*: new store region score function for balance (tikv#1014)

Author: Connor1996

conf/config.toml

@@ -61,7 +61,9 @@ leader-schedule-limit = 4
 region-schedule-limit = 4
 replica-schedule-limit = 8
 merge-schedule-limit = 8
-tolerant-size-ratio = 2.5
+tolerant-size-ratio = 5
+high-space-reatio = 0.8
+low-space-ratio = 0.6
 
 # customized schedulers, the format is as below
 # if empty, it will use balance-leader, balance-region, hot-region as default

server/api/label.go

@@ -72,8 +72,6 @@ func (h *labelsHandler) GetStores(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	maxDownTime := h.svr.GetScheduleConfig().MaxStoreDownTime.Duration
-
 	stores := cluster.GetStores()
 	storesInfo := &StoresInfo{
 		Stores: make([]*StoreInfo, 0, len(stores)),
@@ -87,7 +85,7 @@ func (h *labelsHandler) GetStores(w http.ResponseWriter, r *http.Request) {
 			return
 		}
 
-		storeInfo := newStoreInfo(store, maxDownTime)
+		storeInfo := newStoreInfo(h.svr.GetScheduleConfig(), store)
 		storesInfo.Stores = append(storesInfo.Stores, storeInfo)
 	}
 	storesInfo.Count = len(storesInfo.Stores)

server/api/store.go

@@ -66,7 +66,7 @@ const (
 	downStateName    = "Down"
 )
 
-func newStoreInfo(store *core.StoreInfo, maxStoreDownTime time.Duration) *StoreInfo {
+func newStoreInfo(opt *server.ScheduleConfig, store *core.StoreInfo) *StoreInfo {
 	s := &StoreInfo{
 		Store: &MetaStore{
 			Store:     store.Store,
@@ -77,11 +77,11 @@ func newStoreInfo(store *core.StoreInfo, maxStoreDownTime time.Duration) *StoreI
 			Available:          typeutil.ByteSize(store.Stats.GetAvailable()),
 			LeaderCount:        store.LeaderCount,
 			LeaderWeight:       store.LeaderWeight,
-			LeaderScore:        store.LeaderScore(),
+			LeaderScore:        store.LeaderScore(0),
 			LeaderSize:         store.LeaderSize,
 			RegionCount:        store.RegionCount,
 			RegionWeight:       store.RegionWeight,
-			RegionScore:        store.RegionScore(),
+			RegionScore:        store.RegionScore(opt.HighSpaceRatio, opt.LowSpaceRatio, 0),
 			RegionSize:         store.RegionSize,
 			SendingSnapCount:   store.Stats.GetSendingSnapCount(),
 			ReceivingSnapCount: store.Stats.GetReceivingSnapCount(),
@@ -103,7 +103,7 @@ func newStoreInfo(store *core.StoreInfo, maxStoreDownTime time.Duration) *StoreI
 	}
 
 	if store.State == metapb.StoreState_Up {
-		if store.DownTime() > maxStoreDownTime {
+		if store.DownTime() > opt.MaxStoreDownTime.Duration {
 			s.Store.StateName = downStateName
 		} else if store.IsDisconnected() {
 			s.Store.StateName = disconnectedName
@@ -137,8 +137,6 @@ func (h *storeHandler) Get(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	maxStoreDownTime := h.svr.GetScheduleConfig().MaxStoreDownTime.Duration
-
 	vars := mux.Vars(r)
 	storeIDStr := vars["id"]
 	storeID, err := strconv.ParseUint(storeIDStr, 10, 64)
@@ -153,7 +151,7 @@ func (h *storeHandler) Get(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	storeInfo := newStoreInfo(store, maxStoreDownTime)
+	storeInfo := newStoreInfo(h.svr.GetScheduleConfig(), store)
 	h.rd.JSON(w, http.StatusOK, storeInfo)
 }
 
@@ -324,8 +322,6 @@ func (h *storesHandler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	maxStoreDownTime := h.svr.GetScheduleConfig().MaxStoreDownTime.Duration
-
 	stores := cluster.GetStores()
 	StoresInfo := &StoresInfo{
 		Stores: make([]*StoreInfo, 0, len(stores)),
@@ -345,7 +341,7 @@ func (h *storesHandler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 			return
 		}
 
-		storeInfo := newStoreInfo(store, maxStoreDownTime)
+		storeInfo := newStoreInfo(h.svr.GetScheduleConfig(), store)
 		StoresInfo.Stores = append(StoresInfo.Stores, storeInfo)
 	}
 	StoresInfo.Count = len(StoresInfo.Stores)

server/api/store_test.go

@@ -270,14 +270,14 @@ func (s *testStoreSuite) TestDownState(c *C) {
 		Stats:           &pdpb.StoreStats{},
 		LastHeartbeatTS: time.Now(),
 	}
-	storeInfo := newStoreInfo(store, time.Hour)
+	storeInfo := newStoreInfo(s.svr.GetScheduleConfig(), store)
 	c.Assert(storeInfo.Store.StateName, Equals, metapb.StoreState_Up.String())
 
 	store.LastHeartbeatTS = time.Now().Add(-time.Minute * 2)
-	storeInfo = newStoreInfo(store, time.Hour)
+	storeInfo = newStoreInfo(s.svr.GetScheduleConfig(), store)
 	c.Assert(storeInfo.Store.StateName, Equals, disconnectedName)
 
 	store.LastHeartbeatTS = time.Now().Add(-time.Hour * 2)
-	storeInfo = newStoreInfo(store, time.Hour)
+	storeInfo = newStoreInfo(s.svr.GetScheduleConfig(), store)
 	c.Assert(storeInfo.Store.StateName, Equals, downStateName)
 }

server/api/trend.go

@@ -107,8 +107,6 @@ func (h *trendHandler) Handle(w http.ResponseWriter, r *http.Request) {
 }
 
 func (h *trendHandler) getTrendStores() ([]trendStore, error) {
-	maxStoreDownTime := h.svr.GetScheduleConfig().MaxStoreDownTime.Duration
-
 	var readStats, writeStats core.StoreHotRegionsStat
 	if hotRead := h.GetHotReadRegions(); hotRead != nil {
 		readStats = hotRead.AsLeader
@@ -123,7 +121,7 @@ func (h *trendHandler) getTrendStores() ([]trendStore, error) {
 
 	trendStores := make([]trendStore, 0, len(stores))
 	for _, store := range stores {
-		info := newStoreInfo(store, maxStoreDownTime)
+		info := newStoreInfo(h.svr.GetScheduleConfig(), store)
 		s := trendStore{
 			ID:              info.Store.GetId(),
 			Address:         info.Store.GetAddress(),

server/cache.go

@@ -167,28 +167,6 @@ func (c *clusterInfo) GetStores() []*core.StoreInfo {
 	return c.BasicCluster.GetStores()
 }
 
-// GetStoresAverageScore returns the total resource score of all unfiltered stores.
-func (c *clusterInfo) GetStoresAverageScore(kind core.ResourceKind, filters ...schedule.Filter) float64 {
-	c.RLock()
-	defer c.RUnlock()
-
-	var totalResourceSize int64
-	var totalResourceWeight float64
-	for _, s := range c.BasicCluster.GetStores() {
-		if schedule.FilterSource(c, s, filters) {
-			continue
-		}
-
-		totalResourceWeight += s.ResourceWeight(kind)
-		totalResourceSize += s.ResourceSize(kind)
-	}
-
-	if totalResourceWeight == 0 {
-		return 0
-	}
-	return float64(totalResourceSize) / totalResourceWeight
-}
-
 func (c *clusterInfo) getMetaStores() []*metapb.Store {
 	c.RLock()
 	defer c.RUnlock()
@@ -564,6 +542,14 @@ func (c *clusterInfo) GetTolerantSizeRatio() float64 {
 	return c.opt.GetTolerantSizeRatio()
 }
 
+func (c *clusterInfo) GetLowSpaceRatio() float64 {
+	return c.opt.GetLowSpaceRatio()
+}
+
+func (c *clusterInfo) GetHighSpaceRatio() float64 {
+	return c.opt.GetHighSpaceRatio()
+}
+
 func (c *clusterInfo) GetMaxSnapshotCount() uint64 {
 	return c.opt.GetMaxSnapshotCount()
 }

server/config.go

@@ -352,6 +352,18 @@ type ScheduleConfig struct {
 	MergeScheduleLimit uint64 `toml:"merge-schedule-limit,omitempty" json:"merge-schedule-limit"`
 	// TolerantSizeRatio is the ratio of buffer size for balance scheduler.
 	TolerantSizeRatio float64 `toml:"tolerant-size-ratio,omitempty" json:"tolerant-size-ratio"`
+	//
+	//      high space stage         transition stage           low space stage
+	//   |--------------------|-----------------------------|-------------------------|
+	//   ^                    ^                             ^                         ^
+	//   0       HighSpaceRatio * capacity       LowSpaceRatio * capacity          capacity
+	//
+	// LowSpaceRatio is the lowest usage ratio of store which regraded as low space.
+	// When in low space, store region score increases to very large and varies inversely with available size.
+	LowSpaceRatio float64 `toml:"low-space-ratio,omitempty" json:"low-space-ratio"`
+	// HighSpaceRatio is the highest usage ratio of store which regraded as high space.
+	// High space means there is a lot of spare capacity, and store region score varies directly with used size.
+	HighSpaceRatio float64 `toml:"high-space-ratio,omitempty" json:"high-space-ratio"`
 	// EnableRaftLearner is the option for using AddLearnerNode instead of AddNode
 	EnableRaftLearner bool `toml:"enable-raft-learner" json:"enable-raft-learner,string"`
 	// Schedulers support for loding customized schedulers
@@ -371,6 +383,8 @@ func (c *ScheduleConfig) clone() *ScheduleConfig {
 		ReplicaScheduleLimit: c.ReplicaScheduleLimit,
 		MergeScheduleLimit:   c.MergeScheduleLimit,
 		TolerantSizeRatio:    c.TolerantSizeRatio,
+		LowSpaceRatio:        c.LowSpaceRatio,
+		HighSpaceRatio:       c.HighSpaceRatio,
 		EnableRaftLearner:    c.EnableRaftLearner,
 		Schedulers:           schedulers,
 	}
@@ -396,7 +410,9 @@ const (
 	defaultRegionScheduleLimit  = 4
 	defaultReplicaScheduleLimit = 8
 	defaultMergeScheduleLimit   = 8
-	defaultTolerantSizeRatio    = 2.5
+	defaultTolerantSizeRatio    = 5
+	defaultLowSpaceRatio        = 0.8
+	defaultHighSpaceRatio       = 0.6
 )
 
 var defaultSchedulers = SchedulerConfigs{
@@ -416,6 +432,8 @@ func (c *ScheduleConfig) adjust() {
 	adjustUint64(&c.ReplicaScheduleLimit, defaultReplicaScheduleLimit)
 	adjustUint64(&c.MergeScheduleLimit, defaultMergeScheduleLimit)
 	adjustFloat64(&c.TolerantSizeRatio, defaultTolerantSizeRatio)
+	adjustFloat64(&c.LowSpaceRatio, defaultLowSpaceRatio)
+	adjustFloat64(&c.HighSpaceRatio, defaultHighSpaceRatio)
 	adjustSchedulers(&c.Schedulers, defaultSchedulers)
 }
 

server/coordinator.go

@@ -436,6 +436,7 @@ func (c *coordinator) addOperatorLocked(op *schedule.Operator) bool {
 	if old, ok := c.operators[regionID]; ok {
 		if !isHigherPriorityOperator(op, old) {
 			log.Infof("[region %v] cancel add operator, old: %s", regionID, old)
+			operatorCounter.WithLabelValues(op.Desc(), "canceled").Inc()
 			return false
 		}
 		log.Infof("[region %v] replace old operator: %s", regionID, old)
@@ -470,6 +471,7 @@ func (c *coordinator) addOperators(ops ...*schedule.Operator) bool {
 	for _, op := range ops {
 		if old := c.operators[op.RegionID()]; old != nil && !isHigherPriorityOperator(op, old) {
 			log.Infof("[region %v] cancel add operators, old: %s", op.RegionID(), old)
+			operatorCounter.WithLabelValues(op.Desc(), "canceled").Inc()
 			return false
 		}
 	}

server/coordinator_test.go

@@ -65,8 +65,8 @@ func (c *testClusterInfo) addRegionStore(storeID uint64, regionCount int) {
 	store.LastHeartbeatTS = time.Now()
 	store.RegionCount = regionCount
 	store.RegionSize = int64(regionCount) * 10
-	store.Stats.Capacity = uint64(1024)
-	store.Stats.Available = store.Stats.Capacity
+	store.Stats.Capacity = 1000 * (1 << 20)
+	store.Stats.Available = store.Stats.Capacity - uint64(store.RegionSize)
 	c.putStore(store)
 }
 

server/core/store.go

@@ -45,6 +45,7 @@ type StoreInfo struct {
 func NewStoreInfo(store *metapb.Store) *StoreInfo {
 	return &StoreInfo{
 		Store:        store,
+		Stats:        &pdpb.StoreStats{},
 		LeaderWeight: 1.0,
 		RegionWeight: 1.0,
 	}
@@ -103,15 +104,46 @@ func (s *StoreInfo) DownTime() time.Duration {
 }
 
 const minWeight = 1e-6
+const maxScore = 1024 * 1024 * 1024
 
-// LeaderScore returns the store's leader score: leaderCount / leaderWeight.
-func (s *StoreInfo) LeaderScore() float64 {
-	return float64(s.LeaderSize) / math.Max(s.LeaderWeight, minWeight)
+// LeaderScore returns the store's leader score: leaderSize / leaderWeight.
+func (s *StoreInfo) LeaderScore(delta int64) float64 {
+	return float64(s.LeaderSize+delta) / math.Max(s.LeaderWeight, minWeight)
 }
 
-// RegionScore returns the store's region score: regionSize / regionWeight.
-func (s *StoreInfo) RegionScore() float64 {
-	return float64(s.RegionSize) / math.Max(s.RegionWeight, minWeight)
+// RegionScore returns the store's region score.
+func (s *StoreInfo) RegionScore(highSpaceRatio, lowSpaceRatio float64, delta int64) float64 {
+	if s.RegionSize == 0 {
+		return float64(delta)
+	}
+
+	capacity := float64(s.Stats.GetCapacity()) / (1 << 20)
+	available := float64(s.Stats.GetAvailable()) / (1 << 20)
+
+	var score float64
+
+	// because of rocksdb compression, region size is larger than actual used size
+	amplification := float64(s.RegionSize) / (float64(s.Stats.GetUsedSize()) / (1 << 20))
+
+	if available-float64(delta)/amplification >= (1-highSpaceRatio)*capacity {
+		score = float64(s.RegionSize + delta)
+	} else if available-float64(delta)/amplification <= (1-lowSpaceRatio)*capacity {
+		score = maxScore - (available - float64(delta)/amplification)
+	} else {
+		// to make the score function continuous, we use linear function y = k * x + b as transition period
+		// from above we know that there are two points must on the function image
+		// p1(highSpaceRatio*capacity*amplification, highSpaceRatio*capacity*amplification) and
+		// p2(lowSpaceRatio*capacity*amplification, maxScore-(1-lowSpaceRatio)*capacity)
+		// so k = (y2 - y1) / (x2 - x1)
+		x1, y1 := highSpaceRatio*capacity*amplification, highSpaceRatio*capacity*amplification
+		x2, y2 := lowSpaceRatio*capacity*amplification, maxScore-(1-lowSpaceRatio)*capacity
+
+		k := (y2 - y1) / (x2 - x1)
+		b := y1 - k*x1
+		score = k*float64(s.RegionSize+delta) + b
+	}
+
+	return score / math.Max(s.RegionWeight, minWeight)
 }
 
 // StorageSize returns store's used storage size reported from tikv.
@@ -127,11 +159,9 @@ func (s *StoreInfo) AvailableRatio() float64 {
 	return float64(s.Stats.GetAvailable()) / float64(s.Stats.GetCapacity())
 }
 
-const storeLowSpaceThreshold = 0.2
-
 // IsLowSpace checks if the store is lack of space.
-func (s *StoreInfo) IsLowSpace() bool {
-	return s.Stats != nil && s.AvailableRatio() < storeLowSpaceThreshold
+func (s *StoreInfo) IsLowSpace(lowSpaceRatio float64) bool {
+	return s.Stats != nil && s.AvailableRatio() < 1-lowSpaceRatio
 }
 
 // ResourceCount reutrns count of leader/region in the store.
@@ -159,12 +189,12 @@ func (s *StoreInfo) ResourceSize(kind ResourceKind) int64 {
 }
 
 // ResourceScore reutrns score of leader/region in the store.
-func (s *StoreInfo) ResourceScore(kind ResourceKind) float64 {
+func (s *StoreInfo) ResourceScore(kind ResourceKind, highSpaceRatio, lowSpaceRatio float64, delta int64) float64 {
 	switch kind {
 	case LeaderKind:
-		return s.LeaderScore()
+		return s.LeaderScore(delta)
 	case RegionKind:
-		return s.RegionScore()
+		return s.RegionScore(highSpaceRatio, lowSpaceRatio, delta)
 	default:
 		return 0
 	}