I should have also said: There is a new HA module in Qpid that will eventually replace the older cluster module. You should take a look at that as well, if you're starting out on a new project.
http://qpid.apache.org/books/trunk/AMQP-Messaging-Broker-CPP-Book/html/chap-Messaging_User_Guide-Active_Passive_Cluster.html On Fri, 2012-07-27 at 01:02 +1000, Gavin Alexander wrote: > Hi All, > > I'm looking to set up a durable HA broker and had a few questions. > First off, sorry if these questions are more appropriate on a Linux-HA > forum... just trying to see if other people have come across the same > issues. > > After setting up a two node cluster and being quite happy with the > ease of config - I realised that I would have to solve the split-brain > problem. > So, using 3 virtual machines, I followed the guide @ > https://cwiki.apache.org/confluence/display/qpid/Configuring+qpidd+with+Cluster+Manager > which works reasonable well - (I've had a few issues with rgmanager > hanging occasionally during failover... but I digress...) > > The environment that I'm using at the moment is on a bunch of Linux > KVM's, but it will eventually be moved to physical infrastructure. > I have some reasonable performance requirements and the hardware I'm > using is quite expensive - so adding another node for HA just to > maintain a quorum seems "wasteful". > > So, I'm investigating if a 2 node setup is possible. > > Is it possible to do HA with two nodes (without a "real" qdisk)? > I've tried adding a 3rd node (a cheap virtual machine) for quorum only > - I.e. it doesn't run qpidd and is only used for quorum. It seems to > work - although I haven't thoroughly tested all failure scenarios yet. > Other idea's I've had are > - Using Qdisk heuristics to generate more votes - everything I read > seems to suggest not to do this... > - Integrating with a load balancer (there's an existing LVS cluster > on the same network - that I can piggy-back on) to protect against > split-brain "somehow" :) > - Leave it as a two node cluster and rely on corosync totem protocol, > or interface bonding for fault tolerance. Then, if a split brain does > occur, it's pretty certain (about 99.9% :) that clients wouldn't be > able to access the cluster anyway. > > > A couple of questions... > Do cluster messages require acknowledgement from cluster nodes before > sending ack's back to the client (if the clients require > acknowledgement)? And hence, does adding more nodes degrade > performance? > Has anyone ever tried running qpid over something like drbd? > Is it possible to specify a bind address for qpidd? > > Look forward to any replies/guidance/commentary > > Thanks, > > /gav > > --------------------------------------------------------------------- > To unsubscribe, e-mail: [email protected] > For additional commands, e-mail: [email protected] > --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
