1 Reply Latest reply on Dec 6, 2011 9:56 AM by galder.zamarreno

State propagation to a rejoined node in repl mode

sebastiandr Nov 22, 2011 11:49 AM

We have an scenario where we have 3 nodes running Infinispan 5.0.1.FINAL in an embedded way, in replication mode and with no cacheloader. We have a load balancer distributing put operations among the three nodes.

We have detected some pauses in the nodes (due to GC activity, although this is not the focus of the question) long enough for them to be taken out of the cluster. When this happens to a given node, the node becomes desynchronized, because the balancer keeps pushing items in the other two. After the pause, the node is merged again in the cluster, but the items added to the other nodes are not propagated to this newly merged one. The question is: Is this the correct behavior? Do we have a way to make this node to be synchronized with the others?

Take into account that I'm not talking about when a node is shut down and started up again, situation in which the state is correctly propagated from the other nodes, but when the node is taken out momentarily from the cluster and later rejoined (having changed state in the other nodes).

Thanks.

1. Re: State propagation to a rejoined node in repl mode

galder.zamarreno Dec 6, 2011 9:56 AM (in response to sebastiandr)

You can listen for such merge events via a Infinispan listener (@Listener) that has a method annotation with @Merged. This will give you a MergeEvent that would tell you the groups that are being merged, but unfortunately, we do not yet provide a way to handle merging of the state (https://issues.jboss.org/browse/ISPN-263). In spite of not handling state merges, you could listen for merge events and if those happen, you could restart one or several nodes based on the info you have from the load balancer?
Actions