Achieving high availability in distributed databases

An approach is presented for managing distributed database systems in the face of communication failures and network partitions. The approach is based on the idea of dividing the database into fragments and assigning each fragment a controlling entity called an agent. The goals achieved by this approach include high data availability and the ability to operate without promptly and correctly detecting partitions. A correctness criterion for transaction execution, called fragmentwise serializability, is introduced. It is less strict than the conventional serializability, but provides a valuable alternative for some applications