Hashicorp/raft introduction and source code analysis (III): introduction to cluster node recovery

Cluster node recovery

In all cases , After re selecting the master successfully , backward follower Need to catch up leader The state of ：

• follower It has landed log , Need proofreading . And leader inconsistent , Direct discarding
• follower From scratch , Or restore from the most recent local mirror , And catch up with the application log The state machine

We analyze several abnormal situations , How does the cluster automatically return to normal ：

• 1 individual follower It's restarted , The follower How to finally make your state machine catch up leader
• leader It's restarted , And become follower, How to finally make your state machine catch up leader
• leader It's restarted , And reselect as leader , How to restore your state machine correctly

Scenario analysis 1 - follower restart

follower Restart actually has 2 In this case ：

• No mirror
• Restore from mirror

In either case , Can get their own applyIndex applyTerm

then ,follower By appending the log protocol , hold applyIndex applyTerm, issue leader：

leader Will issue their own applyIndex and commitIndex , And according to follower Of applyIndex applyTerm , Issue follow-up log（ Or mirror image + follow-up log）

• follower according to applyIndex Field , Know that you are behind leader How many? log
• follower according to commi