Two machines both say they are the leader
Three replicas serve a hypothetical ledger. A network partition isolates A from B and C, but some clients can still reach A. B and C elect B. Both A and B believe they lead.
The useful question is not whether two processes can believe the same thing. It is whether both can commit conflicting histories.
Votes must survive the voter
For three members, a majority is two. Any two majorities overlap. That intersection matters only if a voter follows durable rules: one vote per term, and no acceptance behavior that forgets committed history after restart.
If a replica votes for A, crashes, forgets its vote, and votes for B in the same term, what happened to the majority argument?
Raft separates election, log replication, and safety rules. Terms order leadership attempts. Candidates must have sufficiently up-to-date logs; followers check preceding log identity before appending a suffix. Conflicting uncommitted suffixes may be removed when a valid leader repairs them. Committed history must not be discarded.
Two copies do not always mean committed
A Raft leader cannot commit an old-term entry solely by counting how many replicas contain it. It advances the commit index by this counting rule for an entry from its current term; committing that entry also commits the preceding log prefix. A newly elected leader often appends a current-term no-op to establish that boundary.
This restriction and the election rule work together. Removing either turns āmajority replicatedā into an incomplete safety argument. Use an established consensus implementation; a paragraph of pseudocode is not a replacement for its protocol and storage tests. Raft paper, sections 5.4 and 8.
pseudocodeleader receives command append current-term entry and replicate with log-prefix checks wait for required durable acknowledgements under protocol rules advance commit index for a qualifying current-term entry apply committed prefix in order reply according to API contract
An isolated minority leader can receive requests and perhaps append locally, but cannot commit by itself. Its clients may receive timeouts and need idempotency when retrying elsewhere.
A read is also a claim about authority
A former leader can return stale data very quickly. Linearizable leader reads must establish current authority, such as a quorum-confirmed read index followed by waiting for local application, or a rigorously bounded lease protocol. āIt was leader a second agoā is insufficient.
Consensus provides a committed order; linearizable client operations also need correct acknowledgement and read paths, state-machine application, and duplicate handling. Serializability across multi-row operations is a separate database question.
Paxos reaches agreement on chosen values through proposal/acceptance rounds; Multi-Paxos uses a stable leader to amortize repeated decisions. Raft expresses a similar replicated-log goal with explicit leader/log rules. Neither makes the minority available for conflicting authoritative writes during a partition. Safety can hold during arbitrarily long delays; progress needs adequate communication and timing conditions.
Changing the voters changes the proof
Replacing an entire old membership with a new disjoint one permits two separate majorities unless the transition protocol prevents it. Joint consensus uses overlapping configuration requirements during transition; other safe reconfiguration protocols have their own constraints. Do not edit the member list as if it were a load-balancer pool.
Snapshots compact the log but retain the state and log identity needed for subsequent replication. Add new replicas carefully, catch them up, and account for snapshot traffic. Three replicas tolerate one unavailable voter; five tolerate two. Four still need three votes and tolerate only one, though placement and other requirements may justify that configuration.
Watch election churn, fsync latency, quorum response time, apply lag, and leader changes. Timeout tuning cannot repair a saturated disk or a missing quorum. Pre-vote mechanisms can reduce disruption by a returning isolated node, but are protocol-specific.
Mental model: consensus makes authority survive conflicting beliefs. The proof depends on durable state, intersecting quorums, and log/election rules together.
Transfer: the fast minority
Two nodes in a five-node group are isolated from the other three. Their local reads and disks are healthy. May they accept authoritative writes to improve availability?
Source trail: 09-distributed-consensus-raft-and-paxos, 06-replication-quorums-and-read-your-writes. See the source map for consolidation and reference-retained material.