What Lies Beyond Structured Data? A Comparison Study For Metric Data Storage

The handling of massive data requires the retrieval procedures to be aligned with the storage model. Similarity searching is an established paradigm for querying large datasets by content, in which data elements are compared by means of metric distance functions. Although several strategies have been proposed for the storage of data queried by metrics into relational schemas, no empirical assessment on the suitability of such strategies for similarity searching has been conducted. In this study, we aim at filling this gap by providing an indepth evaluation of storage models for Relational Database Management Systems (RDBMS) in standard SQL. Accordingly, we propose a taxonomy, which divides approaches into four categories, Binary, Relational, Object-Relational, and Semistructured, and implement a representative storage model for each category within a common framework. We carried out extensive experiments on the four implemented strategies, and results indicate the Relational and Object-Relational storage models outperform the other competitors in most scenarios, whereas the Binary storage model reaches a good performance for queries with costly comparisons. Finally, the Object-Relational approach showed the best compromise between performance and representation, since its behavior is similar to the Relational storage model with a cleaner representation.