Accelerating approximate aggregation queries with expensive predicates

Researchers and industry analysts are increasingly interested in computing aggregation queries over large, unstructured datasets with selective predicates that computed using expensive deep neural networks (DNNs). As these DNNs because many applications can tolerate approximate answers, accelerating via approximations. Unfortunately, standard query processing techniques to accelerate such not applicable they assume the result of available ahead time. Furthermore, recent work cheap approximations (i.e., proxies) do support predicates. To predicates, we develop analyze a algorithm leverages proxies (ABae). ABae must account for key challenge it may sample records satisfy predicate. address this challenge, first use proxy group into strata so satisfying predicate ideally grouped few strata. Given strata, uses pilot sampling plugin estimates according optimal allocation. We show converges at an rate novel analysis stratified draws further outperforms on baselines six real-world datasets, reducing labeling costs by up 2.3x.