// FixShardReplication will fix the first problem it encounters within
// a ShardReplication object
func FixShardReplication(ctx context.Context, ts Server, logger logutil.Logger, cell, keyspace, shard string) error {
sri, err := ts.GetShardReplication(ctx, cell, keyspace, shard)
if err != nil {
return err
}
for _, node := range sri.Nodes {
ti, err := ts.GetTablet(ctx, node.TabletAlias)
if err == ErrNoNode {
logger.Warningf("Tablet %v is in the replication graph, but does not exist, removing it", node.TabletAlias)
return RemoveShardReplicationRecord(ctx, ts, cell, keyspace, shard, node.TabletAlias)
}
if err != nil {
// unknown error, we probably don't want to continue
return err
}
if ti.Type == pb.TabletType_SCRAP {
logger.Warningf("Tablet %v is in the replication graph, but is scrapped, removing it", node.TabletAlias)
return RemoveShardReplicationRecord(ctx, ts, cell, keyspace, shard, node.TabletAlias)
}
logger.Infof("Keeping tablet %v in the replication graph", node.TabletAlias)
}
logger.Infof("All entries in replication graph are valid")
return nil
}
// RestartSlavesExternal will tell all the slaves in the provided list
// that they have a new master, and also tell all the masters. The
// masters will be scrapped if they don't answer.
// We execute all the actions in parallel.
func RestartSlavesExternal(ts topo.Server, log logutil.Logger, slaveTabletMap, masterTabletMap map[topo.TabletAlias]*topo.TabletInfo, masterElectTabletAlias topo.TabletAlias, slaveWasRestarted func(*topo.TabletInfo, *actionnode.SlaveWasRestartedArgs) error) {
wg := sync.WaitGroup{}
swrd := actionnode.SlaveWasRestartedArgs{
Parent: masterElectTabletAlias,
}
log.Infof("Updating individual tablets with the right master...")
// do all the slaves
for _, ti := range slaveTabletMap {
wg.Add(1)
go func(ti *topo.TabletInfo) {
if err := slaveWasRestarted(ti, &swrd); err != nil {
log.Warningf("Slave %v had an error: %v", ti.Alias, err)
}
wg.Done()
}(ti)
}
// and do the old master and any straggler, if possible.
for _, ti := range masterTabletMap {
wg.Add(1)
go func(ti *topo.TabletInfo) {
err := slaveWasRestarted(ti, &swrd)
if err != nil {
// the old master can be annoying if left
// around in the replication graph, so if we
// can't restart it, we just scrap it.
// We don't rebuild the Shard just yet though.
log.Warningf("Old master %v is not restarting in time, forcing it to spare: %v", ti.Alias, err)
ti.Type = topo.TYPE_SPARE
ti.Parent = masterElectTabletAlias
if err := topo.UpdateTablet(ts, ti); err != nil {
log.Warningf("Failed to change old master %v to spare: %v", ti.Alias, err)
}
}
wg.Done()
}(ti)
}
wg.Wait()
}
// RestartSlavesExternal will tell all the slaves in the provided list
// that they have a new master, and also tell all the masters. The
// masters will be scrapped if they don't answer.
// We execute all the actions in parallel.
func RestartSlavesExternal(ts topo.Server, log logutil.Logger, slaveTabletMap, masterTabletMap map[topodatapb.TabletAlias]*topo.TabletInfo, masterElectTabletAlias *topodatapb.TabletAlias, slaveWasRestarted func(*topo.TabletInfo, *topodatapb.TabletAlias) error) {
wg := sync.WaitGroup{}
log.Infof("Updating individual tablets with the right master...")
// do all the slaves
for _, ti := range slaveTabletMap {
wg.Add(1)
go func(ti *topo.TabletInfo) {
if err := slaveWasRestarted(ti, masterElectTabletAlias); err != nil {
log.Warningf("Slave %v had an error: %v", ti.Alias, err)
}
wg.Done()
}(ti)
}
// and do the old master and any straggler, if possible.
for _, ti := range masterTabletMap {
wg.Add(1)
go func(ti *topo.TabletInfo) {
err := slaveWasRestarted(ti, masterElectTabletAlias)
if err != nil {
// the old master can be annoying if left
// around in the replication graph, so if we
// can't restart it, we just make it spare.
log.Warningf("Old master %v is not restarting in time, forcing it to spare: %v", ti.Alias, err)
ti.Type = topodatapb.TabletType_SPARE
if err := ts.UpdateTablet(context.TODO(), ti); err != nil {
log.Warningf("Failed to change old master %v to spare: %v", ti.Alias, err)
}
}
wg.Done()
}(ti)
}
wg.Wait()
}
// rebuildCellSrvShard computes and writes the serving graph data to a
// single cell
func rebuildCellSrvShard(ctx context.Context, log logutil.Logger, ts topo.Server, si *topo.ShardInfo, cell string) (err error) {
log.Infof("rebuildCellSrvShard %v/%v in cell %v", si.Keyspace(), si.ShardName(), cell)
for {
select {
case <-ctx.Done():
return ctx.Err()
default:
}
// Read existing EndPoints node versions, so we know if any
// changes sneak in after we read the tablets.
versions, err := getEndPointsVersions(ctx, ts, cell, si.Keyspace(), si.ShardName())
// Get all tablets in this cell/shard.
tablets, err := ts.GetTabletMapForShardByCell(ctx, si.Keyspace(), si.ShardName(), []string{cell})
if err != nil {
if err != topo.ErrPartialResult {
return err
}
log.Warningf("Got ErrPartialResult from topo.GetTabletMapForShardByCell(%v), some tablets may not be added properly to serving graph", cell)
}
// Build up the serving graph from scratch.
serving := make(map[pb.TabletType]*pb.EndPoints)
for _, tablet := range tablets {
// Only add serving types.
if !tablet.IsInServingGraph() {
continue
}
// Check the Keyspace and Shard for the tablet are right.
if tablet.Keyspace != si.Keyspace() || tablet.Shard != si.ShardName() {
return fmt.Errorf("CRITICAL: tablet %v is in replication graph for shard %v/%v but belongs to shard %v:%v", tablet.Alias, si.Keyspace(), si.ShardName(), tablet.Keyspace, tablet.Shard)
}
// Add the tablet to the list.
endpoints, ok := serving[tablet.Type]
if !ok {
endpoints = topo.NewEndPoints()
serving[tablet.Type] = endpoints
}
entry, err := topo.TabletEndPoint(tablet.Tablet)
if err != nil {
log.Warningf("EndPointForTablet failed for tablet %v: %v", tablet.Alias, err)
continue
}
endpoints.Entries = append(endpoints.Entries, entry)
}
wg := sync.WaitGroup{}
fatalErrs := concurrency.AllErrorRecorder{}
retryErrs := concurrency.AllErrorRecorder{}
// Write nodes that should exist.
for tabletType, endpoints := range serving {
wg.Add(1)
go func(tabletType pb.TabletType, endpoints *pb.EndPoints) {
defer wg.Done()
log.Infof("saving serving graph for cell %v shard %v/%v tabletType %v", cell, si.Keyspace(), si.ShardName(), tabletType)
version, ok := versions[tabletType]
if !ok {
// This type didn't exist when we first checked.
// Try to create, but only if it still doesn't exist.
if err := ts.CreateEndPoints(ctx, cell, si.Keyspace(), si.ShardName(), tabletType, endpoints); err != nil {
log.Warningf("CreateEndPoints(%v, %v, %v) failed during rebuild: %v", cell, si, tabletType, err)
switch err {
case topo.ErrNodeExists:
retryErrs.RecordError(err)
default:
fatalErrs.RecordError(err)
}
}
return
}
// Update only if the version matches.
if err := ts.UpdateEndPoints(ctx, cell, si.Keyspace(), si.ShardName(), tabletType, endpoints, version); err != nil {
log.Warningf("UpdateEndPoints(%v, %v, %v) failed during rebuild: %v", cell, si, tabletType, err)
switch err {
case topo.ErrBadVersion, topo.ErrNoNode:
retryErrs.RecordError(err)
default:
fatalErrs.RecordError(err)
}
}
}(tabletType, endpoints)
}
// Delete nodes that shouldn't exist.
for tabletType, version := range versions {
if _, ok := serving[tabletType]; !ok {
wg.Add(1)
go func(tabletType pb.TabletType, version int64) {
defer wg.Done()
log.Infof("removing stale db type from serving graph: %v", tabletType)
if err := ts.DeleteEndPoints(ctx, cell, si.Keyspace(), si.ShardName(), tabletType, version); err != nil && err != topo.ErrNoNode {
log.Warningf("DeleteEndPoints(%v, %v, %v) failed during rebuild: %v", cell, si, tabletType, err)
switch err {
case topo.ErrNoNode:
// Someone else deleted it, which is fine.
case topo.ErrBadVersion:
retryErrs.RecordError(err)
default:
fatalErrs.RecordError(err)
}
}
}(tabletType, version)
}
}
// Update srvShard object
wg.Add(1)
go func() {
defer wg.Done()
log.Infof("updating shard serving graph in cell %v for %v/%v", cell, si.Keyspace(), si.ShardName())
if err := UpdateSrvShard(ctx, ts, cell, si); err != nil {
fatalErrs.RecordError(err)
log.Warningf("writing serving data in cell %v for %v/%v failed: %v", cell, si.Keyspace(), si.ShardName(), err)
}
}()
wg.Wait()
// If there are any fatal errors, give up.
if fatalErrs.HasErrors() {
return fatalErrs.Error()
}
// If there are any retry errors, try again.
if retryErrs.HasErrors() {
continue
}
// Otherwise, success!
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()
}
// rebuildCellSrvShard computes and writes the serving graph data to a
// single cell
func rebuildCellSrvShard(log logutil.Logger, ts topo.Server, shardInfo *topo.ShardInfo, cell string, tablets map[topo.TabletAlias]*topo.TabletInfo) error {
log.Infof("rebuildCellSrvShard %v/%v in cell %v", shardInfo.Keyspace(), shardInfo.ShardName(), cell)
// Get all existing db types so they can be removed if nothing
// had been edited.
existingTabletTypes, err := ts.GetSrvTabletTypesPerShard(cell, shardInfo.Keyspace(), shardInfo.ShardName())
if err != nil {
if err != topo.ErrNoNode {
return err
}
}
// Update db type addresses in the serving graph
//
// locationAddrsMap is a map:
// key: tabletType
// value: EndPoints (list of server records)
locationAddrsMap := make(map[topo.TabletType]*topo.EndPoints)
for _, tablet := range tablets {
if !tablet.IsInReplicationGraph() {
// only valid case is a scrapped master in the
// catastrophic reparent case
if tablet.Parent.Uid != topo.NO_TABLET {
log.Warningf("Tablet %v should not be in the replication graph, please investigate (it is being ignored in the rebuild)", tablet.Alias)
}
continue
}
// Check IsInServingGraph, we don't want to add tablets that
// are not serving
if !tablet.IsInServingGraph() {
continue
}
// Check the Keyspace and Shard for the tablet are right
if tablet.Keyspace != shardInfo.Keyspace() || tablet.Shard != shardInfo.ShardName() {
return fmt.Errorf("CRITICAL: tablet %v is in replication graph for shard %v/%v but belongs to shard %v:%v", tablet.Alias, shardInfo.Keyspace(), shardInfo.ShardName(), tablet.Keyspace, tablet.Shard)
}
// Add the tablet to the list
addrs, ok := locationAddrsMap[tablet.Type]
if !ok {
addrs = topo.NewEndPoints()
locationAddrsMap[tablet.Type] = addrs
}
entry, err := tablet.Tablet.EndPoint()
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:
// - writing all the tabletTypes records
// - removing the unused records
// - writing SrvShard
rec := concurrency.AllErrorRecorder{}
wg := sync.WaitGroup{}
// write all the EndPoints nodes everywhere we want them
for tabletType, addrs := range locationAddrsMap {
wg.Add(1)
go func(tabletType topo.TabletType, addrs *topo.EndPoints) {
log.Infof("saving serving graph for cell %v shard %v/%v tabletType %v", cell, shardInfo.Keyspace(), shardInfo.ShardName(), tabletType)
if err := ts.UpdateEndPoints(cell, shardInfo.Keyspace(), shardInfo.ShardName(), tabletType, addrs); err != nil {
rec.RecordError(fmt.Errorf("writing endpoints for cell %v shard %v/%v tabletType %v failed: %v", cell, shardInfo.Keyspace(), shardInfo.ShardName(), tabletType, err))
}
wg.Done()
}(tabletType, addrs)
}
// Delete any pre-existing paths that were not updated by this process.
// That's the existingTabletTypes - locationAddrsMap
for _, tabletType := range existingTabletTypes {
if _, ok := locationAddrsMap[tabletType]; !ok {
wg.Add(1)
go func(tabletType topo.TabletType) {
log.Infof("removing stale db type from serving graph: %v", tabletType)
if err := ts.DeleteEndPoints(cell, shardInfo.Keyspace(), shardInfo.ShardName(), tabletType); err != nil {
log.Warningf("unable to remove stale db type %v from serving graph: %v", tabletType, err)
}
wg.Done()
}(tabletType)
}
}
// Update srvShard object
wg.Add(1)
go func() {
log.Infof("updating shard serving graph in cell %v for %v/%v", cell, shardInfo.Keyspace(), shardInfo.ShardName())
srvShard := &topo.SrvShard{
Name: shardInfo.ShardName(),
KeyRange: shardInfo.KeyRange,
ServedTypes: shardInfo.ServedTypes,
MasterCell: shardInfo.MasterAlias.Cell,
TabletTypes: make([]topo.TabletType, 0, len(locationAddrsMap)),
}
for tabletType := range locationAddrsMap {
srvShard.TabletTypes = append(srvShard.TabletTypes, tabletType)
}
if err := ts.UpdateSrvShard(cell, shardInfo.Keyspace(), shardInfo.ShardName(), srvShard); err != nil {
rec.RecordError(fmt.Errorf("writing serving data in cell %v for %v/%v failed: %v", cell, shardInfo.Keyspace(), shardInfo.ShardName(), err))
}
wg.Done()
}()
wg.Wait()
return rec.Error()
}
