Running #37 the following scenario may result in lost writes: * All 3 replicas down * replica 0 goes up, starts state transfer to 1 and 2. state transfer succeeds * replica 0 and 1 goes down again * clients write to replica 2 * Split brain, manual scale down required * replica 0 goes up as "new cluster", begins state transfer to 1 and again * writes written to replica 1 or 2 are lost