Leader Election 选举

• Select one server to act as leader

• Detect crashes, choose new leader

Introduction

Raft uses a heartbeat mechanism to trigger leader election. When server start up, they begin as followers.

Raft 使用心跳机制来触发 leader 选举. 当服务启动时, 其状态为 follower.

To begin an election, a follower increment its current term and transitions to candidate state. It then votes for itself and issues RequestVote RPC in parallel to each of the other servers in the cluster.

• A candidate wins an election if it receivers votes from a majority of the servers if the full cluster for the same term

• Each server will vote for at most one candidate in a given term, on a first-come-first-served basis

• Raft uses randomized election timeouts to ensure that split votes are rare and that they are resolved quickly

在 Raft 中有二个超时设置控制着 Leader 选举. 第一个是 election timeout, 决定了一个 follower 等待成为 candidate 的时间. election timeout 在某个区间内取随机值 (通常是 150ms, 300ms).

Follower 在等待完 election timeout 时间后, follower 会开始一轮新的选举: follower 将当前 term 加 1, 并且切换到 candidate 状态. 它选举自己并发送 RequestVote RPC 请求给集群的其他节点.

Election Restriction

In any leader-based consensus algorithm, the leader must eventually store all of committed log entries.

Raft uses the voting process to prevent a candidate from winning an election unless its log contains all committed entries.