Raft KV Store

Failure is the normal case

A distributed system is not a faster single machine. Messages arrive late, nodes restart halfway through a write, and a leader can lose contact with the majority without knowing it. The design begins from those conditions rather than treating them as exceptional branches.

The core safety rule is simple to state:

$$committed\_entry \implies majority\_replication$$

The implementation is organised so that this invariant remains visible in the code. Protocol transitions are explicit, persistent state is small, and each mutation can be replayed in the simulator.

Testing the history, not the endpoint

An apparently correct final value says very little. The Jepsen workload records every invocation and response, then checks whether the whole history admits a legal linearizable ordering. Network partitions, process pauses, restarts, and clock skew are applied while clients continue to read and write.

The result is not merely a passing test suite. It is a body of evidence explaining which failures the system tolerates and where its guarantees end.