I was actually asking what it actually does, not what it does conceptually. My main issue with these advanced "easy" solutions is that while they're easy to setup, often that comes at the cost of missing understanding of behaviour under load / failure scenarios.

I'm not saying PostgreSQL couldn't improve (it definitely can), but when it comes to data, I advocate simplicity over ease.

Very quick reading gives me the impression that MemSQL's replication is actually per-partition logical replication (might be physical, if partitions are physically separate databases, but they don't seem to be) and uses an algorithm (didn't find which) to decide which node to failover to when something occurs. It also seems to differentiate between metadata nodes and storage/worker nodes.