Multiple JBoss clusters and mod_jk

[meghachhabra]

HI

We are using JBoss 4.0.3SP1 and clustering 9 JBoss instances into 3 clusters of 3 instances each.

For each cluster we have a separate apache wihch talks to the 3 instances in the cluster.

In front of these 3 apaches is a H/W loadbalanacer.

Sticky sessions are turned on in apache and the H/W load balancer.

We are using the -b switch to run multiple jboss on a single machine and each machine has 3 instances 1 from each cluster.

Things are working fine in our tests.

However the problem we are running into is that when a particular cluster goes down, the apache for that cluster is still up and due to sticky session the request from a user who was on that cluster continues to go there and we get a 503 despite the other 2 clusters being up.

Now we know that state cannot be replicated across clusters so we are not looking for session state to be replicated in this scenario, but it would nice if apache mod_jk could figure out that the cluster is down and pass that information  on to the H/W loadbalance which in turn would hit another apache talking to a cluster which is still up.

Does anybody have a simliar setup and has done something to prevent this issue?

(We could run a 9 instance cluster on our network, but don't know about if there are any performance implications with that many instances in a cluster. Which of the 2 cluster configurations will give us the best performance?)

Does anybody see any red flags with our setup?

Any help will be appreciated.

Thanks

[btlfedgolf]

Is the application crasshing but jboss staying up?  I have found that the sessinos will only move over if the jboss instance is shut down normally, or the java process crashes suddenly.  If the app simply fails, or gets undeployed it most likely wont fail over.

[brian.stansberry]

If your architecture requires that each apache only handles a single cluster, then whether there is any solution depends on your H/W loadbalancer; e.g. if it could detect the 503 and trigger a failover to another apache, that could work.

If all apache servers can handle all clusters, then the issue can be handled at the apache layer by using the mod_jk's "domain" feature:

http://community.jboss.org/docs/DOC-13247