Positional Delta Trees to Reconcile Updates With Read-Optimized Data Storage
Source: Centrum Wiskunde & Informatica
The authors investigate techniques that marry the high read-only analytical query performance of compressed, replicated column storage ("Read-optimized" databases) with the ability to handle a high-throughput update workload. Today's large RAM sizes and the growing gap between sequential vs. random IO disk throughput, bring this once elusive goal in reach, as it has become possible to buffer enough updates in memory to allow background migration of these updates to disk, where efficient sequential IO is amortized among many updates. The key goal is that read-only queries always see the latest database state, yet are not (significantly) slowed down by the update processing.