Netsplit is IRC-specific jargon. Wikipedia:Netsplit shows pretty well what happens. When a TCP connection on an interserver link breaks down, servers may try to recover quick enough not to lose the current state. If that fails (or your dialect of ircd doesn't know how to do that), the complete directory of present users and channels needs to be retransmitted: The so-called netburst. This procedure is highly visible and annoying to each user on the network, all people from the other side seem to log out and back in.

PSYC doesn't share a distributed directory, so a breaking link never causes as much harm. When a TCP circuit breaks, the queues kick in and several attempts to reconnect are made. Only when all of these fail, delivery failures are returned, appropriately tagged, so your client can sort out what went wrong. Additionally, echoes give you an immediate feel for when a network link is in trouble.

If the incoming information was organized in a multicast context, which in most cases it should be, then the recovery process could try to get that context's data from a different node in the tree. We're not doing this as yet, since we never ran into the problem, even in large scale applications.

See Special:Whatlinkshere/Netsplit for more details on this.