Important note about cluster configuration
------------------------------------------

The best is to separate actual cluster nodes from the ones
that run the pgreplicate process, i.e. have dedicated machines
for the replication servers.

The reason is twofold.

First, the replication process makes notes about the replicated nodes'
IP addresses and compares this list with the actual incoming replication
packets. If a node is on the same machine as the replication then
it's likely that the packets comes from "localhost", i.e. it comes from
IP address 127.0.0.1 instead of the IP address of the machine's primary
network address. In this case, you may notice that replication and
recovery may stall. Take this into your consideration when designing
your network of nodes.

The other reason is that the replication process is multithreaded,
and (at least the current Linux pthreads implementation) keeps
stackframe and TLS allocated for the highest number of threads that was
created during the process' lifetime. With very high loads, it may be
considerable amount of actually used memory, not only sparsely allocated.
