Leader election for paxos-based replicated key value store

后端 未结 2 472
天涯浪人
天涯浪人 2021-02-09 14:58

I am going to implement a key value store with multi Paxos. I would have several nodes, one of which is the primary node. This primary node receive update requests and replicate

2条回答
  •  不思量自难忘°
    2021-02-09 15:30

    Before I get to the actual question, I would suggest that for a paxos-like system, you don't think of it as a master-slave relationship, but rather an equal-peer relationship. Basic Paxos doesn't even have a leader concept. Multi-paxos tacks on a leader as a performance optimization, electing that leader is part of the protocol.

    Multi-Paxos boils down to Paxos underneath: there is a prepare phase and an accept phase. The insight of Multi-Paxos is that once a node wins an accept round, it has simultaneously won leader election and after that the prepare phase isn't necessary from that leader until it detects that another node has taken over leadership.


    And now some practical advise. I have many years of experience working on several paxos, multi-paxos, and other consensus systems.

    I first suggest not implementing either Paxos or Multi-paxos. Optimizing Paxos systems for performance while keeping it correct is very hard—especially if you are having these types of questions. I would instead look into implementing the Raft protocol.

    Taking both protocols as is right off the paper, the Raft protocol can have much better throughput than Multi-Paxos. The Raft authors (and others) suggest that Raft is easier to understand, and implement.

    You may also look into using one of the open-source Raft systems. I don't have experience with any of them to tell you how easy it is to maintain. I have heard, though, of pain in maintaining Zookeeper instances. (I have also heard complaints about Zookeeper's correctness proof.)

    Next, it has been proven that every consensus protocol can loop forever. Build into your system a time-out mechanism, and randomized backoffs where appropriate. This is how practical engineers get around theoretical impossibilities.

    Finally, examine your throughput needs. If your throughput is high enough, you will need to figure out how to partition across several consensus-clusters. And that's a whole 'nother ball of wax.

提交回复
热议问题