|
Abstract : |
We present a new architecture and algorithm for distributed replicated database systems. The replication algorithm operates in the presence of message omission faults, processor crashes and recoveries, and network partitions and remerges. The architecture exploits a group communication service to minimize communication costs and to eliminate forced disk writes in the critical path, while preserving complete and consistent operation. End-to-end agreement is required only after a change in the membership of the connected servers, rather than on a per action basis. The updates are globally ordered and, if the system has partitioned, they are applied to the database when they become known to the primary component of the partitioned system. An application may, however, read data and initiate updates at any time, even in a component that is not the primary component. This approach renders replication more efficient and more scalable and, therefore, applicable to many more systems. 1, |
