDistinctFunc() bool {
+	for _, fun := range p.AggFuncs {
+		if fun.HasDistinct {
+			return true
+		}
+	}
+	return false
+}
+
 // PhysicalHashAgg is hash operator of aggregate.
 type PhysicalHashAgg struct {
 	basePhysicalAgg

plan/task.go

@@ -520,10 +520,12 @@ func (p *PhysicalStreamAgg) attach2Task(tasks ...task) task {
 		}
 		t = finishCopTask(p.ctx, cop)
 		attachPlan2Task(finalAgg, t)
-		t.addCost(t.count() * cpuFactor)
 	} else {
 		attachPlan2Task(p, t)
-		t.addCost(t.count() * cpuFactor)
+	}
+	t.addCost(t.count() * cpuFactor)
+	if p.hasDistinctFunc() {
+		t.addCost(t.count() * cpuFactor * distinctAggFactor)
 	}
 	return t
 }
@@ -534,8 +536,8 @@ func (p *PhysicalHashAgg) attach2Task(tasks ...task) task {
 		return invalidTask
 	}
 	cardinality := p.StatsInfo().count
-	task := tasks[0].copy()
-	if cop, ok := task.(*copTask); ok {
+	t := tasks[0].copy()
+	if cop, ok := t.(*copTask); ok {
 		partialAgg, finalAgg := p.newPartialAggregate()
 		if partialAgg != nil {
 			if cop.tablePlan != nil {
@@ -547,12 +549,14 @@ func (p *PhysicalHashAgg) attach2Task(tasks ...task) task {
 				cop.indexPlan = partialAgg
 			}
 		}
-		task = finishCopTask(p.ctx, cop)
-		attachPlan2Task(finalAgg, task)
-		task.addCost(task.count()*cpuFactor + cardinality*hashAggMemFactor)
+		t = finishCopTask(p.ctx, cop)
+		attachPlan2Task(finalAgg, t)
 	} else {
-		attachPlan2Task(p, task)
-		task.addCost(task.count()*cpuFactor + cardinality*hashAggMemFactor)
+		attachPlan2Task(p, t)
 	}
-	return task
+	t.addCost(t.count()*cpuFactor*hashAggFactor + cardinality*createAggCtxFactor)
+	if p.hasDistinctFunc() {
+		t.addCost(t.count() * cpuFactor * distinctAggFactor)
+	}
+	return t
 }