コード例 #1
0
ファイル: raft.go プロジェクト: Mic92/docker
// updateMember submits a configuration change to change a member's address.
func (n *Node) updateMember(ctx context.Context, addr string, raftID uint64, nodeID string) error {
	node := api.RaftMember{
		RaftID: raftID,
		NodeID: nodeID,
		Addr:   addr,
	}

	meta, err := node.Marshal()
	if err != nil {
		return err
	}

	cc := raftpb.ConfChange{
		Type:    raftpb.ConfChangeUpdateNode,
		NodeID:  raftID,
		Context: meta,
	}

	// Wait for a raft round to process the configuration change
	return n.configure(ctx, cc)
}
コード例 #2
0
ファイル: storage.go プロジェクト: jfrazelle/docker
// bootstraps a node's raft store from the raft logs and snapshots on disk
func (n *Node) loadAndStart(ctx context.Context, forceNewCluster bool) error {
	snapshot, waldata, err := n.readFromDisk(ctx)
	if err != nil {
		return err
	}

	// Read logs to fully catch up store
	var raftNode api.RaftMember
	if err := raftNode.Unmarshal(waldata.Metadata); err != nil {
		return errors.Wrap(err, "failed to unmarshal WAL metadata")
	}
	n.Config.ID = raftNode.RaftID

	if snapshot != nil {
		snapCluster, err := n.clusterSnapshot(snapshot.Data)
		if err != nil {
			return err
		}
		var bootstrapMembers []*api.RaftMember
		if forceNewCluster {
			for _, m := range snapCluster.Members {
				if m.RaftID != n.Config.ID {
					n.cluster.RemoveMember(m.RaftID)
					continue
				}
				bootstrapMembers = append(bootstrapMembers, m)
			}
		} else {
			bootstrapMembers = snapCluster.Members
		}
		n.bootstrapMembers = bootstrapMembers
		for _, removedMember := range snapCluster.Removed {
			n.cluster.RemoveMember(removedMember)
		}
	}

	ents, st := waldata.Entries, waldata.HardState

	// All members that are no longer part of the cluster must be added to
	// the removed list right away, so that we don't try to connect to them
	// before processing the configuration change entries, which could make
	// us get stuck.
	for _, ent := range ents {
		if ent.Index <= st.Commit && ent.Type == raftpb.EntryConfChange {
			var cc raftpb.ConfChange
			if err := cc.Unmarshal(ent.Data); err != nil {
				return errors.Wrap(err, "failed to unmarshal config change")
			}
			if cc.Type == raftpb.ConfChangeRemoveNode {
				n.cluster.RemoveMember(cc.NodeID)
			}
		}
	}

	if forceNewCluster {
		// discard the previously uncommitted entries
		for i, ent := range ents {
			if ent.Index > st.Commit {
				log.G(ctx).Infof("discarding %d uncommitted WAL entries", len(ents)-i)
				ents = ents[:i]
				break
			}
		}

		// force append the configuration change entries
		toAppEnts := createConfigChangeEnts(getIDs(snapshot, ents), n.Config.ID, st.Term, st.Commit)

		// All members that are being removed as part of the
		// force-new-cluster process must be added to the
		// removed list right away, so that we don't try to
		// connect to them before processing the configuration
		// change entries, which could make us get stuck.
		for _, ccEnt := range toAppEnts {
			if ccEnt.Type == raftpb.EntryConfChange {
				var cc raftpb.ConfChange
				if err := cc.Unmarshal(ccEnt.Data); err != nil {
					return errors.Wrap(err, "error unmarshalling force-new-cluster config change")
				}
				if cc.Type == raftpb.ConfChangeRemoveNode {
					n.cluster.RemoveMember(cc.NodeID)
				}
			}
		}
		ents = append(ents, toAppEnts...)

		// force commit newly appended entries
		err := n.raftLogger.SaveEntries(st, toAppEnts)
		if err != nil {
			log.G(ctx).WithError(err).Fatalf("failed to save WAL while forcing new cluster")
		}
		if len(toAppEnts) != 0 {
			st.Commit = toAppEnts[len(toAppEnts)-1].Index
		}
	}

	if snapshot != nil {
		if err := n.raftStore.ApplySnapshot(*snapshot); err != nil {
			return err
		}
	}
	if err := n.raftStore.SetHardState(st); err != nil {
		return err
	}
	return n.raftStore.Append(ents)
}
コード例 #3
0
func (n *Node) readWAL(ctx context.Context, snapshot *raftpb.Snapshot, forceNewCluster bool) (err error) {
	var (
		walsnap  walpb.Snapshot
		metadata []byte
		st       raftpb.HardState
		ents     []raftpb.Entry
	)

	if snapshot != nil {
		walsnap.Index = snapshot.Metadata.Index
		walsnap.Term = snapshot.Metadata.Term
	}

	repaired := false
	for {
		if n.wal, err = wal.Open(n.walDir(), walsnap); err != nil {
			return fmt.Errorf("open WAL error: %v", err)
		}
		if metadata, st, ents, err = n.wal.ReadAll(); err != nil {
			if err := n.wal.Close(); err != nil {
				return err
			}
			// we can only repair ErrUnexpectedEOF and we never repair twice.
			if repaired || err != io.ErrUnexpectedEOF {
				return fmt.Errorf("read WAL error (%v) and cannot be repaired", err)
			}
			if !wal.Repair(n.walDir()) {
				return fmt.Errorf("WAL error (%v) cannot be repaired", err)
			}
			log.G(ctx).Infof("repaired WAL error (%v)", err)
			repaired = true
			continue
		}
		break
	}

	defer func() {
		if err != nil {
			if walErr := n.wal.Close(); walErr != nil {
				n.Config.Logger.Errorf("error closing raft WAL: %v", walErr)
			}
		}
	}()

	var raftNode api.RaftMember
	if err := raftNode.Unmarshal(metadata); err != nil {
		return fmt.Errorf("error unmarshalling WAL metadata: %v", err)
	}
	n.Config.ID = raftNode.RaftID

	// All members that are no longer part of the cluster must be added to
	// the removed list right away, so that we don't try to connect to them
	// before processing the configuration change entries, which could make
	// us get stuck.
	for _, ent := range ents {
		if ent.Index <= st.Commit && ent.Type == raftpb.EntryConfChange {
			var cc raftpb.ConfChange
			if err := cc.Unmarshal(ent.Data); err != nil {
				return fmt.Errorf("error unmarshalling config change: %v", err)
			}
			if cc.Type == raftpb.ConfChangeRemoveNode {
				n.cluster.RemoveMember(cc.NodeID)
			}
		}
	}

	if forceNewCluster {
		// discard the previously uncommitted entries
		for i, ent := range ents {
			if ent.Index > st.Commit {
				log.G(context.Background()).Infof("discarding %d uncommitted WAL entries ", len(ents)-i)
				ents = ents[:i]
				break
			}
		}

		// force append the configuration change entries
		toAppEnts := createConfigChangeEnts(getIDs(snapshot, ents), uint64(n.Config.ID), st.Term, st.Commit)

		// All members that are being removed as part of the
		// force-new-cluster process must be added to the
		// removed list right away, so that we don't try to
		// connect to them before processing the configuration
		// change entries, which could make us get stuck.
		for _, ccEnt := range toAppEnts {
			if ccEnt.Type == raftpb.EntryConfChange {
				var cc raftpb.ConfChange
				if err := cc.Unmarshal(ccEnt.Data); err != nil {
					return fmt.Errorf("error unmarshalling force-new-cluster config change: %v", err)
				}
				if cc.Type == raftpb.ConfChangeRemoveNode {
					n.cluster.RemoveMember(cc.NodeID)
				}
			}
		}
		ents = append(ents, toAppEnts...)

		// force commit newly appended entries
		err := n.wal.Save(st, toAppEnts)
		if err != nil {
			log.G(context.Background()).Fatalf("%v", err)
		}
		if len(toAppEnts) != 0 {
			st.Commit = toAppEnts[len(toAppEnts)-1].Index
		}
	}

	if snapshot != nil {
		if err := n.raftStore.ApplySnapshot(*snapshot); err != nil {
			return err
		}
	}
	if err := n.raftStore.SetHardState(st); err != nil {
		return err
	}
	if err := n.raftStore.Append(ents); err != nil {
		return err
	}

	return nil
}
コード例 #4
0
ファイル: storage.go プロジェクト: amitshukla/docker
func (n *Node) readWAL(ctx context.Context, snapshot *raftpb.Snapshot, forceNewCluster bool) (err error) {
	var (
		walsnap  walpb.Snapshot
		metadata []byte
		st       raftpb.HardState
		ents     []raftpb.Entry
	)

	if snapshot != nil {
		walsnap.Index = snapshot.Metadata.Index
		walsnap.Term = snapshot.Metadata.Term
	}

	repaired := false
	for {
		if n.wal, err = wal.Open(n.walDir(), walsnap); err != nil {
			return fmt.Errorf("open wal error: %v", err)
		}
		if metadata, st, ents, err = n.wal.ReadAll(); err != nil {
			if err := n.wal.Close(); err != nil {
				return err
			}
			// we can only repair ErrUnexpectedEOF and we never repair twice.
			if repaired || err != io.ErrUnexpectedEOF {
				return fmt.Errorf("read wal error (%v) and cannot be repaired", err)
			}
			if !wal.Repair(n.walDir()) {
				return fmt.Errorf("WAL error (%v) cannot be repaired", err)
			}
			log.G(ctx).Infof("repaired WAL error (%v)", err)
			repaired = true
			continue
		}
		break
	}

	defer func() {
		if err != nil {
			if walErr := n.wal.Close(); walErr != nil {
				n.Config.Logger.Errorf("error closing raft WAL: %v", walErr)
			}
		}
	}()

	var raftNode api.RaftMember
	if err := raftNode.Unmarshal(metadata); err != nil {
		return fmt.Errorf("error unmarshalling wal metadata: %v", err)
	}
	n.Config.ID = raftNode.RaftID

	if forceNewCluster {
		// discard the previously uncommitted entries
		for i, ent := range ents {
			if ent.Index > st.Commit {
				log.G(context.Background()).Infof("discarding %d uncommitted WAL entries ", len(ents)-i)
				ents = ents[:i]
				break
			}
		}

		// force append the configuration change entries
		toAppEnts := createConfigChangeEnts(getIDs(snapshot, ents), uint64(n.Config.ID), st.Term, st.Commit)
		ents = append(ents, toAppEnts...)

		// force commit newly appended entries
		err := n.wal.Save(st, toAppEnts)
		if err != nil {
			log.G(context.Background()).Fatalf("%v", err)
		}
		if len(toAppEnts) != 0 {
			st.Commit = toAppEnts[len(toAppEnts)-1].Index
		}
	}

	if snapshot != nil {
		if err := n.raftStore.ApplySnapshot(*snapshot); err != nil {
			return err
		}
	}
	if err := n.raftStore.SetHardState(st); err != nil {
		return err
	}
	if err := n.raftStore.Append(ents); err != nil {
		return err
	}

	return nil
}