With the ever-increasing gap between generated data amount and compute performance, moving data becomes the elephant in the room. We propose a solution to this issue where we rethink how we can offload computations and pre-filter data efficiently.
In this talk, we present our vision of a novel sub-operator-centered system architecture. It is developed around existing database engine components that can also leverage heterogeneous hardware. Our concepts expose already present sub-components of a database engine to shape a more versatile interface, which allows mapping and processing various data flows. We use our existing research on database operators on a single machine and put the findings in the context of sub-operators. We look into the question, of when partitioning pays off in join execution and present the results of our apples-to-apples comparison of comparing fingerprint and bloom filter variants.