An OSS scalable distributed message queue with automatic failover. It kicks ass
code.google.com
code.google.com
it seems to me that imatix guys have been there, done that and saw that mq-as-library was the way to go for a more efficient and reliable distributed transaction processing system. http://www.zeromq.org/whitepapers:brokerless is an interesting read in this regard.
We use an AMQP stack and the complexity seems somewhat unjustified. On the other hand, it's good to have a standard, even if it's not perfect. Django+celery is pretty kick-ass, and I would have to give it up if I moved away from AMQP.
Besides, there are cases where you would want to continue operating even if your clients went down.
"The pymq cluster is not designed to survive a master crash. If the master is to crash, all bets are off. However, clients that are already connected to the individual nodes can continue to operate normally. It's just that automatic failover and queue discovery won't work (which is basically what the master does)."
This is not fail-safe, unlike most enterprise messaging systems. Though you could stick the master on a VMware cluster with FT enabled and survive hardware failures, there is nothing preventing data loss from software or networking failures.
If the software is bad, you can't help it.
As far as network failures are concerned, I think those are handled. If your network fails, you basically won't be able to send a message or won't get confirmation of it being sent. Either ways, if the queuing system works in at least once mode, you are guaranteed message delivery. If there is an error, the library reports an error and the application must re-try.