*: add hot region schedule

server/api/hot_status.go

@@ -0,0 +1,41 @@
+// Copyright 2017 PingCAP, Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//	   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package api
+
+import (
+	"net/http"
+
+	"github.com/pingcap/pd/server"
+	"github.com/unrolled/render"
+)
+
+type hotStatusHandler struct {
+	svr *server.Server
+	rd  *render.Render
+}
+
+func newHotStatusHandler(svr *server.Server, rd *render.Render) *hotStatusHandler {
+	return &hotStatusHandler{
+		svr: svr,
+		rd:  rd,
+	}
+}
+
+func (h *hotStatusHandler) GetHotRegions(w http.ResponseWriter, r *http.Request) {
+	h.rd.JSON(w, http.StatusOK, h.svr.GetHotRegions())
+}
+
+func (h *hotStatusHandler) GetHotStores(w http.ResponseWriter, r *http.Request) {
+	h.rd.JSON(w, http.StatusOK, h.svr.GetHotStores())
+}

server/api/router.go

@@ -60,6 +60,10 @@ func createRouter(prefix string, svr *server.Server) *mux.Router {
 	router.HandleFunc("/api/v1/labels", labelsHandler.Get).Methods("GET")
 	router.HandleFunc("/api/v1/labels/stores", labelsHandler.GetStores).Methods("GET")
 
+	hotStatusHandler := newHotStatusHandler(svr, rd)
+	router.HandleFunc("/api/v1/hotspot/regions", hotStatusHandler.GetHotRegions).Methods("GET")
+	router.HandleFunc("/api/v1/hotspot/stores", hotStatusHandler.GetHotStores).Methods("GET")
+
 	router.Handle("/api/v1/events", newEventsHandler(svr, rd)).Methods("GET")
 	router.Handle("/api/v1/feed", newFeedHandler(svr, rd)).Methods("GET")
 

server/balancer.go

@@ -15,6 +15,8 @@ package server
 
 import (
 	"math"
+	"math/rand"
+	"sort"
 	"time"
 
 	"github.com/montanaflynn/stats"
@@ -370,3 +372,262 @@ func (r *replicaChecker) checkBestReplacement(region *RegionInfo) Operator {
 	}
 	return newTransferPeer(region, oldPeer, newPeer)
 }
+
+// RegionStat records each hot region's statistics
+type RegionStat struct {
+	RegionID       uint64    `json:"region_id"`
+	WrittenBytes   uint64    `json:"written_bytes"`
+	HotDegree      int       `json:"hot_degree"`
+	LastUpdateTime time.Time `json:"last_update_time"`
+	StoreID        uint64    `json:"-"`
+	antiCount      int
+	version        uint64
+}
+
+// ListRegionsStat is a list of a group region state type
+type ListRegionsStat []RegionStat
+
+func (m ListRegionsStat) Len() int           { return len(m) }
+func (m ListRegionsStat) Swap(i, j int)      { m[i], m[j] = m[j], m[i] }
+func (m ListRegionsStat) Less(i, j int) bool { return m[i].WrittenBytes < m[j].WrittenBytes }
+
+// StoreHotRegions records all hot regions in one store with sequence
+type StoreHotRegions struct {
+	TotalWrittenBytes uint64          `json:"total_written"`
+	RegionCount       int             `json:"region_count"`
+	RegionsStat       ListRegionsStat `json:"status"`
+}
+
+type balanceHotRegionScheduler struct {
+	opt         *scheduleOption
+	limit       uint64
+	scoreStatus map[uint64]*StoreHotRegions // store id -> regions status in this store
+	r           *rand.Rand
+}
+
+func newBalanceHotRegionScheduler(opt *scheduleOption) *balanceHotRegionScheduler {
+	return &balanceHotRegionScheduler{
+		opt:         opt,
+		limit:       1,
+		scoreStatus: make(map[uint64]*StoreHotRegions),
+		r:           rand.New(rand.NewSource(time.Now().UnixNano())),
+	}
+}
+
+func (l *balanceHotRegionScheduler) GetName() string {
+	return "balance-hot-region-scheduler"
+}
+
+func (l *balanceHotRegionScheduler) GetResourceKind() ResourceKind {
+	return priorityKind
+}
+
+func (l *balanceHotRegionScheduler) GetResourceLimit() uint64 {
+	return l.limit
+}
+
+func (l *balanceHotRegionScheduler) Prepare(cluster *clusterInfo) error { return nil }
+
+func (l *balanceHotRegionScheduler) Cleanup(cluster *clusterInfo) {}
+
+func (l *balanceHotRegionScheduler) Schedule(cluster *clusterInfo) Operator {
+	l.calculateScore(cluster)
+	region := l.SelectSourceRegion(cluster)
+	if region == nil {
+		return nil
+	}
+	newLeader := l.selectTransferLeader(region)
+	if newLeader != nil {
+		return newPriorityTransferLeader(region, newLeader)
+	}
+	peer := l.selectTransferPeer(region, cluster)
+	if peer != nil {
+		return newPriorityTransferPeer(region, region.Leader, peer)
+	}
+	return nil
+}
+
+func (l *balanceHotRegionScheduler) clearScore() {
+	l.scoreStatus = make(map[uint64]*StoreHotRegions)
+}
+func (l *balanceHotRegionScheduler) calculateScore(cluster *clusterInfo) {
+	l.clearScore()
+	items := cluster.writeStatistics.elems()
+	for _, item := range items {
+		r, ok := item.value.(RegionStat)
+		if !ok {
+			continue
+		}
+		if r.HotDegree < hotRegionLowThreshold {
+			continue
+		}
+
+		regionInfo := cluster.getRegion(r.RegionID)
+		storeID := regionInfo.Leader.GetStoreId()
+		status, ok := l.scoreStatus[storeID]
+		if !ok {
+			status = &StoreHotRegions{
+				RegionsStat: make(ListRegionsStat, 0, 100),
+			}
+			l.scoreStatus[storeID] = status
+		}
+		status.TotalWrittenBytes += r.WrittenBytes
+		status.RegionsStat = append(status.RegionsStat, RegionStat{
+			RegionID:       r.RegionID,
+			WrittenBytes:   r.WrittenBytes,
+			HotDegree:      r.HotDegree,
+			LastUpdateTime: r.LastUpdateTime,
+			StoreID:        storeID,
+			antiCount:      r.antiCount,
+			version:        r.version,
+		})
+		status.RegionCount++
+	}
+
+	for _, rs := range l.scoreStatus {
+		sort.Sort(sort.Reverse(rs.RegionsStat))
+	}
+}
+
+func (l *balanceHotRegionScheduler) SelectSourceRegion(cluster *clusterInfo) *RegionInfo {
+	var (
+		maxWritten             uint64
+		sourceStore            uint64
+		maxHotStoreRegionCount int
+	)
+	// choose a hot store as transfer source
+	// the numbers of the hot regions in that store has higher priority than TotalWrittenBytes
+	for sid, s := range l.scoreStatus {
+		if s.RegionsStat.Len() < 2 {
+			continue
+		}
+
+		if maxHotStoreRegionCount < s.RegionsStat.Len() {
+			maxHotStoreRegionCount = s.RegionsStat.Len()
+			maxWritten = s.TotalWrittenBytes
+			sourceStore = sid
+			continue
+		}
+
+		if maxHotStoreRegionCount == s.RegionsStat.Len() && maxWritten < s.TotalWrittenBytes {
+			maxWritten = s.TotalWrittenBytes
+			sourceStore = sid
+		}
+	}
+
+	if sourceStore == 0 {
+		return nil
+	}
+
+	length := l.scoreStatus[sourceStore].RegionsStat.Len()
+	// the hottest region in the store not move
+	// radmonly pick a region from 1 .. length-1
+	// TODO: consider hot degree when pick
+	rr := l.r.Int31n(int32(length-1)) + 1
+	pickedRegionStat := l.scoreStatus[sourceStore].RegionsStat[rr]
+	if pickedRegionStat.antiCount < hotRegionAntiCount {
+		return nil
+	}
+	sourceRegion := cluster.getRegion(pickedRegionStat.RegionID)
+	if len(sourceRegion.DownPeers) != 0 || len(sourceRegion.PendingPeers) != 0 {
+		return nil
+	}
+	// use written bytes per second
+	sourceRegion.WrittenBytes = pickedRegionStat.WrittenBytes
+	l.adjustBalanceLimit(sourceStore)
+	return sourceRegion
+}
+
+func (l *balanceHotRegionScheduler) adjustBalanceLimit(storeID uint64) {
+	s := l.scoreStatus[storeID]
+	var hotRegionTotalCount float64
+	for _, m := range l.scoreStatus {
+		hotRegionTotalCount += float64(m.RegionsStat.Len())
+	}
+
+	avgRegionCount := hotRegionTotalCount / float64(len(l.scoreStatus))
+	// Multiplied by 0.75 to avoid transfer back and forth
+	limit := uint64((float64(s.RegionsStat.Len()) - avgRegionCount) * 0.75)
+	l.limit = maxUint64(1, limit)
+}
+
+func (l *balanceHotRegionScheduler) GetStatus() map[uint64]*StoreHotRegions {
+	return l.scoreStatus
+}
+
+func (l *balanceHotRegionScheduler) selectTransferLeader(sourceRegion *RegionInfo) *metapb.Peer {
+	followPeers := sourceRegion.GetFollowers()
+	storeIDs := make([]uint64, 0, len(followPeers))
+	for _, peer := range followPeers {
+		storeIDs = append(storeIDs, peer.GetStoreId())
+	}
+
+	targetStoreID := l.selectBestStore(storeIDs, sourceRegion)
+	var targetPeer *metapb.Peer
+	for _, peer := range followPeers {
+		if peer.GetStoreId() == targetStoreID {
+			targetPeer = peer
+		}
+	}
+	return targetPeer
+}
+
+func (l *balanceHotRegionScheduler) selectTransferPeer(sourceRegion *RegionInfo, cluster *clusterInfo) *metapb.Peer {
+	filter := newExcludedFilter(sourceRegion.GetStoreIds(), sourceRegion.GetStoreIds())
+	stores := cluster.getStores()
+
+	storeIDs := make([]uint64, 0, len(stores))
+	for _, store := range stores {
+		if filter.FilterTarget(store) {
+			continue
+		}
+		storeIDs = append(storeIDs, store.GetId())
+	}
+	targetStoreID := l.selectBestStore(storeIDs, sourceRegion)
+	targetStore := cluster.getStore(targetStoreID)
+	if targetStore == nil {
+		return nil
+	}
+	newPeer, err := cluster.allocPeer(targetStore.GetId())
+	if err != nil {
+		log.Errorf("failed to allocate peer: %v", err)
+		return nil
+	}
+
+	return newPeer
+}
+
+// select a store to transfer peer
+// preferred to the store that with the least number of regions
+// and then choose the least total written bytes store
+func (l *balanceHotRegionScheduler) selectBestStore(stores []uint64, sourceRegion *RegionInfo) uint64 {
+	sr := l.scoreStatus[sourceRegion.Leader.GetStoreId()]
+	sourceStoreWrittenBytes := sr.TotalWrittenBytes
+	sourceStoreHotRegionCount := sr.RegionsStat.Len()
+
+	var (
+		targetStore     uint64
+		minWrittenBytes uint64 = math.MaxUint64
+	)
+	minRegionCount := int(math.MaxInt32)
+	for _, store := range stores {
+		if s, ok := l.scoreStatus[store]; ok {
+			if sourceStoreHotRegionCount-s.RegionsStat.Len() > 1 && minRegionCount > s.RegionsStat.Len() {
+				targetStore = store
+				minWrittenBytes = s.TotalWrittenBytes
+				minRegionCount = s.RegionsStat.Len()
+				continue
+			}
+			if minRegionCount == s.RegionsStat.Len() && minWrittenBytes > s.TotalWrittenBytes &&
+				uint64(float64(sourceStoreWrittenBytes)*hotRegionScheduleFactor) > s.TotalWrittenBytes+2*sourceRegion.WrittenBytes {
+				minWrittenBytes = s.TotalWrittenBytes
+				targetStore = store
+			}
+
+		} else {
+			targetStore = store
+			break
+		}
+	}
+	return targetStore
+}