recall

← recall

consensus concept

Multiple nodes agreeing on a single value (or sequence of values) despite some being slow, crashed, or temporarily partitioned.

Multiple nodes agreeing on a single value (or sequence of values) despite some being slow, crashed, or temporarily partitioned. The foundation of distributed coordination — leader election, distributed locks, configuration management, distributed transactions all reduce to consensus. FLP impossibility shows you can't have it deterministically in a fully asynchronous network with even one failure; practical systems work around this with timeouts and randomization. Algorithms: Paxos, Raft, Viewstamped Replication, Zab. Cost: every committed decision needs a round trip to a majority of nodes.

Raft and Paxos are the famous algorithms. Used for leader election, replicated state machines, distributed locks.

see also

aliases: agreement, raft, paxos

topics: distributed-systems

references: