etcd raft library设计原理和使用

zeromax

早在2013年11月份，在raft论文还只能在网上下载到草稿版时，我曾经写过一篇blog对其进行简要分析。4年过去了，各种raft协议的讲解铺天盖地，raft也确实得到了广泛的应用。其中最知名的应用莫过于etcd。etcd将raft协议本身实现为一个library，位于https://github.com/coreos/etcd/tree/master/raft，然后本身作为一个应用使用它。
本文不讲解raft协议核心内容，而是站在一个etcd raft library使用者的角度，讲解要用上这个library需要了解的东西。
这个library使用起来相对来说还是有点麻烦。官方有一个使用示例在 https://github.com/coreos/etcd/tree/master/contrib/raftexample。整体来说，这个库实现了raft协议核心的内容，比如append log的逻辑，选主逻辑，snapshot，成员变更等逻辑。需要明确的是：library没有实现消息的网络传输和接收，库只会把一些待发送的消息保存在内存中，用户自定义的网络传输层取出消息并发送出去，并且在网络接收端，需要调一个library的函数，用于将收到的消息传入library，后面会详细说明。同时，library定义了一个Storage接口，需要library的使用者自行实现。
Storage接口如下:
// Storage is an interface that may be implemented by the application// to retrieve log entries from storage.//// If any Storage method returns an error, the raft instance will// become inoperable and refuse to participate in elections; the// application is responsible for cleanup and recovery in this case.type Storage interface {    // InitialState returns the saved HardState and ConfState information.
    InitialState() (pb.HardState, pb.ConfState, error)    // Entries returns a slice of log entries in the range [lo,hi).
    // MaxSize limits the total size of the log entries returned, but
    // Entries returns at least one entry if any.
    Entries(lo, hi, maxSize uint64) ([]pb.Entry, error)    // Term returns the term of entry i, which must be in the range
    // [FirstIndex()-1, LastIndex()]. The term of the entry before
    // FirstIndex is retained for matching purposes even though the
    // rest of that entry may not be available.
    Term(i uint64) (uint64, error)    // LastIndex returns the index of the last entry in the log.
    LastIndex() (uint64, error)    // FirstIndex returns the index of the first log entry that is
    // possibly available via Entries (older entries have been incorporated
    // into the latest Snapshot; if storage only contains the dummy entry the
    // first log entry is not available).
    FirstIndex() (uint64, error)    // Snapshot returns the most recent snapshot.
    // If snapshot is temporarily unavailable, it should return ErrSnapshotTemporarilyUnavailable,
    // so raft state machine could know that Storage needs some time to prepare
    // snapshot and call Snapshot later.
    Snapshot() (pb.Snapshot, error)
}这些接口在library中会被用到。熟悉raft协议的人不难理解。上面提到的官方示例https://github.com/coreos/etcd/tree/master/contrib/raftexample中使用了library自带的MemoryStorage，和etcd的wal和snap包做持久化，重启的时候从wal和snap中获取日志恢复MemoryStorage。
要提供这种IO/网络密集型的东西，提高吞吐最好的手段就是batch加批处理了。etcd raft library正是这么做的。
下面看一下为了做这事，etcd提供的核心抽象Ready结构体：
// Ready encapsulates the entries and messages that are ready to read,// be saved to stable storage, committed or sent to other peers.// All fields in Ready are read-only.type Ready struct {    // The current volatile state of a Node.
    // SoftState will be nil if there is no update.
    // It is not required to consume or store SoftState.
    *SoftState    // The current state of a Node to be saved to stable storage BEFORE
    // Messages are sent.
    // HardState will be equal to empty state if there is no update.
    pb.HardState    // ReadStates can be used for node to serve linearizable read requests locally
    // when its applied index is greater than the index in ReadState.
    // Note that the readState will be returned when raft receives msgReadIndex.
    // The returned is only valid for the request that requested to read.
    ReadStates []ReadState    // Entries specifies entries to be saved to stable storage BEFORE
    // Messages are sent.
    Entries []pb.Entry    // Snapshot specifies the snapshot to be saved to stable storage.
    Snapshot pb.Snapshot    // CommittedEntries specifies entries to be committed to a
    // store/state-machine. These have previously been committed to stable
    // store.
    CommittedEntries []pb.Entry    // Messages specifies outbound messages to be sent AFTER Entries are
    // committed to stable storage.
    // If it contains a MsgSnap message, the application MUST report back to raft
    // when the snapshot has been received or has failed by calling ReportSnapshot.
    Messages []pb.Message    // MustSync indicates whether the HardState and Entries must be synchronously
    // written to disk or if an asynchronous write is permissible.
    MustSync bool}可以说，这个Ready结构体封装了一批更新，这些更新包括：

• 　　pb.HardState: 包含当前节点见过的最大的term，以及在这个term给谁投过票，已经当前节点知道的commit index
  
• 　　Messages: 需要广播给所有peers的消息
  
• 　　CommittedEntries:已经commit了，还没有apply到状态机的日志
  
• 　　Snapshot:需要持久化的快照
  

库的使用者从node结构体提供的一个ready channel中不断的pop出一个个的Ready进行处理，库使用者通过如下方法拿到Ready channel：
func (n *node) Ready()