Optimizing machine learning inference queries with correlative proxy models

We consider accelerating machine learning (ML) inference queries on unstructured datasets. Expensive operators such as feature extractors and classifiers are deployed user-defined functions (UDFs), which not penetrable with classic query optimization techniques predicate push-down. Recent schemes (e.g., Probabilistic Predicates or PP) assume independence among the predicates, build a proxy model for each offline, rewrite new by injecting these cheap models in front of expensive ML UDFs. In manner, unlikely inputs that do satisfy predicates filtered early to bypass show enforcing assumption this context may result sub-optimal plans. paper, we propose CORE, optimizer better exploits correlations accelerates queries. Our solution builds online leverages branch-and-bound search process reduce building costs. Results three real-world text, image video datasets CORE improves throughput up 63% compared PP 80% running it is.