Agreement is not free
Consensus protocols such as Paxos and Raft let a set of nodes agree on a single ordered log of operations even when some fail. This guarantee comes at a real latency cost that shapes system design.
The quorum round trip
To commit an entry, a leader must replicate it to a majority quorum and wait for their acknowledgements. With five nodes a leader needs two followers to agree, so the commit latency is at least one network round trip to the slower of those followers.
- Latency is bound by the median replica, since a majority must respond.
- Geographically spread replicas add wide area round trips, often tens of milliseconds each.
- A leader change adds an election delay before progress resumes.
Living with the cost
Systems reduce the pain by batching many operations into one consensus round, keeping replicas close for latency sensitive data, and reading from leases to avoid a quorum on reads. The fundamental floor remains, since a durable agreed write needs a majority to acknowledge.
Key idea
Consensus requires a majority quorum to acknowledge each write, so every committed operation pays at least one round trip to the median replica, a cost batching and locality can soften but not remove.