Example #1
0
// ReparentTablet tells a tablet to reparent this tablet to the current
// master, based on the current replication position. If there is no
// match, it will fail.
func (wr *Wrangler) ReparentTablet(ctx context.Context, tabletAlias *pb.TabletAlias) error {
	// Get specified tablet.
	// Get current shard master tablet.
	// Sanity check they are in the same keyspace/shard.
	// Issue a SetMaster to the tablet.
	ti, err := wr.ts.GetTablet(ctx, tabletAlias)
	if err != nil {
		return err
	}

	shardInfo, err := wr.ts.GetShard(ctx, ti.Keyspace, ti.Shard)
	if err != nil {
		return err
	}
	if topo.TabletAliasIsZero(shardInfo.MasterAlias) {
		return fmt.Errorf("no master tablet for shard %v/%v", ti.Keyspace, ti.Shard)
	}

	masterTi, err := wr.ts.GetTablet(ctx, shardInfo.MasterAlias)
	if err != nil {
		return err
	}

	// Basic sanity checking.
	if masterTi.Type != pb.TabletType_MASTER {
		return fmt.Errorf("TopologyServer has inconsistent state for shard master %v", shardInfo.MasterAlias)
	}
	if masterTi.Keyspace != ti.Keyspace || masterTi.Shard != ti.Shard {
		return fmt.Errorf("master %v and potential slave not in same keyspace/shard", shardInfo.MasterAlias)
	}

	// and do the remote command
	return wr.TabletManagerClient().SetMaster(ctx, ti, shardInfo.MasterAlias, 0, false)
}
Example #2
0
// init phase:
// - read the shard info, make sure it has sources
func (vsdw *VerticalSplitDiffWorker) init(ctx context.Context) error {
	vsdw.SetState(WorkerStateInit)

	var err error

	// read the keyspace and validate it
	vsdw.keyspaceInfo, err = vsdw.wr.TopoServer().GetKeyspace(ctx, vsdw.keyspace)
	if err != nil {
		return fmt.Errorf("cannot read keyspace %v: %v", vsdw.keyspace, err)
	}
	if len(vsdw.keyspaceInfo.ServedFroms) == 0 {
		return fmt.Errorf("keyspace %v has no KeyspaceServedFrom", vsdw.keyspace)
	}

	// read the shardinfo and validate it
	vsdw.shardInfo, err = vsdw.wr.TopoServer().GetShard(ctx, vsdw.keyspace, vsdw.shard)
	if err != nil {
		return fmt.Errorf("cannot read shard %v/%v: %v", vsdw.keyspace, vsdw.shard, err)
	}
	if len(vsdw.shardInfo.SourceShards) != 1 {
		return fmt.Errorf("shard %v/%v has bad number of source shards", vsdw.keyspace, vsdw.shard)
	}
	if len(vsdw.shardInfo.SourceShards[0].Tables) == 0 {
		return fmt.Errorf("shard %v/%v has no tables in source shard[0]", vsdw.keyspace, vsdw.shard)
	}
	if topo.TabletAliasIsZero(vsdw.shardInfo.MasterAlias) {
		return fmt.Errorf("shard %v/%v has no master", vsdw.keyspace, vsdw.shard)
	}

	return nil
}
Example #3
0
// ValidateVersionKeyspace validates all versions are the same in all
// tablets in a keyspace
func (wr *Wrangler) ValidateVersionKeyspace(ctx context.Context, keyspace string) error {
	// find all the shards
	shards, err := wr.ts.GetShardNames(ctx, keyspace)
	if err != nil {
		return err
	}

	// corner cases
	if len(shards) == 0 {
		return fmt.Errorf("No shards in keyspace %v", keyspace)
	}
	sort.Strings(shards)
	if len(shards) == 1 {
		return wr.ValidateVersionShard(ctx, keyspace, shards[0])
	}

	// find the reference version using the first shard's master
	si, err := wr.ts.GetShard(ctx, keyspace, shards[0])
	if err != nil {
		return err
	}
	if topo.TabletAliasIsZero(si.MasterAlias) {
		return fmt.Errorf("No master in shard %v/%v", keyspace, shards[0])
	}
	referenceAlias := si.MasterAlias
	log.Infof("Gathering version for reference master %v", topo.TabletAliasString(referenceAlias))
	referenceVersion, err := wr.GetVersion(ctx, referenceAlias)
	if err != nil {
		return err
	}

	// then diff with all tablets but master 0
	er := concurrency.AllErrorRecorder{}
	wg := sync.WaitGroup{}
	for _, shard := range shards {
		aliases, err := topo.FindAllTabletAliasesInShard(ctx, wr.ts, keyspace, shard)
		if err != nil {
			er.RecordError(err)
			continue
		}

		for _, alias := range aliases {
			if topo.TabletAliasEqual(alias, si.MasterAlias) {
				continue
			}

			wg.Add(1)
			go wr.diffVersion(ctx, referenceVersion, referenceAlias, alias, &wg, &er)
		}
	}
	wg.Wait()
	if er.HasErrors() {
		return fmt.Errorf("Version diffs:\n%v", er.Error().Error())
	}
	return nil
}
Example #4
0
// updateShardCellsAndMaster will update the 'Cells' and possibly
// MasterAlias records for the shard, if needed.
func (wr *Wrangler) updateShardCellsAndMaster(ctx context.Context, si *topo.ShardInfo, tabletAlias *pb.TabletAlias, tabletType pb.TabletType, force bool) error {
	// See if we need to update the Shard:
	// - add the tablet's cell to the shard's Cells if needed
	// - change the master if needed
	shardUpdateRequired := false
	if !si.HasCell(tabletAlias.Cell) {
		shardUpdateRequired = true
	}
	if tabletType == pb.TabletType_MASTER && !topo.TabletAliasEqual(si.MasterAlias, tabletAlias) {
		shardUpdateRequired = true
	}
	if !shardUpdateRequired {
		return nil
	}

	actionNode := actionnode.UpdateShard()
	keyspace := si.Keyspace()
	shard := si.ShardName()
	lockPath, err := wr.lockShard(ctx, keyspace, shard, actionNode)
	if err != nil {
		return err
	}

	// re-read the shard with the lock
	si, err = wr.ts.GetShard(ctx, keyspace, shard)
	if err != nil {
		return wr.unlockShard(ctx, keyspace, shard, actionNode, lockPath, err)
	}

	// update it
	wasUpdated := false
	if !si.HasCell(tabletAlias.Cell) {
		si.Cells = append(si.Cells, tabletAlias.Cell)
		wasUpdated = true
	}
	if tabletType == pb.TabletType_MASTER && !topo.TabletAliasEqual(si.MasterAlias, tabletAlias) {
		if !topo.TabletAliasIsZero(si.MasterAlias) && !force {
			return wr.unlockShard(ctx, keyspace, shard, actionNode, lockPath, fmt.Errorf("creating this tablet would override old master %v in shard %v/%v", topo.TabletAliasString(si.MasterAlias), keyspace, shard))
		}
		si.MasterAlias = tabletAlias
		wasUpdated = true
	}

	if wasUpdated {
		// write it back
		if err := topo.UpdateShard(ctx, wr.ts, si); err != nil {
			return wr.unlockShard(ctx, keyspace, shard, actionNode, lockPath, err)
		}
	}

	// and unlock
	return wr.unlockShard(ctx, keyspace, shard, actionNode, lockPath, err)
}
Example #5
0
// ValidateVersionShard validates all versions are the same in all
// tablets in a shard
func (wr *Wrangler) ValidateVersionShard(ctx context.Context, keyspace, shard string) error {
	si, err := wr.ts.GetShard(ctx, keyspace, shard)
	if err != nil {
		return err
	}

	// get version from the master, or error
	if topo.TabletAliasIsZero(si.MasterAlias) {
		return fmt.Errorf("No master in shard %v/%v", keyspace, shard)
	}
	log.Infof("Gathering version for master %v", topo.TabletAliasString(si.MasterAlias))
	masterVersion, err := wr.GetVersion(ctx, si.MasterAlias)
	if err != nil {
		return err
	}

	// read all the aliases in the shard, that is all tablets that are
	// replicating from the master
	aliases, err := topo.FindAllTabletAliasesInShard(ctx, wr.ts, keyspace, shard)
	if err != nil {
		return err
	}

	// then diff with all slaves
	er := concurrency.AllErrorRecorder{}
	wg := sync.WaitGroup{}
	for _, alias := range aliases {
		if topo.TabletAliasEqual(alias, si.MasterAlias) {
			continue
		}

		wg.Add(1)
		go wr.diffVersion(ctx, masterVersion, si.MasterAlias, alias, &wg, &er)
	}
	wg.Wait()
	if er.HasErrors() {
		return fmt.Errorf("Version diffs:\n%v", er.Error().Error())
	}
	return nil
}
Example #6
0
// ValidateSchemaShard will diff the schema from all the tablets in the shard.
func (wr *Wrangler) ValidateSchemaShard(ctx context.Context, keyspace, shard string, excludeTables []string, includeViews bool) error {
	si, err := wr.ts.GetShard(ctx, keyspace, shard)
	if err != nil {
		return err
	}

	// get schema from the master, or error
	if topo.TabletAliasIsZero(si.MasterAlias) {
		return fmt.Errorf("No master in shard %v/%v", keyspace, shard)
	}
	log.Infof("Gathering schema for master %v", si.MasterAlias)
	masterSchema, err := wr.GetSchema(ctx, topo.ProtoToTabletAlias(si.MasterAlias), nil, excludeTables, includeViews)
	if err != nil {
		return err
	}

	// read all the aliases in the shard, that is all tablets that are
	// replicating from the master
	aliases, err := topo.FindAllTabletAliasesInShard(ctx, wr.ts, keyspace, shard)
	if err != nil {
		return err
	}

	// then diff with all slaves
	er := concurrency.AllErrorRecorder{}
	wg := sync.WaitGroup{}
	for _, alias := range aliases {
		if alias == topo.ProtoToTabletAlias(si.MasterAlias) {
			continue
		}

		wg.Add(1)
		go wr.diffSchema(ctx, masterSchema, topo.ProtoToTabletAlias(si.MasterAlias), alias, excludeTables, includeViews, &wg, &er)
	}
	wg.Wait()
	if er.HasErrors() {
		return fmt.Errorf("Schema diffs:\n%v", er.Error().Error())
	}
	return nil
}
Example #7
0
// init phase:
// - read the shard info, make sure it has sources
func (sdw *SplitDiffWorker) init(ctx context.Context) error {
	sdw.SetState(WorkerStateInit)

	var err error
	sdw.keyspaceInfo, err = sdw.wr.TopoServer().GetKeyspace(ctx, sdw.keyspace)
	if err != nil {
		return fmt.Errorf("cannot read keyspace %v: %v", sdw.keyspace, err)
	}
	sdw.shardInfo, err = sdw.wr.TopoServer().GetShard(ctx, sdw.keyspace, sdw.shard)
	if err != nil {
		return fmt.Errorf("cannot read shard %v/%v: %v", sdw.keyspace, sdw.shard, err)
	}

	if len(sdw.shardInfo.SourceShards) == 0 {
		return fmt.Errorf("shard %v/%v has no source shard", sdw.keyspace, sdw.shard)
	}
	if topo.TabletAliasIsZero(sdw.shardInfo.MasterAlias) {
		return fmt.Errorf("shard %v/%v has no master", sdw.keyspace, sdw.shard)
	}

	return nil
}
Example #8
0
// ValidatePermissionsShard validates all the permissions are the same
// in a shard
func (wr *Wrangler) ValidatePermissionsShard(ctx context.Context, keyspace, shard string) error {
	si, err := wr.ts.GetShard(ctx, keyspace, shard)
	if err != nil {
		return err
	}

	// get permissions from the master, or error
	if topo.TabletAliasIsZero(si.MasterAlias) {
		return fmt.Errorf("No master in shard %v/%v", keyspace, shard)
	}
	log.Infof("Gathering permissions for master %v", si.MasterAlias)
	masterPermissions, err := wr.GetPermissions(ctx, topo.ProtoToTabletAlias(si.MasterAlias))
	if err != nil {
		return err
	}

	// read all the aliases in the shard, that is all tablets that are
	// replicating from the master
	aliases, err := topo.FindAllTabletAliasesInShard(ctx, wr.ts, keyspace, shard)
	if err != nil {
		return err
	}

	// then diff all of them, except master
	er := concurrency.AllErrorRecorder{}
	wg := sync.WaitGroup{}
	for _, alias := range aliases {
		if alias == topo.ProtoToTabletAlias(si.MasterAlias) {
			continue
		}
		wg.Add(1)
		go wr.diffPermissions(ctx, masterPermissions, topo.ProtoToTabletAlias(si.MasterAlias), alias, &wg, &er)
	}
	wg.Wait()
	if er.HasErrors() {
		return fmt.Errorf("Permissions diffs:\n%v", er.Error().Error())
	}
	return nil
}
Example #9
0
// Does a topo lookup for a single shard, and returns the tablet record of the master tablet.
func resolveDestinationShardMaster(ctx context.Context, keyspace, shard string, wr *wrangler.Wrangler) (*topo.TabletInfo, error) {
	var ti *topo.TabletInfo
	shortCtx, cancel := context.WithTimeout(ctx, *remoteActionsTimeout)
	si, err := topo.GetShard(shortCtx, wr.TopoServer(), keyspace, shard)
	cancel()
	if err != nil {
		return ti, fmt.Errorf("unable to resolve destination shard %v/%v", keyspace, shard)
	}

	if topo.TabletAliasIsZero(si.MasterAlias) {
		return ti, fmt.Errorf("no master in destination shard %v/%v", keyspace, shard)
	}

	wr.Logger().Infof("Found target master alias %v in shard %v/%v", si.MasterAlias, keyspace, shard)

	shortCtx, cancel = context.WithTimeout(ctx, *remoteActionsTimeout)
	ti, err = topo.GetTablet(shortCtx, wr.TopoServer(), topo.ProtoToTabletAlias(si.MasterAlias))
	cancel()
	if err != nil {
		return ti, fmt.Errorf("unable to get master tablet from alias %v in shard %v/%v",
			si.MasterAlias, keyspace, shard)
	}
	return ti, nil
}
Example #10
0
// InitTablet creates or updates a tablet. If no parent is specified
// in the tablet, and the tablet has a slave type, we will find the
// appropriate parent. If createShardAndKeyspace is true and the
// parent keyspace or shard don't exist, they will be created.  If
// update is true, and a tablet with the same ID exists, update it.
// If Force is true, and a tablet with the same ID already exists, it
// will be scrapped and deleted, and then recreated.
func (wr *Wrangler) InitTablet(ctx context.Context, tablet *topo.Tablet, force, createShardAndKeyspace, update bool) error {
	if err := topo.TabletComplete(tablet); err != nil {
		return err
	}

	if topo.IsInReplicationGraph(tablet.Type) {
		// get the shard, possibly creating it
		var err error
		var si *topo.ShardInfo

		if createShardAndKeyspace {
			// create the parent keyspace and shard if needed
			si, err = topotools.GetOrCreateShard(ctx, wr.ts, tablet.Keyspace, tablet.Shard)
		} else {
			si, err = wr.ts.GetShard(ctx, tablet.Keyspace, tablet.Shard)
			if err == topo.ErrNoNode {
				return fmt.Errorf("missing parent shard, use -parent option to create it, or CreateKeyspace / CreateShard")
			}
		}

		// get the shard, checks a couple things
		if err != nil {
			return fmt.Errorf("cannot get (or create) shard %v/%v: %v", tablet.Keyspace, tablet.Shard, err)
		}
		if key.ProtoToKeyRange(si.KeyRange) != tablet.KeyRange {
			return fmt.Errorf("shard %v/%v has a different KeyRange: %v != %v", tablet.Keyspace, tablet.Shard, si.KeyRange, tablet.KeyRange)
		}
		if tablet.Type == topo.TYPE_MASTER && !topo.TabletAliasIsZero(si.MasterAlias) && topo.ProtoToTabletAlias(si.MasterAlias) != tablet.Alias && !force {
			return fmt.Errorf("creating this tablet would override old master %v in shard %v/%v", si.MasterAlias, tablet.Keyspace, tablet.Shard)
		}

		// update the shard record if needed
		if err := wr.updateShardCellsAndMaster(ctx, si, topo.TabletAliasToProto(tablet.Alias), topo.TabletTypeToProto(tablet.Type), force); err != nil {
			return err
		}
	}

	err := topo.CreateTablet(ctx, wr.ts, tablet)
	if err != nil && err == topo.ErrNodeExists {
		// Try to update nicely, but if it fails fall back to force behavior.
		if update || force {
			oldTablet, err := wr.ts.GetTablet(ctx, tablet.Alias)
			if err != nil {
				wr.Logger().Warningf("failed reading tablet %v: %v", tablet.Alias, err)
			} else {
				if oldTablet.Keyspace == tablet.Keyspace && oldTablet.Shard == tablet.Shard {
					*(oldTablet.Tablet) = *tablet
					if err := topo.UpdateTablet(ctx, wr.ts, oldTablet); err != nil {
						wr.Logger().Warningf("failed updating tablet %v: %v", tablet.Alias, err)
						// now fall through the Scrap case
					} else {
						if !topo.IsInReplicationGraph(tablet.Type) {
							return nil
						}

						if err := topo.UpdateTabletReplicationData(ctx, wr.ts, tablet); err != nil {
							wr.Logger().Warningf("failed updating tablet replication data for %v: %v", tablet.Alias, err)
							// now fall through the Scrap case
						} else {
							return nil
						}
					}
				}
			}
		}
		if force {
			if err = wr.Scrap(ctx, tablet.Alias, force, false); err != nil {
				wr.Logger().Errorf("failed scrapping tablet %v: %v", tablet.Alias, err)
				return err
			}
			if err := wr.ts.DeleteTablet(ctx, tablet.Alias); err != nil {
				// we ignore this
				wr.Logger().Errorf("failed deleting tablet %v: %v", tablet.Alias, err)
			}
			return topo.CreateTablet(ctx, wr.ts, tablet)
		}
	}
	return err
}
Example #11
0
func (wr *Wrangler) applySchemaShardComplex(ctx context.Context, statusArray []*tabletStatus, shardInfo *topo.ShardInfo, preflight *myproto.SchemaChangeResult, masterTabletAlias *pb.TabletAlias, change string, newParentTabletAlias *pb.TabletAlias, force bool, waitSlaveTimeout time.Duration) (*myproto.SchemaChangeResult, error) {
	// apply the schema change to all replica / slave tablets
	for _, status := range statusArray {
		// if already applied, we skip this guy
		diffs := myproto.DiffSchemaToArray("after", preflight.AfterSchema, topo.TabletAliasString(status.ti.Alias), status.beforeSchema)
		if len(diffs) == 0 {
			log.Infof("Tablet %v already has the AfterSchema, skipping", status.ti.Alias)
			continue
		}

		// make sure the before schema matches
		diffs = myproto.DiffSchemaToArray("master", preflight.BeforeSchema, topo.TabletAliasString(status.ti.Alias), status.beforeSchema)
		if len(diffs) > 0 {
			if force {
				log.Warningf("Tablet %v has inconsistent schema, ignoring: %v", status.ti.Alias, strings.Join(diffs, "\n"))
			} else {
				return nil, fmt.Errorf("Tablet %v has inconsistent schema: %v", status.ti.Alias, strings.Join(diffs, "\n"))
			}
		}

		// take this guy out of the serving graph if necessary
		ti, err := wr.ts.GetTablet(ctx, status.ti.Alias)
		if err != nil {
			return nil, err
		}
		typeChangeRequired := ti.IsInServingGraph()
		if typeChangeRequired {
			// note we want to update the serving graph there
			err = wr.changeTypeInternal(ctx, ti.Alias, pb.TabletType_SCHEMA_UPGRADE)
			if err != nil {
				return nil, err
			}
		}

		// apply the schema change
		log.Infof("Applying schema change to slave %v in complex mode", status.ti.Alias)
		sc := &myproto.SchemaChange{Sql: change, Force: force, AllowReplication: false, BeforeSchema: preflight.BeforeSchema, AfterSchema: preflight.AfterSchema}
		_, err = wr.ApplySchema(ctx, status.ti.Alias, sc)
		if err != nil {
			return nil, err
		}

		// put this guy back into the serving graph
		if typeChangeRequired {
			err = wr.changeTypeInternal(ctx, ti.Alias, ti.Tablet.Type)
			if err != nil {
				return nil, err
			}
		}
	}

	// if newParentTabletAlias is passed in, use that as the new master
	if !topo.TabletAliasIsZero(newParentTabletAlias) {
		log.Infof("Reparenting with new master set to %v", newParentTabletAlias)
		oldMasterAlias := shardInfo.MasterAlias

		// Create reusable Reparent event with available info
		ev := &events.Reparent{}

		if err := wr.plannedReparentShardLocked(ctx, ev, shardInfo.Keyspace(), shardInfo.ShardName(), newParentTabletAlias, waitSlaveTimeout); err != nil {
			return nil, err
		}

		// Here we would apply the schema change to the old
		// master, but we just scrap it, to be consistent
		// with the previous implementation of the reparent.
		// (this code will be refactored at some point anyway)
		if err := wr.Scrap(ctx, oldMasterAlias, false, false); err != nil {
			wr.Logger().Warningf("Scrapping old master %v from shard %v/%v failed: %v", oldMasterAlias, shardInfo.Keyspace(), shardInfo.ShardName(), err)
		}
	}
	return &myproto.SchemaChangeResult{BeforeSchema: preflight.BeforeSchema, AfterSchema: preflight.AfterSchema}, nil
}
Example #12
0
// ValidateSchemaKeyspace will diff the schema from all the tablets in
// the keyspace.
func (wr *Wrangler) ValidateSchemaKeyspace(ctx context.Context, keyspace string, excludeTables []string, includeViews bool) error {
	// find all the shards
	shards, err := wr.ts.GetShardNames(ctx, keyspace)
	if err != nil {
		return err
	}

	// corner cases
	if len(shards) == 0 {
		return fmt.Errorf("No shards in keyspace %v", keyspace)
	}
	sort.Strings(shards)
	if len(shards) == 1 {
		return wr.ValidateSchemaShard(ctx, keyspace, shards[0], excludeTables, includeViews)
	}

	// find the reference schema using the first shard's master
	si, err := wr.ts.GetShard(ctx, keyspace, shards[0])
	if err != nil {
		return err
	}
	if topo.TabletAliasIsZero(si.MasterAlias) {
		return fmt.Errorf("No master in shard %v/%v", keyspace, shards[0])
	}
	referenceAlias := si.MasterAlias
	log.Infof("Gathering schema for reference master %v", referenceAlias)
	referenceSchema, err := wr.GetSchema(ctx, referenceAlias, nil, excludeTables, includeViews)
	if err != nil {
		return err
	}

	// then diff with all other tablets everywhere
	er := concurrency.AllErrorRecorder{}
	wg := sync.WaitGroup{}

	// first diff the slaves in the reference shard 0
	aliases, err := topo.FindAllTabletAliasesInShard(ctx, wr.ts, keyspace, shards[0])
	if err != nil {
		return err
	}

	for _, alias := range aliases {
		if topo.TabletAliasEqual(alias, si.MasterAlias) {
			continue
		}

		wg.Add(1)
		go wr.diffSchema(ctx, referenceSchema, referenceAlias, alias, excludeTables, includeViews, &wg, &er)
	}

	// then diffs all tablets in the other shards
	for _, shard := range shards[1:] {
		si, err := wr.ts.GetShard(ctx, keyspace, shard)
		if err != nil {
			er.RecordError(err)
			continue
		}

		if topo.TabletAliasIsZero(si.MasterAlias) {
			er.RecordError(fmt.Errorf("No master in shard %v/%v", keyspace, shard))
			continue
		}

		aliases, err := topo.FindAllTabletAliasesInShard(ctx, wr.ts, keyspace, shard)
		if err != nil {
			er.RecordError(err)
			continue
		}

		for _, alias := range aliases {
			wg.Add(1)
			go wr.diffSchema(ctx, referenceSchema, referenceAlias, alias, excludeTables, includeViews, &wg, &er)
		}
	}
	wg.Wait()
	if er.HasErrors() {
		return fmt.Errorf("Schema diffs:\n%v", er.Error().Error())
	}
	return nil
}
Example #13
0
// finalizeTabletExternallyReparented performs slow, synchronized reconciliation
// tasks that ensure topology is self-consistent, and then marks the reparent as
// finished by updating the global shard record.
func (agent *ActionAgent) finalizeTabletExternallyReparented(ctx context.Context, si *topo.ShardInfo, ev *events.Reparent) (err error) {
	var wg sync.WaitGroup
	var errs concurrency.AllErrorRecorder
	oldMasterAlias := si.MasterAlias

	// Update the tablet records and serving graph for the old and new master concurrently.
	event.DispatchUpdate(ev, "updating old and new master tablet records")
	log.Infof("finalizeTabletExternallyReparented: updating tablet records")
	wg.Add(1)
	go func() {
		defer wg.Done()
		// Update our own record to master.
		var updatedTablet *topo.Tablet
		err := topo.UpdateTabletFields(ctx, agent.TopoServer, agent.TabletAlias,
			func(tablet *topo.Tablet) error {
				tablet.Type = topo.TYPE_MASTER
				tablet.Health = nil
				updatedTablet = tablet
				return nil
			})
		if err != nil {
			errs.RecordError(err)
			return
		}

		// Update the serving graph for the tablet.
		if updatedTablet != nil {
			errs.RecordError(
				topotools.UpdateTabletEndpoints(ctx, agent.TopoServer, updatedTablet))
		}
	}()

	if !topo.TabletAliasIsZero(oldMasterAlias) {
		wg.Add(1)
		go func() {
			// Force the old master to spare.
			var oldMasterTablet *topo.Tablet
			err := topo.UpdateTabletFields(ctx, agent.TopoServer, topo.ProtoToTabletAlias(oldMasterAlias),
				func(tablet *topo.Tablet) error {
					tablet.Type = topo.TYPE_SPARE
					oldMasterTablet = tablet
					return nil
				})
			if err != nil {
				errs.RecordError(err)
				wg.Done()
				return
			}

			if oldMasterTablet != nil {
				// We now know more about the old master, so add it to event data.
				ev.OldMaster = *oldMasterTablet

				// Update the serving graph.
				errs.RecordError(
					topotools.UpdateTabletEndpoints(ctx, agent.TopoServer, oldMasterTablet))
				wg.Done()

				// Tell the old master to refresh its state. We don't need to wait for it.
				tmc := tmclient.NewTabletManagerClient()
				tmc.RefreshState(ctx, topo.NewTabletInfo(oldMasterTablet, -1))
			}
		}()
	}

	tablet := agent.Tablet()

	// Wait for the tablet records to be updated. At that point, any rebuild will
	// see the new master, so we're ready to mark the reparent as done in the
	// global shard record.
	wg.Wait()
	if errs.HasErrors() {
		return errs.Error()
	}

	// Update the master field in the global shard record. We don't use a lock
	// here anymore. The lock was only to ensure that the global shard record
	// didn't get modified between the time when we read it and the time when we
	// write it back. Now we use an update loop pattern to do that instead.
	event.DispatchUpdate(ev, "updating global shard record")
	log.Infof("finalizeTabletExternallyReparented: updating global shard record")
	si, err = topo.UpdateShardFields(ctx, agent.TopoServer, tablet.Keyspace, tablet.Shard, func(shard *pb.Shard) error {
		shard.MasterAlias = topo.TabletAliasToProto(tablet.Alias)
		return nil
	})
	if err != nil {
		return err
	}

	// We already took care of updating the serving graph for the old and new masters.
	// All that's left now is in case of a cross-cell reparent, we need to update the
	// master cell setting in the SrvShard records of all cells.
	if oldMasterAlias == nil || oldMasterAlias.Cell != tablet.Alias.Cell {
		event.DispatchUpdate(ev, "rebuilding shard serving graph")
		log.Infof("finalizeTabletExternallyReparented: updating SrvShard in all cells for cross-cell reparent")
		if err := topotools.UpdateAllSrvShards(ctx, agent.TopoServer, si); err != nil {
			return err
		}
	}

	event.DispatchUpdate(ev, "finished")
	return nil
}
Example #14
0
func (wr *Wrangler) emergencyReparentShardLocked(ctx context.Context, ev *events.Reparent, keyspace, shard string, masterElectTabletAlias *pb.TabletAlias, waitSlaveTimeout time.Duration) error {
	shardInfo, err := wr.ts.GetShard(ctx, keyspace, shard)
	if err != nil {
		return err
	}
	ev.ShardInfo = *shardInfo

	event.DispatchUpdate(ev, "reading all tablets")
	tabletMap, err := topo.GetTabletMapForShard(ctx, wr.ts, keyspace, shard)
	if err != nil {
		return err
	}

	// Check corner cases we're going to depend on
	masterElectTabletInfo, ok := tabletMap[*masterElectTabletAlias]
	if !ok {
		return fmt.Errorf("master-elect tablet %v is not in the shard", masterElectTabletAlias)
	}
	ev.NewMaster = *masterElectTabletInfo.Tablet
	if topo.TabletAliasEqual(shardInfo.MasterAlias, masterElectTabletAlias) {
		return fmt.Errorf("master-elect tablet %v is already the master", masterElectTabletAlias)
	}

	// Deal with the old master: try to remote-scrap it, if it's
	// truely dead we force-scrap it. Remove it from our map in any case.
	if !topo.TabletAliasIsZero(shardInfo.MasterAlias) {
		scrapOldMaster := true
		oldMasterTabletInfo, ok := tabletMap[*shardInfo.MasterAlias]
		if ok {
			delete(tabletMap, *shardInfo.MasterAlias)
		} else {
			oldMasterTabletInfo, err = wr.ts.GetTablet(ctx, shardInfo.MasterAlias)
			if err != nil {
				wr.logger.Warningf("cannot read old master tablet %v, won't touch it: %v", shardInfo.MasterAlias, err)
				scrapOldMaster = false
			}
		}

		if scrapOldMaster {
			ev.OldMaster = *oldMasterTabletInfo.Tablet
			wr.logger.Infof("scrapping old master %v", shardInfo.MasterAlias)

			ctx, cancel := context.WithTimeout(ctx, waitSlaveTimeout)
			defer cancel()

			if err := wr.tmc.Scrap(ctx, oldMasterTabletInfo); err != nil {
				wr.logger.Warningf("remote scrapping failed master failed, will force the scrap: %v", err)

				if err := topotools.Scrap(ctx, wr.ts, shardInfo.MasterAlias, true); err != nil {
					wr.logger.Warningf("old master topo scrapping failed, continuing anyway: %v", err)
				}
			}
		}
	}

	// Stop replication on all slaves, get their current
	// replication position
	event.DispatchUpdate(ev, "stop replication on all slaves")
	wg := sync.WaitGroup{}
	mu := sync.Mutex{}
	statusMap := make(map[pb.TabletAlias]myproto.ReplicationStatus)
	for alias, tabletInfo := range tabletMap {
		wg.Add(1)
		go func(alias pb.TabletAlias, tabletInfo *topo.TabletInfo) {
			defer wg.Done()
			wr.logger.Infof("getting replication position from %v", alias)
			ctx, cancel := context.WithTimeout(ctx, waitSlaveTimeout)
			defer cancel()
			rp, err := wr.TabletManagerClient().StopReplicationAndGetStatus(ctx, tabletInfo)
			if err != nil {
				wr.logger.Warningf("failed to get replication status from %v, ignoring tablet: %v", alias, err)
				return
			}
			mu.Lock()
			statusMap[alias] = rp
			mu.Unlock()
		}(alias, tabletInfo)
	}
	wg.Wait()

	// Verify masterElect is alive and has the most advanced position
	masterElectStatus, ok := statusMap[*masterElectTabletAlias]
	if !ok {
		return fmt.Errorf("couldn't get master elect %v replication position", masterElectTabletAlias)
	}
	for alias, status := range statusMap {
		if topo.TabletAliasEqual(&alias, masterElectTabletAlias) {
			continue
		}
		if !masterElectStatus.Position.AtLeast(status.Position) {
			return fmt.Errorf("tablet %v is more advanced than master elect tablet %v: %v > %v", alias, masterElectTabletAlias, status.Position, masterElectStatus)
		}
	}

	// Promote the masterElect
	wr.logger.Infof("promote slave %v", masterElectTabletAlias)
	event.DispatchUpdate(ev, "promoting slave")
	rp, err := wr.tmc.PromoteSlave(ctx, masterElectTabletInfo)
	if err != nil {
		return fmt.Errorf("master-elect tablet %v failed to be upgraded to master: %v", masterElectTabletAlias, err)
	}

	// Reset replication on all slaves to point to the new master, and
	// insert test row in the new master.
	// Go through all the tablets:
	// - new master: populate the reparent journal
	// - everybody else: reparent to new master, wait for row
	event.DispatchUpdate(ev, "reparenting all tablets")
	now := time.Now().UnixNano()
	wgMaster := sync.WaitGroup{}
	wgSlaves := sync.WaitGroup{}
	rec := concurrency.AllErrorRecorder{}
	var masterErr error
	for alias, tabletInfo := range tabletMap {
		if topo.TabletAliasEqual(&alias, masterElectTabletAlias) {
			wgMaster.Add(1)
			go func(alias pb.TabletAlias, tabletInfo *topo.TabletInfo) {
				defer wgMaster.Done()
				wr.logger.Infof("populating reparent journal on new master %v", alias)
				masterErr = wr.TabletManagerClient().PopulateReparentJournal(ctx, tabletInfo, now, emergencyReparentShardOperation, &alias, rp)
			}(alias, tabletInfo)
		} else {
			wgSlaves.Add(1)
			go func(alias pb.TabletAlias, tabletInfo *topo.TabletInfo) {
				defer wgSlaves.Done()
				wr.logger.Infof("setting new master on slave %v", alias)
				forceStartSlave := false
				if status, ok := statusMap[alias]; ok {
					forceStartSlave = status.SlaveIORunning || status.SlaveSQLRunning
				}
				if err := wr.TabletManagerClient().SetMaster(ctx, tabletInfo, masterElectTabletAlias, now, forceStartSlave); err != nil {
					rec.RecordError(fmt.Errorf("Tablet %v SetMaster failed: %v", alias, err))
				}
			}(alias, tabletInfo)
		}
	}

	// After the master is done, we can update the shard record
	// (note with semi-sync, it also means at least one slave is done)
	wgMaster.Wait()
	if masterErr != nil {
		wgSlaves.Wait()
		return fmt.Errorf("failed to PopulateReparentJournal on master: %v", masterErr)
	}
	wr.logger.Infof("updating shard record with new master %v", masterElectTabletAlias)
	shardInfo.MasterAlias = masterElectTabletAlias
	if err := topo.UpdateShard(ctx, wr.ts, shardInfo); err != nil {
		wgSlaves.Wait()
		return fmt.Errorf("failed to update shard master record: %v", err)
	}

	// Wait for the slaves to complete. If some of them fail, we
	// will rebuild the shard serving graph anyway
	wgSlaves.Wait()
	if err := rec.Error(); err != nil {
		wr.Logger().Errorf("Some slaves failed to reparent: %v", err)
		return err
	}

	// Then we rebuild the entire serving graph for the shard,
	// to account for all changes.
	wr.logger.Infof("rebuilding shard graph")
	event.DispatchUpdate(ev, "rebuilding shard serving graph")
	_, err = wr.RebuildShardGraph(ctx, keyspace, shard, nil)
	return err
}