No single best protocol
The consensus algorithms differ less in raw capability than in engineering tradeoffs. Choosing one means weighing fault model, latency, throughput, and how understandable the protocol is to the team operating it.
The main axes
- Fault model: crash tolerant protocols like Paxos and Raft assume nodes only stop. Byzantine protocols like PBFT and Tendermint also tolerate lies but need 3f plus 1 nodes and more messages.
- Understandability: Raft was explicitly designed to be teachable, which lowers operational risk versus the famously subtle Paxos.
- Latency and load: a single leader simplifies reasoning but bottlenecks throughput. Leaderless designs like EPaxos spread load at the cost of complexity.
- Quorum tuning: Flexible Paxos lets you trade fast commits against costlier recovery.
Practical guidance
Most internal datacenter systems pick Raft or Multi Paxos because crash faults dominate and operability matters most. Open or adversarial settings, like public blockchains, justify the heavier Byzantine protocols. The right choice follows the threat model and the team rather than any abstract notion of the best algorithm.
Key idea
Consensus protocol choice is a tradeoff among fault model, latency, throughput, and understandability, so match the protocol to your threat model and operating team.