Positional Delta Trees to reconcile updates with read-optimized data storage

Positional Delta Trees to reconcile updates with read-optimized data storage
Sa ́ndor He ́man, Niels Nes, Marcin Zukowski, Peter Boncz
Centrum voor Wiskunde en Informatica Kruislaan 413, 1098 SJ Amsterdam, The Netherlands

Abstract— We 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. Our key goal is that read-only queries always see the latest database state, yet are not (significantly) slowed down by the update processing. To this end, we propose the Positional Delta Tree (PDT), that is designed to minimize the overhead of on-the-fly merging of differential updates into (index) scans on stale disk-based data. We describe the PDT data structure and its basic operations (lookup, insert, delete, modify) and provide an in-detail study of their performance. Further, we propose a storage architecture called Replicated Mirrors, that replicates tables in multiple orders, storing each table copy mirrored in both column- and row-wise data formats, and uses PDTs to handle updates. Experiments in the MonetDB/X100 system show that this integrated architecture is able to achieve our main goals.

Download article (.PDF)PositionalDeltaTrees

You may also like...