Descripción
|
|
---|---|
Database replication replicates data to different databases for the purposes of low response time, scalability, and fault tolerance. The challenge of database replication is how to correctly replicate data so that the whole replicated database system behaves as if there is only one copy of database. The standard correctness criterion is one-copy-serializability. In this thesis, I propose a new criterion, one-copy-snapshot-isolation, due to the popularity of snapshot isolation over serializability in major database management systems. Snapshot isolation is weaker than serializability at many aspects, one of which is not garanteeing integrity constraints. I further propose a criterion for both snapshot isolation and integrity constraints. The thesis provides the formulism based on previous work. At the second step, I investigate how to design replication protocols which can be used practically. Some existing protocols put severe limitations on their applications such as read-only transactions must be identified at their start time. Other protocols do not work with databases with integrity constraints, a very important characteristics of databases. I propose a solution which overcomes all the limitations with a centralized middleware scheduler. The third step of my thesis is to investigate replication protocols which work well in both local area networks and wide area networks. Most of existing protocols require serveral message rounds between different databases within a transaction response time and/or they use multicast messages provided by group communication systems. They do not work well in wide area networks due to high latency of message delivery. I extend my practical solution above by using a decentralized architecture. The new solution keeps the number of message within a transaction as constant (e.g., one round trip messages). Fault-tolerance issue is also discussed. To evaluate my solutions, I designed a middleware-based framework which integrates the implementation of my protocols and other typical protocols with a real database system, PostgreSQL. Experiments are performed in different setup, e.g., local area networks and wide area networks. The results are satisfactory. In summary, this thesis is to present you a practical database replication solution which works well in both local area network and wide area networks. It has been implemented in a real system and verified to outperform existing protocols. | |
Internacional
|
Si |
ISBN
|
|
Tipo de Tesis
|
|
Calificación
|
Sobresaliente cum laude |
Fecha
|
05/12/2007 |