Maintaining consistency in client-server database systems with client-side caching

Abstract

Caching has been used in client-server database systems to improve the performance of applications. Much of the current work has concentrated on caching techniques at the server side, since the underlying assumption has been that clients are “thin” with application level processing taking place mainly at the server side. There are also a new class of “thick client” applications where clients need to access the database at the server but also perform substantial amount of processing at the client side; here client-side caching is needed to provide good performance for applications. This thesis presents a transactional cache consistency scheme suitable for systems with client-side caching. The scheme is based on the optimistic approach to concurrency control. The scheme provides serializability for committed transactions. This is in contrast to many modern systems that only provide the snapshot isolation property which is weaker than serializability. A novel feature is that the processing load for validating transactions at commit time is shared between clients and the database server, thereby reducing the load at the server. Read-only transactions can be validated at the client-side, without communicating with the server. Another feature is that the scheme permits disconnected operation, allowing clients with cached objects to work offline. The performance of the scheme is evaluated using simulation experiments. The experiments demonstrate that for mostly read only transaction load – for which caching is most effective - the scheme outperforms the existing concurrency control scheme with client-side caching considered to be the best, and matches the performance of the widely used scheme that only provides snapshot isolation. The results also show that the scheme in a disconnected environment provides reasonable performance.