func (wr *Wrangler) finishReparent(si *topo.ShardInfo, masterElect *topo.TabletInfo, majorityRestart, leaveMasterReadOnly bool) error {
// If the majority of slaves restarted, move ahead.
if majorityRestart {
if leaveMasterReadOnly {
wr.logger.Warningf("leaving master-elect read-only, change with: vtctl SetReadWrite %v", masterElect.Alias)
} else {
wr.logger.Infof("marking master-elect read-write %v", masterElect.Alias)
if err := wr.tmc.SetReadWrite(wr.ctx, masterElect); err != nil {
wr.logger.Warningf("master master-elect read-write failed, leaving master-elect read-only, change with: vtctl SetReadWrite %v", masterElect.Alias)
}
}
} else {
wr.logger.Warningf("minority reparent, manual fixes are needed, leaving master-elect read-only, change with: vtctl SetReadWrite %v", masterElect.Alias)
}
// save the new master in the shard info
si.MasterAlias = masterElect.Alias
if err := topo.UpdateShard(context.TODO(), wr.ts, si); err != nil {
wr.logger.Errorf("Failed to save new master into shard: %v", err)
return err
}
// We rebuild all the cells, as we may have taken tablets in and
// out of the graph.
wr.logger.Infof("rebuilding shard serving graph data")
_, err := topotools.RebuildShard(context.TODO(), wr.logger, wr.ts, masterElect.Keyspace, masterElect.Shard, nil, wr.lockTimeout, interrupted)
return err
}
// rebuildShardIfNeeded will rebuild the serving graph if we need to
func (agent *ActionAgent) rebuildShardIfNeeded(tablet *topo.TabletInfo, targetTabletType topo.TabletType) error {
if topo.IsInServingGraph(targetTabletType) {
// TODO: interrupted may need to be a global one closed when we exit
interrupted := make(chan struct{})
// no need to take the shard lock in this case
if _, err := topotools.RebuildShard(context.TODO(), logutil.NewConsoleLogger(), agent.TopoServer, tablet.Keyspace, tablet.Shard, []string{tablet.Alias.Cell}, agent.LockTimeout, interrupted); err != nil {
return fmt.Errorf("topotools.RebuildShard returned an error: %v", err)
}
}
return nil
}
// same as ChangeType, but assume we already have the shard lock,
// and do not have the option to force anything.
func (wr *Wrangler) changeTypeInternal(tabletAlias topo.TabletAlias, dbType topo.TabletType) error {
ti, err := wr.ts.GetTablet(tabletAlias)
if err != nil {
return err
}
rebuildRequired := ti.Tablet.IsInServingGraph()
// change the type
if err := wr.tmc.ChangeType(wr.ctx, ti, dbType); err != nil {
return err
}
// rebuild if necessary
if rebuildRequired {
_, err = topotools.RebuildShard(context.TODO(), wr.logger, wr.ts, ti.Keyspace, ti.Shard, []string{ti.Alias.Cell}, wr.lockTimeout, interrupted)
if err != nil {
return err
}
}
return nil
}
// Rebuild the serving and replication rollup data data while locking
// out other changes.
func (wr *Wrangler) RebuildShardGraph(keyspace, shard string, cells []string) (*topo.ShardInfo, error) {
return topotools.RebuildShard(context.TODO(), wr.logger, wr.ts, keyspace, shard, cells, wr.lockTimeout, interrupted)
}
// tabletExternallyReparentedLocked is called with the shard lock.
// It returns if agent.refreshTablet should be called, and the error.
// Note both are set independently (can have both true and an error).
func (agent *ActionAgent) tabletExternallyReparentedLocked(ctx context.Context, externalID string, interrupted chan struct{}) (bool, error) {
// re-read the tablet record to be sure we have the latest version
tablet, err := agent.TopoServer.GetTablet(agent.TabletAlias)
if err != nil {
return false, err
}
// read the shard, make sure again the master is not already good.
shardInfo, err := agent.TopoServer.GetShard(tablet.Keyspace, tablet.Shard)
if err != nil {
return false, err
}
if shardInfo.MasterAlias == tablet.Alias {
log.Infof("TabletExternallyReparented: tablet became the master before we get the lock?")
return false, nil
}
log.Infof("TabletExternallyReparented called and we're not the master, doing the work")
// Read the tablets, make sure the master elect is known to the shard
// (it's this tablet, so it better be!).
// Note we will keep going with a partial tablet map, which usually
// happens when a cell is not reachable. After these checks, the
// guarantees we'll have are:
// - global cell is reachable (we just locked and read the shard)
// - the local cell that contains the new master is reachable
// (as we're going to check the new master is in the list)
// That should be enough.
tabletMap, err := topo.GetTabletMapForShard(ctx, agent.TopoServer, tablet.Keyspace, tablet.Shard)
switch err {
case nil:
// keep going
case topo.ErrPartialResult:
log.Warningf("Got topo.ErrPartialResult from GetTabletMapForShard, may need to re-init some tablets")
default:
return false, err
}
masterElectTablet, ok := tabletMap[tablet.Alias]
if !ok {
return false, fmt.Errorf("this master-elect tablet %v not found in replication graph %v/%v %v", tablet.Alias, tablet.Keyspace, tablet.Shard, topotools.MapKeys(tabletMap))
}
// Create reusable Reparent event with available info
ev := &events.Reparent{
ShardInfo: *shardInfo,
NewMaster: *tablet.Tablet,
ExternalID: externalID,
}
if oldMasterTablet, ok := tabletMap[shardInfo.MasterAlias]; ok {
ev.OldMaster = *oldMasterTablet.Tablet
}
defer func() {
if err != nil {
event.DispatchUpdate(ev, "failed: "+err.Error())
}
}()
// sort the tablets, and handle them
slaveTabletMap, masterTabletMap := topotools.SortedTabletMap(tabletMap)
event.DispatchUpdate(ev, "starting external from tablet")
// We fix the new master in the replication graph.
// Note after this call, we may have changed the tablet record,
// so we will always return true, so the tablet record is re-read
// by the agent.
event.DispatchUpdate(ev, "mark ourself as new master")
err = agent.updateReplicationGraphForPromotedSlave(ctx, tablet)
if err != nil {
// This suggests we can't talk to topo server. This is bad.
return true, fmt.Errorf("updateReplicationGraphForPromotedSlave failed: %v", err)
}
// Once this tablet is promoted, remove it from our maps
delete(slaveTabletMap, tablet.Alias)
delete(masterTabletMap, tablet.Alias)
// Then fix all the slaves, including the old master. This
// last step is very likely to time out for some tablets (one
// random guy is dead, the old master is dead, ...). We
// execute them all in parallel until we get to
// wr.ActionTimeout(). After this, no other action with a
// timeout is executed, so even if we got to the timeout,
// we're still good.
event.DispatchUpdate(ev, "restarting slaves")
logger := logutil.NewConsoleLogger()
tmc := tmclient.NewTabletManagerClient()
topotools.RestartSlavesExternal(agent.TopoServer, logger, slaveTabletMap, masterTabletMap, masterElectTablet.Alias, func(ti *topo.TabletInfo, swrd *actionnode.SlaveWasRestartedArgs) error {
return tmc.SlaveWasRestarted(ctx, ti, swrd)
})
// Compute the list of Cells we need to rebuild: old master and
// all other cells if reparenting to another cell.
cells := []string{shardInfo.MasterAlias.Cell}
if shardInfo.MasterAlias.Cell != tablet.Alias.Cell {
cells = nil
}
// now update the master record in the shard object
event.DispatchUpdate(ev, "updating shard record")
log.Infof("Updating Shard's MasterAlias record")
shardInfo.MasterAlias = tablet.Alias
if err = topo.UpdateShard(ctx, agent.TopoServer, shardInfo); err != nil {
return true, err
}
// and rebuild the shard serving graph
event.DispatchUpdate(ev, "rebuilding shard serving graph")
log.Infof("Rebuilding shard serving graph data")
if _, err = topotools.RebuildShard(ctx, logger, agent.TopoServer, tablet.Keyspace, tablet.Shard, cells, agent.LockTimeout, interrupted); err != nil {
return true, err
}
event.DispatchUpdate(ev, "finished")
return true, nil
}