> Raft on the other hand... I've implemented versions of both paxos and raft, and by golly was the latter more difficult to get right.

Could you clarify what you mean? "Paxos" as defined in the literature (and as presented in this article) is a single-shot consensus algorithm. Raft is a leader-based replicated state machine. They are not comparable at all and it's not surprising if you find Raft more difficult.

However, in practice, when people say "Paxos" they usually mean some variant of MultiPaxos and "Raft vs Paxos" is typically interpreted as "Raft vs MultiPaxos". It's also well-established that MultiPaxos and Raft have roughly the same complexity.

(In the context of embedding paxos in some sort of proof-of-concept replication system. Something like Multi-paxos.)

You're right, I guess we generally talk about state machine replication instead of one-shot systems.