HYBRIDJOIN for Near-Real-Time Data Warehousing

An important component of near-real-time data warehouses is the near-real-time integration layer. One important element in near-real-time data integration is the join of a continuous input data stream with a disk-based relation. For high-throughput streams, stream-based algorithms, such as Mesh Join (MESHJOIN), can be used. However, in MESHJOIN the performance of the algorithm is inversely proportional to the size of disk-based relation. Also, MESHJOIN cannot deal with intermittent streams efficiently, because tuples could wait for an undetermined time, thus defying the near-real-time character of the stream. The Index Nested Loop Join (INLJ) can be set up so that it processes stream input, and can deal with intermittences in the update stream but it has low throughput. In this paper we introduce a robust stream-based join algorithm called Hybrid Join (HYBRIDJOIN) which combines the two approaches. As a theoretical result we show that HYBRIDJOIN is asymptotically as fast as the fastest of both algorithms. We present performance measurements of our implementation. We use synthetic data that we base on a Zipfian distribution, which is widely accepted as a plausible distribution for real world identifier sets in many domains. In our experiments, HYBRIDJOIN performs significantly better for typical parameters of the Zipfian distribution, and in general performs in accordance with the theoretical model while the other two algorithms are unacceptably slow under different settings. Hence HYBRIDJOIN is a robust algorithm that generally performs at an acceptable speed.