// findCellsForRebuild will find all the cells in the given keyspace
// and create an entry if the map for them
func (wr *Wrangler) findCellsForRebuild(ki *topo.KeyspaceInfo, keyspace string, shards []string, cells []string, srvKeyspaceMap map[string]*topo.SrvKeyspace) error {
	er := concurrency.FirstErrorRecorder{}
	mu := sync.Mutex{}
	wg := sync.WaitGroup{}
	for _, shard := range shards {
		wg.Add(1)
		go func(shard string) {
			if si, err := wr.ts.GetShard(keyspace, shard); err != nil {
				er.RecordError(fmt.Errorf("GetShard(%v,%v) failed: %v", keyspace, shard, err))
			} else {
				mu.Lock()
				for _, cell := range si.Cells {
					if !topo.InCellList(cell, cells) {
						continue
					}
					if _, ok := srvKeyspaceMap[cell]; !ok {
						srvKeyspaceMap[cell] = &topo.SrvKeyspace{
							Shards:             make([]topo.SrvShard, 0, 16),
							ShardingColumnName: ki.ShardingColumnName,
							ShardingColumnType: ki.ShardingColumnType,
							ServedFrom:         ki.ServedFrom,
						}
					}
				}
				mu.Unlock()
			}
			wg.Done()
		}(shard)
	}
	wg.Wait()
	return er.Error()
}

// RebuildShard updates the SrvShard objects and underlying serving graph.
//
// Re-read from TopologyServer to make sure we are using the side
// effects of all actions.
//
// This function will start each cell over from the beginning on ErrBadVersion,
// so it doesn't need a lock on the shard.
func RebuildShard(ctx context.Context, log logutil.Logger, ts topo.Server, keyspace, shard string, cells []string, lockTimeout time.Duration) (*topo.ShardInfo, error) {
	log.Infof("RebuildShard %v/%v", keyspace, shard)

	span := trace.NewSpanFromContext(ctx)
	span.StartLocal("topotools.RebuildShard")
	defer span.Finish()
	ctx = trace.NewContext(ctx, span)

	// read the existing shard info. It has to exist.
	shardInfo, err := ts.GetShard(ctx, keyspace, shard)
	if err != nil {
		return nil, err
	}

	// rebuild all cells in parallel
	wg := sync.WaitGroup{}
	rec := concurrency.AllErrorRecorder{}
	for _, cell := range shardInfo.Cells {
		// skip this cell if we shouldn't rebuild it
		if !topo.InCellList(cell, cells) {
			continue
		}

		wg.Add(1)
		go func(cell string) {
			defer wg.Done()
			rec.RecordError(rebuildCellSrvShard(ctx, log, ts, shardInfo, cell))
		}(cell)
	}
	wg.Wait()

	return shardInfo, rec.Error()
}

func (wr *Wrangler) removeShardCell(keyspace, shard, cell string, force bool) error {
	shardInfo, err := wr.ts.GetShardCritical(keyspace, shard)
	if err != nil {
		return err
	}

	// check the cell is in the list already
	if !topo.InCellList(cell, shardInfo.Cells) {
		return fmt.Errorf("cell %v in not in shard info", cell)
	}

	// check the master alias is not in the cell
	if shardInfo.MasterAlias.Cell == cell {
		return fmt.Errorf("master %v is in the cell '%v' we want to remove", shardInfo.MasterAlias, cell)
	}

	// get the ShardReplication object in the cell
	sri, err := wr.ts.GetShardReplication(cell, keyspace, shard)
	switch err {
	case nil:
		if len(sri.ReplicationLinks) > 0 {
			return fmt.Errorf("cell %v has %v possible tablets in replication graph", cell, len(sri.ReplicationLinks))
		}

		// ShardReplication object is now useless, remove it
		if err := wr.ts.DeleteShardReplication(cell, keyspace, shard); err != nil {
			return fmt.Errorf("error deleting ShardReplication object in cell %v: %v", cell, err)
		}

		// we keep going
	case topo.ErrNoNode:
		// no ShardReplication object, we keep going
	default:
		// we can't get the object, assume topo server is down there,
		// so we look at force flag
		if !force {
			return err
		}
		log.Warningf("Cannot get ShardReplication from cell %v, assuming cell topo server is down, and forcing the removal", cell)
	}

	// now we can update the shard
	log.Infof("Removing cell %v from shard %v/%v", cell, keyspace, shard)
	newCells := make([]string, 0, len(shardInfo.Cells)-1)
	for _, c := range shardInfo.Cells {
		if c != cell {
			newCells = append(newCells, c)
		}
	}
	shardInfo.Cells = newCells

	return wr.ts.UpdateShard(shardInfo)
}

// findCellsForRebuild will find all the cells in the given keyspace
// and create an entry if the map for them
func findCellsForRebuild(ki *topo.KeyspaceInfo, shardMap map[string]*topo.ShardInfo, cells []string, srvKeyspaceMap map[string]*topodatapb.SrvKeyspace) {
	for _, si := range shardMap {
		for _, cell := range si.Cells {
			if !topo.InCellList(cell, cells) {
				continue
			}
			if _, ok := srvKeyspaceMap[cell]; !ok {
				srvKeyspaceMap[cell] = &topodatapb.SrvKeyspace{
					ShardingColumnName: ki.ShardingColumnName,
					ShardingColumnType: ki.ShardingColumnType,
					ServedFrom:         ki.ComputeCellServedFrom(cell),
				}
			}
		}
	}
}

// findCellsForRebuild will find all the cells in the given keyspace
// and create an entry if the map for them
func (wr *Wrangler) findCellsForRebuild(ki *topo.KeyspaceInfo, shardMap map[string]*topo.ShardInfo, cells []string, srvKeyspaceMap map[string]*topo.SrvKeyspace) {
	for _, si := range shardMap {
		for _, cell := range si.Cells {
			if !topo.InCellList(cell, cells) {
				continue
			}
			if _, ok := srvKeyspaceMap[cell]; !ok {
				srvKeyspaceMap[cell] = &topo.SrvKeyspace{
					ShardingColumnName: ki.ShardingColumnName,
					ShardingColumnType: key.ProtoToKeyspaceIdType(ki.ShardingColumnType),
					ServedFrom:         ki.ComputeCellServedFrom(cell),
					SplitShardCount:    ki.SplitShardCount,
				}
			}
		}
	}
}

// changeCallback is run after every action that might
// have changed something in the tablet record.
func (agent *ActionAgent) changeCallback(oldTablet, newTablet *topo.Tablet) error {
	allowQuery := newTablet.IsRunningQueryService()

	// Read the shard to get SourceShards / TabletControlMap if
	// we're going to use it.
	var shardInfo *topo.ShardInfo
	var tabletControl *topo.TabletControl
	var blacklistedTables []string
	var err error
	if allowQuery {
		shardInfo, err = agent.TopoServer.GetShard(newTablet.Keyspace, newTablet.Shard)
		if err != nil {
			log.Errorf("Cannot read shard for this tablet %v, might have inaccurate SourceShards and TabletControls: %v", newTablet.Alias, err)
		} else {
			if newTablet.Type == topo.TYPE_MASTER {
				allowQuery = len(shardInfo.SourceShards) == 0
			}
			if tc, ok := shardInfo.TabletControlMap[newTablet.Type]; ok {
				if topo.InCellList(newTablet.Alias.Cell, tc.Cells) {
					if tc.DisableQueryService {
						allowQuery = false
					}
					blacklistedTables = tc.BlacklistedTables
					tabletControl = tc
				}
			}
		}
	}

	// Read the keyspace on masters to get ShardingColumnType,
	// for binlog replication, only if source shards are set.
	var keyspaceInfo *topo.KeyspaceInfo
	if newTablet.Type == topo.TYPE_MASTER && shardInfo != nil && len(shardInfo.SourceShards) > 0 {
		keyspaceInfo, err = agent.TopoServer.GetKeyspace(newTablet.Keyspace)
		if err != nil {
			log.Errorf("Cannot read keyspace for this tablet %v: %v", newTablet.Alias, err)
			keyspaceInfo = nil
		}
	}

	if allowQuery {
		// There are a few transitions when we're
		// going to need to restart the query service:
		// - transitioning from replica to master, so clients
		//   that were already connected don't keep on using
		//   the master as replica or rdonly.
		// - having different parameters for the query
		//   service. It needs to stop and restart with the
		//   new parameters. That includes:
		//   - changing KeyRange
		//   - changing the BlacklistedTables list
		if (newTablet.Type == topo.TYPE_MASTER &&
			oldTablet.Type != topo.TYPE_MASTER) ||
			(newTablet.KeyRange != oldTablet.KeyRange) ||
			!reflect.DeepEqual(blacklistedTables, agent.BlacklistedTables()) {
			agent.disallowQueries()
		}
		if err := agent.allowQueries(newTablet, blacklistedTables); err != nil {
			log.Errorf("Cannot start query service: %v", err)
		}
	} else {
		agent.disallowQueries()
	}

	// save the tabletControl we've been using, so the background
	// healthcheck makes the same decisions as we've been making.
	agent.setTabletControl(tabletControl)

	// update stream needs to be started or stopped too
	if agent.DBConfigs != nil {
		if topo.IsRunningUpdateStream(newTablet.Type) {
			binlog.EnableUpdateStreamService(agent.DBConfigs.App.DbName, agent.Mysqld)
		} else {
			binlog.DisableUpdateStreamService()
		}
	}

	statsType.Set(string(newTablet.Type))
	statsKeyspace.Set(newTablet.Keyspace)
	statsShard.Set(newTablet.Shard)
	statsKeyRangeStart.Set(string(newTablet.KeyRange.Start.Hex()))
	statsKeyRangeEnd.Set(string(newTablet.KeyRange.End.Hex()))

	// See if we need to start or stop any binlog player
	if agent.BinlogPlayerMap != nil {
		if newTablet.Type == topo.TYPE_MASTER {
			agent.BinlogPlayerMap.RefreshMap(newTablet, keyspaceInfo, shardInfo)
		} else {
			agent.BinlogPlayerMap.StopAllPlayersAndReset()
		}
	}
	return nil
}

func (wr *Wrangler) removeShardCell(ctx context.Context, keyspace, shard, cell string, force, recursive bool) error {
	shardInfo, err := wr.ts.GetShard(ctx, keyspace, shard)
	if err != nil {
		return err
	}

	// check the cell is in the list already
	if !topo.InCellList(cell, shardInfo.Cells) {
		return fmt.Errorf("cell %v in not in shard info", cell)
	}

	// check the master alias is not in the cell
	if shardInfo.MasterAlias.Cell == cell {
		return fmt.Errorf("master %v is in the cell '%v' we want to remove", shardInfo.MasterAlias, cell)
	}

	// get the ShardReplication object in the cell
	sri, err := wr.ts.GetShardReplication(ctx, cell, keyspace, shard)
	switch err {
	case nil:
		if recursive {
			wr.Logger().Infof("Deleting all tablets in shard %v/%v", keyspace, shard)
			for _, node := range sri.Nodes {
				// We don't care about scrapping or updating the replication graph,
				// because we're about to delete the entire replication graph.
				wr.Logger().Infof("Deleting tablet %v", node.TabletAlias)
				if err := wr.TopoServer().DeleteTablet(ctx, topo.ProtoToTabletAlias(node.TabletAlias)); err != nil && err != topo.ErrNoNode {
					return fmt.Errorf("can't delete tablet %v: %v", node.TabletAlias, err)
				}
			}
		} else if len(sri.Nodes) > 0 {
			return fmt.Errorf("cell %v has %v possible tablets in replication graph", cell, len(sri.Nodes))
		}

		// ShardReplication object is now useless, remove it
		if err := wr.ts.DeleteShardReplication(ctx, cell, keyspace, shard); err != nil && err != topo.ErrNoNode {
			return fmt.Errorf("error deleting ShardReplication object in cell %v: %v", cell, err)
		}

		// Rebuild the shard serving graph to reflect the tablets we deleted.
		// This must be done before removing the cell from the global shard record,
		// since this cell will be skipped by all future rebuilds.
		if _, err := wr.RebuildShardGraph(ctx, keyspace, shard, []string{cell}); err != nil {
			return fmt.Errorf("can't rebuild serving graph for shard %v/%v in cell %v: %v", keyspace, shard, cell, err)
		}

		// we keep going
	case topo.ErrNoNode:
		// no ShardReplication object, we keep going
	default:
		// we can't get the object, assume topo server is down there,
		// so we look at force flag
		if !force {
			return err
		}
		wr.Logger().Warningf("Cannot get ShardReplication from cell %v, assuming cell topo server is down, and forcing the removal", cell)
	}

	// now we can update the shard
	wr.Logger().Infof("Removing cell %v from shard %v/%v", cell, keyspace, shard)
	newCells := make([]string, 0, len(shardInfo.Cells)-1)
	for _, c := range shardInfo.Cells {
		if c != cell {
			newCells = append(newCells, c)
		}
	}
	shardInfo.Cells = newCells

	return topo.UpdateShard(ctx, wr.ts, shardInfo)
}

// Write serving graph data to the cells
func (wr *Wrangler) rebuildShardSrvGraph(shardInfo *topo.ShardInfo, tablets []*topo.TabletInfo, cells []string) error {
	log.Infof("rebuildShardSrvGraph %v/%v", shardInfo.Keyspace(), shardInfo.ShardName())

	// Get all existing db types so they can be removed if nothing
	// had been editted.  This applies to all cells, which can't
	// be determined until you walk through all the tablets.
	//
	// existingDbTypeLocations is a map:
	//   key: {cell,keyspace,shard,tabletType}
	//   value: true
	existingDbTypeLocations := make(map[cellKeyspaceShardType]bool)

	// Update db type addresses in the serving graph
	//
	// locationAddrsMap is a map:
	//   key: {cell,keyspace,shard,tabletType}
	//   value: topo.EndPoints (list of server records)
	locationAddrsMap := make(map[cellKeyspaceShardType]*topo.EndPoints)

	// we keep track of the existingDbTypeLocations we've already looked at
	knownShardLocations := make(map[cellKeyspaceShard]bool)

	for _, tablet := range tablets {
		// only look at tablets in the cells we want to rebuild
		if !topo.InCellList(tablet.Tablet.Alias.Cell, cells) {
			continue
		}

		// this is {cell,keyspace,shard}
		// we'll get the children to find the existing types
		shardLocation := cellKeyspaceShard{tablet.Tablet.Alias.Cell, tablet.Tablet.Keyspace, tablet.Shard}
		// only need to do this once per cell
		if !knownShardLocations[shardLocation] {
			log.Infof("Getting tablet types on cell %v for %v/%v", tablet.Tablet.Alias.Cell, tablet.Tablet.Keyspace, tablet.Shard)
			tabletTypes, err := wr.ts.GetSrvTabletTypesPerShard(tablet.Tablet.Alias.Cell, tablet.Tablet.Keyspace, tablet.Shard)
			if err != nil {
				if err != topo.ErrNoNode {
					return err
				}
			} else {
				for _, tabletType := range tabletTypes {
					existingDbTypeLocations[cellKeyspaceShardType{tablet.Tablet.Alias.Cell, tablet.Tablet.Keyspace, tablet.Shard, tabletType}] = true
				}
			}
			knownShardLocations[shardLocation] = true
		}

		// Check IsInServingGraph after we have populated
		// existingDbTypeLocations so we properly prune data
		// if the definition of serving type changes.
		if !tablet.IsInServingGraph() {
			continue
		}

		location := cellKeyspaceShardType{tablet.Tablet.Alias.Cell, tablet.Keyspace, tablet.Shard, tablet.Type}
		addrs, ok := locationAddrsMap[location]
		if !ok {
			addrs = topo.NewEndPoints()
			locationAddrsMap[location] = addrs
		}

		entry, err := tabletmanager.EndPointForTablet(tablet.Tablet)
		if err != nil {
			log.Warningf("EndPointForTablet failed for tablet %v: %v", tablet.Alias, err)
			continue
		}
		addrs.Entries = append(addrs.Entries, *entry)
	}

	// we're gonna parallelize a lot here
	rec := concurrency.AllErrorRecorder{}
	wg := sync.WaitGroup{}

	// write all the {cell,keyspace,shard,type}
	// nodes everywhere we want them
	for location, addrs := range locationAddrsMap {
		wg.Add(1)
		go func(location cellKeyspaceShardType, addrs *topo.EndPoints) {
			log.Infof("saving serving graph for cell %v shard %v/%v tabletType %v", location.cell, location.keyspace, location.shard, location.tabletType)
			if err := wr.ts.UpdateEndPoints(location.cell, location.keyspace, location.shard, location.tabletType, addrs); err != nil {
				rec.RecordError(fmt.Errorf("writing endpoints for cell %v shard %v/%v tabletType %v failed: %v", location.cell, location.keyspace, location.shard, location.tabletType, err))
			}
			wg.Done()
		}(location, addrs)
	}

	// Delete any pre-existing paths that were not updated by this process.
	// That's the existingDbTypeLocations - locationAddrsMap
	for dbTypeLocation := range existingDbTypeLocations {
		if _, ok := locationAddrsMap[dbTypeLocation]; !ok {
			cell := dbTypeLocation.cell
			if !topo.InCellList(cell, cells) {
				continue
			}

			wg.Add(1)
			go func(dbTypeLocation cellKeyspaceShardType) {
				log.Infof("removing stale db type from serving graph: %v", dbTypeLocation)
				if err := wr.ts.DeleteSrvTabletType(dbTypeLocation.cell, dbTypeLocation.keyspace, dbTypeLocation.shard, dbTypeLocation.tabletType); err != nil {
//.........这里部分代码省略.........

// changeCallback is run after every action that might
// have changed something in the tablet record.
func (agent *ActionAgent) changeCallback(ctx context.Context, oldTablet, newTablet *pbt.Tablet) error {
	span := trace.NewSpanFromContext(ctx)
	span.StartLocal("ActionAgent.changeCallback")
	defer span.Finish()

	allowQuery := topo.IsRunningQueryService(newTablet.Type)

	// Read the shard to get SourceShards / TabletControlMap if
	// we're going to use it.
	var shardInfo *topo.ShardInfo
	var tabletControl *pbt.Shard_TabletControl
	var blacklistedTables []string
	var err error
	var disallowQueryReason string
	if allowQuery {
		shardInfo, err = agent.TopoServer.GetShard(ctx, newTablet.Keyspace, newTablet.Shard)
		if err != nil {
			log.Errorf("Cannot read shard for this tablet %v, might have inaccurate SourceShards and TabletControls: %v", newTablet.Alias, err)
		} else {
			if newTablet.Type == pbt.TabletType_MASTER {
				if len(shardInfo.SourceShards) > 0 {
					allowQuery = false
					disallowQueryReason = "old master is still in shard info"
				}
			}
			if tc := shardInfo.GetTabletControl(newTablet.Type); tc != nil {
				if topo.InCellList(newTablet.Alias.Cell, tc.Cells) {
					if tc.DisableQueryService {
						allowQuery = false
						disallowQueryReason = "query service disabled by tablet control"
					}
					blacklistedTables = tc.BlacklistedTables
					tabletControl = tc
				}
			}
		}
	} else {
		disallowQueryReason = fmt.Sprintf("not a serving tablet type(%v)", newTablet.Type)
	}

	// Read the keyspace on masters to get ShardingColumnType,
	// for binlog replication, only if source shards are set.
	var keyspaceInfo *topo.KeyspaceInfo
	if newTablet.Type == pbt.TabletType_MASTER && shardInfo != nil && len(shardInfo.SourceShards) > 0 {
		keyspaceInfo, err = agent.TopoServer.GetKeyspace(ctx, newTablet.Keyspace)
		if err != nil {
			log.Errorf("Cannot read keyspace for this tablet %v: %v", newTablet.Alias, err)
			keyspaceInfo = nil
		}
	}

	if allowQuery {
		// There are a few transitions when we need to restart the query service:
		switch {
		// If either InitMaster or InitSlave was called, because those calls
		// (or a prior call to ResetReplication) may have silently broken the
		// rowcache invalidator by executing RESET MASTER.
		// Note that we don't care about fixing it after ResetReplication itself
		// since that call breaks everything on purpose, and we don't expect
		// anything to start working until either InitMaster or InitSlave.
		case agent.initReplication:
			agent.initReplication = false
			agent.stopQueryService("initialize replication")

		// Transitioning from replica to master, so clients that were already
		// connected don't keep on using the master as replica or rdonly.
		case newTablet.Type == pbt.TabletType_MASTER && oldTablet.Type != pbt.TabletType_MASTER:
			agent.stopQueryService("tablet promoted to master")

		// Having different parameters for the query service.
		// It needs to stop and restart with the new parameters.
		// That includes:
		//   - changing KeyRange
		//   - changing the BlacklistedTables list
		case (newTablet.KeyRange != oldTablet.KeyRange),
			!reflect.DeepEqual(blacklistedTables, agent.BlacklistedTables()):
			agent.stopQueryService("keyrange/blacklistedtables changed")
		}

		if err := agent.allowQueries(newTablet, blacklistedTables); err != nil {
			log.Errorf("Cannot start query service: %v", err)
		}
	} else {
		agent.stopQueryService(disallowQueryReason)
	}

	// save the tabletControl we've been using, so the background
	// healthcheck makes the same decisions as we've been making.
	agent.setTabletControl(tabletControl)

	// update stream needs to be started or stopped too
	if agent.DBConfigs != nil {
		if topo.IsRunningUpdateStream(newTablet.Type) {
			binlog.EnableUpdateStreamService(agent.DBConfigs.App.DbName, agent.MysqlDaemon)
		} else {
			binlog.DisableUpdateStreamService()
		}
	}
//.........这里部分代码省略.........

// changeCallback is run after every action that might
// have changed something in the tablet record or in the topology.
//
// It owns making changes to the BinlogPlayerMap. The input for this is the
// tablet type (has to be master), and the shard's SourceShards.
//
// It owns updating the blacklisted tables.
//
// It owns updating the stats record for 'TabletType'.
//
// It owns starting and stopping the update stream service.
//
// It owns reading the TabletControl for the current tablet, and storing it.
func (agent *ActionAgent) changeCallback(ctx context.Context, oldTablet, newTablet *topodatapb.Tablet) {
	span := trace.NewSpanFromContext(ctx)
	span.StartLocal("ActionAgent.changeCallback")
	defer span.Finish()

	allowQuery := topo.IsRunningQueryService(newTablet.Type)
	broadcastHealth := false
	runUpdateStream := allowQuery

	// Read the shard to get SourceShards / TabletControlMap if
	// we're going to use it.
	var shardInfo *topo.ShardInfo
	var err error
	var disallowQueryReason string
	var blacklistedTables []string
	updateBlacklistedTables := true
	if allowQuery {
		shardInfo, err = agent.TopoServer.GetShard(ctx, newTablet.Keyspace, newTablet.Shard)
		if err != nil {
			log.Errorf("Cannot read shard for this tablet %v, might have inaccurate SourceShards and TabletControls: %v", newTablet.Alias, err)
			updateBlacklistedTables = false
		} else {
			if newTablet.Type == topodatapb.TabletType_MASTER {
				if len(shardInfo.SourceShards) > 0 {
					allowQuery = false
					disallowQueryReason = "master tablet with filtered replication on"
				}
			}
			if tc := shardInfo.GetTabletControl(newTablet.Type); tc != nil {
				if topo.InCellList(newTablet.Alias.Cell, tc.Cells) {
					if tc.DisableQueryService {
						allowQuery = false
						disallowQueryReason = "TabletControl.DisableQueryService set"
					}
					blacklistedTables = tc.BlacklistedTables
				}
			}
		}
	} else {
		disallowQueryReason = fmt.Sprintf("not a serving tablet type(%v)", newTablet.Type)
	}
	agent.setServicesDesiredState(disallowQueryReason, runUpdateStream)
	if updateBlacklistedTables {
		if err := agent.loadBlacklistRules(newTablet, blacklistedTables); err != nil {
			// FIXME(alainjobart) how to handle this error?
			log.Errorf("Cannot update blacklisted tables rule: %v", err)
		} else {
			agent.setBlacklistedTables(blacklistedTables)
		}
	}

	if allowQuery {
		// Query service should be running.
		if oldTablet.Type == topodatapb.TabletType_REPLICA &&
			newTablet.Type == topodatapb.TabletType_MASTER {
			// When promoting from replica to master, allow both master and replica
			// queries to be served during gracePeriod.
			if _, err := agent.QueryServiceControl.SetServingType(newTablet.Type,
				true, []topodatapb.TabletType{oldTablet.Type}); err == nil {
				// If successful, broadcast to vtgate and then wait.
				agent.broadcastHealth()
				time.Sleep(*gracePeriod)
			} else {
				log.Errorf("Can't start query service for MASTER+REPLICA mode: %v", err)
			}
		}

		if stateChanged, err := agent.QueryServiceControl.SetServingType(newTablet.Type, true, nil); err == nil {
			// If the state changed, broadcast to vtgate.
			// (e.g. this happens when the tablet was already master, but it just
			// changed from NOT_SERVING to SERVING due to
			// "vtctl MigrateServedFrom ... master".)
			if stateChanged {
				broadcastHealth = true
			}
		} else {
			runUpdateStream = false
			log.Errorf("Cannot start query service: %v", err)
		}
	} else {
		// Query service should be stopped.
		if topo.IsSubjectToLameduck(oldTablet.Type) &&
			newTablet.Type == topodatapb.TabletType_SPARE &&
			*gracePeriod > 0 {
			// When a non-MASTER serving type is going SPARE,
			// put query service in lameduck during gracePeriod.
			agent.lameduck(disallowQueryReason)
//.........这里部分代码省略.........

// changeCallback is run after every action that might
// have changed something in the tablet record or in the topology.
func (agent *ActionAgent) changeCallback(ctx context.Context, oldTablet, newTablet *pbt.Tablet) error {
	span := trace.NewSpanFromContext(ctx)
	span.StartLocal("ActionAgent.changeCallback")
	defer span.Finish()

	allowQuery := topo.IsRunningQueryService(newTablet.Type)

	// Read the shard to get SourceShards / TabletControlMap if
	// we're going to use it.
	var shardInfo *topo.ShardInfo
	var tabletControl *pbt.Shard_TabletControl
	var blacklistedTables []string
	var err error
	var disallowQueryReason string
	if allowQuery {
		shardInfo, err = agent.TopoServer.GetShard(ctx, newTablet.Keyspace, newTablet.Shard)
		if err != nil {
			log.Errorf("Cannot read shard for this tablet %v, might have inaccurate SourceShards and TabletControls: %v", newTablet.Alias, err)
		} else {
			if newTablet.Type == pbt.TabletType_MASTER {
				if len(shardInfo.SourceShards) > 0 {
					allowQuery = false
					disallowQueryReason = "master tablet with filtered replication on"
				}
			}
			if tc := shardInfo.GetTabletControl(newTablet.Type); tc != nil {
				if topo.InCellList(newTablet.Alias.Cell, tc.Cells) {
					if tc.DisableQueryService {
						allowQuery = false
						disallowQueryReason = "query service disabled by tablet control"
					}
					blacklistedTables = tc.BlacklistedTables
					tabletControl = tc
				}
			}
		}
	} else {
		disallowQueryReason = fmt.Sprintf("not a serving tablet type(%v)", newTablet.Type)
	}

	if allowQuery {
		if err := agent.allowQueries(newTablet, blacklistedTables); err != nil {
			log.Errorf("Cannot start query service: %v", err)
		}
	} else {
		agent.disallowQueries(newTablet, disallowQueryReason)
	}

	// save the tabletControl we've been using, so the background
	// healthcheck makes the same decisions as we've been making.
	agent.setTabletControl(tabletControl)

	// update stream needs to be started or stopped too
	if topo.IsRunningUpdateStream(newTablet.Type) {
		agent.UpdateStream.Enable()
	} else {
		agent.UpdateStream.Disable()
	}

	statsType.Set(strings.ToLower(newTablet.Type.String()))
	statsKeyspace.Set(newTablet.Keyspace)
	statsShard.Set(newTablet.Shard)
	if newTablet.KeyRange != nil {
		statsKeyRangeStart.Set(hex.EncodeToString(newTablet.KeyRange.Start))
		statsKeyRangeEnd.Set(hex.EncodeToString(newTablet.KeyRange.End))
	} else {
		statsKeyRangeStart.Set("")
		statsKeyRangeEnd.Set("")
	}

	// See if we need to start or stop any binlog player
	if agent.BinlogPlayerMap != nil {
		if newTablet.Type == pbt.TabletType_MASTER {
			// Read the keyspace on masters to get
			// ShardingColumnType, for binlog replication,
			// only if source shards are set.
			var keyspaceInfo *topo.KeyspaceInfo
			if shardInfo != nil && len(shardInfo.SourceShards) > 0 {
				keyspaceInfo, err = agent.TopoServer.GetKeyspace(ctx, newTablet.Keyspace)
				if err != nil {
					keyspaceInfo = nil
				}
			}
			agent.BinlogPlayerMap.RefreshMap(agent.batchCtx, newTablet, keyspaceInfo, shardInfo)
		} else {
			agent.BinlogPlayerMap.StopAllPlayersAndReset()
		}
	}
	return nil
}

// Update shard file with new master, replicas, etc.
//
// Re-read from TopologyServer to make sure we are using the side
// effects of all actions.
//
// This function locks individual SvrShard paths, so it doesn't need a lock
// on the shard.
func RebuildShard(log logutil.Logger, ts topo.Server, keyspace, shard string, cells []string, timeout time.Duration, interrupted chan struct{}) error {
	log.Infof("RebuildShard %v/%v", keyspace, shard)

	// read the existing shard info. It has to exist.
	shardInfo, err := ts.GetShard(keyspace, shard)
	if err != nil {
		return err
	}

	// rebuild all cells in parallel
	wg := sync.WaitGroup{}
	rec := concurrency.AllErrorRecorder{}
	for _, cell := range shardInfo.Cells {
		// skip this cell if we shouldn't rebuild it
		if !topo.InCellList(cell, cells) {
			continue
		}

		// start with the master if it's in the current cell
		tabletsAsMap := make(map[topo.TabletAlias]bool)
		if shardInfo.MasterAlias.Cell == cell {
			tabletsAsMap[shardInfo.MasterAlias] = true
		}

		wg.Add(1)
		go func(cell string) {
			defer wg.Done()

			// read the ShardReplication object to find tablets
			sri, err := ts.GetShardReplication(cell, keyspace, shard)
			if err != nil {
				rec.RecordError(fmt.Errorf("GetShardReplication(%v, %v, %v) failed: %v", cell, keyspace, shard, err))
				return
			}

			// add all relevant tablets to the map
			for _, rl := range sri.ReplicationLinks {
				tabletsAsMap[rl.TabletAlias] = true
				if rl.Parent.Cell == cell {
					tabletsAsMap[rl.Parent] = true
				}
			}

			// convert the map to a list
			aliases := make([]topo.TabletAlias, 0, len(tabletsAsMap))
			for a := range tabletsAsMap {
				aliases = append(aliases, a)
			}

			// read all the Tablet records
			tablets, err := topo.GetTabletMap(ts, aliases)
			switch err {
			case nil:
				// keep going, we're good
			case topo.ErrPartialResult:
				log.Warningf("Got ErrPartialResult from topo.GetTabletMap in cell %v, some tablets may not be added properly to serving graph", cell)
			default:
				rec.RecordError(fmt.Errorf("GetTabletMap in cell %v failed: %v", cell, err))
				return
			}

			// Lock the SrvShard so we write a consistent data set.
			actionNode := actionnode.RebuildSrvShard()
			lockPath, err := actionNode.LockSrvShard(ts, cell, keyspace, shard, timeout, interrupted)
			if err != nil {
				rec.RecordError(err)
				return
			}

			// write the data we need to
			rebuildErr := rebuildCellSrvShard(log, ts, shardInfo, cell, tablets)

			// and unlock
			if err := actionNode.UnlockSrvShard(ts, cell, keyspace, shard, lockPath, rebuildErr); err != nil {
				rec.RecordError(err)
			}
		}(cell)
	}
	wg.Wait()

	return rec.Error()
}

// RemoveShardCell will remove a cell from the Cells list in a shard.
//
// It will first check the shard has no tablets there. If 'force' is
// specified, it will remove the cell even when the tablet map cannot
// be retrieved. This is intended to be used when a cell is completely
// down and its topology server cannot even be reached.
//
// If 'recursive' is specified, it will delete any tablets in the cell/shard,
// with the assumption that the tablet processes have already been terminated.
func (wr *Wrangler) RemoveShardCell(ctx context.Context, keyspace, shard, cell string, force, recursive bool) error {
	shardInfo, err := wr.ts.GetShard(ctx, keyspace, shard)
	if err != nil {
		return err
	}

	// check the cell is in the list already
	if !topo.InCellList(cell, shardInfo.Cells) {
		return fmt.Errorf("cell %v in not in shard info", cell)
	}

	// check the master alias is not in the cell
	if shardInfo.MasterAlias != nil && shardInfo.MasterAlias.Cell == cell {
		return fmt.Errorf("master %v is in the cell '%v' we want to remove", topoproto.TabletAliasString(shardInfo.MasterAlias), cell)
	}

	// get the ShardReplication object in the cell
	sri, err := wr.ts.GetShardReplication(ctx, cell, keyspace, shard)
	switch err {
	case nil:
		if recursive {
			wr.Logger().Infof("Deleting all tablets in shard %v/%v", keyspace, shard)
			for _, node := range sri.Nodes {
				// We don't care about scrapping or updating the replication graph,
				// because we're about to delete the entire replication graph.
				wr.Logger().Infof("Deleting tablet %v", topoproto.TabletAliasString(node.TabletAlias))
				if err := wr.TopoServer().DeleteTablet(ctx, node.TabletAlias); err != nil && err != topo.ErrNoNode {
					return fmt.Errorf("can't delete tablet %v: %v", topoproto.TabletAliasString(node.TabletAlias), err)
				}
			}
		} else if len(sri.Nodes) > 0 {
			return fmt.Errorf("cell %v has %v possible tablets in replication graph", cell, len(sri.Nodes))
		}

		// ShardReplication object is now useless, remove it
		if err := wr.ts.DeleteShardReplication(ctx, cell, keyspace, shard); err != nil && err != topo.ErrNoNode {
			return fmt.Errorf("error deleting ShardReplication object in cell %v: %v", cell, err)
		}

		// we keep going
	case topo.ErrNoNode:
		// no ShardReplication object, we keep going
	default:
		// we can't get the object, assume topo server is down there,
		// so we look at force flag
		if !force {
			return err
		}
		wr.Logger().Warningf("Cannot get ShardReplication from cell %v, assuming cell topo server is down, and forcing the removal", cell)
	}

	// now we can update the shard
	wr.Logger().Infof("Removing cell %v from shard %v/%v", cell, keyspace, shard)
	_, err = wr.ts.UpdateShardFields(ctx, keyspace, shard, func(si *topo.ShardInfo) error {
		// since no lock is taken, protect against corner cases.
		if len(si.Cells) == 0 {
			return topo.ErrNoUpdateNeeded
		}
		var newCells []string
		for _, c := range si.Cells {
			if c != cell {
				newCells = append(newCells, c)
			}
		}
		si.Cells = newCells
		return nil
	})
	return err
}

// changeCallback is run after every action that might
// have changed something in the tablet record or in the topology.
//
// It owns making changes to the BinlogPlayerMap. The input for this is the
// tablet type (has to be master), and the shard's SourceShards.
//
// It owns updating the blacklisted tables.
//
// It owns updating the stats record for 'TabletType'.
//
// It owns starting and stopping the update stream service.
//
// It owns reading the TabletControl for the current tablet, and storing it.
func (agent *ActionAgent) changeCallback(ctx context.Context, oldTablet, newTablet *topodatapb.Tablet) error {
	span := trace.NewSpanFromContext(ctx)
	span.StartLocal("ActionAgent.changeCallback")
	defer span.Finish()

	allowQuery := topo.IsRunningQueryService(newTablet.Type)

	// Read the shard to get SourceShards / TabletControlMap if
	// we're going to use it.
	var shardInfo *topo.ShardInfo
	var tabletControl *topodatapb.Shard_TabletControl
	var err error
	var disallowQueryReason string
	var blacklistedTables []string
	updateBlacklistedTables := true
	if allowQuery {
		shardInfo, err = agent.TopoServer.GetShard(ctx, newTablet.Keyspace, newTablet.Shard)
		if err != nil {
			log.Errorf("Cannot read shard for this tablet %v, might have inaccurate SourceShards and TabletControls: %v", newTablet.Alias, err)
			updateBlacklistedTables = false
		} else {
			if newTablet.Type == topodatapb.TabletType_MASTER {
				if len(shardInfo.SourceShards) > 0 {
					allowQuery = false
					disallowQueryReason = "master tablet with filtered replication on"
				}
			}
			if tc := shardInfo.GetTabletControl(newTablet.Type); tc != nil {
				if topo.InCellList(newTablet.Alias.Cell, tc.Cells) {
					if tc.DisableQueryService {
						allowQuery = false
						disallowQueryReason = "query service disabled by tablet control"
					}
					blacklistedTables = tc.BlacklistedTables
					tabletControl = tc
				}
			}
		}
	} else {
		disallowQueryReason = fmt.Sprintf("not a serving tablet type(%v)", newTablet.Type)
	}
	if updateBlacklistedTables {
		if err := agent.loadBlacklistRules(newTablet, blacklistedTables); err != nil {
			// FIXME(alainjobart) how to handle this error?
			log.Errorf("Cannot update blacklisted tables rule: %v", err)
		}
	}

	if allowQuery {
		// Query service should be running.
		if oldTablet.Type == topodatapb.TabletType_REPLICA &&
			newTablet.Type == topodatapb.TabletType_MASTER {
			// When promoting from replica to master, allow both master and replica
			// queries to be served during gracePeriod.
			if err := agent.QueryServiceControl.SetServingType(newTablet.Type, true,
				[]topodatapb.TabletType{oldTablet.Type}); err != nil {
				log.Errorf("Can't start query service for MASTER+REPLICA mode: %v", err)
			} else {
				// If successful, broadcast to vtgate and then wait.
				agent.broadcastHealth()
				time.Sleep(*gracePeriod)
			}
		}
		if err := agent.allowQueries(newTablet.Type); err != nil {
			log.Errorf("Cannot start query service: %v", err)
		}
	} else {
		// Query service should be stopped.
		if (oldTablet.Type == topodatapb.TabletType_REPLICA ||
			oldTablet.Type == topodatapb.TabletType_RDONLY) &&
			newTablet.Type == topodatapb.TabletType_SPARE {
			// When a non-MASTER serving type is going SPARE,
			// put query service in lameduck during gracePeriod.
			agent.enterLameduck(disallowQueryReason)
			agent.broadcastHealth()
			time.Sleep(*gracePeriod)
		}
		agent.disallowQueries(newTablet.Type, disallowQueryReason)
	}

	// save the tabletControl we've been using, so the background
	// healthcheck makes the same decisions as we've been making.
	agent.setTabletControl(tabletControl)

	// update stream needs to be started or stopped too
	if topo.IsRunningUpdateStream(newTablet.Type) {
		agent.UpdateStream.Enable()
//.........这里部分代码省略.........