There is a method in DocumentTranslator that can optimize how the child references are persisted in the underlying documents. This method is not currently being used, but should be called. Perhaps it's possible to distribute the execution of this method to Infinispan to operate on each document in each process in its cluster (rather than pull documents locally and do everything within the local process). Some coordination will have to be done to ensure that multiple ModeShape processes are not simultaneously performing the optimization.