Consensus gives you one leader at a time

Consensus protocols such as Raft and Paxos let a group of servers agree on one leader and one order of writes, and keep working while a minority of them fail. Elections and commits both need a majority, and that single rule is what stops two leaders from both committing writes.

In Raft a candidate needs votes from a majority of the full cluster and each server votes once per term, so at most one leader wins a term. An entry commits once a majority stores it. Terms only rise and act as a logical clock: servers reject requests with a stale term, and a leader that sees a newer one steps down. That term is what a fencing token needs.

The guarantee is per term. A superseded leader may not yet know it, so the paper has it confirm with a majority before answering reads. The majority keeps split brain out of writes; reads need that extra check.

The costs are latency and arithmetic. Five servers tolerate two failures. Once voters span regions, each write waits on a cross-region round trip, and a majority must outlive the outage, which two regions cannot arrange; quorums reuse the same overlap, loosely.