Exemple #1
0
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 {
			log.Warningf("leaving master-elect read-only, change with: vtctl SetReadWrite %v", masterElect.Alias)
		} else {
			log.Infof("marking master-elect read-write %v", masterElect.Alias)
			actionPath, err := wr.ai.SetReadWrite(masterElect.Alias)
			if err == nil {
				err = wr.ai.WaitForCompletion(actionPath, wr.actionTimeout())
			}
			if err != nil {
				log.Warningf("master master-elect read-write failed, leaving master-elect read-only, change with: vtctl SetReadWrite %v", masterElect.Alias)
			}
		}
	} else {
		log.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 := wr.ts.UpdateShard(si); err != nil {
		log.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.
	log.Infof("rebuilding shard serving graph data")
	return topotools.RebuildShard(wr.ts, masterElect.Keyspace, masterElect.Shard, topotools.RebuildShardOptions{IgnorePartialResult: false}, wr.lockTimeout, interrupted)
}
Exemple #2
0
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(masterElect, wr.ActionTimeout()); 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(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(wr.logger, wr.ts, masterElect.Keyspace, masterElect.Shard, nil, wr.lockTimeout, interrupted)
	return err
}
Exemple #3
0
// 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 topo.TabletAlias, tabletType topo.TabletType, force 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 == topo.TYPE_MASTER && 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 == topo.TYPE_MASTER && si.MasterAlias != tabletAlias {
		if !si.MasterAlias.IsZero() && !force {
			return wr.unlockShard(ctx, keyspace, shard, actionNode, lockPath, fmt.Errorf("creating this tablet would override old master %v in shard %v/%v", si.MasterAlias, keyspace, shard))
		}
		si.MasterAlias = tabletAlias
		wasUpdated = true
	}

	if wasUpdated {
		// write it back
		if err := topo.UpdateShard(ctx, wr.ts, si); err != nil {
			return wr.unlockShard(ctx, keyspace, shard, actionNode, lockPath, err)
		}
	}

	// and unlock
	return wr.unlockShard(ctx, keyspace, shard, actionNode, lockPath, err)
}
// finalizeTabletExternallyReparented performs slow, synchronized reconciliation
// tasks that ensure topology is self-consistent, and then marks the reparent as
// finished by updating the global shard record.
func (agent *ActionAgent) finalizeTabletExternallyReparented(ctx context.Context, si *topo.ShardInfo, ev *events.Reparent) (err error) {
	var wg sync.WaitGroup
	var errs concurrency.AllErrorRecorder
	oldMasterAlias := si.MasterAlias

	// Update the tablet records concurrently.
	event.DispatchUpdate(ev, "updating old and new master tablet records")
	log.Infof("finalizeTabletExternallyReparented: updating tablet records")
	wg.Add(1)
	go func() {
		defer wg.Done()
		// Update our own record to master.
		_, err := agent.TopoServer.UpdateTabletFields(ctx, agent.TabletAlias,
			func(tablet *topodatapb.Tablet) error {
				tablet.Type = topodatapb.TabletType_MASTER
				return nil
			})
		if err != nil {
			errs.RecordError(err)
		}
	}()

	if !topoproto.TabletAliasIsZero(oldMasterAlias) {
		wg.Add(1)
		go func() {
			// Forcibly demote the old master in topology, since we can't rely on the
			// old master to be up to change its own record.
			oldMasterTablet, err := agent.TopoServer.UpdateTabletFields(ctx, oldMasterAlias,
				func(tablet *topodatapb.Tablet) error {
					tablet.Type = topodatapb.TabletType_REPLICA
					return nil
				})
			if err != nil {
				errs.RecordError(err)
				wg.Done()
				return
			}

			// We now know more about the old master, so add it to event data.
			ev.OldMaster = *oldMasterTablet
			wg.Done()

			// Tell the old master to re-read its tablet record and change its state.
			// We don't need to wait for it.
			tmc := tmclient.NewTabletManagerClient()
			tmc.RefreshState(ctx, oldMasterTablet)
		}()
	}

	tablet := agent.Tablet()

	// Wait for the tablet records to be updated. At that point, any rebuild will
	// see the new master, so we're ready to mark the reparent as done in the
	// global shard record.
	wg.Wait()
	if errs.HasErrors() {
		return errs.Error()
	}

	// Update the master field in the global shard record. We don't use a lock
	// here anymore. The lock was only to ensure that the global shard record
	// didn't get modified between the time when we read it and the time when we
	// write it back. Now we use an update loop pattern to do that instead.
	event.DispatchUpdate(ev, "updating global shard record")
	log.Infof("finalizeTabletExternallyReparented: updating global shard record if needed")
	_, err = agent.TopoServer.UpdateShardFields(ctx, tablet.Keyspace, tablet.Shard, func(si *topo.ShardInfo) error {
		if topoproto.TabletAliasEqual(si.MasterAlias, tablet.Alias) {
			return topo.ErrNoUpdateNeeded
		}
		si.MasterAlias = tablet.Alias
		return nil
	})
	if err != nil {
		return err
	}

	event.DispatchUpdate(ev, "finished")
	return nil
}