diff --git a/components/cluster/command/scale_in.go b/components/cluster/command/scale_in.go
index bcee9cacca..e4083d4b91 100644
--- a/components/cluster/command/scale_in.go
+++ b/components/cluster/command/scale_in.go
@@ -37,25 +37,14 @@ func newScaleInCmd() *cobra.Command {
 			scale := func(b *task.Builder, imetadata spec.Metadata, tlsCfg *tls.Config) {
 				metadata := imetadata.(*spec.ClusterMeta)
 
+				nodes := gOpt.Nodes
 				if !gOpt.Force {
-					b.ClusterOperate(metadata.Topology, operator.ScaleInOperation, gOpt, tlsCfg).
-						UpdateMeta(clusterName, metadata, operator.AsyncNodes(metadata.Topology, gOpt.Nodes, false)).
-						UpdateTopology(
-							clusterName,
-							tidbSpec.Path(clusterName),
-							metadata,
-							operator.AsyncNodes(metadata.Topology, gOpt.Nodes, false), /* deleteNodeIds */
-						)
-				} else {
-					b.ClusterOperate(metadata.Topology, operator.ScaleInOperation, gOpt, tlsCfg).
-						UpdateMeta(clusterName, metadata, gOpt.Nodes).
-						UpdateTopology(
-							clusterName,
-							tidbSpec.Path(clusterName),
-							metadata,
-							gOpt.Nodes,
-						)
+					nodes = operator.AsyncNodes(metadata.Topology, nodes, false)
 				}
+
+				b.ClusterOperate(metadata.Topology, operator.ScaleInOperation, gOpt, tlsCfg).
+					UpdateMeta(clusterName, metadata, nodes).
+					UpdateTopology(clusterName, tidbSpec.Path(clusterName), metadata, nodes)
 			}
 
 			return manager.ScaleIn(
diff --git a/pkg/cluster/operation/scale_in.go b/pkg/cluster/operation/scale_in.go
index 277cc77302..0ddbacccd7 100644
--- a/pkg/cluster/operation/scale_in.go
+++ b/pkg/cluster/operation/scale_in.go
@@ -243,6 +243,16 @@ func ScaleInCluster(
 		}
 	}
 
+	pdServers := make([]spec.PDSpec, 0, len(cluster.PDServers))
+	for i := 0; i < len(cluster.PDServers); i++ {
+		s := cluster.PDServers[i]
+		id := s.Host + ":" + strconv.Itoa(s.ClientPort)
+		if !deletedNodes.Exist(id) {
+			pdServers = append(pdServers, s)
+		}
+	}
+	cluster.PDServers = pdServers
+
 	for i := 0; i < len(cluster.TiKVServers); i++ {
 		s := cluster.TiKVServers[i]
 		id := s.Host + ":" + strconv.Itoa(s.Port)
diff --git a/tests/tiup-cluster/run.sh b/tests/tiup-cluster/run.sh
index 0ce27816e4..cd390aa75b 100755
--- a/tests/tiup-cluster/run.sh
+++ b/tests/tiup-cluster/run.sh
@@ -21,7 +21,7 @@ cp -r ./local /tmp/local
 ls -lh /tmp/local
 
 function tiup-cluster() {
-    mkdir -p "~/.tiup/bin" && cp -f ./root.json ~/.tiup/bin/
+    mkdir -p ~/.tiup/bin && cp -f ./root.json ~/.tiup/bin/
     # echo "in function"
     if [ -f "./bin/tiup-cluster.test" ]; then
         ./bin/tiup-cluster.test -test.coverprofile=./cover/cov.itest-$(date +'%s')-$RANDOM.out __DEVEL--i-heard-you-like-tests "$@"
diff --git a/tests/tiup-cluster/script/scale_core.sh b/tests/tiup-cluster/script/scale_core.sh
index 718415f637..b932bec0ac 100755
--- a/tests/tiup-cluster/script/scale_core.sh
+++ b/tests/tiup-cluster/script/scale_core.sh
@@ -58,10 +58,17 @@ function scale_core() {
     echo "start scale in pd"
     tiup-cluster $client --yes scale-in $name -N $ipprefix.103:2379
     wait_instance_num_reach $name $total_sub_one $native_ssh
+
+    # validate https://github.com/pingcap/tiup/issues/786
+    # ensure that this instance is removed from the startup scripts of other components that need to rely on PD
+    ! tiup-cluster $client exec $name -N $ipprefix.101 --command "grep -q $ipprefix.103:2379 /home/tidb/deploy/tidb-4000/scripts/run_tidb.sh"
     echo "start scale out pd"
     topo=./topo/full_scale_in_pd.yaml
     sed "s/__IPPREFIX__/$ipprefix/g" $topo.tpl > $topo
     tiup-cluster $client --yes scale-out $name $topo
 
+    # after scalue-out, ensure this instance come back
+    tiup-cluster $client exec $name -N $ipprefix.101 --command "grep -q $ipprefix.103:2379 /home/tidb/deploy/tidb-4000/scripts/run_tidb.sh"
+
     tiup-cluster $client _test $name writable
 }