| 摘要 |
Hybrid main-memory database systems-capable of storing data both in a horizontal (tuple store) and in a vertical (column store) manner, are gaining traction due to their flexibility and high efficiency. The problem of how to store data in such systems, i.e., of picking the right physical layout for the data and workload at hand, is central to those systems and is usually solved offline. While determining the optimal layout in advance for a given database workload is possible, it typically leads to sub-optimal layouts for a large fraction of enterprise application databases where the workload is constantly changing or exhibits some periodic behavior. In this paper, we present a system that can dynamically adapt to the workload changes and adopts the best possible physical layout on the fly - while allowing simultaneous updates to the table. We present an algorithm that is able to continuously and incrementally compute the optimal physical layout based on workload changes and to determine whether or not switching to this new layout would be beneficial. In addition, we present an efficient solution to the process of rewriting the tables taking into account both read and write-optimized partitions. We experimentally show that our research prototype HYRISE can perform online reorganization of hybrid main memory databases with a negligible overhead, leading up to three orders of magnitude performance gains when determining the optimal layout of dynamic workloads and providing guarantees on the worst case performance of our system. |
