*: fix region stats check

pkg/cluster/cluster.go

@@ -55,7 +55,7 @@ func HandleOverlaps(c Cluster, overlaps []*core.RegionInfo) {
 }
 
 // Collect collects the cluster information.
-func Collect(c Cluster, region *core.RegionInfo, stores []*core.StoreInfo, hasRegionStats, isNew, isPrepared bool) {
+func Collect(c Cluster, region *core.RegionInfo, stores []*core.StoreInfo, hasRegionStats bool) {
 	if hasRegionStats {
 		c.GetRegionStats().Observe(region, stores)
 	}

pkg/core/region.go

@@ -98,6 +98,12 @@ func (r *RegionInfo) LoadedFromStorage() bool {
 	return r.source == Storage
 }
 
+// LoadedFromSync means this region's meta info loaded from region syncer.
+// Only used for test.
+func (r *RegionInfo) LoadedFromSync() bool {
+	return r.source == Sync
+}
+
 // NewRegionInfo creates RegionInfo with region's meta and leader peer.
 func NewRegionInfo(region *metapb.Region, leader *metapb.Peer, opts ...RegionCreateOption) *RegionInfo {
 	regionInfo := &RegionInfo{
@@ -705,7 +711,7 @@ func (r *RegionInfo) isRegionRecreated() bool {
 
 // RegionGuideFunc is a function that determines which follow-up operations need to be performed based on the origin
 // and new region information.
-type RegionGuideFunc func(region, origin *RegionInfo) (isNew, saveKV, saveCache, needSync bool)
+type RegionGuideFunc func(region, origin *RegionInfo) (saveKV, saveCache, needSync bool)
 
 // GenerateRegionGuideFunc is used to generate a RegionGuideFunc. Control the log output by specifying the log function.
 // nil means do not print the log.
@@ -718,19 +724,16 @@ func GenerateRegionGuideFunc(enableLog bool) RegionGuideFunc {
 	}
 	// Save to storage if meta is updated.
 	// Save to cache if meta or leader is updated, or contains any down/pending peer.
-	// Mark isNew if the region in cache does not have leader.
-	return func(region, origin *RegionInfo) (isNew, saveKV, saveCache, needSync bool) {
+	// Now, no need to hols `isNew` because `saveCache` must be true if if the region in cache does not have leader.
+	return func(region, origin *RegionInfo) (saveKV, saveCache, needSync bool) {
 		if origin == nil {
 			if log.GetLevel() <= zap.DebugLevel {
 				debug("insert new region",
 					zap.Uint64("region-id", region.GetID()),
 					logutil.ZapRedactStringer("meta-region", RegionToHexMeta(region.GetMeta())))
 			}
-			saveKV, saveCache, isNew = true, true, true
+			saveKV, saveCache = true, true
 		} else {
-			if origin.LoadedFromStorage() {
-				isNew = true
-			}
 			r := region.GetRegionEpoch()
 			o := origin.GetRegionEpoch()
 			if r.GetVersion() > o.GetVersion() {
@@ -756,9 +759,7 @@ func GenerateRegionGuideFunc(enableLog bool) RegionGuideFunc {
 				saveKV, saveCache = true, true
 			}
 			if region.GetLeader().GetId() != origin.GetLeader().GetId() {
-				if origin.GetLeader().GetId() == 0 {
-					isNew = true
-				} else if log.GetLevel() <= zap.InfoLevel {
+				if log.GetLevel() <= zap.InfoLevel {
 					info("leader changed",
 						zap.Uint64("region-id", region.GetID()),
 						zap.Uint64("from", origin.GetLeader().GetStoreId()),

pkg/core/region_test.go

@@ -363,7 +363,7 @@ func TestNeedSync(t *testing.T) {
 	for _, testCase := range testCases {
 		regionA := region.Clone(testCase.optionsA...)
 		regionB := region.Clone(testCase.optionsB...)
-		_, _, _, needSync := RegionGuide(regionA, regionB)
+		_, _, needSync := RegionGuide(regionA, regionB)
 		re.Equal(testCase.needSync, needSync)
 	}
 }

pkg/mcs/scheduling/server/cluster.go

@@ -552,9 +552,8 @@ func (c *Cluster) processRegionHeartbeat(region *core.RegionInfo) error {
 	hasRegionStats := c.regionStats != nil
 	// Save to storage if meta is updated, except for flashback.
 	// Save to cache if meta or leader is updated, or contains any down/pending peer.
-	// Mark isNew if the region in cache does not have leader.
-	isNew, _, saveCache, _ := core.GenerateRegionGuideFunc(true)(region, origin)
-	if !saveCache && !isNew {
+	_, saveCache, _ := core.GenerateRegionGuideFunc(true)(region, origin)
+	if !saveCache {
 		// Due to some config changes need to update the region stats as well,
 		// so we do some extra checks here.
 		if hasRegionStats && c.regionStats.RegionStatsNeedUpdate(region) {
@@ -576,7 +575,7 @@ func (c *Cluster) processRegionHeartbeat(region *core.RegionInfo) error {
 		cluster.HandleOverlaps(c, overlaps)
 	}
 
-	cluster.Collect(c, region, c.GetRegionStores(region), hasRegionStats, isNew, c.IsPrepared())
+	cluster.Collect(c, region, c.GetRegionStores(region), hasRegionStats)
 	return nil
 }
 

pkg/mcs/scheduling/server/grpc_service.go

@@ -158,7 +158,7 @@ func (s *Service) RegionHeartbeat(stream schedulingpb.Scheduling_RegionHeartbeat
 			s.hbStreams.BindStream(storeID, server)
 			lastBind = time.Now()
 		}
-		region := core.RegionFromHeartbeat(request, core.SetSource(core.Heartbeat))
+		region := core.RegionFromHeartbeat(request)
 		err = c.HandleRegionHeartbeat(region)
 		if err != nil {
 			// TODO: if we need to send the error back to API server.

pkg/statistics/region_collection.go

@@ -148,6 +148,9 @@
 // due to some special state types.
 func (r *RegionStatistics) RegionStatsNeedUpdate(region *core.RegionInfo) bool {
 	regionID := region.GetID()
+	if !r.isObserved(regionID) {
+		return true
+	}
 	if r.IsRegionStatsType(regionID, OversizedRegion) !=
 		region.IsOversized(int64(r.conf.GetRegionMaxSize()), int64(r.conf.GetRegionMaxKeys())) {
 		return true
@@ -156,6 +159,14 @@
 		region.NeedMerge(int64(r.conf.GetMaxMergeRegionSize()), int64(r.conf.GetMaxMergeRegionKeys()))
 }
 
+// isObserved returns whether the region is observed. And it also shows whether PD recived heartbeat of this region.
+func (r *RegionStatistics) isObserved(id uint64) bool {
+	r.RLock()
+	defer r.RUnlock()
+	_, ok := r.index[id]
+	return ok
+}
+
 // Observe records the current regions' status.
 func (r *RegionStatistics) Observe(region *core.RegionInfo, stores []*core.StoreInfo) {
 	r.Lock()
@@ -164,7 +175,6 @@
 		desiredReplicas = r.conf.GetMaxReplicas()
 		desiredVoters   = desiredReplicas
 		peerTypeIndex   RegionStatisticType
-		deleteIndex     RegionStatisticType
 	)
 	// Check if the region meets count requirements of its rules.
 	if r.conf.IsPlacementRulesEnabled() {
@@ -240,10 +250,10 @@
 		}
 	}
 	// Remove the info if any of the conditions are not met any more.
-	if oldIndex, ok := r.index[regionID]; ok {
-		deleteIndex = oldIndex &^ peerTypeIndex
+	if oldIndex, ok := r.index[regionID]; ok && oldIndex > 0 {
+		deleteIndex := oldIndex &^ peerTypeIndex
+		r.deleteEntry(deleteIndex, regionID)
 	}
-	r.deleteEntry(deleteIndex, regionID)
 	r.index[regionID] = peerTypeIndex
 }
 
@@ -252,7 +262,10 @@
 	r.Lock()
 	defer r.Unlock()
 	if oldIndex, ok := r.index[regionID]; ok {
-		r.deleteEntry(oldIndex, regionID)
+		delete(r.index, regionID)
+		if oldIndex > 0 {
+			r.deleteEntry(oldIndex, regionID)
+		}
 	}
 }
 

pkg/syncer/client.go

@@ -205,7 +205,7 @@ func (s *RegionSyncer) StartSyncWithLeader(addr string) {
 						log.Debug("region is stale", zap.Stringer("origin", origin.GetMeta()), errs.ZapError(err))
 						continue
 					}
-					_, saveKV, _, _ := regionGuide(region, origin)
+					saveKV, _, _ := regionGuide(region, origin)
 					overlaps := bc.PutRegion(region)
 
 					if hasBuckets {

server/cluster/cluster.go

@@ -1004,9 +1004,8 @@ func (c *RaftCluster) processRegionHeartbeat(region *core.RegionInfo) error {
 	hasRegionStats := c.regionStats != nil
 	// Save to storage if meta is updated, except for flashback.
 	// Save to cache if meta or leader is updated, or contains any down/pending peer.
-	// Mark isNew if the region in cache does not have leader.
-	isNew, saveKV, saveCache, needSync := regionGuide(region, origin)
-	if !saveKV && !saveCache && !isNew {
+	saveKV, saveCache, needSync := regionGuide(region, origin)
+	if !saveKV && !saveCache {
 		// Due to some config changes need to update the region stats as well,
 		// so we do some extra checks here.
 		if hasRegionStats && c.regionStats.RegionStatsNeedUpdate(region) {
@@ -1037,11 +1036,7 @@ func (c *RaftCluster) processRegionHeartbeat(region *core.RegionInfo) error {
 		regionUpdateCacheEventCounter.Inc()
 	}
 
-	isPrepared := true
-	if !c.IsServiceIndependent(mcsutils.SchedulingServiceName) {
-		isPrepared = c.IsPrepared()
-	}
-	cluster.Collect(c, region, c.GetRegionStores(region), hasRegionStats, isNew, isPrepared)
+	cluster.Collect(c, region, c.GetRegionStores(region), hasRegionStats)
 
 	if c.storage != nil {
 		// If there are concurrent heartbeats from the same region, the last write will win even if

tests/server/cluster/cluster_test.go

@@ -38,6 +38,7 @@
 	sc "github.com/tikv/pd/pkg/schedule/config"
 	"github.com/tikv/pd/pkg/schedule/operator"
 	"github.com/tikv/pd/pkg/schedule/schedulers"
+	"github.com/tikv/pd/pkg/statistics"
 	"github.com/tikv/pd/pkg/storage"
 	"github.com/tikv/pd/pkg/syncer"
 	"github.com/tikv/pd/pkg/tso"
@@ -181,6 +182,99 @@
 	re.Equal(uint64(1), rc.GetOperatorController().OperatorCount(operator.OpAdmin))
 }
 
+func TestRegionStatistics(t *testing.T) {
+	re := require.New(t)
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	tc, err := tests.NewTestCluster(ctx, 2)
+	defer tc.Destroy()
+	re.NoError(err)
+
+	err = tc.RunInitialServers()
+	re.NoError(err)
+
+	leaderName := tc.WaitLeader()
+	leaderServer := tc.GetLeaderServer()
+	grpcPDClient := testutil.MustNewGrpcClient(re, leaderServer.GetAddr())
+	clusterID := leaderServer.GetClusterID()
+	bootstrapCluster(re, clusterID, grpcPDClient)
+	rc := leaderServer.GetRaftCluster()
+
+	region := &metapb.Region{
+		Id:       10,
+		StartKey: []byte("abc"),
+		EndKey:   []byte("xyz"),
+		Peers: []*metapb.Peer{
+			{Id: 101, StoreId: 1},
+			{Id: 102, StoreId: 2},
+			{Id: 103, StoreId: 3},
+			{Id: 104, StoreId: 4, Role: metapb.PeerRole_Learner},
+		},
+	}
+
+	// To put region.
+	regionInfo := core.NewRegionInfo(region, region.Peers[0], core.SetApproximateSize(0))
+	err = tc.HandleRegionHeartbeat(regionInfo)
+	re.NoError(err)
+	regions := rc.GetRegionStatsByType(statistics.LearnerPeer)
+	re.Len(regions, 1)
+
+	// wait for sync region
+	time.Sleep(1000 * time.Millisecond)
+
+	leaderServer.ResignLeader()
+	newLeaderName := tc.WaitLeader()
+	re.NotEqual(newLeaderName, leaderName)
+	leaderServer = tc.GetLeaderServer()
+	rc = leaderServer.GetRaftCluster()
+	r := rc.GetRegion(region.Id)
+	re.NotNil(r)
+	re.True(r.LoadedFromSync())
+	regions = rc.GetRegionStatsByType(statistics.LearnerPeer)
+	re.Len(regions, 0)
+	err = tc.HandleRegionHeartbeat(regionInfo)
+	re.NoError(err)
+	regions = rc.GetRegionStatsByType(statistics.LearnerPeer)
+	re.Len(regions, 1)
+
+	leaderServer.ResignLeader()
+	newLeaderName = tc.WaitLeader()
+	re.Equal(newLeaderName, leaderName)
+	leaderServer = tc.GetLeaderServer()
+	rc = leaderServer.GetRaftCluster()
+	re.NotNil(r)
+	re.True(r.LoadedFromStorage() || r.LoadedFromSync())
+	regions = rc.GetRegionStatsByType(statistics.LearnerPeer)
+	re.Len(regions, 0)
+	regionInfo = regionInfo.Clone(core.SetSource(core.Heartbeat), core.SetApproximateSize(30))
+	err = tc.HandleRegionHeartbeat(regionInfo)
+	re.NoError(err)
+	rc = leaderServer.GetRaftCluster()
+	r = rc.GetRegion(region.Id)
+	re.NotNil(r)
+	re.False(r.LoadedFromStorage() && r.LoadedFromSync())
+
+	leaderServer.ResignLeader()
+	newLeaderName = tc.WaitLeader()
+	re.NotEqual(newLeaderName, leaderName)
+	leaderServer.ResignLeader()
+	newLeaderName = tc.WaitLeader()
+	re.Equal(newLeaderName, leaderName)
+	leaderServer = tc.GetLeaderServer()
+	rc = leaderServer.GetRaftCluster()
+	r = rc.GetRegion(region.Id)
+	re.NotNil(r)
+	re.False(r.LoadedFromStorage() && r.LoadedFromSync())
+	regions = rc.GetRegionStatsByType(statistics.LearnerPeer)
+	re.Len(regions, 0)
+
+	regionInfo = regionInfo.Clone(core.SetSource(core.Heartbeat), core.SetApproximateSize(30))
+	err = tc.HandleRegionHeartbeat(regionInfo)
+	re.NoError(err)
+	regions = rc.GetRegionStatsByType(statistics.LearnerPeer)
+	re.Len(regions, 1)
+}
+
 func TestStaleRegion(t *testing.T) {
 	re := require.New(t)
 	ctx, cancel := context.WithCancel(context.Background())