OctopusDB: Flexible and Scalable Storage Management for Arbitrary Database Engines

Abstract

We live in a dynamic age with the economy, the technology, and the people around us changing faster than ever before. Consequently, the data management needs in our modern world are much different than those envisioned by the early database inventors in the 70s. Today, enterprises face the challenge of managing ever-growing dataset sizes with dynamically changing query workloads. As a result, modern data managing systems, including relational as well as big data management systems, can no longer afford to be carved-in-stone solutions. Instead, data managing systems must inherently provide flexible data management techniques in order to cope with the constantly changing business needs. The current practice to deal with changing query workloads is to have a different specialized product for each workload type, e.g. row stores for OLTP workload, column stores for OLAP workload, streaming systems for streaming workload, and scan-oriented systems for shared query processing. However, this means that the enterprises have to now glue different data managing products together and copy data from one product to another, in order to support several query workloads. This has the additional penalty of managing a zoo of data managing systems in the first place, which is tedious, expensive, as well as counter-productive for modern enterprises. This thesis presents an alternative approach to supporting several query workloads in a data managing system. We observe that each specialized database product has a different data store, indicating that different query workloads work well with different data layouts. Therefore, a key requirement for supporting several query workloads is to support several data layouts. Therefore, in this thesis, we study ways to inject different data layouts into existing (and familiar) data managing systems. The goal is to develop a flexible storage layer which can support several query workloads in a single data managing system. We present a set of non-invasive techniques, coined Trojan Techniques, to inject different data layouts into a data managing system. The core idea of Trojan Techniques is to drop the assumption of having one fixed data store per data managing system. Trojan Techniques are non-invasive in the sense that they do not make heavy untenable changes to the system. Rather, they affect the data managing system from inside, almost at the core. As a result, Trojan Techniques bring significant improvements in query performance. It is interesting to note that in our approach we follow a design pattern that has been used in other non-invasive research works as well, such as PAX, fractal prefetching B+-trees, and RowCol. We propose four Trojan Techniques. First, Trojan Indexes add an additional index access path in Hadoop MapReduce. Second, Trojan Joins allow for co-partitioned joins in Hadoop MapReduce. Third, Trojan Layouts allow for row, column, or column-grouped layouts in Hadoop MapReduce. Together, these three techniques provide a highly flexible data storage layer for Hadoop MapReduce. Our final proposal, Trojan Columns, introduces columnar functionality in row-oriented relational databases, including closed source commercial databases, thus bridging the gap between row and column oriented databases. Our experimental results show that Trojan Techniques can improve the performance of Hadoop MapReduce by a factor of up to 18, and that of a top-notch commercial database product by a factor of up to 17.