// 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 *topodatapb.TabletAlias, tabletType topodatapb.TabletType, allowMasterOverride 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 == topodatapb.TabletType_MASTER && !topoproto.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 == topodatapb.TabletType_MASTER && !topoproto.TabletAliasEqual(si.MasterAlias, tabletAlias) {
if si.HasMaster() && !allowMasterOverride {
return wr.unlockShard(ctx, keyspace, shard, actionNode, lockPath, fmt.Errorf("creating this tablet would override old master %v in shard %v/%v", topoproto.TabletAliasString(si.MasterAlias), keyspace, shard))
}
si.MasterAlias = tabletAlias
wasUpdated = true
}
if wasUpdated {
// write it back
if err := wr.ts.UpdateShard(ctx, si); err != nil {
return wr.unlockShard(ctx, keyspace, shard, actionNode, lockPath, err)
}
}
// and unlock
return wr.unlockShard(ctx, keyspace, shard, actionNode, lockPath, err)
}
// 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 *topodatapb.TabletAlias, tabletType topodatapb.TabletType, allowMasterOverride 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 == topodatapb.TabletType_MASTER && !topoproto.TabletAliasEqual(si.MasterAlias, tabletAlias) {
shardUpdateRequired = true
}
if !shardUpdateRequired {
return nil
}
// we do need to update the shard, lock it to not interfere with
// reparenting operations.
actionNode := actionnode.UpdateShard()
keyspace := si.Keyspace()
shard := si.ShardName()
lockPath, err := wr.lockShard(ctx, keyspace, shard, actionNode)
if err != nil {
return err
}
// run the update
_, err = wr.ts.UpdateShardFields(ctx, keyspace, shard, func(s *topodatapb.Shard) error {
wasUpdated := false
if !topoproto.ShardHasCell(s, tabletAlias.Cell) {
s.Cells = append(s.Cells, tabletAlias.Cell)
wasUpdated = true
}
if tabletType == topodatapb.TabletType_MASTER && !topoproto.TabletAliasEqual(s.MasterAlias, tabletAlias) {
if !topoproto.TabletAliasIsZero(s.MasterAlias) && !allowMasterOverride {
return fmt.Errorf("creating this tablet would override old master %v in shard %v/%v", topoproto.TabletAliasString(s.MasterAlias), keyspace, shard)
}
s.MasterAlias = tabletAlias
wasUpdated = true
}
if !wasUpdated {
return topo.ErrNoUpdateNeeded
}
return nil
})
return wr.unlockShard(ctx, keyspace, shard, actionNode, lockPath, err)
}
// Validate makes sure a tablet is represented correctly in the topology server.
func Validate(ctx context.Context, ts Server, tabletAlias *topodatapb.TabletAlias) error {
// read the tablet record, make sure it parses
tablet, err := ts.GetTablet(ctx, tabletAlias)
if err != nil {
return err
}
if !topoproto.TabletAliasEqual(tablet.Alias, tabletAlias) {
return fmt.Errorf("bad tablet alias data for tablet %v: %#v", topoproto.TabletAliasString(tabletAlias), tablet.Alias)
}
// Validate the entry in the shard replication nodes
if err = ts.ValidateShard(ctx, tablet.Keyspace, tablet.Shard); err != nil {
return err
}
si, err := ts.GetShardReplication(ctx, tablet.Alias.Cell, tablet.Keyspace, tablet.Shard)
if err != nil {
return err
}
if _, err = si.GetShardReplicationNode(tabletAlias); err != nil {
return fmt.Errorf("tablet %v not found in cell %v shard replication: %v", tabletAlias, tablet.Alias.Cell, err)
}
return nil
}
// remove takes in an array and returns it with the specified element removed
// (leaves the array unchanged if element isn't in the array).
func remove(tablets []*discovery.TabletStats, tabletAlias *topodata.TabletAlias) []*discovery.TabletStats {
filteredTablets := tablets[:0]
for _, tablet := range tablets {
if !topoproto.TabletAliasEqual(tablet.Tablet.Alias, tabletAlias) {
filteredTablets = append(filteredTablets, tablet)
}
}
return filteredTablets
}
// ValidatePermissionsKeyspace validates all the permissions are the same
// in a keyspace
func (wr *Wrangler) ValidatePermissionsKeyspace(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.ValidatePermissionsShard(ctx, keyspace, shards[0])
}
// find the reference permissions using the first shard's master
si, err := wr.ts.GetShard(ctx, keyspace, shards[0])
if err != nil {
return err
}
if !si.HasMaster() {
return fmt.Errorf("No master in shard %v/%v", keyspace, shards[0])
}
referenceAlias := si.MasterAlias
log.Infof("Gathering permissions for reference master %v", topoproto.TabletAliasString(referenceAlias))
referencePermissions, err := wr.GetPermissions(ctx, si.MasterAlias)
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 := wr.ts.FindAllTabletAliasesInShard(ctx, keyspace, shard)
if err != nil {
er.RecordError(err)
continue
}
for _, alias := range aliases {
if topoproto.TabletAliasEqual(alias, si.MasterAlias) {
continue
}
wg.Add(1)
go wr.diffPermissions(ctx, referencePermissions, referenceAlias, alias, &wg, &er)
}
}
wg.Wait()
if er.HasErrors() {
return fmt.Errorf("Permissions diffs:\n%v", er.Error().Error())
}
return nil
}
// 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
// allowUpdate is true, and a tablet with the same ID exists, just update it.
// If a tablet is created as master, and there is already a different
// master in the shard, allowMasterOverride must be set.
func (wr *Wrangler) InitTablet(ctx context.Context, tablet *topodatapb.Tablet, allowMasterOverride, createShardAndKeyspace, allowUpdate bool) error {
if err := topo.TabletComplete(tablet); err != nil {
return err
}
// 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 = wr.ts.GetOrCreateShard(ctx, 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.KeyRangeEqual(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 == topodatapb.TabletType_MASTER && si.HasMaster() && !topoproto.TabletAliasEqual(si.MasterAlias, tablet.Alias) && !allowMasterOverride {
return fmt.Errorf("creating this tablet would override old master %v in shard %v/%v, use allow_master_override flag", topoproto.TabletAliasString(si.MasterAlias), tablet.Keyspace, tablet.Shard)
}
// update the shard record if needed
if err := wr.updateShardCellsAndMaster(ctx, si, tablet.Alias, tablet.Type, allowMasterOverride); err != nil {
return err
}
err = wr.ts.CreateTablet(ctx, tablet)
if err == topo.ErrNodeExists && allowUpdate {
// Try to update then
oldTablet, err := wr.ts.GetTablet(ctx, tablet.Alias)
if err != nil {
return fmt.Errorf("failed reading existing tablet %v: %v", topoproto.TabletAliasString(tablet.Alias), err)
}
// Check we have the same keyspace / shard, and if not,
// require the allowDifferentShard flag.
if oldTablet.Keyspace != tablet.Keyspace || oldTablet.Shard != tablet.Shard {
return fmt.Errorf("old tablet has shard %v/%v. Cannot override with shard %v/%v. Delete and re-add tablet if you want to change the tablet's keyspace/shard", oldTablet.Keyspace, oldTablet.Shard, tablet.Keyspace, tablet.Shard)
}
*(oldTablet.Tablet) = *tablet
if err := wr.ts.UpdateTablet(ctx, oldTablet); err != nil {
return fmt.Errorf("failed updating tablet %v: %v", topoproto.TabletAliasString(tablet.Alias), err)
}
}
return nil
}
func (agent *ActionAgent) startReplication(ctx context.Context, pos replication.Position) error {
// Set the position at which to resume from the master.
cmds, err := agent.MysqlDaemon.SetSlavePositionCommands(pos)
if err != nil {
return err
}
if err := agent.MysqlDaemon.ExecuteSuperQueryList(ctx, cmds); err != nil {
return fmt.Errorf("failed to set slave position: %v", err)
}
// Read the shard to find the current master, and its location.
tablet := agent.Tablet()
si, err := agent.TopoServer.GetShard(ctx, tablet.Keyspace, tablet.Shard)
if err != nil {
return fmt.Errorf("can't read shard: %v", err)
}
if si.MasterAlias == nil {
// We've restored, but there's no master. This is fine, since we've
// already set the position at which to resume when we're later reparented.
// If we had instead considered this fatal, all tablets would crash-loop
// until a master appears, which would make it impossible to elect a master.
log.Warningf("Can't start replication after restore: shard %v/%v has no master.", tablet.Keyspace, tablet.Shard)
return nil
}
if topoproto.TabletAliasEqual(si.MasterAlias, tablet.Alias) {
// We used to be the master before we got restarted in an empty data dir,
// and no other master has been elected in the meantime.
// This shouldn't happen, so we'll let the operator decide which tablet
// should actually be promoted to master.
log.Warningf("Can't start replication after restore: master record still points to this tablet.")
return nil
}
ti, err := agent.TopoServer.GetTablet(ctx, si.MasterAlias)
if err != nil {
return fmt.Errorf("Cannot read master tablet %v: %v", si.MasterAlias, err)
}
// If using semi-sync, we need to enable it before connecting to master.
if *enableSemiSync {
if err := agent.enableSemiSync(false); err != nil {
return err
}
}
// Set master and start slave.
cmds, err = agent.MysqlDaemon.SetMasterCommands(ti.Hostname, int(ti.PortMap["mysql"]))
if err != nil {
return fmt.Errorf("MysqlDaemon.SetMasterCommands failed: %v", err)
}
cmds = append(cmds, "START SLAVE")
if err := agent.MysqlDaemon.ExecuteSuperQueryList(ctx, cmds); err != nil {
return fmt.Errorf("failed to start replication: %v", err)
}
return nil
}
// DeleteTablet removes a tablet from a shard.
// - if allowMaster is set, we can Delete a master tablet (and clear
// its record from the Shard record if it was the master).
// - if skipRebuild is set, we do not rebuild the serving graph.
func (wr *Wrangler) DeleteTablet(ctx context.Context, tabletAlias *topodatapb.TabletAlias, allowMaster, skipRebuild bool) error {
// load the tablet, see if we'll need to rebuild
ti, err := wr.ts.GetTablet(ctx, tabletAlias)
if err != nil {
return err
}
rebuildRequired := ti.IsInServingGraph()
wasMaster := ti.Type == topodatapb.TabletType_MASTER
if wasMaster && !allowMaster {
return fmt.Errorf("cannot delete tablet %v as it is a master, use allow_master flag", topoproto.TabletAliasString(tabletAlias))
}
// remove the record and its replication graph entry
if err := topotools.DeleteTablet(ctx, wr.ts, ti.Tablet); err != nil {
return err
}
// update the Shard object if the master was scrapped.
// we lock the shard to not conflict with reparent operations.
if wasMaster {
actionNode := actionnode.UpdateShard()
lockPath, err := wr.lockShard(ctx, ti.Keyspace, ti.Shard, actionNode)
if err != nil {
return err
}
// update the shard record's master
if _, err := wr.ts.UpdateShardFields(ctx, ti.Keyspace, ti.Shard, func(s *topodatapb.Shard) error {
if !topoproto.TabletAliasEqual(s.MasterAlias, tabletAlias) {
wr.Logger().Warningf("Deleting master %v from shard %v/%v but master in Shard object was %v", topoproto.TabletAliasString(tabletAlias), ti.Keyspace, ti.Shard, topoproto.TabletAliasString(s.MasterAlias))
return topo.ErrNoUpdateNeeded
}
s.MasterAlias = nil
return nil
}); err != nil {
return wr.unlockShard(ctx, ti.Keyspace, ti.Shard, actionNode, lockPath, err)
}
// and unlock
if err := wr.unlockShard(ctx, ti.Keyspace, ti.Shard, actionNode, lockPath, err); err != nil {
return err
}
}
// and rebuild the original shard if needed
if !rebuildRequired {
wr.Logger().Infof("Rebuild not required")
return nil
}
if skipRebuild {
wr.Logger().Warningf("Rebuild required, but skipping it")
return nil
}
_, err = wr.RebuildShardGraph(ctx, ti.Keyspace, ti.Shard, []string{ti.Alias.Cell})
return err
}
// FIXME(msolomon) This validate presumes the master is up and running.
// Even when that isn't true, there are validation processes that might be valuable.
func (wr *Wrangler) validateShard(ctx context.Context, keyspace, shard string, pingTablets bool, wg *sync.WaitGroup, results chan<- error) {
shardInfo, err := wr.ts.GetShard(ctx, keyspace, shard)
if err != nil {
results <- fmt.Errorf("TopologyServer.GetShard(%v, %v) failed: %v", keyspace, shard, err)
return
}
aliases, err := wr.ts.FindAllTabletAliasesInShard(ctx, keyspace, shard)
if err != nil {
results <- fmt.Errorf("TopologyServer.FindAllTabletAliasesInShard(%v, %v) failed: %v", keyspace, shard, err)
return
}
tabletMap, _ := wr.ts.GetTabletMap(ctx, aliases)
var masterAlias *pb.TabletAlias
for _, alias := range aliases {
tabletInfo, ok := tabletMap[*alias]
if !ok {
results <- fmt.Errorf("tablet %v not found in map", topoproto.TabletAliasString(alias))
continue
}
if tabletInfo.Type == pb.TabletType_MASTER {
if masterAlias != nil {
results <- fmt.Errorf("shard %v/%v already has master %v but found other master %v", keyspace, shard, topoproto.TabletAliasString(masterAlias), topoproto.TabletAliasString(alias))
} else {
masterAlias = alias
}
}
}
if masterAlias == nil {
results <- fmt.Errorf("no master for shard %v/%v", keyspace, shard)
} else if !topoproto.TabletAliasEqual(shardInfo.MasterAlias, masterAlias) {
results <- fmt.Errorf("master mismatch for shard %v/%v: found %v, expected %v", keyspace, shard, topoproto.TabletAliasString(masterAlias), topoproto.TabletAliasString(shardInfo.MasterAlias))
}
for _, alias := range aliases {
wg.Add(1)
go func(alias *pb.TabletAlias) {
defer wg.Done()
if err := topo.Validate(ctx, wr.ts, alias); err != nil {
results <- fmt.Errorf("Validate(%v) failed: %v", topoproto.TabletAliasString(alias), err)
} else {
wr.Logger().Infof("tablet %v is valid", topoproto.TabletAliasString(alias))
}
}(alias)
}
if pingTablets {
wr.validateReplication(ctx, shardInfo, tabletMap, results)
wr.pingTablets(ctx, tabletMap, wg, results)
}
return
}
// Validate makes sure a tablet is represented correctly in the topology server.
func Validate(ctx context.Context, ts Server, tabletAlias *pb.TabletAlias) error {
// read the tablet record, make sure it parses
tablet, err := ts.GetTablet(ctx, tabletAlias)
if err != nil {
return err
}
if !topoproto.TabletAliasEqual(tablet.Alias, tabletAlias) {
return fmt.Errorf("bad tablet alias data for tablet %v: %#v", topoproto.TabletAliasString(tabletAlias), tablet.Alias)
}
// Some tablets have no information to generate valid replication paths.
// We have three cases to handle:
// - we are a tablet in the replication graph, and should have
// replication data (first case below)
// - we are in scrap mode but used to be assigned in the graph
// somewhere (second case below)
// Idle tablets are just not in any graph at all, we don't even know
// their keyspace / shard to know where to check.
if tablet.IsInReplicationGraph() {
if err = ts.ValidateShard(ctx, tablet.Keyspace, tablet.Shard); err != nil {
return err
}
si, err := ts.GetShardReplication(ctx, tablet.Alias.Cell, tablet.Keyspace, tablet.Shard)
if err != nil {
return err
}
_, err = si.GetShardReplicationNode(tabletAlias)
if err != nil {
return fmt.Errorf("tablet %v not found in cell %v shard replication: %v", tabletAlias, tablet.Alias.Cell, err)
}
} else if tablet.IsAssigned() {
// this case is to make sure a scrap node that used to be in
// a replication graph doesn't leave a node behind.
// However, while an action is running, there is some
// time where this might be inconsistent.
si, err := ts.GetShardReplication(ctx, tablet.Alias.Cell, tablet.Keyspace, tablet.Shard)
if err != nil {
return err
}
node, err := si.GetShardReplicationNode(tabletAlias)
if err != ErrNoNode {
return fmt.Errorf("unexpected replication data found(possible pending action?): %v (%v)", node, tablet.Type)
}
}
return nil
}
// 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 *topodatapb.TabletAlias, tabletType topodatapb.TabletType, allowMasterOverride 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 == topodatapb.TabletType_MASTER && !topoproto.TabletAliasEqual(si.MasterAlias, tabletAlias) {
shardUpdateRequired = true
}
if !shardUpdateRequired {
return nil
}
// run the update
_, err := wr.ts.UpdateShardFields(ctx, si.Keyspace(), si.ShardName(), func(s *topo.ShardInfo) error {
wasUpdated := false
if !s.HasCell(tabletAlias.Cell) {
s.Cells = append(s.Cells, tabletAlias.Cell)
wasUpdated = true
}
if tabletType == topodatapb.TabletType_MASTER && !topoproto.TabletAliasEqual(s.MasterAlias, tabletAlias) {
if !topoproto.TabletAliasIsZero(s.MasterAlias) && !allowMasterOverride {
return fmt.Errorf("creating this tablet would override old master %v in shard %v/%v", topoproto.TabletAliasString(s.MasterAlias), si.Keyspace(), si.ShardName())
}
s.MasterAlias = tabletAlias
wasUpdated = true
}
if !wasUpdated {
return topo.ErrNoUpdateNeeded
}
return nil
})
return err
}
// TabletEquality returns true iff two Tablet are representing the same tablet
// process: same uid/cell, running on the same host / ports.
func TabletEquality(left, right *topodatapb.Tablet) bool {
if !topoproto.TabletAliasEqual(left.Alias, right.Alias) {
return false
}
if left.Hostname != right.Hostname {
return false
}
if len(left.PortMap) != len(right.PortMap) {
return false
}
for key, lvalue := range left.PortMap {
rvalue, ok := right.PortMap[key]
if !ok {
return false
}
if lvalue != rvalue {
return false
}
}
return true
}
// 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 !si.HasMaster() {
return fmt.Errorf("No master in shard %v/%v", keyspace, shard)
}
log.Infof("Gathering schema for master %v", topoproto.TabletAliasString(si.MasterAlias))
masterSchema, err := wr.GetSchema(ctx, 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 := wr.ts.FindAllTabletAliasesInShard(ctx, keyspace, shard)
if err != nil {
return err
}
// then diff with all slaves
er := concurrency.AllErrorRecorder{}
wg := sync.WaitGroup{}
for _, alias := range aliases {
if topoproto.TabletAliasEqual(alias, si.MasterAlias) {
continue
}
wg.Add(1)
go wr.diffSchema(ctx, masterSchema, si.MasterAlias, alias, excludeTables, includeViews, &wg, &er)
}
wg.Wait()
if er.HasErrors() {
return fmt.Errorf("Schema diffs: %v", er.Error().Error())
}
return nil
}
// DeleteTablet removes a tablet from a shard.
// - if allowMaster is set, we can Delete a master tablet (and clear
// its record from the Shard record if it was the master).
func (wr *Wrangler) DeleteTablet(ctx context.Context, tabletAlias *topodatapb.TabletAlias, allowMaster bool) (err error) {
// load the tablet, see if we'll need to rebuild
ti, err := wr.ts.GetTablet(ctx, tabletAlias)
if err != nil {
return err
}
wasMaster := ti.Type == topodatapb.TabletType_MASTER
if wasMaster && !allowMaster {
return fmt.Errorf("cannot delete tablet %v as it is a master, use allow_master flag", topoproto.TabletAliasString(tabletAlias))
}
// remove the record and its replication graph entry
if err := topotools.DeleteTablet(ctx, wr.ts, ti.Tablet); err != nil {
return err
}
// update the Shard object if the master was scrapped.
if wasMaster {
// We lock the shard to not conflict with reparent operations.
ctx, unlock, lockErr := wr.ts.LockShard(ctx, ti.Keyspace, ti.Shard, fmt.Sprintf("DeleteTablet(%v)", topoproto.TabletAliasString(tabletAlias)))
if lockErr != nil {
return lockErr
}
defer unlock(&err)
// update the shard record's master
_, err = wr.ts.UpdateShardFields(ctx, ti.Keyspace, ti.Shard, func(si *topo.ShardInfo) error {
if !topoproto.TabletAliasEqual(si.MasterAlias, tabletAlias) {
wr.Logger().Warningf("Deleting master %v from shard %v/%v but master in Shard object was %v", topoproto.TabletAliasString(tabletAlias), ti.Keyspace, ti.Shard, topoproto.TabletAliasString(si.MasterAlias))
return topo.ErrNoUpdateNeeded
}
si.MasterAlias = nil
return nil
})
return err
}
return nil
}
// 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 !si.HasMaster() {
return fmt.Errorf("No master in shard %v/%v", keyspace, shard)
}
log.Infof("Gathering permissions for master %v", topoproto.TabletAliasString(si.MasterAlias))
masterPermissions, err := wr.GetPermissions(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 := wr.ts.FindAllTabletAliasesInShard(ctx, 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 topoproto.TabletAliasEqual(alias, si.MasterAlias) {
continue
}
wg.Add(1)
go wr.diffPermissions(ctx, masterPermissions, si.MasterAlias, alias, &wg, &er)
}
wg.Wait()
if er.HasErrors() {
return fmt.Errorf("Permissions diffs:\n%v", er.Error().Error())
}
return nil
}
// chooseNewMaster finds a tablet that is going to become master after reparent. The criterias
// for the new master-elect are (preferably) to be in the same cell as the current master, and
// to be different from avoidMasterTabletAlias. The tablet with the largest replication
// position is chosen to minimize the time of catching up with the master. Note that the search
// for largest replication position will race with transactions being executed on the master at
// the same time, so when all tablets are roughly at the same position then the choice of the
// new master-elect will be somewhat unpredictable.
func (wr *Wrangler) chooseNewMaster(
ctx context.Context,
shardInfo *topo.ShardInfo,
tabletMap map[topodatapb.TabletAlias]*topo.TabletInfo,
avoidMasterTabletAlias *topodatapb.TabletAlias,
waitSlaveTimeout time.Duration) (*topodatapb.TabletAlias, error) {
if avoidMasterTabletAlias == nil {
return nil, fmt.Errorf("tablet to avoid for reparent is not provided, cannot choose new master")
}
var masterCell string
if shardInfo.MasterAlias != nil {
masterCell = shardInfo.MasterAlias.Cell
}
maxPosSearch := maxReplPosSearch{
wrangler: wr,
ctx: ctx,
waitSlaveTimeout: waitSlaveTimeout,
waitGroup: sync.WaitGroup{},
maxPosLock: sync.Mutex{},
}
for tabletAlias, tabletInfo := range tabletMap {
if (masterCell != "" && tabletAlias.Cell != masterCell) ||
topoproto.TabletAliasEqual(&tabletAlias, avoidMasterTabletAlias) {
continue
}
maxPosSearch.waitGroup.Add(1)
go maxPosSearch.processTablet(tabletInfo.Tablet)
}
maxPosSearch.waitGroup.Wait()
if maxPosSearch.maxPosTablet == nil {
return nil, nil
}
return maxPosSearch.maxPosTablet.Alias, nil
}
func (wr *Wrangler) plannedReparentShardLocked(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 tablet map")
tabletMap, err := wr.ts.GetTabletMapForShard(ctx, 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", topoproto.TabletAliasString(masterElectTabletAlias))
}
ev.NewMaster = *masterElectTabletInfo.Tablet
if topoproto.TabletAliasEqual(shardInfo.MasterAlias, masterElectTabletAlias) {
return fmt.Errorf("master-elect tablet %v is already the master", topoproto.TabletAliasString(masterElectTabletAlias))
}
oldMasterTabletInfo, ok := tabletMap[*shardInfo.MasterAlias]
if !ok {
return fmt.Errorf("old master tablet %v is not in the shard", topoproto.TabletAliasString(shardInfo.MasterAlias))
}
ev.OldMaster = *oldMasterTabletInfo.Tablet
// Demote the current master, get its replication position
wr.logger.Infof("demote current master %v", shardInfo.MasterAlias)
event.DispatchUpdate(ev, "demoting old master")
rp, err := wr.tmc.DemoteMaster(ctx, oldMasterTabletInfo)
if err != nil {
return fmt.Errorf("old master tablet %v DemoteMaster failed: %v", topoproto.TabletAliasString(shardInfo.MasterAlias), err)
}
// Wait on the master-elect tablet until it reaches that position,
// then promote it
wr.logger.Infof("promote slave %v", topoproto.TabletAliasString(masterElectTabletAlias))
event.DispatchUpdate(ev, "promoting slave")
rp, err = wr.tmc.PromoteSlaveWhenCaughtUp(ctx, masterElectTabletInfo, rp)
if err != nil {
return fmt.Errorf("master-elect tablet %v failed to catch up with replication or be upgraded to master: %v", topoproto.TabletAliasString(masterElectTabletAlias), err)
}
// 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 topoproto.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", topoproto.TabletAliasString(&alias))
masterErr = wr.TabletManagerClient().PopulateReparentJournal(ctx, tabletInfo, now, plannedReparentShardOperation, &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", topoproto.TabletAliasString(&alias))
// also restart replication on old master
forceStartSlave := topoproto.TabletAliasEqual(&alias, oldMasterTabletInfo.Alias)
if err := wr.TabletManagerClient().SetMaster(ctx, tabletInfo, masterElectTabletAlias, now, forceStartSlave); err != nil {
rec.RecordError(fmt.Errorf("Tablet %v SetMaster failed: %v", topoproto.TabletAliasString(&alias), err))
return
}
}(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 := wr.ts.UpdateShard(ctx, 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
}
func (wr *Wrangler) initShardMasterLocked(ctx context.Context, ev *events.Reparent, keyspace, shard string, masterElectTabletAlias *pb.TabletAlias, force bool, waitSlaveTimeout time.Duration) error {
shardInfo, err := wr.ts.GetShard(ctx, keyspace, shard)
if err != nil {
return err
}
ev.ShardInfo = *shardInfo
event.DispatchUpdate(ev, "reading tablet map")
tabletMap, err := wr.ts.GetTabletMapForShard(ctx, keyspace, shard)
if err != nil {
return err
}
// Check the master elect is in tabletMap
masterElectTabletInfo, ok := tabletMap[*masterElectTabletAlias]
if !ok {
return fmt.Errorf("master-elect tablet %v is not in the shard", topoproto.TabletAliasString(masterElectTabletAlias))
}
ev.NewMaster = *masterElectTabletInfo.Tablet
// Check the master is the only master is the shard, or -force was used.
_, masterTabletMap := topotools.SortedTabletMap(tabletMap)
if !topoproto.TabletAliasEqual(shardInfo.MasterAlias, masterElectTabletAlias) {
if !force {
return fmt.Errorf("master-elect tablet %v is not the shard master, use -force to proceed anyway", topoproto.TabletAliasString(masterElectTabletAlias))
}
wr.logger.Warningf("master-elect tablet %v is not the shard master, proceeding anyway as -force was used", topoproto.TabletAliasString(masterElectTabletAlias))
}
if _, ok := masterTabletMap[*masterElectTabletAlias]; !ok {
if !force {
return fmt.Errorf("master-elect tablet %v is not a master in the shard, use -force to proceed anyway", topoproto.TabletAliasString(masterElectTabletAlias))
}
wr.logger.Warningf("master-elect tablet %v is not a master in the shard, proceeding anyway as -force was used", topoproto.TabletAliasString(masterElectTabletAlias))
}
haveOtherMaster := false
for alias, ti := range masterTabletMap {
if !topoproto.TabletAliasEqual(&alias, masterElectTabletAlias) && ti.Type != pb.TabletType_SCRAP {
haveOtherMaster = true
}
}
if haveOtherMaster {
if !force {
return fmt.Errorf("master-elect tablet %v is not the only master in the shard, use -force to proceed anyway", topoproto.TabletAliasString(masterElectTabletAlias))
}
wr.logger.Warningf("master-elect tablet %v is not the only master in the shard, proceeding anyway as -force was used", topoproto.TabletAliasString(masterElectTabletAlias))
}
// First phase: reset replication on all tablets. If anyone fails,
// we stop. It is probably because it is unreachable, and may leave
// an unstable database process in the mix, with a database daemon
// at a wrong replication spot.
event.DispatchUpdate(ev, "resetting replication on all tablets")
wg := sync.WaitGroup{}
rec := concurrency.AllErrorRecorder{}
for alias, tabletInfo := range tabletMap {
wg.Add(1)
go func(alias pb.TabletAlias, tabletInfo *topo.TabletInfo) {
defer wg.Done()
wr.logger.Infof("resetting replication on tablet %v", topoproto.TabletAliasString(&alias))
if err := wr.TabletManagerClient().ResetReplication(ctx, tabletInfo); err != nil {
rec.RecordError(fmt.Errorf("Tablet %v ResetReplication failed (either fix it, or Scrap it): %v", topoproto.TabletAliasString(&alias), err))
}
}(alias, tabletInfo)
}
wg.Wait()
if err := rec.Error(); err != nil {
return err
}
// Tell the new master to break its slaves, return its replication
// position
wr.logger.Infof("initializing master on %v", topoproto.TabletAliasString(masterElectTabletAlias))
event.DispatchUpdate(ev, "initializing master")
rp, err := wr.TabletManagerClient().InitMaster(ctx, masterElectTabletInfo)
if err != nil {
return err
}
// Now tell the new master to insert the reparent_journal row,
// and tell everybody else to become a slave of the new master,
// and wait for the row in the reparent_journal table.
// We start all these in parallel, to handle the semi-sync
// case: for the master to be able to commit its row in the
// reparent_journal table, it needs connected slaves.
event.DispatchUpdate(ev, "reparenting all tablets")
now := time.Now().UnixNano()
wgMaster := sync.WaitGroup{}
wgSlaves := sync.WaitGroup{}
var masterErr error
for alias, tabletInfo := range tabletMap {
if topoproto.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", topoproto.TabletAliasString(&alias))
masterErr = wr.TabletManagerClient().PopulateReparentJournal(ctx, tabletInfo, now, initShardMasterOperation, &alias, rp)
}(alias, tabletInfo)
} else {
wgSlaves.Add(1)
go func(alias pb.TabletAlias, tabletInfo *topo.TabletInfo) {
defer wgSlaves.Done()
wr.logger.Infof("initializing slave %v", topoproto.TabletAliasString(&alias))
if err := wr.TabletManagerClient().InitSlave(ctx, tabletInfo, masterElectTabletAlias, rp, now); err != nil {
rec.RecordError(fmt.Errorf("Tablet %v InitSlave failed: %v", topoproto.TabletAliasString(&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)
}
if !topoproto.TabletAliasEqual(shardInfo.MasterAlias, masterElectTabletAlias) {
shardInfo.MasterAlias = masterElectTabletAlias
if err := wr.ts.UpdateShard(ctx, 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
// don't want to rebuild the shard serving graph (the failure
// will most likely be a timeout, and our context will be
// expired, so the rebuild will fail anyway)
wgSlaves.Wait()
if err := rec.Error(); err != nil {
return err
}
// Then we rebuild the entire serving graph for the shard,
// to account for all changes.
event.DispatchUpdate(ev, "rebuilding shard graph")
_, err = wr.RebuildShardGraph(ctx, keyspace, shard, nil)
return err
}
func TestTabletExternallyReparented(t *testing.T) {
tabletmanager.SetReparentFlags(time.Minute /* finalizeTimeout */)
ctx := context.Background()
db := fakesqldb.Register()
ts := zktopo.NewTestServer(t, []string{"cell1", "cell2"})
wr := wrangler.New(logutil.NewConsoleLogger(), ts, tmclient.NewTabletManagerClient(), time.Second)
vp := NewVtctlPipe(t, ts)
defer vp.Close()
// Create an old master, a new master, two good slaves, one bad slave
oldMaster := NewFakeTablet(t, wr, "cell1", 0, pb.TabletType_MASTER, db)
newMaster := NewFakeTablet(t, wr, "cell1", 1, pb.TabletType_REPLICA, db)
goodSlave1 := NewFakeTablet(t, wr, "cell1", 2, pb.TabletType_REPLICA, db)
goodSlave2 := NewFakeTablet(t, wr, "cell2", 3, pb.TabletType_REPLICA, db)
badSlave := NewFakeTablet(t, wr, "cell1", 4, pb.TabletType_REPLICA, db)
// Add a new Cell to the Shard, that doesn't map to any read topo cell,
// to simulate a data center being unreachable.
si, err := ts.GetShard(ctx, "test_keyspace", "0")
if err != nil {
t.Fatalf("GetShard failed: %v", err)
}
si.Cells = append(si.Cells, "cell666")
if err := ts.UpdateShard(ctx, si); err != nil {
t.Fatalf("UpdateShard failed: %v", err)
}
// Slightly unrelated test: make sure we can find the tablets
// even with a datacenter being down.
tabletMap, err := ts.GetTabletMapForShardByCell(ctx, "test_keyspace", "0", []string{"cell1"})
if err != nil {
t.Fatalf("GetTabletMapForShardByCell should have worked but got: %v", err)
}
master, err := topotools.FindTabletByIPAddrAndPort(tabletMap, oldMaster.Tablet.Ip, "vt", oldMaster.Tablet.PortMap["vt"])
if err != nil || !topoproto.TabletAliasEqual(&master, oldMaster.Tablet.Alias) {
t.Fatalf("FindTabletByIPAddrAndPort(master) failed: %v %v", err, master)
}
slave1, err := topotools.FindTabletByIPAddrAndPort(tabletMap, goodSlave1.Tablet.Ip, "vt", goodSlave1.Tablet.PortMap["vt"])
if err != nil || !topoproto.TabletAliasEqual(&slave1, goodSlave1.Tablet.Alias) {
t.Fatalf("FindTabletByIPAddrAndPort(slave1) failed: %v %v", err, master)
}
slave2, err := topotools.FindTabletByIPAddrAndPort(tabletMap, goodSlave2.Tablet.Ip, "vt", goodSlave2.Tablet.PortMap["vt"])
if err != topo.ErrNoNode {
t.Fatalf("FindTabletByIPAddrAndPort(slave2) worked: %v %v", err, slave2)
}
// Make sure the master is not exported in other cells
tabletMap, err = ts.GetTabletMapForShardByCell(ctx, "test_keyspace", "0", []string{"cell2"})
master, err = topotools.FindTabletByIPAddrAndPort(tabletMap, oldMaster.Tablet.Ip, "vt", oldMaster.Tablet.PortMap["vt"])
if err != topo.ErrNoNode {
t.Fatalf("FindTabletByIPAddrAndPort(master) worked in cell2: %v %v", err, master)
}
tabletMap, err = ts.GetTabletMapForShard(ctx, "test_keyspace", "0")
if err != topo.ErrPartialResult {
t.Fatalf("GetTabletMapForShard should have returned ErrPartialResult but got: %v", err)
}
master, err = topotools.FindTabletByIPAddrAndPort(tabletMap, oldMaster.Tablet.Ip, "vt", oldMaster.Tablet.PortMap["vt"])
if err != nil || !topoproto.TabletAliasEqual(&master, oldMaster.Tablet.Alias) {
t.Fatalf("FindTabletByIPAddrAndPort(master) failed: %v %v", err, master)
}
// On the elected master, we will respond to
// TabletActionSlaveWasPromoted
newMaster.StartActionLoop(t, wr)
defer newMaster.StopActionLoop(t)
// On the old master, we will only respond to
// TabletActionSlaveWasRestarted.
oldMaster.StartActionLoop(t, wr)
defer oldMaster.StopActionLoop(t)
// On the good slaves, we will respond to
// TabletActionSlaveWasRestarted.
goodSlave1.StartActionLoop(t, wr)
defer goodSlave1.StopActionLoop(t)
goodSlave2.StartActionLoop(t, wr)
defer goodSlave2.StopActionLoop(t)
// On the bad slave, we will respond to
// TabletActionSlaveWasRestarted with bad data.
badSlave.StartActionLoop(t, wr)
defer badSlave.StopActionLoop(t)
// First test: reparent to the same master, make sure it works
// as expected.
tmc := tmclient.NewTabletManagerClient()
ti, err := ts.GetTablet(ctx, oldMaster.Tablet.Alias)
if err != nil {
t.Fatalf("GetTablet failed: %v", err)
}
if err := vp.Run([]string{"TabletExternallyReparented", topoproto.TabletAliasString(oldMaster.Tablet.Alias)}); err != nil {
t.Fatalf("TabletExternallyReparented(same master) should have worked: %v", err)
}
// Second test: reparent to a replica, and pretend the old
// master is still good to go.
// This tests a bad case; the new designated master is a slave,
// but we should do what we're told anyway
ti, err = ts.GetTablet(ctx, goodSlave1.Tablet.Alias)
if err != nil {
t.Fatalf("GetTablet failed: %v", err)
}
waitID := makeWaitID()
if err := tmc.TabletExternallyReparented(context.Background(), ti, waitID); err != nil {
t.Fatalf("TabletExternallyReparented(slave) error: %v", err)
}
waitForExternalReparent(t, waitID)
// This tests the good case, where everything works as planned
t.Logf("TabletExternallyReparented(new master) expecting success")
ti, err = ts.GetTablet(ctx, newMaster.Tablet.Alias)
if err != nil {
t.Fatalf("GetTablet failed: %v", err)
}
waitID = makeWaitID()
if err := tmc.TabletExternallyReparented(context.Background(), ti, waitID); err != nil {
t.Fatalf("TabletExternallyReparented(replica) failed: %v", err)
}
waitForExternalReparent(t, waitID)
// Now double-check the serving graph is good.
// Should only have one good replica left.
addrs, _, err := ts.GetEndPoints(ctx, "cell1", "test_keyspace", "0", pb.TabletType_REPLICA)
if err != nil {
t.Fatalf("GetEndPoints failed at the end: %v", err)
}
if len(addrs.Entries) != 1 {
t.Fatalf("GetEndPoints has too many entries: %v", addrs)
}
}
// TestInitMasterShard is the good scenario test, where everything
// works as planned
func TestInitMasterShard(t *testing.T) {
ctx := context.Background()
db := fakesqldb.Register()
ts := zktestserver.New(t, []string{"cell1", "cell2"})
wr := wrangler.New(logutil.NewConsoleLogger(), ts, tmclient.NewTabletManagerClient())
vp := NewVtctlPipe(t, ts)
defer vp.Close()
db.AddQuery("CREATE DATABASE IF NOT EXISTS `vt_test_keyspace`", &sqltypes.Result{})
// Create a master, a couple good slaves
master := NewFakeTablet(t, wr, "cell1", 0, topodatapb.TabletType_MASTER, db)
goodSlave1 := NewFakeTablet(t, wr, "cell1", 1, topodatapb.TabletType_REPLICA, db)
goodSlave2 := NewFakeTablet(t, wr, "cell2", 2, topodatapb.TabletType_REPLICA, db)
// Master: set a plausible ReplicationPosition to return,
// and expect to add entry in _vt.reparent_journal
master.FakeMysqlDaemon.CurrentMasterPosition = replication.Position{
GTIDSet: replication.MariadbGTID{
Domain: 5,
Server: 456,
Sequence: 890,
},
}
master.FakeMysqlDaemon.ReadOnly = true
master.FakeMysqlDaemon.ResetReplicationResult = []string{"reset rep 1"}
master.FakeMysqlDaemon.SetSlavePositionCommandsResult = []string{"new master shouldn't use this"}
master.FakeMysqlDaemon.SetMasterCommandsResult = []string{"new master shouldn't use this"}
master.FakeMysqlDaemon.ExpectedExecuteSuperQueryList = []string{
"reset rep 1",
"CREATE DATABASE IF NOT EXISTS _vt",
"SUBCREATE TABLE IF NOT EXISTS _vt.reparent_journal",
"CREATE DATABASE IF NOT EXISTS _vt",
"SUBCREATE TABLE IF NOT EXISTS _vt.reparent_journal",
"SUBINSERT INTO _vt.reparent_journal (time_created_ns, action_name, master_alias, replication_position) VALUES",
}
master.StartActionLoop(t, wr)
defer master.StopActionLoop(t)
// Slave1: expect to be reset and re-parented
goodSlave1.FakeMysqlDaemon.ReadOnly = true
goodSlave1.FakeMysqlDaemon.ResetReplicationResult = []string{"reset rep 1"}
goodSlave1.FakeMysqlDaemon.SetSlavePositionCommandsPos = master.FakeMysqlDaemon.CurrentMasterPosition
goodSlave1.FakeMysqlDaemon.SetSlavePositionCommandsResult = []string{"cmd1"}
goodSlave1.FakeMysqlDaemon.SetMasterCommandsInput = fmt.Sprintf("%v:%v", master.Tablet.Hostname, master.Tablet.PortMap["mysql"])
goodSlave1.FakeMysqlDaemon.SetMasterCommandsResult = []string{"set master cmd 1"}
goodSlave1.FakeMysqlDaemon.ExpectedExecuteSuperQueryList = []string{
"reset rep 1",
"cmd1",
"set master cmd 1",
"START SLAVE",
}
goodSlave1.StartActionLoop(t, wr)
defer goodSlave1.StopActionLoop(t)
// Slave2: expect to be re-parented
goodSlave2.FakeMysqlDaemon.ReadOnly = true
goodSlave2.FakeMysqlDaemon.ResetReplicationResult = []string{"reset rep 2"}
goodSlave2.FakeMysqlDaemon.SetSlavePositionCommandsPos = master.FakeMysqlDaemon.CurrentMasterPosition
goodSlave2.FakeMysqlDaemon.SetSlavePositionCommandsResult = []string{"cmd1", "cmd2"}
goodSlave2.FakeMysqlDaemon.SetMasterCommandsInput = fmt.Sprintf("%v:%v", master.Tablet.Hostname, master.Tablet.PortMap["mysql"])
goodSlave2.FakeMysqlDaemon.SetMasterCommandsResult = []string{"set master cmd 1"}
goodSlave2.FakeMysqlDaemon.ExpectedExecuteSuperQueryList = []string{
"reset rep 2",
"cmd1",
"cmd2",
"set master cmd 1",
"START SLAVE",
}
goodSlave2.StartActionLoop(t, wr)
defer goodSlave2.StopActionLoop(t)
// run InitShardMaster
if err := vp.Run([]string{"InitShardMaster", "-wait_slave_timeout", "10s", master.Tablet.Keyspace + "/" + master.Tablet.Shard, topoproto.TabletAliasString(master.Tablet.Alias)}); err != nil {
t.Fatalf("InitShardMaster failed: %v", err)
}
// check what was run
if master.FakeMysqlDaemon.ReadOnly {
t.Errorf("master was not turned read-write")
}
si, err := ts.GetShard(ctx, master.Tablet.Keyspace, master.Tablet.Shard)
if err != nil {
t.Fatalf("GetShard failed: %v", err)
}
if !topoproto.TabletAliasEqual(si.MasterAlias, master.Tablet.Alias) {
t.Errorf("unexpected shard master alias, got %v expected %v", si.MasterAlias, master.Tablet.Alias)
}
if err := master.FakeMysqlDaemon.CheckSuperQueryList(); err != nil {
t.Fatalf("master.FakeMysqlDaemon.CheckSuperQueryList failed: %v", err)
}
if err := goodSlave1.FakeMysqlDaemon.CheckSuperQueryList(); err != nil {
t.Fatalf("goodSlave1.FakeMysqlDaemon.CheckSuperQueryList failed: %v", err)
}
if err := goodSlave2.FakeMysqlDaemon.CheckSuperQueryList(); err != nil {
t.Fatalf("goodSlave2.FakeMysqlDaemon.CheckSuperQueryList failed: %v", err)
}
}
// TestInitMasterShardOneSlaveFails makes sure that if one slave fails to
// proceed, the action completes anyway
func TestInitMasterShardOneSlaveFails(t *testing.T) {
ctx := context.Background()
db := fakesqldb.Register()
ts := zktestserver.New(t, []string{"cell1", "cell2"})
wr := wrangler.New(logutil.NewConsoleLogger(), ts, tmclient.NewTabletManagerClient())
// Create a master, a couple slaves
master := NewFakeTablet(t, wr, "cell1", 0, topodatapb.TabletType_MASTER, db)
goodSlave := NewFakeTablet(t, wr, "cell1", 1, topodatapb.TabletType_REPLICA, db)
badSlave := NewFakeTablet(t, wr, "cell2", 2, topodatapb.TabletType_REPLICA, db)
// Master: set a plausible ReplicationPosition to return,
// and expect to add entry in _vt.reparent_journal
master.FakeMysqlDaemon.CurrentMasterPosition = replication.Position{
GTIDSet: replication.MariadbGTID{
Domain: 5,
Server: 456,
Sequence: 890,
},
}
master.FakeMysqlDaemon.ReadOnly = true
master.FakeMysqlDaemon.ExpectedExecuteSuperQueryList = []string{
"CREATE DATABASE IF NOT EXISTS _vt",
"SUBCREATE TABLE IF NOT EXISTS _vt.reparent_journal",
"CREATE DATABASE IF NOT EXISTS _vt",
"SUBCREATE TABLE IF NOT EXISTS _vt.reparent_journal",
"SUBINSERT INTO _vt.reparent_journal (time_created_ns, action_name, master_alias, replication_position) VALUES",
}
master.StartActionLoop(t, wr)
defer master.StopActionLoop(t)
// goodSlave: expect to be re-parented
goodSlave.FakeMysqlDaemon.ReadOnly = true
goodSlave.FakeMysqlDaemon.SetSlavePositionCommandsPos = master.FakeMysqlDaemon.CurrentMasterPosition
goodSlave.FakeMysqlDaemon.SetSlavePositionCommandsResult = []string{"cmd1"}
goodSlave.FakeMysqlDaemon.SetMasterCommandsInput = fmt.Sprintf("%v:%v", master.Tablet.Hostname, master.Tablet.PortMap["mysql"])
goodSlave.FakeMysqlDaemon.SetMasterCommandsResult = []string{"set master cmd 1"}
goodSlave.FakeMysqlDaemon.ExpectedExecuteSuperQueryList = []string{
"cmd1",
"set master cmd 1",
"START SLAVE",
}
goodSlave.StartActionLoop(t, wr)
defer goodSlave.StopActionLoop(t)
// badSlave: insert an error by failing the master hostname input
// on purpose
badSlave.FakeMysqlDaemon.ReadOnly = true
badSlave.FakeMysqlDaemon.SetSlavePositionCommandsPos = master.FakeMysqlDaemon.CurrentMasterPosition
badSlave.FakeMysqlDaemon.SetMasterCommandsInput = fmt.Sprintf("%v:%v", "", master.Tablet.PortMap["mysql"])
badSlave.StartActionLoop(t, wr)
defer badSlave.StopActionLoop(t)
// also change the master alias in the Shard object, to make sure it
// is set back.
si, err := ts.GetShard(ctx, master.Tablet.Keyspace, master.Tablet.Shard)
if err != nil {
t.Fatalf("GetShard failed: %v", err)
}
si.MasterAlias.Uid++
if err := ts.UpdateShard(ctx, si); err != nil {
t.Fatalf("UpdateShard failed: %v", err)
}
// run InitShardMaster without force, it fails because master is
// changing.
if err := wr.InitShardMaster(ctx, master.Tablet.Keyspace, master.Tablet.Shard, master.Tablet.Alias, false /*force*/, 10*time.Second); err == nil || !strings.Contains(err.Error(), "is not the shard master") {
t.Errorf("InitShardMaster with mismatched new master returned wrong error: %v", err)
}
// run InitShardMaster
if err := wr.InitShardMaster(ctx, master.Tablet.Keyspace, master.Tablet.Shard, master.Tablet.Alias, true /*force*/, 10*time.Second); err == nil || !strings.Contains(err.Error(), "wrong input for SetMasterCommands") {
t.Errorf("InitShardMaster with one failed slave returned wrong error: %v", err)
}
// check what was run: master should still be good
if master.FakeMysqlDaemon.ReadOnly {
t.Errorf("master was not turned read-write")
}
si, err = ts.GetShard(ctx, master.Tablet.Keyspace, master.Tablet.Shard)
if err != nil {
t.Fatalf("GetShard failed: %v", err)
}
if !topoproto.TabletAliasEqual(si.MasterAlias, master.Tablet.Alias) {
t.Errorf("unexpected shard master alias, got %v expected %v", si.MasterAlias, master.Tablet.Alias)
}
}
// TabletExternallyReparented updates all topo records so the current
// tablet is the new master for this shard.
// Should be called under RPCWrapLock.
func (agent *ActionAgent) TabletExternallyReparented(ctx context.Context, externalID string) error {
startTime := time.Now()
// If there is a finalize step running, wait for it to finish or time out
// before checking the global shard record again.
if agent.finalizeReparentCtx != nil {
select {
case <-agent.finalizeReparentCtx.Done():
agent.finalizeReparentCtx = nil
case <-ctx.Done():
return ctx.Err()
}
}
tablet := agent.Tablet()
// Check the global shard record.
si, err := agent.TopoServer.GetShard(ctx, tablet.Keyspace, tablet.Shard)
if err != nil {
log.Warningf("fastTabletExternallyReparented: failed to read global shard record for %v/%v: %v", tablet.Keyspace, tablet.Shard, err)
return err
}
if topoproto.TabletAliasEqual(si.MasterAlias, tablet.Alias) {
// We may get called on the current master even when nothing has changed.
// If the global shard record is already updated, it means we successfully
// finished a previous reparent to this tablet.
return nil
}
// Create a reusable Reparent event with available info.
ev := &events.Reparent{
ShardInfo: *si,
NewMaster: *tablet.Tablet,
OldMaster: pb.Tablet{
Alias: si.MasterAlias,
Type: pb.TabletType_MASTER,
},
ExternalID: externalID,
}
defer func() {
if err != nil {
event.DispatchUpdate(ev, "failed: "+err.Error())
}
}()
event.DispatchUpdate(ev, "starting external from tablet (fast)")
// Execute state change to master by force-updating only the local copy of the
// tablet record. The actual record in topo will be updated later.
log.Infof("fastTabletExternallyReparented: executing change callback for state change to MASTER")
oldTablet := *tablet.Tablet
newTablet := oldTablet
newTablet.Type = pb.TabletType_MASTER
newTablet.HealthMap = nil
agent.setTablet(topo.NewTabletInfo(&newTablet, -1))
if err := agent.updateState(ctx, &oldTablet, "fastTabletExternallyReparented"); err != nil {
return fmt.Errorf("fastTabletExternallyReparented: failed to change tablet state to MASTER: %v", err)
}
agent.mutex.Lock()
agent._tabletExternallyReparentedTime = time.Now()
agent.mutex.Unlock()
// Directly write the new master endpoint in the serving graph.
// We will do a true rebuild in the background soon, but in the meantime,
// this will be enough for clients to re-resolve the new master.
event.DispatchUpdate(ev, "writing new master endpoint")
log.Infof("fastTabletExternallyReparented: writing new master endpoint to serving graph")
ep, err := topo.TabletEndPoint(tablet.Tablet)
if err != nil {
return fmt.Errorf("fastTabletExternallyReparented: failed to generate EndPoint for tablet %v: %v", tablet.Alias, err)
}
err = topo.UpdateEndPoints(ctx, agent.TopoServer, tablet.Alias.Cell,
si.Keyspace(), si.ShardName(), pb.TabletType_MASTER,
&pb.EndPoints{Entries: []*pb.EndPoint{ep}}, -1)
if err != nil {
return fmt.Errorf("fastTabletExternallyReparented: failed to update master endpoint: %v", err)
}
externalReparentStats.Record("NewMasterVisible", startTime)
// Start the finalize stage with a background context, but connect the trace.
bgCtx, cancel := context.WithTimeout(agent.batchCtx, *finalizeReparentTimeout)
bgCtx = trace.CopySpan(bgCtx, ctx)
agent.finalizeReparentCtx = bgCtx
go func() {
err := agent.finalizeTabletExternallyReparented(bgCtx, si, ev)
cancel()
if err != nil {
log.Warningf("finalizeTabletExternallyReparented error: %v", err)
event.DispatchUpdate(ev, "failed: "+err.Error())
return
}
externalReparentStats.Record("FullRebuild", startTime)
}()
return nil
}
// 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 !si.HasMaster() {
return fmt.Errorf("No master in shard %v/%v", keyspace, shards[0])
}
referenceAlias := si.MasterAlias
log.Infof("Gathering schema for reference master %v", topoproto.TabletAliasString(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 := wr.ts.FindAllTabletAliasesInShard(ctx, keyspace, shards[0])
if err != nil {
return err
}
for _, alias := range aliases {
if topoproto.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 !si.HasMaster() {
er.RecordError(fmt.Errorf("No master in shard %v/%v", keyspace, shard))
continue
}
aliases, err := wr.ts.FindAllTabletAliasesInShard(ctx, 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: %v", er.Error().Error())
}
return nil
}
// InitTablet initializes the tablet record if necessary.
func (agent *ActionAgent) InitTablet(port, gRPCPort int32) error {
// it should be either we have all three of init_keyspace,
// init_shard and init_tablet_type, or none.
if *initKeyspace == "" && *initShard == "" && *initTabletType == "" {
// not initializing the record
return nil
}
if *initKeyspace == "" || *initShard == "" || *initTabletType == "" {
return fmt.Errorf("either need all of init_keyspace, init_shard and init_tablet_type, or none")
}
// parse init_tablet_type
tabletType, err := topoproto.ParseTabletType(*initTabletType)
if err != nil {
return fmt.Errorf("invalid init_tablet_type %v: %v", *initTabletType, err)
}
if tabletType == topodatapb.TabletType_MASTER {
// We disallow MASTER, so we don't have to change
// shard.MasterAlias, and deal with the corner cases.
return fmt.Errorf("init_tablet_type cannot be master, use replica instead")
}
// parse and validate shard name
shard, _, err := topo.ValidateShardName(*initShard)
if err != nil {
return fmt.Errorf("cannot validate shard name %v: %v", *initShard, err)
}
// create a context for this whole operation
ctx, cancel := context.WithTimeout(agent.batchCtx, *initTimeout)
defer cancel()
// read the shard, create it if necessary
log.Infof("Reading shard record %v/%v", *initKeyspace, shard)
si, err := agent.TopoServer.GetOrCreateShard(ctx, *initKeyspace, shard)
if err != nil {
return fmt.Errorf("InitTablet cannot GetOrCreateShard shard: %v", err)
}
if si.MasterAlias != nil && topoproto.TabletAliasEqual(si.MasterAlias, agent.TabletAlias) {
// We're marked as master in the shard record, which could mean the master
// tablet process was just restarted. However, we need to check if a new
// master is in the process of taking over. In that case, it will let us
// know by forcibly updating the old master's tablet record.
oldTablet, err := agent.TopoServer.GetTablet(ctx, agent.TabletAlias)
switch err {
case topo.ErrNoNode:
// There's no existing tablet record, so we can assume
// no one has left us a message to step down.
tabletType = topodatapb.TabletType_MASTER
case nil:
if oldTablet.Type == topodatapb.TabletType_MASTER {
// We're marked as master in the shard record,
// and our existing tablet record agrees.
tabletType = topodatapb.TabletType_MASTER
}
default:
return fmt.Errorf("InitTablet failed to read existing tablet record: %v", err)
}
}
// See if we need to add the tablet's cell to the shard's cell list.
if !si.HasCell(agent.TabletAlias.Cell) {
si, err = agent.TopoServer.UpdateShardFields(ctx, *initKeyspace, shard, func(si *topo.ShardInfo) error {
if si.HasCell(agent.TabletAlias.Cell) {
// Someone else already did it.
return topo.ErrNoUpdateNeeded
}
si.Cells = append(si.Cells, agent.TabletAlias.Cell)
return nil
})
if err != nil {
return fmt.Errorf("couldn't add tablet's cell to shard record: %v", err)
}
}
log.Infof("Initializing the tablet for type %v", tabletType)
// figure out the hostname
hostname := *tabletHostname
if hostname != "" {
log.Infof("Using hostname: %v from -tablet_hostname flag.", hostname)
} else {
hostname, err := netutil.FullyQualifiedHostname()
if err != nil {
return err
}
log.Infof("Using detected machine hostname: %v To change this, fix your machine network configuration or override it with -tablet_hostname.", hostname)
}
// create and populate tablet record
tablet := &topodatapb.Tablet{
Alias: agent.TabletAlias,
Hostname: hostname,
PortMap: make(map[string]int32),
Keyspace: *initKeyspace,
Shard: *initShard,
Type: tabletType,
DbNameOverride: *initDbNameOverride,
Tags: initTags,
}
if port != 0 {
tablet.PortMap["vt"] = port
}
if gRPCPort != 0 {
tablet.PortMap["grpc"] = gRPCPort
}
if err := topo.TabletComplete(tablet); err != nil {
return fmt.Errorf("InitTablet TabletComplete failed: %v", err)
}
// Now try to create the record (it will also fix up the
// ShardReplication record if necessary).
err = agent.TopoServer.CreateTablet(ctx, tablet)
switch err {
case nil:
// It worked, we're good.
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 {
return 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)
}
// Then overwrite everything, ignoring version mismatch.
if err := agent.TopoServer.UpdateTablet(ctx, topo.NewTabletInfo(tablet, -1)); err != nil {
return fmt.Errorf("UpdateTablet failed: %v", err)
}
default:
return fmt.Errorf("CreateTablet failed: %v", err)
}
return nil
}
func (wr *Wrangler) initShardMasterLocked(ctx context.Context, ev *events.Reparent, keyspace, shard string, masterElectTabletAlias *topodatapb.TabletAlias, force bool, waitSlaveTimeout time.Duration) error {
shardInfo, err := wr.ts.GetShard(ctx, keyspace, shard)
if err != nil {
return err
}
ev.ShardInfo = *shardInfo
event.DispatchUpdate(ev, "reading tablet map")
tabletMap, err := wr.ts.GetTabletMapForShard(ctx, keyspace, shard)
if err != nil {
return err
}
// Check the master elect is in tabletMap
masterElectTabletInfo, ok := tabletMap[*masterElectTabletAlias]
if !ok {
return fmt.Errorf("master-elect tablet %v is not in the shard", topoproto.TabletAliasString(masterElectTabletAlias))
}
ev.NewMaster = *masterElectTabletInfo.Tablet
// Check the master is the only master is the shard, or -force was used.
_, masterTabletMap := topotools.SortedTabletMap(tabletMap)
if !topoproto.TabletAliasEqual(shardInfo.MasterAlias, masterElectTabletAlias) {
if !force {
return fmt.Errorf("master-elect tablet %v is not the shard master, use -force to proceed anyway", topoproto.TabletAliasString(masterElectTabletAlias))
}
wr.logger.Warningf("master-elect tablet %v is not the shard master, proceeding anyway as -force was used", topoproto.TabletAliasString(masterElectTabletAlias))
}
if _, ok := masterTabletMap[*masterElectTabletAlias]; !ok {
if !force {
return fmt.Errorf("master-elect tablet %v is not a master in the shard, use -force to proceed anyway", topoproto.TabletAliasString(masterElectTabletAlias))
}
wr.logger.Warningf("master-elect tablet %v is not a master in the shard, proceeding anyway as -force was used", topoproto.TabletAliasString(masterElectTabletAlias))
}
haveOtherMaster := false
for alias := range masterTabletMap {
if !topoproto.TabletAliasEqual(&alias, masterElectTabletAlias) {
haveOtherMaster = true
}
}
if haveOtherMaster {
if !force {
return fmt.Errorf("master-elect tablet %v is not the only master in the shard, use -force to proceed anyway", topoproto.TabletAliasString(masterElectTabletAlias))
}
wr.logger.Warningf("master-elect tablet %v is not the only master in the shard, proceeding anyway as -force was used", topoproto.TabletAliasString(masterElectTabletAlias))
}
// First phase: reset replication on all tablets. If anyone fails,
// we stop. It is probably because it is unreachable, and may leave
// an unstable database process in the mix, with a database daemon
// at a wrong replication spot.
event.DispatchUpdate(ev, "resetting replication on all tablets")
wg := sync.WaitGroup{}
rec := concurrency.AllErrorRecorder{}
for alias, tabletInfo := range tabletMap {
wg.Add(1)
go func(alias topodatapb.TabletAlias, tabletInfo *topo.TabletInfo) {
defer wg.Done()
wr.logger.Infof("resetting replication on tablet %v", topoproto.TabletAliasString(&alias))
if err := wr.tmc.ResetReplication(ctx, tabletInfo.Tablet); err != nil {
rec.RecordError(fmt.Errorf("Tablet %v ResetReplication failed (either fix it, or Scrap it): %v", topoproto.TabletAliasString(&alias), err))
}
}(alias, tabletInfo)
}
wg.Wait()
if err := rec.Error(); err != nil {
return err
}
// Tell the new master to break its slaves, return its replication
// position
wr.logger.Infof("initializing master on %v", topoproto.TabletAliasString(masterElectTabletAlias))
event.DispatchUpdate(ev, "initializing master")
rp, err := wr.tmc.InitMaster(ctx, masterElectTabletInfo.Tablet)
if err != nil {
return err
}
// Now tell the new master to insert the reparent_journal row,
// and tell everybody else to become a slave of the new master,
// and wait for the row in the reparent_journal table.
// We start all these in parallel, to handle the semi-sync
// case: for the master to be able to commit its row in the
// reparent_journal table, it needs connected slaves.
event.DispatchUpdate(ev, "reparenting all tablets")
now := time.Now().UnixNano()
wgMaster := sync.WaitGroup{}
wgSlaves := sync.WaitGroup{}
var masterErr error
for alias, tabletInfo := range tabletMap {
if topoproto.TabletAliasEqual(&alias, masterElectTabletAlias) {
wgMaster.Add(1)
go func(alias topodatapb.TabletAlias, tabletInfo *topo.TabletInfo) {
defer wgMaster.Done()
wr.logger.Infof("populating reparent journal on new master %v", topoproto.TabletAliasString(&alias))
masterErr = wr.tmc.PopulateReparentJournal(ctx, tabletInfo.Tablet, now, initShardMasterOperation, &alias, rp)
}(alias, tabletInfo)
} else {
wgSlaves.Add(1)
go func(alias topodatapb.TabletAlias, tabletInfo *topo.TabletInfo) {
defer wgSlaves.Done()
wr.logger.Infof("initializing slave %v", topoproto.TabletAliasString(&alias))
if err := wr.tmc.InitSlave(ctx, tabletInfo.Tablet, masterElectTabletAlias, rp, now); err != nil {
rec.RecordError(fmt.Errorf("Tablet %v InitSlave failed: %v", topoproto.TabletAliasString(&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)
}
if !topoproto.TabletAliasEqual(shardInfo.MasterAlias, masterElectTabletAlias) {
if _, err := wr.ts.UpdateShardFields(ctx, keyspace, shard, func(si *topo.ShardInfo) error {
si.MasterAlias = masterElectTabletAlias
return nil
}); 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
// don't want to rebuild the shard serving graph (the failure
// will most likely be a timeout, and our context will be
// expired, so the rebuild will fail anyway)
wgSlaves.Wait()
if err := rec.Error(); err != nil {
return err
}
// Create database if necessary on the master. Slaves will get it too through
// replication. Since the user called InitShardMaster, they've told us to
// assume that whatever data is on all the slaves is what they intended.
// If the database doesn't exist, it means the user intends for these tablets
// to begin serving with no data (i.e. first time initialization).
createDB := fmt.Sprintf("CREATE DATABASE IF NOT EXISTS `%s`", topoproto.TabletDbName(masterElectTabletInfo.Tablet))
if _, err := wr.tmc.ExecuteFetchAsDba(ctx, masterElectTabletInfo.Tablet, false, []byte(createDB), 1, false, true); err != nil {
return fmt.Errorf("failed to create database: %v", err)
}
return nil
}
// TabletExternallyReparented updates all topo records so the current
// tablet is the new master for this shard.
// Should be called under RPCWrapLock.
func (agent *ActionAgent) TabletExternallyReparented(ctx context.Context, externalID string) error {
startTime := time.Now()
// If there is a finalize step running, wait for it to finish or time out
// before checking the global shard record again.
if agent.finalizeReparentCtx != nil {
select {
case <-agent.finalizeReparentCtx.Done():
agent.finalizeReparentCtx = nil
case <-ctx.Done():
return ctx.Err()
}
}
tablet := agent.Tablet()
// Check the global shard record.
si, err := agent.TopoServer.GetShard(ctx, tablet.Keyspace, tablet.Shard)
if err != nil {
log.Warningf("fastTabletExternallyReparented: failed to read global shard record for %v/%v: %v", tablet.Keyspace, tablet.Shard, err)
return err
}
if topoproto.TabletAliasEqual(si.MasterAlias, tablet.Alias) {
// We may get called on the current master even when nothing has changed.
// If the global shard record is already updated, it means we successfully
// finished a previous reparent to this tablet.
return nil
}
// Remember when we were first told we're the master.
// If another tablet claims to be master and offers a more recent time,
// that tablet will be trusted over us.
agent.mutex.Lock()
agent._tabletExternallyReparentedTime = startTime
agent._replicationDelay = 0
agent.mutex.Unlock()
// Create a reusable Reparent event with available info.
ev := &events.Reparent{
ShardInfo: *si,
NewMaster: *tablet,
OldMaster: topodatapb.Tablet{
Alias: si.MasterAlias,
Type: topodatapb.TabletType_MASTER,
},
ExternalID: externalID,
}
defer func() {
if err != nil {
event.DispatchUpdate(ev, "failed: "+err.Error())
}
}()
event.DispatchUpdate(ev, "starting external from tablet (fast)")
var wg sync.WaitGroup
var errs concurrency.AllErrorRecorder
// Execute state change to master by force-updating only the local copy of the
// tablet record. The actual record in topo will be updated later.
log.Infof("fastTabletExternallyReparented: executing change callback for state change to MASTER")
oldTablet := proto.Clone(tablet).(*topodatapb.Tablet)
tablet.Type = topodatapb.TabletType_MASTER
tablet.HealthMap = nil
agent.setTablet(tablet)
wg.Add(1)
go func() {
defer wg.Done()
// This is where updateState will block for gracePeriod, while it gives
// vtgate a chance to stop sending replica queries.
if err := agent.updateState(ctx, oldTablet, "fastTabletExternallyReparented"); err != nil {
errs.RecordError(fmt.Errorf("fastTabletExternallyReparented: failed to change tablet state to MASTER: %v", err))
}
}()
wg.Add(1)
go func() {
defer wg.Done()
// Directly write the new master endpoint in the serving graph.
// We will do a true rebuild in the background soon, but in the meantime,
// this will be enough for clients to re-resolve the new master.
event.DispatchUpdate(ev, "writing new master endpoint")
log.Infof("fastTabletExternallyReparented: writing new master endpoint to serving graph")
ep, err := topo.TabletEndPoint(tablet)
if err != nil {
errs.RecordError(fmt.Errorf("fastTabletExternallyReparented: failed to generate EndPoint for tablet %v: %v", tablet.Alias, err))
return
}
err = topo.UpdateEndPoints(ctx, agent.TopoServer, tablet.Alias.Cell,
si.Keyspace(), si.ShardName(), topodatapb.TabletType_MASTER,
&topodatapb.EndPoints{Entries: []*topodatapb.EndPoint{ep}}, -1)
if err != nil {
errs.RecordError(fmt.Errorf("fastTabletExternallyReparented: failed to update master endpoint: %v", err))
return
}
externalReparentStats.Record("NewMasterVisible", startTime)
}()
// Wait for serving state grace period and serving graph update.
wg.Wait()
if errs.HasErrors() {
return errs.Error()
}
// Start the finalize stage with a background context, but connect the trace.
bgCtx, cancel := context.WithTimeout(agent.batchCtx, *finalizeReparentTimeout)
bgCtx = trace.CopySpan(bgCtx, ctx)
agent.finalizeReparentCtx = bgCtx
go func() {
err := agent.finalizeTabletExternallyReparented(bgCtx, si, ev)
cancel()
if err != nil {
log.Warningf("finalizeTabletExternallyReparented error: %v", err)
event.DispatchUpdate(ev, "failed: "+err.Error())
return
}
externalReparentStats.Record("FullRebuild", startTime)
}()
return nil
}
func TestShardReplicationStatuses(t *testing.T) {
ctx := context.Background()
ts := zktopo.NewTestServer(t, []string{"cell1", "cell2"})
wr := wrangler.New(logutil.NewConsoleLogger(), ts, tmclient.NewTabletManagerClient(), time.Second)
// create shard and tablets
if err := ts.CreateShard(ctx, "test_keyspace", "0"); err != nil {
t.Fatalf("CreateShard failed: %v", err)
}
master := NewFakeTablet(t, wr, "cell1", 1, pb.TabletType_MASTER)
slave := NewFakeTablet(t, wr, "cell1", 2, pb.TabletType_REPLICA)
// mark the master inside the shard
si, err := ts.GetShard(ctx, "test_keyspace", "0")
if err != nil {
t.Fatalf("GetShard failed: %v", err)
}
si.MasterAlias = master.Tablet.Alias
if err := ts.UpdateShard(ctx, si); err != nil {
t.Fatalf("UpdateShard failed: %v", err)
}
// master action loop (to initialize host and port)
master.FakeMysqlDaemon.CurrentMasterPosition = myproto.ReplicationPosition{
GTIDSet: myproto.MariadbGTID{
Domain: 5,
Server: 456,
Sequence: 892,
},
}
master.StartActionLoop(t, wr)
defer master.StopActionLoop(t)
// slave loop
slave.FakeMysqlDaemon.CurrentMasterPosition = myproto.ReplicationPosition{
GTIDSet: myproto.MariadbGTID{
Domain: 5,
Server: 456,
Sequence: 890,
},
}
slave.FakeMysqlDaemon.CurrentMasterHost = master.Tablet.Hostname
slave.FakeMysqlDaemon.CurrentMasterPort = int(master.Tablet.PortMap["mysql"])
slave.StartActionLoop(t, wr)
defer slave.StopActionLoop(t)
// run ShardReplicationStatuses
ti, rs, err := wr.ShardReplicationStatuses(ctx, "test_keyspace", "0")
if err != nil {
t.Fatalf("ShardReplicationStatuses failed: %v", err)
}
// check result (make master first in the array)
if len(ti) != 2 || len(rs) != 2 {
t.Fatalf("ShardReplicationStatuses returned wrong results: %v %v", ti, rs)
}
if topoproto.TabletAliasEqual(ti[0].Alias, slave.Tablet.Alias) {
ti[0], ti[1] = ti[1], ti[0]
rs[0], rs[1] = rs[1], rs[0]
}
if !topoproto.TabletAliasEqual(ti[0].Alias, master.Tablet.Alias) ||
!topoproto.TabletAliasEqual(ti[1].Alias, slave.Tablet.Alias) ||
rs[0].MasterHost != "" ||
rs[1].MasterHost != master.Tablet.Hostname {
t.Fatalf("ShardReplicationStatuses returend wrong results: %v %v", ti, rs)
}
}
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 := wr.ts.GetTabletMapForShard(ctx, 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", topoproto.TabletAliasString(masterElectTabletAlias))
}
ev.NewMaster = *masterElectTabletInfo.Tablet
if topoproto.TabletAliasEqual(shardInfo.MasterAlias, masterElectTabletAlias) {
return fmt.Errorf("master-elect tablet %v is already the master", topoproto.TabletAliasString(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 shardInfo.HasMaster() {
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", topoproto.TabletAliasString(shardInfo.MasterAlias), err)
scrapOldMaster = false
}
}
if scrapOldMaster {
ev.OldMaster = *oldMasterTabletInfo.Tablet
wr.logger.Infof("scrapping old master %v", topoproto.TabletAliasString(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", topoproto.TabletAliasString(&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", topoproto.TabletAliasString(&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", topoproto.TabletAliasString(masterElectTabletAlias))
}
for alias, status := range statusMap {
if topoproto.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", topoproto.TabletAliasString(&alias), topoproto.TabletAliasString(masterElectTabletAlias), status.Position, masterElectStatus)
}
}
// Promote the masterElect
wr.logger.Infof("promote slave %v", topoproto.TabletAliasString(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", topoproto.TabletAliasString(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 topoproto.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", topoproto.TabletAliasString(&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", topoproto.TabletAliasString(&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", topoproto.TabletAliasString(&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", topoproto.TabletAliasString(masterElectTabletAlias))
shardInfo.MasterAlias = masterElectTabletAlias
if err := wr.ts.UpdateShard(ctx, 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
}
// InitTablet initializes the tablet record if necessary.
func (agent *ActionAgent) InitTablet(port, gRPCPort int32) 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 topodatapb.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
var err error
tabletType, err = topoproto.ParseTabletType(*initTabletType)
if err != nil {
log.Fatalf("Invalid init tablet type %v: %v", *initTabletType, err)
}
if tabletType == topodatapb.TabletType_MASTER {
// We disallow MASTER, so we don't have to change
// shard.MasterAlias, and deal with the corner cases.
log.Fatalf("init_tablet_type cannot be %v", tabletType)
}
} else if *targetTabletType != "" {
if strings.ToUpper(*targetTabletType) == topodatapb.TabletType_name[int32(topodatapb.TabletType_MASTER)] {
log.Fatalf("target_tablet_type cannot be '%v'. Use '%v' instead.", tabletType, topodatapb.TabletType_REPLICA)
}
// use spare, the healthcheck will turn us into what
// we need to be eventually
tabletType = topodatapb.TabletType_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()
// since we're assigned to a shard, make sure it exists, see if
// we are its master, and update its cells list if necessary
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 != nil && topoproto.TabletAliasEqual(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 = topodatapb.TabletType_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 := agent.TopoServer.UpdateShard(ctx, 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 != "" {
log.Infof("Using hostname: %v from -tablet_hostname flag.", hostname)
} else {
hostname, err := netutil.FullyQualifiedHostname()
if err != nil {
return err
}
log.Infof("Using detected machine hostname: %v To change this, fix your machine network configuration or override it with -tablet_hostname.", hostname)
}
// create and populate tablet record
tablet := &topodatapb.Tablet{
Alias: agent.TabletAlias,
Hostname: hostname,
PortMap: make(map[string]int32),
Keyspace: *initKeyspace,
Shard: *initShard,
Type: tabletType,
DbNameOverride: *initDbNameOverride,
Tags: initTags,
}
if port != 0 {
tablet.PortMap["vt"] = port
}
if gRPCPort != 0 {
tablet.PortMap["grpc"] = gRPCPort
}
if err := topo.TabletComplete(tablet); err != nil {
return fmt.Errorf("InitTablet TabletComplete failed: %v", err)
}
// now try to create the record
err = agent.TopoServer.CreateTablet(ctx, tablet)
switch err {
case nil:
// it worked, we're good, can update the replication graph
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 {
return 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 := agent.TopoServer.UpdateTablet(ctx, 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 err := topotools.UpdateTabletEndpoints(ctx, agent.TopoServer, tablet); err != nil {
return fmt.Errorf("UpdateTabletEndpoints failed: %v", err)
}
return nil
}
// TabletExternallyReparented updates all topo records so the current
// tablet is the new master for this shard.
func (agent *ActionAgent) TabletExternallyReparented(ctx context.Context, externalID string) error {
if err := agent.lock(ctx); err != nil {
return err
}
defer agent.unlock()
startTime := time.Now()
// If there is a finalize step running, wait for it to finish or time out
// before checking the global shard record again.
if agent.finalizeReparentCtx != nil {
select {
case <-agent.finalizeReparentCtx.Done():
agent.finalizeReparentCtx = nil
case <-ctx.Done():
return ctx.Err()
}
}
tablet := agent.Tablet()
// Check the global shard record.
si, err := agent.TopoServer.GetShard(ctx, tablet.Keyspace, tablet.Shard)
if err != nil {
log.Warningf("fastTabletExternallyReparented: failed to read global shard record for %v/%v: %v", tablet.Keyspace, tablet.Shard, err)
return err
}
if topoproto.TabletAliasEqual(si.MasterAlias, tablet.Alias) {
// We may get called on the current master even when nothing has changed.
// If the global shard record is already updated, it means we successfully
// finished a previous reparent to this tablet.
return nil
}
// Remember when we were first told we're the master.
// If another tablet claims to be master and offers a more recent time,
// that tablet will be trusted over us.
agent.mutex.Lock()
agent._tabletExternallyReparentedTime = startTime
agent._replicationDelay = 0
agent.mutex.Unlock()
// Create a reusable Reparent event with available info.
ev := &events.Reparent{
ShardInfo: *si,
NewMaster: *tablet,
OldMaster: topodatapb.Tablet{
Alias: si.MasterAlias,
Type: topodatapb.TabletType_MASTER,
},
ExternalID: externalID,
}
defer func() {
if err != nil {
event.DispatchUpdate(ev, "failed: "+err.Error())
}
}()
event.DispatchUpdate(ev, "starting external from tablet (fast)")
// Execute state change to master by force-updating only the local copy of the
// tablet record. The actual record in topo will be updated later.
log.Infof("fastTabletExternallyReparented: executing change callback for state change to MASTER")
oldTablet := proto.Clone(tablet).(*topodatapb.Tablet)
tablet.Type = topodatapb.TabletType_MASTER
agent.setTablet(tablet)
// This is where updateState will block for gracePeriod, while it gives
// vtgate a chance to stop sending replica queries.
agent.updateState(ctx, oldTablet, "fastTabletExternallyReparented")
// Start the finalize stage with a background context, but connect the trace.
bgCtx, cancel := context.WithTimeout(agent.batchCtx, *finalizeReparentTimeout)
bgCtx = trace.CopySpan(bgCtx, ctx)
agent.finalizeReparentCtx = bgCtx
go func() {
err := agent.finalizeTabletExternallyReparented(bgCtx, si, ev)
cancel()
if err != nil {
log.Warningf("finalizeTabletExternallyReparented error: %v", err)
event.DispatchUpdate(ev, "failed: "+err.Error())
return
}
externalReparentStats.Record("FullRebuild", startTime)
}()
return nil
}