The balancer does an OID check to determine if all shards are unique processes, but this check happens before it starts and requires all shards be reachable at the same time. After that time, the balancer will act as usual. A cluster should to function well even when partially unavailable when mongos starts, so we should at least remember which hosts we've contacted already successfully and not retry those hosts, or better, allow migrations between hosts that we know are fine.