Genomic-based personalized medicine encompasses storing, analysing and interpreting genomic variants as its central issues. At a time when thousands of patientss sequenced exomes and genomes are becoming available, there is a growing need for efficient database storage and querying. The answer could be the application of modern distributed storage systems and query engines. However, the application of large genomic variant databases to this problem has not been sufficiently far explored so far in the literature. To investigate the effectiveness of modern columnar storage [column-oriented Database Management System (DBMS)] and query engines, we have developed a prototypic genomic variant data warehouse, populated with large generated content of genomic variants and phenotypic data. Next, we have benchmarked performance of a number of combinations of distributed storages and query engines on a set of SQL queries that address biological questions essential for both research and medical applications. In addition, a non-distributed, analytical database (MonetDB) has been used as a baseline. Comparison of query execution times confirms that distributed data warehousing solutions outperform classic relational DBMSs. Moreover, pre-aggregation and further denormalization of data, which reduce the number of distributed join operations, significantly improve query performance by several orders of magnitude. Most of distributed back-ends offer a good performance for complex analytical queries, while the Optimized Row Columnar (ORC) format paired with Presto and Parquet with Spark 2 query engines provide, on average, the lowest execution times. Apache Kudu on the other hand, is the only solution that guarantees a sub-second performance for simple genome range queries returning a small subset of data, where low-latency response is expected, while still offering decent performance for running analytical queries. In summary, research and clinical applications that require the storage and analysis of variants from thousands of samples can benefit from the scalability and performance of distributed data warehouse solutions.

Database, 2017

ISSN:1758-0463