Because the initial cache topology is installed in a single phase, it's possible to receive a StateRequestCommand from another node before receiving the JOIN response from the coordinator and installing the initial cache topology. As such, commands received before we have a cache topology should block instead of being ignored.
This bug is especially visible in the Map/Reduce tests (e.g. ReplicatedTwoNodesMapReduceTest) because the M/R task sends the CreateCacheCommand to the other nodes before it's executed on the originator. Since the originator happens to be the coordinator, it's possible for the coordinator to start the rebalance before the first member installed the initial topology.