planner: introduce a new variable to adjust the penalty double read c…

…ost of IndexJoin (#40573)

close #40556

planner/core/plan_cost_ver2.go

@@ -559,6 +559,7 @@ func (p *PhysicalIndexJoin) getIndexJoinCostVer2(taskType property.TaskType, opt
 	probeConcurrency := float64(p.ctx.GetSessionVars().IndexLookupJoinConcurrency())
 	cpuFactor := getTaskCPUFactorVer2(p, taskType)
 	memFactor := getTaskMemFactorVer2(p, taskType)
+	requestFactor := getTaskRequestFactorVer2(p, taskType)
 
 	buildFilterCost := filterCostVer2(option, buildRows, buildFilters, cpuFactor)
 	buildChildCost, err := build.getPlanCostVer2(taskType, option)
@@ -595,7 +596,18 @@ func (p *PhysicalIndexJoin) getIndexJoinCostVer2(taskType property.TaskType, opt
 	batchRatio := 6.0
 	probeCost := divCostVer2(mulCostVer2(probeChildCost, buildRows), batchRatio)
 
-	p.planCostVer2 = sumCostVer2(startCost, buildChildCost, buildFilterCost, buildTaskCost, divCostVer2(sumCostVer2(probeCost, probeFilterCost, hashTableCost), probeConcurrency))
+	// Double Read Cost
+	doubleReadCost := newZeroCostVer2(traceCost(option))
+	if p.ctx.GetSessionVars().IndexJoinDoubleReadPenaltyCostRate > 0 &&
+		buildRows > 10000 { // ignore double-read costs for small IndexJoin
+		batchSize := float64(p.ctx.GetSessionVars().IndexJoinBatchSize)
+		taskPerBatch := 1024.0 // TODO: remove this magic number
+		doubleReadTasks := buildRows / batchSize * taskPerBatch
+		doubleReadCost = doubleReadCostVer2(option, doubleReadTasks, requestFactor)
+		doubleReadCost = mulCostVer2(doubleReadCost, p.ctx.GetSessionVars().IndexJoinDoubleReadPenaltyCostRate)
+	}
+
+	p.planCostVer2 = sumCostVer2(startCost, buildChildCost, buildFilterCost, buildTaskCost, divCostVer2(sumCostVer2(doubleReadCost, probeCost, probeFilterCost, hashTableCost), probeConcurrency))
 	p.planCostInit = true
 	return p.planCostVer2, nil
 }

sessionctx/variable/session.go

@@ -1207,6 +1207,9 @@ type SessionVars struct {
 	EnableNewCostInterface bool
 	// CostModelVersion is a internal switch to indicates the Cost Model Version.
 	CostModelVersion int
+	// IndexJoinDoubleReadPenaltyCostRate indicates whether to add some penalty cost to IndexJoin and how much of it.
+	IndexJoinDoubleReadPenaltyCostRate float64
+
 	// BatchPendingTiFlashCount shows the threshold of pending TiFlash tables when batch adding.
 	BatchPendingTiFlashCount int
 	// RcWriteCheckTS indicates whether some special write statements don't get latest tso from PD at RC

sessionctx/variable/sysvar.go

@@ -1981,6 +1981,12 @@ var defaultSysVars = []*SysVar{
 			return nil
 		},
 	},
+	{Scope: ScopeGlobal | ScopeSession, Name: TiDBIndexJoinDoubleReadPenaltyCostRate, Value: strconv.Itoa(1), Hidden: false, Type: TypeFloat, MinValue: 0, MaxValue: math.MaxUint64,
+		SetSession: func(vars *SessionVars, s string) error {
+			vars.IndexJoinDoubleReadPenaltyCostRate = tidbOptFloat64(s, 0)
+			return nil
+		},
+	},
 	{Scope: ScopeGlobal | ScopeSession, Name: TiDBRCWriteCheckTs, Type: TypeBool, Value: BoolToOnOff(DefTiDBRcWriteCheckTs), SetSession: func(s *SessionVars, val string) error {
 		s.RcWriteCheckTS = TiDBOptOn(val)
 		return nil

sessionctx/variable/tidb_vars.go

@@ -700,6 +700,12 @@ const (
 	// TiDBCostModelVersion is a internal switch to indicates the cost model version.
 	TiDBCostModelVersion = "tidb_cost_model_version"
 
+	// TiDBIndexJoinDoubleReadPenaltyCostRate indicates whether to add some penalty cost to IndexJoin and how much of it.
+	// IndexJoin can cause plenty of extra double read tasks, which consume lots of resources and take a long time.
+	// Since the number of double read tasks is hard to estimated accurately, we leave this variable to let us can adjust this
+	// part of cost manually.
+	TiDBIndexJoinDoubleReadPenaltyCostRate = "tidb_index_join_double_read_penalty_cost_rate"
+
 	// TiDBBatchPendingTiFlashCount indicates the maximum count of non-available TiFlash tables.
 	TiDBBatchPendingTiFlashCount = "tidb_batch_pending_tiflash_count"