No single leader
In leaderless replication, used by systems like Dynamo style stores, there is no elected leader. A client sends a write to several replicas directly and considers it successful once enough of them acknowledge. Reads also query several replicas.
The quorum condition
Let N be the number of replicas, W the writes that must acknowledge, and R the reads that must respond. If W plus R is greater than N, every read overlaps at least one replica that saw the latest write, so reads can find an up to date value.
- Choosing W and R high favours consistency but reduces availability.
- Choosing them low favours availability and speed but risks stale reads.
Healing stale replicas
Because some replicas may miss a write, the system repairs them. Read repair updates a stale replica when a read detects it lags. Anti entropy background processes compare replicas and copy missing data. Conflicts from concurrent writes are resolved with vector clocks or last write wins.
Key idea
Leaderless replication lets any replica accept writes and uses the rule W plus R greater than N for overlap, trading tunable consistency against availability and repairing stale replicas in the background.