recall

← recall

leader election pattern

Picking one node out of a cluster to be the coordinator/primary. Done via consensus (Raft, Paxos) or via lease-based primitives (ZooKeeper ephemeral nodes). Failures of the leader trigger a re-election; the cluster is briefly leaderless.

Picking one node out of a cluster to be the coordinator/primary. Done via consensus (Raft, Paxos) or via lease-based primitives (ZooKeeper ephemeral nodes). Failures of the leader trigger a re-election; the cluster is briefly leaderless.

symptoms

  • need a single coordinator that can survive node failures
  • split-brain during partitions
  • leadership churn

causes

  • need consistent decisions
  • leader is bottleneck or SPOF

fixes

  • Raft / Paxos for strong election
  • ZooKeeper / etcd ephemeral leases
  • fencing tokens to detect stale leaders

you might say

  • leader election
  • who owns the lock
  • fail over to the next leader

related

topics: distributed-systems, consensus

references: