Inclusão de Operadores Físicos de Junção por Similaridade em um SGBD Comercial

Complex data, such as video, images and audio, require particular forms of querying, storing and indexing that commercial DBMSs still do not provide. One of the solutions for DBMS to o↵er support for complex data is to extend relational operators to represent similarity queries. Similarity queries retrieve data based on similarity relations among stored data, which are derived from the intrinsic data content. One important type of similarity query is the similarity join that returns pairs of elements from two input datasets that satisfy the stated join condition, which can be for instance if they are closer to each other than a given threshold (Range join) or if one element is among the k-nearest neighbors of the other (k-NN join). Existing algorithms to execute similarity joins essentially consider input datasets/relations are directly read from disk. However, join is one of the most time-consuming operations in a query, therefore delaying it in the execution plan and join filtered data in memory usually results in performance gain. In this paper, we present range join algorithms developed to perform in memory on filtered data on top of a commecial DBMS. Our proposal is to allow executing similarity joins in di↵erent positions in the query plan and evaluate how distinct algorithms behave according to varied situations. Presented results show that best developed options are the state-of-the-art DBSimJoin algorithm for high selective filtered inputs and an indexbased similarity join for queries in which the join selectivity is high and a proper data index is available.