The real threshold is multi-region writes on a hard latency budget, and even then a single-region primary plus accepting cross-region read latency beats eating a consensus round-trip on every write for a lot of teams. Curious what workload pushed you past single-primary - usually a better story than the impl itself.