Raft Consensus with a Minority of Nodes

The key correctness insight is this: any two majorities of nodes must overlap in at least one node. So between any two consecutive global state changes [...] at least one node participated in both. This single overlapping node carries forward the knowledge of what was previously committed, preventing conflicts and ensuring consistency.

There is another side to this, it must not be possible for two »majorities« to coexist, otherwise they could independently move on in case of a split cluster. This also rules out allowing consensus by majority in addition to majority by a bloc. In the seven node example, there could be a { 1, 2, 3 } and { 4, 5, 6, 7 } split, the first partition being a bloc and the second one being a majority but not containing a bloc.

I really enjoy when it when someone injects a dose of "wacky" into something that is taken more or less for granted (Raft) to challenge the standard way of thinking about it.

This article flipped my understanding of split-brain or network partitions on its head: You don't actually have to have a majority to ensure progress, you just have to design your quorum selection criteria in such a way that no other partition believes they are authoritative, and these finite projection planes are an interesting way of proving that (with caveats).

excellent article.

> The key correctness insight is this: any two majorities of nodes must overlap in at least one node. So between any two consecutive global state changes — whether two commits, two leader elections, or one of each — at least one node participated in both.

intuitively makes sense, but would be nice to see this result explicitly derived or illustrated the same way the fano planes were.