I scratched my head about this too. Each node has the same code. A set of ears each with their own spring context defining the entity manager and scanning the entities. I had the same thought as you, not every context scanned the complete set entities. It would only take a one table difference and the entire TreeMap would be off. I didn't dig deeply into the code but I'm sure that's what was happening. My solution was a bit crude but I could not think of another way other than defining the index in a mapping file. I would imagine others with a similar use case. Ears containing common shared entities as well as custom. The common shared entities appear unique across the cluster. I can't find my notes, but I think in our case we have ~500 HQL statements but we're seeing ~25k unique SQL statements on the backend.