Fabric網路組織與主節點選舉

• 2020 年 6 月 18 日
• 筆記
• 區塊鏈

一、Fabric網路組織

Fabric網路組織按如下結構組成：Fabric網路–>Channel通道–>組織（成員）–>節點。即整個網路由數個通道組成，每個通道都由多個組織構成，而每個組織內部由數個節點組成（可能由功能或其他劃分方式分為多個節點）。如下圖所示：

二、主節點（leader peer）選舉

一個組織（其實是成員）在一個通道上可以有多個Peer節點，這時候為了提高通訊效率，需要選舉出來一個主節點（leader）作為代表和排序服務節點通訊，負責從排序服務節點處獲取最新的區塊並在組織內部同步。有如下兩種方式：

1. 靜態指定

配置文件中配置

    # Gossip related configuration
    gossip:
        # Defines whenever peer will initialize dynamic algorithm for
        # "leader" selection, where leader is the peer to establish
        # connection with ordering service and use delivery protocol
        # to pull ledger blocks from ordering service
        useLeaderElection: false
        # Statically defines peer to be an organization "leader",
        # where this means that current peer will maintain connection
        # with ordering service and disseminate block across peers in
        # its own organization
        orgLeader: true

2. 動態選舉

相關配置：

    # Gossip related configuration
    gossip:
        # Leader election service configuration
        election:
            # Longest time peer wait for stable membership during leader election startup (unit: second)
            startupGracePeriod: 15s
            # Interval gossip membership sampled to check its stability (unit: second)
            membershipSampleInterval: 1s
            # Time pass since last declaration message before peer decide to go to election (unit: second)
            leaderAliveThreshold: 10s
            # Time between peer sends propose message and declare itself as a leader (sends declaration message) (unit: second)
            leaderElectionDuration: 5s

3. leader節點選舉流程

選舉流程（簡要）：

如果當前沒有leader，進入選舉演算法
    如果當前是leader:廣播leadership declearation，如果收到比自己小的leadership declearation，自己變為follower;
    如果當前是follower:指定時間內沒有收到leadership declearation，則認為leader離線了，進入選舉流程

選舉演算法（簡要）:

廣播提議自己為leader消息
各個節點收集選舉消息
比對ID，如果自己ID最小，則自己為leader

詳細過程如下圖所示：

偽程式碼實現：

// Gossip leader election module
// Algorithm properties:
// - Peers break symmetry by comparing IDs
// - Each peer is either a leader or a follower,
//   and the aim is to have exactly 1 leader if the membership view
//   is the same for all peers
// - If the network is partitioned into 2 or more sets, the number of leaders
//   is the number of network partitions, but when the partition heals,
//   only 1 leader should be left eventually
// - Peers communicate by gossiping leadership proposal or declaration messages

// The Algorithm, in pseudo code:
//
//
// variables:
// 	leaderKnown = false
//
// Invariant:
//	Peer listens for messages from remote peers
//	and whenever it receives a leadership declaration,
//	leaderKnown is set to true
//
// Startup():
// 	wait for membership view to stabilize, or for a leadership declaration is received
//      or the startup timeout expires.
//	goto SteadyState()
//
// SteadyState():
// 	while true:
//		If leaderKnown is false:
// 			LeaderElection()
//		If you are the leader:
//			Broadcast leadership declaration
//			If a leadership declaration was received from
// 			a peer with a lower ID,
//			become a follower
//		Else, you're a follower:
//			If haven't received a leadership declaration within
// 			a time threshold:
//				set leaderKnown to false
//
// LeaderElection():
// 	Gossip leadership proposal message
//	Collect messages from other peers sent within a time period
//	If received a leadership declaration:
//		return
//	Iterate over all proposal messages collected.
// 	If a proposal message from a peer with an ID lower
// 	than yourself was received, return.
//	Else, declare yourself a leader

4. 消息定義

// Leadership Message is sent during leader election to inform
// remote peers about intent of peer to proclaim itself as leader
message LeadershipMessage {
    bytes pki_id        = 1;
    PeerTime timestamp = 2;
    bool is_declaration = 3;
}