> One of the interesting things about Raft that doesn't get focused on much is that strictly speaking Raft does not require that all nodes be able to see each o

> One of the interesting things about Raft that doesn't get focused on much is that strictly speaking Raft does not require that all nodes be able to see each other, just that some nodes can see all of the other nodes. The same is true of Paxos. If you have a common view of participant health, then consensus is much easier than in the general async model, so Raft and Paxos (and Viewstamped Replication, etc) are like this for a reason. I don't know about SnowFlake, but there is a very interesting tradeoff here between graph connectedness and resilience to the failure of masters (e.g. if the topology is really a star, then you have no real redundancy and you may as well have a single master rather than using consensus).