I think that the Terracotta integration implementation should provide a way to automatically split single, giant, TerracottaFile instances into smaller ones prior to add them to the TerracottaDirectory hash map. That's because executing certain indexing operations over very large indexes often leads to very large TerracottaFile instances that, when added to the clustered TerracottaDirectory hash map, generate an in-memory Terracotta transaction that takes too much heap space and kills the client with an OutOfMemoryError. Splitting too large TerracottaFile instances into smaller ones would solve the problem because wouldn't lead to big Terracotta in-memory transactions, acting the hash map put() method as a transaction boundary. What do you think?