State-of-the-art fault-tolerant and federated data management systems rely on fully-replicated designs in which all participants have equivalent roles. Consequently, these systems have only limited scalability and are ill-suited for high-performance data management. As an alternative, we propose a hierarchical design in which a Byzantine cluster manages data, while an arbitrary number of learners can reliable learn these updates and use the corresponding data. To realize our design, we propose the delayed-replication algorithm, an efficient solution to the Byzantine learner problem that is central to our design. The delayed-replication algorithm is coordination-free, scalable, and has minimal communication cost for all participants involved. In doing so, the delayed-broadcast algorithm opens the door to new high-performance fault-tolerant and federated data management systems. To illustrate this, we show that the delayed-replication algorithm is not only useful to support specialized learners, but can also be used to reduce the overall communication cost of permissioned blockchains and to improve their storage scalability.