In Search of an Understandable Consensus Algorithm
The consensus algorithm that replaced Paxos across the modern cloud infrastructure.
AUTHORS: Diego Ongaro, John Ousterhout
CORE ARCHITECTURAL BREAKTHROUGH
"Deconstruct consensus into three independent subproblems: Leader Election, Log Replication, and Safety, using a strong single-leader model."
WHY MODERN SYSTEMS STILL DEPEND ON IT
Powering Kubernetes (via etcd), CockroachDB, TiKV, HashiCorp Consul, and Kafka Kraft. Made distributed consensus implementable for mortals.
KEY PROBLEMS SOLVED
01.Multi-Paxos was notoriously incomprehensible and lacked concrete implementation details for practical systems.
02.Prior algorithms allowed multiple nodes to propose competing log entries simultaneously, creating edge-case complexity.
03.Cluster membership changes (adding/removing servers) routinely led to split-brain bugs in production.
DIRECT MODERN SUCCESSORS
etcd (Kubernetes control plane)Uses pure Raft in Go to maintain entire cluster state, pod scheduling, and secret storage.
CockroachDB / TiDBUses Multi-Raft: dividing distributed SQL table ranges into thousands of independent Raft consensus groups.
Apache Kafka (KRaft)Replaced ZooKeeper with an internal Raft quorum controller for metadata replication.