For cross-replication topology setups load_balancing=in_order works best
as nodes handle equal amount of load and usually they hit only 1/n of
data (n = number of replicas), which makes page cache usage more
efficient.
The problem is when one node of the shard goes down. If one replica goes
down, the next one in config will handle twice the usual load while
remaining ones will handle usual traffic.
Closes#4820.
