We're going to revisit the when-to-send_complete() logic. We've had one
case where a slave was stuck doing repeated send_completes.
It's also problematic that send_complete() is synchronous and
ipropd-master does not fork() and is not multi-threaded. We've seen a
case where one send_complete caused other slaves to timeout and
reconnect and missed enough transactions that they then had to... get a
full prop, which then led to a cycle of slaves doing nothing but full
The thing we're likely to do to fix the latter is fork() to
send_complete(), or else fork() per-slave (which will complicate the
writing of the stats file).