schedulers: evict-leader supports schedule the regions with unhealthy…

… peers (#4096) (#4133)

close #4093

Signed-off-by: HunDunDM <hundundm@gmail.com>

Co-authored-by: HunDunDM <hundundm@gmail.com>

server/schedulers/evict_leader.go

@@ -246,12 +246,26 @@ func (s *evictLeaderScheduler) IsScheduleAllowed(cluster opt.Cluster) bool {
 func (s *evictLeaderScheduler) scheduleOnce(cluster opt.Cluster) []*operator.Operator {
 	var ops []*operator.Operator
 	for id, ranges := range s.conf.StoreIDWithRanges {
+		var filters []filter.Filter
 		region := cluster.RandLeaderRegion(id, ranges, opt.HealthRegion(cluster))
 		if region == nil {
-			schedulerCounter.WithLabelValues(s.GetName(), "no-leader").Inc()
-			continue
+			// try to pick unhealthy region
+			region = cluster.RandLeaderRegion(id, ranges)
+			if region == nil {
+				schedulerCounter.WithLabelValues(s.GetName(), "no-leader").Inc()
+				continue
+			}
+			schedulerCounter.WithLabelValues(s.GetName(), "pick-unhealthy-region").Inc()
+			unhealthyPeerStores := make(map[uint64]struct{})
+			for _, peer := range region.GetDownPeers() {
+				unhealthyPeerStores[peer.GetPeer().GetStoreId()] = struct{}{}
+			}
+			for _, peer := range region.GetPendingPeers() {
+				unhealthyPeerStores[peer.GetStoreId()] = struct{}{}
+			}
+			filters = append(filters, filter.NewExcludedFilter(s.GetName(), nil, unhealthyPeerStores))
 		}
-		target := s.selector.SelectTarget(cluster, cluster.GetFollowerStores(region))
+		target := s.selector.SelectTarget(cluster, cluster.GetFollowerStores(region), filters...)
 		if target == nil {
 			schedulerCounter.WithLabelValues(s.GetName(), "no-target-store").Inc()
 			continue

server/schedulers/scheduler_test.go

@@ -19,6 +19,7 @@ import (
 
 	. "github.com/pingcap/check"
 	"github.com/pingcap/kvproto/pkg/metapb"
+	"github.com/pingcap/kvproto/pkg/pdpb"
 	"github.com/tikv/pd/pkg/mock/mockcluster"
 	"github.com/tikv/pd/pkg/mock/mockoption"
 	"github.com/tikv/pd/pkg/testutil"
@@ -359,6 +360,40 @@ func (s *testEvictLeaderSuite) TestEvictLeader(c *C) {
 	testutil.CheckTransferLeader(c, op[0], operator.OpLeader, 1, 2)
 }
 
+func (s *testEvictLeaderSuite) TestEvictLeaderWithUnhealthyPeer(c *C) {
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	opt := mockoption.NewScheduleOptions()
+	tc := mockcluster.NewCluster(opt)
+	sl, err := schedule.CreateScheduler(EvictLeaderType, schedule.NewOperatorController(ctx, nil, nil), core.NewStorage(kv.NewMemoryKV()), schedule.ConfigSliceDecoder(EvictLeaderType, []string{"1"}))
+	c.Assert(err, IsNil)
+
+	// Add stores 1, 2, 3
+	tc.AddLeaderStore(1, 0)
+	tc.AddLeaderStore(2, 0)
+	tc.AddLeaderStore(3, 0)
+	// Add region 1, which has 3 peers. 1 is leader. 2 is healthy or pending, 3 is healthy or down.
+	tc.AddLeaderRegion(1, 1, 2, 3)
+	region := tc.MockRegionInfo(1, 1, []uint64{2, 3}, nil, nil)
+	withDownPeer := core.WithDownPeers([]*pdpb.PeerStats{{
+		Peer:        region.GetPeers()[2],
+		DownSeconds: 1000,
+	}})
+	withPendingPeer := core.WithPendingPeers([]*metapb.Peer{region.GetPeers()[1]})
+
+	// only pending
+	tc.PutRegion(region.Clone(withPendingPeer))
+	op := sl.Schedule(tc)
+	testutil.CheckTransferLeader(c, op[0], operator.OpLeader, 1, 3)
+	// only down
+	tc.PutRegion(region.Clone(withDownPeer))
+	op = sl.Schedule(tc)
+	testutil.CheckTransferLeader(c, op[0], operator.OpLeader, 1, 2)
+	// pending + down
+	tc.PutRegion(region.Clone(withPendingPeer, withDownPeer))
+	c.Assert(sl.Schedule(tc), HasLen, 0)
+}
+
 var _ = Suite(&testShuffleRegionSuite{})
 
 type testShuffleRegionSuite struct{}