server: make leader can lost lease as soon as possible (#3712) (#3719)

* This is an automated cherry-pick of #3712 Signed-off-by: ti-chi-bot <[email protected]> * reslove conflict Signed-off-by: nolouch <[email protected]> Co-authored-by: ShuNing <[email protected]>
tikv · May 28, 2021 · 04ed1ff · 04ed1ff
1 parent 8a2cfe9
commit 04ed1ff
Show file tree

Hide file tree

Showing 3 changed files with 39 additions and 6 deletions.
diff --git a/server/cluster/cluster.go b/server/cluster/cluster.go
@@ -364,6 +364,7 @@ func (c *RaftCluster) Stop() {
 	c.coordinator.stop()
 	c.Unlock()
 	c.wg.Wait()
+	log.Info("raftcluster is stopped")
 }
 
 // IsRunning return if the cluster is running.

diff --git a/server/server.go b/server/server.go
@@ -647,6 +647,7 @@ func (s *Server) createRaftCluster() error {
 }
 
 func (s *Server) stopRaftCluster() {
+	failpoint.Inject("raftclusterIsBusy", func() {})
 	s.cluster.Stop()
 }
 
@@ -1219,9 +1220,13 @@ func (s *Server) campaignLeader() {
 	//   1. lease based approach is not affected by thread pause, slow runtime schedule, etc.
 	//   2. load region could be slow. Based on lease we can recover TSO service faster.
 	ctx, cancel := context.WithCancel(s.serverLoopCtx)
-	defer cancel()
-	defer s.member.ResetLeader()
-	// maintain the PD leader
+	var resetLeaderOnce sync.Once
+	defer resetLeaderOnce.Do(func() {
+		cancel()
+		s.member.ResetLeader()
+	})
+
+	// maintain the PD leadership, after this, TSO can be service.
 	go s.member.KeepLeader(ctx)
 	log.Info("campaign pd leader ok", zap.String("campaign-pd-leader-name", s.Name()))
 
@@ -1236,8 +1241,6 @@ func (s *Server) campaignLeader() {
 		return
 	}
 	defer s.tsoAllocatorManager.ResetAllocatorGroup(tso.GlobalDCLocation)
-	// Check the cluster dc-location after the PD leader is elected
-	go s.tsoAllocatorManager.ClusterDCLocationChecker()
 
 	if err := s.reloadConfigFromKV(); err != nil {
 		log.Error("failed to reload configuration", errs.ZapError(err))
@@ -1263,7 +1266,16 @@ func (s *Server) campaignLeader() {
 		log.Error("failed to sync id from etcd", errs.ZapError(err))
 		return
 	}
+	// EnableLeader to accept the remaining service, such as GetStore, GetRegion.
 	s.member.EnableLeader()
+	// Check the cluster dc-location after the PD leader is elected.
+	go s.tsoAllocatorManager.ClusterDCLocationChecker()
+	defer resetLeaderOnce.Do(func() {
+		// as soon as cancel the leadership keepalive, then other member have chance
+		// to be new leader.
+		cancel()
+		s.member.ResetLeader()
+	})
 
 	CheckPDVersion(s.persistOptions)
 	log.Info("PD cluster leader is ready to serve", zap.String("pd-leader-name", s.Name()))

diff --git a/tests/server/member/member_test.go b/tests/server/member/member_test.go
@@ -26,6 +26,7 @@ import (
 
 	. "github.com/pingcap/check"
 	"github.com/pingcap/errors"
+	"github.com/pingcap/failpoint"
 	"github.com/pingcap/kvproto/pkg/pdpb"
 	"github.com/tikv/pd/pkg/etcdutil"
 	"github.com/tikv/pd/pkg/testutil"
@@ -238,12 +239,31 @@ func (s *serverTestSuite) TestLeaderResign(c *C) {
 	c.Assert(leader3, Equals, leader1)
 }
 
+func (s *serverTestSuite) TestLeaderResignWithBlock(c *C) {
+	cluster, err := tests.NewTestCluster(s.ctx, 3)
+	defer cluster.Destroy()
+	c.Assert(err, IsNil)
+
+	err = cluster.RunInitialServers()
+	c.Assert(err, IsNil)
+
+	leader1 := cluster.WaitLeader()
+	addr1 := cluster.GetServer(leader1).GetConfig().ClientUrls
+
+	err = failpoint.Enable("github.com/tikv/pd/server/raftclusterIsBusy", `pause`)
+	c.Assert(err, IsNil)
+	defer failpoint.Disable("github.com/tikv/pd/server/raftclusterIsBusy")
+	s.post(c, addr1+"/pd/api/v1/leader/resign", "")
+	leader2 := s.waitLeaderChange(c, cluster, leader1)
+	c.Log("leader2:", leader2)
+	c.Assert(leader2, Not(Equals), leader1)
+}
+
 func (s *serverTestSuite) waitLeaderChange(c *C, cluster *tests.TestCluster, old string) string {
 	var leader string
 	testutil.WaitUntil(c, func(c *C) bool {
 		leader = cluster.GetLeader()
 		if leader == old || leader == "" {
-			time.Sleep(time.Second)
 			return false
 		}
 		return true