Deferred Update Protocols for Multi-Site Transactions

Several commercial distributed database systems provide an optional protocol that defers updates of replicas in order to attain higher transaction throughput. Each replicated data item is assigned a primary copy site, and has a set of sites with secondary copies. Typically, in a deferred update protocol, a transaction directly updates only the primary copy of each data item it modiies. After the transaction commits, the updates to primary copies by that transaction are sent transactionally to other sites containing secondary copies of these data items. Thus, the protocol allows the values of the primary and secondary copies of a data item to diier until the secondary copies are updated. In an earlier paper, we investigated the deferred update approach for transactions that operate at only one site and then commit (with updates to secondary copies propagated after commitment). There, we showed that the serializability of transactions can be guaranteed if and only if the replication topology of the system is either a tree or a forest. In this paper, we discuss global transactions that operate at multiple sites in a system utilizing the deferred update approach. In such a system, a global transaction that updates replicated data items can use the underlying deferred update mechanism to ensure replica consistency. Such a transaction may commit after updating primary copies. The deferred update mechanism forwards the updates to sites containing secondary copies as usual. However, it is not clear if global serializability is ensured. We present two protocols to execute global transactions in a replicated database systems with a deferred update mechanism. The rst protocol is applicable to systems whose replication topology is a tree. The second protocol is more general, and is applicable to systems whose replication topology is a forest. The protocols preserve the execution autonomy of the local sites and the underlying deferred update mechanism. They impose minimal overhead on transactions that operate at only one site. We show that both protocols ensure global serializability. We also propose a method for tolerating site failures.