Shinobi : insert-aware partitioning and indexing techniques for skewed database workloads

Abstract

Many data-intensive websites are characterized by a dataset that grows much faster than the rate that users access the data and possibly high insertion rates. In such systems, the growing size of the dataset leads to a larger overhead for maintaining and accessing indexes even while the query workload becomes increasingly skewed. Additionally, the database index update costs can be a non-trivial proportion of the overall system cost. Shinobi introduces a cost model that takes index update costs account, and proposes database design algorithms that optimally partition tables and drop indexes from partitions that are not queried often, and that maintain these partitions as workloads change. We show a 60x performance improvement over traditionally indexed tables using a real-world query workload derived from a traffic monitoring application and over 8x improvement for a Wikipedia workload.