今天我們用配置變更來結(jié)束整個EtcdRaft源碼分析系列恒界。橫向擴展能力是衡量分布式系統(tǒng)優(yōu)劣的決定性指標十酣，而能否輕松，快捷耸采，有效虾宇，及時的變更集群成員是其中的關(guān)鍵。下面我們一起來看看EtcdRaft是怎么實現(xiàn)的好唯。

接口

type Node interface {
    ...
   // ProposeConfChange proposes config change.
   // At most one ConfChange can be in the process of going through consensus.
   // Application needs to call ApplyConfChange when applying EntryConfChange type entry.
   ProposeConfChange(ctx context.Context, cc pb.ConfChange) error
    ...
   // ApplyConfChange applies config change to the local node.
   // Returns an opaque ConfState protobuf which must be recorded
   // in snapshots. Will never return nil; it returns a pointer only
   // to match MemoryStorage.Compact.
   ApplyConfChange(cc pb.ConfChange) *pb.ConfState
    ...
}

可以看到有兩個方法跟配置變更相關(guān)燥翅，看過前面的知道，外部跟Raft打交道的方式靶端。先提案(propose), 然后等內(nèi)部達成一致，再落地(Apply)杨名。

struct

type ConfChange struct {
   ID               uint64         `protobuf:"varint,1,opt,name=ID" json:"ID"`
   Type             ConfChangeType `protobuf:"varint,2,opt,name=Type,enum=raftpb.ConfChangeType" json:"Type"`
   NodeID           uint64         `protobuf:"varint,3,opt,name=NodeID" json:"NodeID"`
   Context          []byte         `protobuf:"bytes,4,opt,name=Context" json:"Context,omitempty"`
   XXX_unrecognized []byte         `json:"-"`
}


const (
    ConfChangeAddNode        ConfChangeType = 0
    ConfChangeRemoveNode     ConfChangeType = 1
    ConfChangeUpdateNode     ConfChangeType = 2
    ConfChangeAddLearnerNode ConfChangeType = 3
)

以上是提案內(nèi)容台谍，很清晰，但有個地方需要注意趁蕊，一次只能變更一個節(jié)點。至于為什么是己，有興趣的可以去看論文哈任柜。

Propose

func (n *node) ProposeConfChange(ctx context.Context, cc pb.ConfChange) error {
   data, err := cc.Marshal()
   if err != nil {
      return err
   }
   return n.Step(ctx, pb.Message{Type: pb.MsgProp, Entries: []pb.Entry{{Type: pb.EntryConfChange, Data: data}}})
}

• 基本就是走的提交數(shù)據(jù)的流程，唯一需要注意的是這里用pb.EntryConfChange將它與其他提案區(qū)別開來摔认。
• 接下來宅粥，我們再走一遍數(shù)據(jù)提交的流程

Leader

...
for i, e := range m.Entries {
   if e.Type == pb.EntryConfChange {
      if r.pendingConfIndex > r.raftLog.applied {
         r.logger.Infof("propose conf %s ignored since pending unapplied configuration [index %d, applied %d]",
            e.String(), r.pendingConfIndex, r.raftLog.applied)
         m.Entries[i] = pb.Entry{Type: pb.EntryNormal}
      } else {
         r.pendingConfIndex = r.raftLog.lastIndex() + uint64(i) + 1
      }
   }
}
...
if !r.appendEntry(m.Entries...) {
    return ErrProposalDropped
}
r.bcastAppend()

• 如果當前還有配置更新沒有處理完粹胯，那么這次新的變更將丟棄，用一個空的entry來替換它
• 如果都處理完了风纠，那么記下這個配置變更的位置到pendingConfIndex
• 后面就一樣了竹观，累加到本地，而且群發(fā)給其他人臭增。
• 問題來了，配置變更都同步給成員了列牺，怎么確認都收到了拗窃，可以開始apply了呢泌辫？我想也猜得到九默，會通過Ready的committedIndex來通知應(yīng)用層驼修。

apply

應(yīng)用層

case raftpb.EntryConfChange:
   var cc raftpb.ConfChange
   if err := cc.Unmarshal(ents[i].Data); err != nil {
      c.logger.Warnf("Failed to unmarshal ConfChange data: %s", err)
      continue
   }

   c.confState = *c.Node.ApplyConfChange(cc)

   switch cc.Type {
   case raftpb.ConfChangeAddNode:
      c.logger.Infof("Applied config change to add node %d, current nodes in channel: %+v", cc.NodeID, c.confState.Nodes)
   case raftpb.ConfChangeRemoveNode:
      c.logger.Infof("Applied config change to remove node %d, current nodes in channel: %+v", cc.NodeID, c.confState.Nodes)
   default:
      c.logger.Panic("Programming error, encountered unsupported raft config change")
   }

   // This ConfChange was introduced by a previously committed config block,
   // we can now unblock submitC to accept envelopes.
   if c.confChangeInProgress != nil &&
      c.confChangeInProgress.NodeID == cc.NodeID &&
      c.confChangeInProgress.Type == cc.Type {

      if err := c.configureComm(); err != nil {
         c.logger.Panicf("Failed to configure communication: %s", err)
      }

      c.confChangeInProgress = nil
      c.configInflight = false
      // report the new cluster size
      c.Metrics.ClusterSize.Set(float64(len(c.opts.RaftMetadata.Consenters)))
   }

   if cc.Type == raftpb.ConfChangeRemoveNode && cc.NodeID == c.raftID {
      c.logger.Infof("Current node removed from replica set for channel %s", c.channelID)
      // calling goroutine, since otherwise it will be blocked
      // trying to write into haltC
      go c.Halt()
   }
}

• 這里乙各，我舉的是Fabric的例子，只關(guān)注關(guān)鍵流程就好
• 收到Raft的ConfChange觅丰，第一件事妇萄，我們就要Node.ApplyConfChange(cc)
• Raft的通訊層是需要應(yīng)用層托管的咬荷，所以不是Raft那邊做完配置變更，就可以收工了懦底。
  • Fabric要根據(jù)最新的集群成員數(shù)據(jù)遂唧，去做grpc的連接
  • 如果有刪除的節(jié)點辫继，還要去停掉這個成員杆查，這個后面會講臀蛛。

Raft

ApplyConfChange

case cc := <-n.confc:
            if cc.NodeID == None {
                select {
                case n.confstatec <- pb.ConfState{
                    Nodes:    r.nodes(),
                    Learners: r.learnerNodes()}:
                case <-n.done:
                }
                break
            }
            switch cc.Type {
            case pb.ConfChangeAddNode:
                r.addNode(cc.NodeID)
            case pb.ConfChangeAddLearnerNode:
                r.addLearner(cc.NodeID)
            case pb.ConfChangeRemoveNode:
                // block incoming proposal when local node is
                // removed
                if cc.NodeID == r.id {
                    propc = nil
                }
                r.removeNode(cc.NodeID)
            case pb.ConfChangeUpdateNode:
            default:
                panic("unexpected conf type")
            }
            select {
            case n.confstatec <- pb.ConfState{
                Nodes:    r.nodes(),
                Learners: r.learnerNodes()}:
            case <-n.done:
            }

• 首先浊仆，這里有個調(diào)用技巧，如果調(diào)用的時候傳入的NodeID為None舔琅，那么會返回當前Raft的成員
• 下面我們具體看下這幾種變更類型具體是在干嘛

addNode&addLearner

func (r *raft) addNodeOrLearnerNode(id uint64, isLearner bool) {
   pr := r.getProgress(id)
   if pr == nil {
      r.setProgress(id, 0, r.raftLog.lastIndex()+1, isLearner)
   } else {
      if isLearner && !pr.IsLearner {
         // can only change Learner to Voter
         r.logger.Infof("%x ignored addLearner: do not support changing %x from raft peer to learner.", r.id, id)
         return
      }

      if isLearner == pr.IsLearner {
         // Ignore any redundant addNode calls (which can happen because the
         // initial bootstrapping entries are applied twice).
         return
      }

      // change Learner to Voter, use origin Learner progress
      delete(r.learnerPrs, id)
      pr.IsLearner = false
      r.prs[id] = pr
   }

   if r.id == id {
      r.isLearner = isLearner
   }

   // When a node is first added, we should mark it as recently active.
   // Otherwise, CheckQuorum may cause us to step down if it is invoked
   // before the added node has a chance to communicate with us.
   pr = r.getProgress(id)
   pr.RecentActive = true
}

• 如果是全新的節(jié)點搏明，初始化Progress，這里match=0星著，next=r.raftLog.lastIndex()+1
• 如果之前是learner虚循，那么從learner里面轉(zhuǎn)移到正常節(jié)點里面

removeNode

func (r *raft) removeNode(id uint64) {
   r.delProgress(id)

   // do not try to commit or abort transferring if there is no nodes in the cluster.
   if len(r.prs) == 0 && len(r.learnerPrs) == 0 {
      return
   }

   // The quorum size is now smaller, so see if any pending entries can
   // be committed.
   if r.maybeCommit() {
      r.bcastAppend()
   }
   // If the removed node is the leadTransferee, then abort the leadership transferring.
   if r.state == StateLeader && r.leadTransferee == id {
      r.abortLeaderTransfer()
   }
}

• 真的刪除Progress就夠了么？想象下整個系統(tǒng)運轉(zhuǎn)是靠什么铺遂？
• 還記得應(yīng)用層會tick么茎刚？Leader靠這個發(fā)心跳，非Leader靠這個選舉超時粮坞。
• 我們再回顧下應(yīng)用層還做了些什么

應(yīng)用層

if cc.Type == raftpb.ConfChangeRemoveNode && cc.NodeID == c.raftID {
    // calling goroutine, since otherwise it will be blocked
    // trying to write into haltC
    go c.Halt()
}

case <-n.chain.haltC:
   ticker.Stop()
   n.Stop()
   n.storage.Close()
   n.logger.Infof("Raft node stopped")
   close(n.chain.doneC) // close after all the artifacts are closed
   return
}

• 可以看到如果是刪除當前節(jié)點的消息初狰，會最終會讓該節(jié)點的ticker.Stop。這也導(dǎo)致該節(jié)點最終會被Raft拋棄筝闹。