Preloading the cache is causing conflicts with region activation, since the preload happens before the region is activated (and thus pollutes the in memory cache, causing the activation to fail.) In any case, preloading doesn't make sense for the SFSB cache; any state on disk would be passivated sessions that had been in use on another node; they are only useful in memory if the client fails over. No point aggressively bringing them into memory.

Fix is to remove the <preload>/</preload> element from the ejb3-clustered-sfsbcache-service.xml file.

An open question is why there are nodes to be preloaded. StatefulTreeCache.stop() should be causing them to be deleted, and starting the cache shouldn't be bringing over any state transfer. Perhaps some file lock is preventing the node deletion?