HPL-2010-70LazyBase: Trading freshness for performance at scale
Morrey III, Charles B.; Keeton, Kimberly; Soules, Craig A. N.; Veitch, Alistair
Abstract: Increasingly, enterprise applications wish to perform data mining analyses on rapidly changing data sets. Many of these applications can perform correctly on varying degrees of stale data. However, traditional systems poorly support simultaneous updates and read queries at the performance desired by data mining systems. Furthermore, they provide limited choices along the freshness axis. To address these challenges, we have implemented LazyBase, a databas e-like system that simultaneously supports both high update throughput and read query throughput, by reducing query result freshness to increase read query throughput and decrease read query latency. Experiments with LazyBase show that if an application can tolerate 30 minutes of stale data it decrease query latency by 2.2-3.8x and increase query throughput by 2.2-4.0x. Furthermore, queries to the freshest data do not degrade ingest performance. LazyBase achieves a 20% improvement in update throughput over and optimistic estimate of a scale-out RDBMS approach despite using 60% fewer resources. Its query performance for stale data almost always exceeds that of a traditional RDBMS.
External Posting Date: June 21, 2010 [Abstract Only]. Approved for External Publication - External Copyright Consideration
Internal Posting Date: June 21, 2010 [Fulltext]