The current queueing mechanism behaves badly when the input rate of messages which should depart on a certain connection is higher than the rate that the connection writer is managing to drain the queue and send them out.
This shows up in some configurations of loadtest.
The right solution is a variable-sized queue which grows up to a maximum size. Message drops after that point should be logged in a $SYS stat.