// init phase:
// - read the destination keyspace, make sure it has 'servedFrom' values
func (scw *SplitCloneWorker) init() error {
scw.setState(stateSCInit)
var err error
// read the keyspace and validate it
scw.keyspaceInfo, err = scw.wr.TopoServer().GetKeyspace(scw.keyspace)
if err != nil {
return fmt.Errorf("cannot read keyspace %v: %v", scw.keyspace, err)
}
// find the OverlappingShards in the keyspace
osList, err := topotools.FindOverlappingShards(scw.wr.TopoServer(), scw.keyspace)
if err != nil {
return fmt.Errorf("cannot FindOverlappingShards in %v: %v", scw.keyspace, err)
}
// find the shard we mentioned in there, if any
os := topotools.OverlappingShardsForShard(osList, scw.shard)
if os == nil {
return fmt.Errorf("the specified shard %v/%v is not in any overlapping shard", scw.keyspace, scw.shard)
}
// one side should have served types, the other one none,
// figure out wich is which, then double check them all
if len(os.Left[0].ServedTypes) > 0 {
scw.sourceShards = os.Left
scw.destinationShards = os.Right
} else {
scw.sourceShards = os.Right
scw.destinationShards = os.Left
}
// validate all serving types
servingTypes := []topo.TabletType{topo.TYPE_MASTER, topo.TYPE_REPLICA, topo.TYPE_RDONLY}
for _, st := range servingTypes {
for _, si := range scw.sourceShards {
if !topo.IsTypeInList(st, si.ServedTypes) {
return fmt.Errorf("source shard %v/%v is not serving type %v", si.Keyspace(), si.ShardName(), st)
}
}
}
for _, si := range scw.destinationShards {
if len(si.ServedTypes) > 0 {
return fmt.Errorf("destination shard %v/%v is serving some types", si.Keyspace(), si.ShardName())
}
}
return nil
}
func (wr *Wrangler) MigrateServedTypes(keyspace, shard string, servedType topo.TabletType, reverse, skipRebuild bool) error {
if servedType == topo.TYPE_MASTER {
// we cannot migrate a master back, since when master migration
// is done, the source shards are dead
if reverse {
return fmt.Errorf("Cannot migrate master back to %v/%v", keyspace, shard)
}
// we cannot skip rebuild for a master
if skipRebuild {
return fmt.Errorf("Cannot skip rebuild for master migration on %v/%v", keyspace, shard)
}
}
// first figure out the destination shards
// TODO(alainjobart) for now we only look in the same keyspace.
// We might want to look elsewhere eventually too, maybe through
// an extra command line parameter?
shardNames, err := wr.ts.GetShardNames(keyspace)
if err != nil {
return nil
}
destinationShards := make([]*topo.ShardInfo, 0, 0)
for _, shardName := range shardNames {
si, err := wr.ts.GetShard(keyspace, shardName)
if err != nil {
return err
}
for _, sourceShard := range si.SourceShards {
if sourceShard.Keyspace == keyspace && sourceShard.Shard == shard {
// this shard is replicating from the source shard we specified
log.Infof("Found %v/%v as a destination shard", si.Keyspace(), si.ShardName())
destinationShards = append(destinationShards, si)
break
}
}
}
if len(destinationShards) == 0 {
return fmt.Errorf("Cannot find any destination shard replicating from %v/%v", keyspace, shard)
}
// TODO(alainjobart) for a reverse split, we also need to find
// more sources. For now, a single source is all we need, but we
// still use a list of sources to not have to change the code later.
sourceShards := make([]*topo.ShardInfo, 0, 0)
// Verify the source has the type we're migrating
si, err := wr.ts.GetShard(keyspace, shard)
if err != nil {
return err
}
foundType := topo.IsTypeInList(servedType, si.ServedTypes)
if reverse {
if foundType {
return fmt.Errorf("Source shard %v/%v is already serving type %v", keyspace, shard, servedType)
}
} else {
if !foundType {
return fmt.Errorf("Source shard %v/%v is not serving type %v", keyspace, shard, servedType)
}
}
if servedType == topo.TYPE_MASTER && len(si.ServedTypes) > 1 {
return fmt.Errorf("Cannot migrate master out of %v/%v until everything else is migrated out", keyspace, shard)
}
sourceShards = append(sourceShards, si)
// lock the shards: sources, then destinations
// (note they're all ordered by shard name)
actionNode := actionnode.MigrateServedTypes(servedType)
sourceLockPath := make([]string, len(sourceShards))
for i, si := range sourceShards {
sourceLockPath[i], err = wr.lockShard(si.Keyspace(), si.ShardName(), actionNode)
if err != nil {
log.Errorf("Failed to lock source shard %v/%v, may need to unlock other shards manually", si.Keyspace(), si.ShardName())
return err
}
}
destinationLockPath := make([]string, len(destinationShards))
for i, si := range destinationShards {
destinationLockPath[i], err = wr.lockShard(si.Keyspace(), si.ShardName(), actionNode)
if err != nil {
log.Errorf("Failed to lock destination shard %v/%v, may need to unlock other shards manually", si.Keyspace(), si.ShardName())
return err
}
}
// record the action error and all unlock errors
rec := concurrency.AllErrorRecorder{}
// execute the migration
shardCache := make(map[string]*topo.ShardInfo)
rec.RecordError(wr.migrateServedTypes(keyspace, sourceShards, destinationShards, servedType, reverse, shardCache))
// unlock the shards, we're done
for i := len(destinationShards) - 1; i >= 0; i-- {
rec.RecordError(wr.unlockShard(destinationShards[i].Keyspace(), destinationShards[i].ShardName(), actionNode, destinationLockPath[i], nil))
}
for i := len(sourceShards) - 1; i >= 0; i-- {
rec.RecordError(wr.unlockShard(sourceShards[i].Keyspace(), sourceShards[i].ShardName(), actionNode, sourceLockPath[i], nil))
}
// rebuild the keyspace serving graph if there was no error
if rec.Error() == nil {
if skipRebuild {
log.Infof("Skipping keyspace rebuild, please run it at earliest convenience")
} else {
rec.RecordError(wr.RebuildKeyspaceGraph(keyspace, nil, shardCache))
}
}
return rec.Error()
}
// migrateServedTypes operates with all concerned shards locked.
func (wr *Wrangler) migrateServedTypes(keyspace string, sourceShards, destinationShards []*topo.ShardInfo, servedType topo.TabletType, reverse bool, shardCache map[string]*topo.ShardInfo) (err error) {
// re-read all the shards so we are up to date
for i, si := range sourceShards {
if sourceShards[i], err = wr.ts.GetShard(si.Keyspace(), si.ShardName()); err != nil {
return err
}
shardCache[si.ShardName()] = sourceShards[i]
}
for i, si := range destinationShards {
if destinationShards[i], err = wr.ts.GetShard(si.Keyspace(), si.ShardName()); err != nil {
return err
}
shardCache[si.ShardName()] = destinationShards[i]
}
ev := &events.MigrateServedTypes{
Keyspace: *topo.NewKeyspaceInfo(keyspace, nil, -1),
SourceShards: sourceShards,
DestinationShards: destinationShards,
ServedType: servedType,
Reverse: reverse,
}
event.DispatchUpdate(ev, "start")
defer func() {
if err != nil {
event.DispatchUpdate(ev, "failed: "+err.Error())
}
}()
// check and update all shard records, in memory only
for _, si := range sourceShards {
if reverse {
// need to add to source
if topo.IsTypeInList(servedType, si.ServedTypes) {
return fmt.Errorf("Source shard %v/%v is already serving type %v", si.Keyspace(), si.ShardName(), servedType)
}
si.ServedTypes = append(si.ServedTypes, servedType)
} else {
// need to remove from source
var found bool
if si.ServedTypes, found = removeType(servedType, si.ServedTypes); !found {
return fmt.Errorf("Source shard %v/%v is not serving type %v", si.Keyspace(), si.ShardName(), servedType)
}
}
}
for _, si := range destinationShards {
if reverse {
// need to remove from destination
var found bool
if si.ServedTypes, found = removeType(servedType, si.ServedTypes); !found {
return fmt.Errorf("Destination shard %v/%v is not serving type %v", si.Keyspace(), si.ShardName(), servedType)
}
} else {
// need to add to destination
if topo.IsTypeInList(servedType, si.ServedTypes) {
return fmt.Errorf("Destination shard %v/%v is already serving type %v", si.Keyspace(), si.ShardName(), servedType)
}
si.ServedTypes = append(si.ServedTypes, servedType)
}
}
// For master type migration, need to:
// - switch the source shards to read-only
// - gather all replication points
// - wait for filtered replication to catch up before we continue
// - disable filtered replication after the fact
if servedType == topo.TYPE_MASTER {
event.DispatchUpdate(ev, "setting all source masters read-only")
err := wr.makeMastersReadOnly(sourceShards)
if err != nil {
return err
}
event.DispatchUpdate(ev, "getting positions of source masters")
masterPositions, err := wr.getMastersPosition(sourceShards)
if err != nil {
return err
}
event.DispatchUpdate(ev, "waiting for destination masters to catch up")
if err := wr.waitForFilteredReplication(masterPositions, destinationShards); err != nil {
return err
}
for _, si := range destinationShards {
si.SourceShards = nil
}
}
// All is good, we can save the shards now
event.DispatchUpdate(ev, "updating source shards")
for _, si := range sourceShards {
if err := topo.UpdateShard(wr.ts, si); err != nil {
return err
}
shardCache[si.ShardName()] = si
}
event.DispatchUpdate(ev, "updating destination shards")
for _, si := range destinationShards {
if err := topo.UpdateShard(wr.ts, si); err != nil {
return err
}
shardCache[si.ShardName()] = si
}
// And tell the new shards masters they can now be read-write.
// Invoking a remote action will also make the tablet stop filtered
// replication.
if servedType == topo.TYPE_MASTER {
event.DispatchUpdate(ev, "setting destination masters read-write")
if err := wr.makeMastersReadWrite(destinationShards); err != nil {
return err
}
}
event.DispatchUpdate(ev, "finished")
return nil
}
// InitTablet initializes the tablet record if necessary.
func (agent *ActionAgent) InitTablet(port, securePort, gRPCPort int) error {
// only enabled if one of init_tablet_type (when healthcheck
// is disabled) or init_keyspace (when healthcheck is enabled)
// is passed in, then check other parameters
if *initTabletType == "" && *initKeyspace == "" {
return nil
}
// figure out our default target type
var tabletType topo.TabletType
if *initTabletType != "" {
if *targetTabletType != "" {
log.Fatalf("cannot specify both target_tablet_type and init_tablet_type parameters (as they might conflict)")
}
// use the type specified on the command line
tabletType = topo.TabletType(*initTabletType)
if !topo.IsTypeInList(tabletType, topo.AllTabletTypes) {
log.Fatalf("InitTablet encountered unknown init_tablet_type '%v'", *initTabletType)
}
if tabletType == topo.TYPE_MASTER || tabletType == topo.TYPE_SCRAP {
// We disallow TYPE_MASTER, so we don't have to change
// shard.MasterAlias, and deal with the corner cases.
// We also disallow TYPE_SCRAP, obviously.
log.Fatalf("init_tablet_type cannot be %v", tabletType)
}
} else if *targetTabletType != "" {
if tabletType := topo.TabletType(*targetTabletType); tabletType == topo.TYPE_MASTER {
log.Fatalf("target_tablet_type cannot be '%v'. Use '%v' instead.", tabletType, topo.TYPE_REPLICA)
}
// use spare, the healthcheck will turn us into what
// we need to be eventually
tabletType = topo.TYPE_SPARE
} else {
log.Fatalf("if init tablet is enabled, one of init_tablet_type or target_tablet_type needs to be specified")
}
// create a context for this whole operation
ctx, cancel := context.WithTimeout(agent.batchCtx, *initTimeout)
defer cancel()
// if we're assigned to a shard, make sure it exists, see if
// we are its master, and update its cells list if necessary
if tabletType != topo.TYPE_IDLE {
if *initKeyspace == "" || *initShard == "" {
log.Fatalf("if init tablet is enabled and the target type is not idle, init_keyspace and init_shard also need to be specified")
}
shard, _, err := topo.ValidateShardName(*initShard)
if err != nil {
log.Fatalf("cannot validate shard name: %v", err)
}
log.Infof("Reading shard record %v/%v", *initKeyspace, shard)
// read the shard, create it if necessary
si, err := topotools.GetOrCreateShard(ctx, agent.TopoServer, *initKeyspace, shard)
if err != nil {
return fmt.Errorf("InitTablet cannot GetOrCreateShard shard: %v", err)
}
if si.MasterAlias == agent.TabletAlias {
// we are the current master for this shard (probably
// means the master tablet process was just restarted),
// so InitTablet as master.
tabletType = topo.TYPE_MASTER
}
// See if we need to add the tablet's cell to the shard's cell
// list. If we do, it has to be under the shard lock.
if !si.HasCell(agent.TabletAlias.Cell) {
actionNode := actionnode.UpdateShard()
lockPath, err := actionNode.LockShard(ctx, agent.TopoServer, *initKeyspace, shard)
if err != nil {
return fmt.Errorf("LockShard(%v/%v) failed: %v", *initKeyspace, shard, err)
}
// re-read the shard with the lock
si, err = agent.TopoServer.GetShard(ctx, *initKeyspace, shard)
if err != nil {
return actionNode.UnlockShard(ctx, agent.TopoServer, *initKeyspace, shard, lockPath, err)
}
// see if we really need to update it now
if !si.HasCell(agent.TabletAlias.Cell) {
si.Cells = append(si.Cells, agent.TabletAlias.Cell)
// write it back
if err := topo.UpdateShard(ctx, agent.TopoServer, si); err != nil {
return actionNode.UnlockShard(ctx, agent.TopoServer, *initKeyspace, shard, lockPath, err)
}
}
// and unlock
if err := actionNode.UnlockShard(ctx, agent.TopoServer, *initKeyspace, shard, lockPath, nil); err != nil {
return err
}
}
}
log.Infof("Initializing the tablet for type %v", tabletType)
// figure out the hostname
hostname := *tabletHostname
if hostname == "" {
var err error
hostname, err = netutil.FullyQualifiedHostname()
if err != nil {
return err
}
}
// create and populate tablet record
tablet := &topo.Tablet{
Alias: agent.TabletAlias,
Hostname: hostname,
Portmap: make(map[string]int),
Keyspace: *initKeyspace,
Shard: *initShard,
Type: tabletType,
DbNameOverride: *initDbNameOverride,
Tags: initTags,
}
if port != 0 {
tablet.Portmap["vt"] = port
}
if securePort != 0 {
tablet.Portmap["vts"] = securePort
}
if gRPCPort != 0 {
tablet.Portmap["grpc"] = gRPCPort
}
if err := tablet.Complete(); err != nil {
return fmt.Errorf("InitTablet tablet.Complete failed: %v", err)
}
// now try to create the record
err := topo.CreateTablet(ctx, agent.TopoServer, tablet)
switch err {
case nil:
// it worked, we're good, can update the replication graph
if tablet.IsInReplicationGraph() {
if err := topo.UpdateTabletReplicationData(ctx, agent.TopoServer, tablet); err != nil {
return fmt.Errorf("UpdateTabletReplicationData failed: %v", err)
}
}
case topo.ErrNodeExists:
// The node already exists, will just try to update
// it. So we read it first.
oldTablet, err := agent.TopoServer.GetTablet(ctx, tablet.Alias)
if err != nil {
fmt.Errorf("InitTablet failed to read existing tablet record: %v", err)
}
// Sanity check the keyspace and shard
if oldTablet.Keyspace != tablet.Keyspace || oldTablet.Shard != tablet.Shard {
return fmt.Errorf("InitTablet failed because existing tablet keyspace and shard %v/%v differ from the provided ones %v/%v", oldTablet.Keyspace, oldTablet.Shard, tablet.Keyspace, tablet.Shard)
}
// And overwrite the rest
*(oldTablet.Tablet) = *tablet
if err := topo.UpdateTablet(ctx, agent.TopoServer, oldTablet); err != nil {
return fmt.Errorf("UpdateTablet failed: %v", err)
}
// Note we don't need to UpdateTabletReplicationData
// as the tablet already existed with the right data
// in the replication graph
default:
return fmt.Errorf("CreateTablet failed: %v", err)
}
// and now update the serving graph. Note we do that in any case,
// to clean any inaccurate record from any part of the serving graph.
if tabletType != topo.TYPE_IDLE {
if err := topotools.UpdateTabletEndpoints(ctx, agent.TopoServer, tablet); err != nil {
return fmt.Errorf("UpdateTabletEndpoints failed: %v", err)
}
}
return nil
}
// migrateServedTypes operates with all concerned shards locked.
func (wr *Wrangler) migrateServedTypes(sourceShards, destinationShards []*topo.ShardInfo, servedType topo.TabletType, reverse bool) error {
// re-read all the shards so we are up to date
var err error
for i, si := range sourceShards {
if sourceShards[i], err = wr.ts.GetShard(si.Keyspace(), si.ShardName()); err != nil {
return err
}
}
for i, si := range destinationShards {
if destinationShards[i], err = wr.ts.GetShard(si.Keyspace(), si.ShardName()); err != nil {
return err
}
}
// check and update all shard records, in memory only
for _, si := range sourceShards {
if reverse {
// need to add to source
if topo.IsTypeInList(servedType, si.ServedTypes) {
return fmt.Errorf("Source shard %v/%v is already serving type %v", si.Keyspace(), si.ShardName(), servedType)
}
si.ServedTypes = append(si.ServedTypes, servedType)
} else {
// need to remove from source
var found bool
if si.ServedTypes, found = removeType(servedType, si.ServedTypes); !found {
return fmt.Errorf("Source shard %v/%v is not serving type %v", si.Keyspace(), si.ShardName(), servedType)
}
}
}
for _, si := range destinationShards {
if reverse {
// need to remove from destination
var found bool
if si.ServedTypes, found = removeType(servedType, si.ServedTypes); !found {
return fmt.Errorf("Destination shard %v/%v is not serving type %v", si.Keyspace(), si.ShardName(), servedType)
}
} else {
// need to add to destination
if topo.IsTypeInList(servedType, si.ServedTypes) {
return fmt.Errorf("Destination shard %v/%v is already serving type %v", si.Keyspace(), si.ShardName(), servedType)
}
si.ServedTypes = append(si.ServedTypes, servedType)
}
}
// For master type migration, need to:
// - switch the source shards to read-only
// - gather all replication points
// - wait for filtered replication to catch up before we continue
// - disable filtered replication after the fact
if servedType == topo.TYPE_MASTER {
err := wr.makeMastersReadOnly(sourceShards)
if err != nil {
return err
}
masterPositions, err := wr.getMastersPosition(sourceShards)
if err != nil {
return err
}
if err := wr.waitForFilteredReplication(masterPositions, destinationShards); err != nil {
return err
}
for _, si := range destinationShards {
si.SourceShards = nil
}
}
// All is good, we can save the shards now
for _, si := range sourceShards {
if err := wr.ts.UpdateShard(si); err != nil {
return err
}
}
for _, si := range destinationShards {
if err := wr.ts.UpdateShard(si); err != nil {
return err
}
}
// And tell the new shards masters they can now be read-write.
// Invoking a remote action will also make the tablet stop filtered
// replication.
if servedType == topo.TYPE_MASTER {
if err := wr.makeMastersReadWrite(destinationShards); err != nil {
return err
}
}
return nil
}
// MigrateServedTypes is used during horizontal splits to migrate a
// served type from a list of shards to another.
func (wr *Wrangler) MigrateServedTypes(keyspace, shard string, servedType topo.TabletType, reverse, skipRebuild bool) error {
if servedType == topo.TYPE_MASTER {
// we cannot migrate a master back, since when master migration
// is done, the source shards are dead
if reverse {
return fmt.Errorf("Cannot migrate master back to %v/%v", keyspace, shard)
}
// we cannot skip rebuild for a master
if skipRebuild {
return fmt.Errorf("Cannot skip rebuild for master migration on %v/%v", keyspace, shard)
}
}
// find overlapping shards in this keyspace
wr.Logger().Infof("Finding the overlapping shards in keyspace %v", keyspace)
osList, err := topotools.FindOverlappingShards(wr.ts, keyspace)
if err != nil {
return fmt.Errorf("FindOverlappingShards failed: %v", err)
}
// find our shard in there
os := topotools.OverlappingShardsForShard(osList, shard)
if os == nil {
return fmt.Errorf("Shard %v is not involved in any overlapping shards", shard)
}
// find which list is which: the sources have no source
// shards, the destination have source shards. We check the
// first entry in the lists, then just check they're
// consistent
var sourceShards []*topo.ShardInfo
var destinationShards []*topo.ShardInfo
if len(os.Left[0].SourceShards) == 0 {
sourceShards = os.Left
destinationShards = os.Right
} else {
sourceShards = os.Right
destinationShards = os.Left
}
// Verify the sources has the type we're migrating (or not if reverse)
for _, si := range sourceShards {
foundType := topo.IsTypeInList(servedType, si.ServedTypes)
if reverse {
if foundType {
return fmt.Errorf("Source shard %v/%v is already serving type %v", si.Keyspace(), si.ShardName(), servedType)
}
} else {
if !foundType {
return fmt.Errorf("Source shard %v/%v is not serving type %v", si.Keyspace, si.ShardName(), servedType)
}
}
if servedType == topo.TYPE_MASTER && len(si.ServedTypes) > 1 {
return fmt.Errorf("Cannot migrate master out of %v/%v until everything else is migrated out", si.Keyspace(), si.ShardName())
}
}
// Verify the destinations do not have the type we're
// migrating (or do if reverse)
for _, si := range destinationShards {
foundType := topo.IsTypeInList(servedType, si.ServedTypes)
if reverse {
if !foundType {
return fmt.Errorf("Destination shard %v/%v is not serving type %v", si.Keyspace, si.ShardName(), servedType)
}
} else {
if foundType {
return fmt.Errorf("Destination shard %v/%v is already serving type %v", si.Keyspace(), si.ShardName(), servedType)
}
}
}
// lock the shards: sources, then destinations
// (note they're all ordered by shard name)
actionNode := actionnode.MigrateServedTypes(servedType)
sourceLockPath := make([]string, len(sourceShards))
for i, si := range sourceShards {
sourceLockPath[i], err = wr.lockShard(si.Keyspace(), si.ShardName(), actionNode)
if err != nil {
wr.Logger().Errorf("Failed to lock source shard %v/%v, may need to unlock other shards manually", si.Keyspace(), si.ShardName())
return err
}
}
destinationLockPath := make([]string, len(destinationShards))
for i, si := range destinationShards {
destinationLockPath[i], err = wr.lockShard(si.Keyspace(), si.ShardName(), actionNode)
if err != nil {
wr.Logger().Errorf("Failed to lock destination shard %v/%v, may need to unlock other shards manually", si.Keyspace(), si.ShardName())
return err
}
}
// record the action error and all unlock errors
rec := concurrency.AllErrorRecorder{}
// execute the migration
shardCache := make(map[string]*topo.ShardInfo)
rec.RecordError(wr.migrateServedTypes(keyspace, sourceShards, destinationShards, servedType, reverse, shardCache))
// unlock the shards, we're done
for i := len(destinationShards) - 1; i >= 0; i-- {
rec.RecordError(wr.unlockShard(destinationShards[i].Keyspace(), destinationShards[i].ShardName(), actionNode, destinationLockPath[i], nil))
}
for i := len(sourceShards) - 1; i >= 0; i-- {
rec.RecordError(wr.unlockShard(sourceShards[i].Keyspace(), sourceShards[i].ShardName(), actionNode, sourceLockPath[i], nil))
}
// rebuild the keyspace serving graph if there was no error
if rec.Error() == nil {
if skipRebuild {
wr.Logger().Infof("Skipping keyspace rebuild, please run it at earliest convenience")
} else {
rec.RecordError(wr.RebuildKeyspaceGraph(keyspace, nil, shardCache))
}
}
return rec.Error()
}