diff --git a/pkg/planner/core/operator/logicalop/logical_join.go b/pkg/planner/core/operator/logicalop/logical_join.go
index 72ea8eb4f5aab..8ce6482ebfccb 100644
--- a/pkg/planner/core/operator/logicalop/logical_join.go
+++ b/pkg/planner/core/operator/logicalop/logical_join.go
@@ -510,6 +510,27 @@ func (p *LogicalJoin) PushDownTopN(topNLogicalPlan base.LogicalPlan) base.Logica
 	case base.RightOuterJoin:
 		p.Children()[1], topnEliminated = p.pushDownTopNToChild(topN, 1)
 		p.Children()[0] = p.Children()[0].PushDownTopN(nil)
+	case base.InnerJoin:
+		if p.canPushLimitDownToCartesianChildren(topN) {
+			limitCount, carry := bits.Add64(topN.Count, topN.Offset, 0)
+			if carry > 0 {
+				limitCount = math.MaxUint64
+			}
+			leftLimit := LogicalTopN{
+				Count:            limitCount,
+				Offset:           0,
+				PreferLimitToCop: topN.PreferLimitToCop,
+			}.Init(topN.SCtx(), topN.QueryBlockOffset())
+			rightLimit := LogicalTopN{
+				Count:            limitCount,
+				Offset:           0,
+				PreferLimitToCop: topN.PreferLimitToCop,
+			}.Init(topN.SCtx(), topN.QueryBlockOffset())
+			p.Children()[0] = p.Children()[0].PushDownTopN(leftLimit)
+			p.Children()[1] = p.Children()[1].PushDownTopN(rightLimit)
+			return topN.AttachChild(p.Self())
+		}
+		return p.BaseLogicalPlan.PushDownTopN(topN)
 	default:
 		return p.BaseLogicalPlan.PushDownTopN(topN)
 	}
@@ -531,6 +552,20 @@ func (p *LogicalJoin) PushDownTopN(topNLogicalPlan base.LogicalPlan) base.Logica
 	return p.Self()
 }
 
+func (p *LogicalJoin) canPushLimitDownToCartesianChildren(topN *LogicalTopN) bool {
+	if topN == nil || !topN.IsLimit() {
+		return false
+	}
+	if _, isApply := p.Self().(*LogicalApply); isApply {
+		return false
+	}
+	return len(p.EqualConditions) == 0 &&
+		len(p.NAEQConditions) == 0 &&
+		len(p.LeftConditions) == 0 &&
+		len(p.RightConditions) == 0 &&
+		len(p.OtherConditions) == 0
+}
+
 // DeriveTopN inherits the BaseLogicalPlan.LogicalPlan.<6th> implementation.
 
 // PredicateSimplification inherits the BaseLogicalPlan.LogicalPlan.<7th> implementation.
diff --git a/pkg/planner/core/testdata/plan_suite_unexported_out.json b/pkg/planner/core/testdata/plan_suite_unexported_out.json
index 673752ce18199..33580446bb20e 100644
--- a/pkg/planner/core/testdata/plan_suite_unexported_out.json
+++ b/pkg/planner/core/testdata/plan_suite_unexported_out.json
@@ -148,7 +148,7 @@
       "DataScan(t)->Aggr(count(test.t.b),firstrow(test.t.a))->Limit->Projection",
       "DataScan(t)->Aggr(count(test.t.b),firstrow(test.t.a),firstrow(test.t.c))->TopN([test.t.c],0,5)->Projection",
       "Join{DataScan(t)->DataScan(s)}->TopN([test.t.a],0,5)->Projection",
-      "Join{DataScan(t)->DataScan(s)}->Limit->Projection",
+      "Join{DataScan(t)->Limit->DataScan(s)->Limit}->Limit->Projection",
       "DataScan(t)->Projection->TopN([Column#14 true],0,1)->Projection",
       "Join{DataScan(t)->TopN([test.t.a],0,5)->DataScan(s)}(test.t.a,test.t.a)->Sort->Projection",
       "Join{DataScan(t)->TopN([test.t.a],5,5)->DataScan(s)}(test.t.a,test.t.a)->Sort->Projection",