// MigrateServedFrom is used during vertical splits to migrate a
// served type from a keyspace to another.
func (wr *Wrangler) MigrateServedFrom(ctx context.Context, keyspace, shard string, servedType topodatapb.TabletType, cells []string, reverse bool, filteredReplicationWaitTime time.Duration) (err error) {
// read the destination keyspace, check it
ki, err := wr.ts.GetKeyspace(ctx, keyspace)
if err != nil {
return err
}
if len(ki.ServedFroms) == 0 {
return fmt.Errorf("Destination keyspace %v is not a vertical split target", keyspace)
}
// read the destination shard, check it
si, err := wr.ts.GetShard(ctx, keyspace, shard)
if err != nil {
return err
}
if len(si.SourceShards) != 1 || len(si.SourceShards[0].Tables) == 0 {
return fmt.Errorf("Destination shard %v/%v is not a vertical split target", keyspace, shard)
}
// check the migration is valid before locking (will also be checked
// after locking to be sure)
sourceKeyspace := si.SourceShards[0].Keyspace
if err := ki.CheckServedFromMigration(servedType, cells, sourceKeyspace, !reverse); err != nil {
return err
}
// lock the keyspaces, source first.
ctx, unlock, lockErr := wr.ts.LockKeyspace(ctx, sourceKeyspace, fmt.Sprintf("MigrateServedFrom(%v)", servedType))
if lockErr != nil {
return lockErr
}
defer unlock(&err)
ctx, unlock, lockErr = wr.ts.LockKeyspace(ctx, keyspace, fmt.Sprintf("MigrateServedFrom(%v)", servedType))
if lockErr != nil {
return lockErr
}
defer unlock(&err)
// execute the migration
err = wr.migrateServedFromLocked(ctx, ki, si, servedType, cells, reverse, filteredReplicationWaitTime)
// rebuild the keyspace serving graph if there was no error
if err == nil {
err = topotools.RebuildKeyspaceLocked(ctx, wr.logger, wr.ts, keyspace, cells)
}
return err
}
// MigrateServedTypes is used during horizontal splits to migrate a
// served type from a list of shards to another.
func (wr *Wrangler) MigrateServedTypes(ctx context.Context, keyspace, shard string, cells []string, servedType topodatapb.TabletType, reverse, skipReFreshState bool, filteredReplicationWaitTime time.Duration) (err error) {
// check input parameters
if servedType == topodatapb.TabletType_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 refresh state for a master
if skipReFreshState {
return fmt.Errorf("Cannot skip refresh state for master migration on %v/%v", keyspace, shard)
}
}
// lock the keyspace
ctx, unlock, lockErr := wr.ts.LockKeyspace(ctx, keyspace, fmt.Sprintf("MigrateServedTypes(%v)", servedType))
if lockErr != nil {
return lockErr
}
defer unlock(&err)
// find overlapping shards in this keyspace
wr.Logger().Infof("Finding the overlapping shards in keyspace %v", keyspace)
osList, err := topotools.FindOverlappingShards(ctx, 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 {
if err := si.CheckServedTypesMigration(servedType, cells, !reverse); err != nil {
return err
}
}
// Verify the destinations do not have the type we're
// migrating (or do if reverse)
for _, si := range destinationShards {
if err := si.CheckServedTypesMigration(servedType, cells, reverse); err != nil {
return err
}
}
// execute the migration
if err = wr.migrateServedTypesLocked(ctx, keyspace, sourceShards, destinationShards, cells, servedType, reverse, filteredReplicationWaitTime); err != nil {
return err
}
// rebuild the keyspace serving graph now that there is no error
if err = topotools.RebuildKeyspaceLocked(ctx, wr.logger, wr.ts, keyspace, cells); err != nil {
return err
}
// Send a refresh to the tablets we just disabled, iff:
// - we're not migrating a master
// - we don't have any errors
// - we're not told to skip the refresh
if servedType != topodatapb.TabletType_MASTER && !skipReFreshState {
rec := concurrency.AllErrorRecorder{}
var refreshShards []*topo.ShardInfo
if reverse {
// For a backwards migration, we just disabled query service on the destination shards
refreshShards = destinationShards
} else {
// For a forwards migration, we just disabled query service on the source shards
refreshShards = sourceShards
}
for _, si := range refreshShards {
rec.RecordError(wr.RefreshTabletsByShard(ctx, si, servedType, cells))
}
return rec.Error()
}
return nil
}
// MigrateServedTypes is used during horizontal splits to migrate a
// served type from a list of shards to another.
func (wr *Wrangler) MigrateServedTypes(ctx context.Context, keyspace, shard string, cells []string, servedType topodatapb.TabletType, reverse, skipReFreshState bool, filteredReplicationWaitTime time.Duration) (err error) {
// check input parameters
if servedType == topodatapb.TabletType_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 refresh state for a master
if skipReFreshState {
return fmt.Errorf("Cannot skip refresh state for master migration on %v/%v", keyspace, shard)
}
}
// lock the keyspace
ctx, unlock, lockErr := wr.ts.LockKeyspace(ctx, keyspace, fmt.Sprintf("MigrateServedTypes(%v)", servedType))
if lockErr != nil {
return lockErr
}
defer unlock(&err)
// find overlapping shards in this keyspace
wr.Logger().Infof("Finding the overlapping shards in keyspace %v", keyspace)
osList, err := topotools.FindOverlappingShards(ctx, 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 {
if len(os.Right[0].SourceShards) == 0 {
return fmt.Errorf("neither Shard '%v' nor Shard '%v' have a 'SourceShards' entry. Did you successfully run vtworker SplitClone before? Or did you already migrate the MASTER type?", os.Left[0].ShardName(), os.Right[0].ShardName())
}
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 {
if err := si.CheckServedTypesMigration(servedType, cells, !reverse); err != nil {
return err
}
}
// Verify the destinations do not have the type we're
// migrating (or do if reverse)
for _, si := range destinationShards {
if err := si.CheckServedTypesMigration(servedType, cells, reverse); err != nil {
return err
}
}
// execute the migration
if err = wr.migrateServedTypesLocked(ctx, keyspace, sourceShards, destinationShards, cells, servedType, reverse, filteredReplicationWaitTime); err != nil {
return err
}
// rebuild the keyspace serving graph now that there is no error
if err = topotools.RebuildKeyspaceLocked(ctx, wr.logger, wr.ts, keyspace, cells); err != nil {
return err
}
// Send a refresh to the tablets we just disabled, iff:
// - we're not migrating a master
// - we don't have any errors
// - we're not told to skip the refresh
if servedType != topodatapb.TabletType_MASTER && !skipReFreshState {
rec := concurrency.AllErrorRecorder{}
var refreshShards []*topo.ShardInfo
if reverse {
// For a backwards migration, we just disabled query service on the destination shards
refreshShards = destinationShards
} else {
// For a forwards migration, we just disabled query service on the source shards
refreshShards = sourceShards
}
// TODO(b/26388813): Integrate vtctl WaitForDrain here instead of just sleeping.
var waitForDrainSleep time.Duration
switch servedType {
case topodatapb.TabletType_RDONLY:
waitForDrainSleep = *waitForDrainSleepRdonly
case topodatapb.TabletType_REPLICA:
waitForDrainSleep = *waitForDrainSleepReplica
default:
wr.Logger().Warningf("invalid TabletType: %v for MigrateServedTypes command", servedType)
}
wr.Logger().Infof("WaitForDrain: Sleeping for %.0f seconds before shutting down query service on old tablets...", waitForDrainSleep.Seconds())
time.Sleep(waitForDrainSleep)
wr.Logger().Infof("WaitForDrain: Sleeping finished. Shutting down queryservice on old tablets now.")
for _, si := range refreshShards {
rec.RecordError(wr.RefreshTabletsByShard(ctx, si, []topodatapb.TabletType{servedType}, cells))
}
return rec.Error()
}
return nil
}