Application-level caching with transactional consistency

Abstract

Distributed in-memory application data caches like memcached are a popular solution for scaling database-driven web sites. These systems increase performance significantly by reducing load on both the database and application servers. Unfortunately, such caches present two challenges for application developers. First, they cannot ensure that the application sees a consistent view of the data within a transaction, violating the isolation properties of the underlying database. Second, they leave the application responsible for locating data in the cache and keeping it up to date, a frequent source of application complexity and programming errors. This thesis addresses both of these problems in a new cache called TxCache. TxCache is a transactional cache: it ensures that any data seen within a transaction, whether from the cache or the database, reflects a slightly stale but consistent snapshot of the database. TxCache also offers a simple programming model. Application developers simply designate certain functions as cacheable, and the system automatically caches their results and invalidates the cached data as the underlying database changes. Our experiments found that TxCache can substantially increase the performance of a web application: on the RUBiS benchmark, it increases throughput by up to 5.2x relative to a system without caching. More importantly, on this application, TxCache achieves performance comparable (within 5%) to that of a non-transactional cache, showing that consistency does not have to come at the price of performance.