Demo of LearnedWMP: Workload Memory Prediction Using Deep Query Template Representations

The database management system (DBMS) relies on a limited amount of working memory for in-memory operations during query execution. Inaccurate estimations can result in suboptimal query performance and potential failures. Achieving precise working memory estimations is therefore crucial to enhance the DBMS efficiency. We present LearnedWMP, a novel approach for simultaneously estimating working memory for a set of queries (i.e., a workload), as opposed to the conventional practice of estimating resources separately for each query. In a series of comprehensive experiments, the LearnedWMP model demonstrates improved accuracy in memory estimation and faster performance during both training and inference when compared to standard methods. The findings highlight the effectiveness and potential impact of the LearnedWMP approach on optimizing query performance. The audience can explore LearnedWMP through various scenarios showcasing its estimation accuracy and internal model workings.