Analysing sequencing data in Hadoop: The road to interactivity via SQL

Abstract

Analysis of high volumes of data has always been performed with distributed computing on computer clusters. But due to rapidly increasing data amounts in, for example, DNA sequencing, new approaches to data analysis are needed. Warehouse-scale computing environments with up to tens of thousands of networked nodes may be necessary to solve future Big Data problems related to sequencing data analysis. And to utilize such systems effectively, specialized software is needed. Hadoop is a collection of software built specifically for Big Data processing, with a core consisting of the Hadoop MapReduce scalable distributed computing platform and the Hadoop Distributed File System, HDFS. This work explains the principles underlying Hadoop MapReduce and HDFS as well as certain prominent higher-level interfaces to them: Pig, Hive, and HBase. An overview of the current state of Hadoop usage in bioinformatics is then provided alongside brief introductions to the Hadoop-BAM and SeqPig projects of the author and his colleagues. Data analysis tasks are often performed interactively, exploring the data sets at hand in order to familiarize oneself with them in preparation for well targeted long-running computations. Hadoop MapReduce is optimized for throughput instead of latency, making it a poor fit for interactive use. This Thesis presents two high-level alternatives designed especially with interactive data analysis in mind: Shark and Impala, both of which are Hive-compatible SQL-based systems. Aside from the computational framework used, the format in which the data sets are stored can greatly affect analytical performance. Thus new file formats are being developed to better cope with the needs of modern and future Big Data sets. This work analyses the current state of the art storage formats used in the worlds of bioinformatics and Hadoop. Finally, this Thesis presents the results of experiments performed by the author with the goal of understanding how well the landscape of available frameworks and storage formats can tackle interactive sequencing data analysis tasks.