It has been suggested that QDataStream has considerable overhead and that build graph loading performance could benefit from using a custom backend instead. There aren't too many direct accesses to QDataStream in persistence.*, to it should be feasible to do a relatively quick evaluation.