Popular approaches to DBMS (Database Management System) management rely on monitoring log files or using third-party statistical tools to observe workload variations, which requires continuous manual intervention from DBAs (Database Administrator) for diagnosis and tuning. Such reactive management is labor-intensive and difficult to scale. To address this limitation, we present a non-intrusive DBMS workload forecasting framework that anticipates future queries and their execution durations directly from SQL-level statistics, rather than from hardware metrics such as CPU or memory utilization. Our method is designed for hybrid transactional/analytical processing (HTAP) workloads, which are increasingly common in data-lakehouse-based applications. Unlike existing approaches that rely on cluster-based workload modeling, we propose a sorting-based, template-level workload forecasting method. By explicitly modeling and forecasting individual SQL templates, our approach replaces coarse-grained cluster representations with fine-grained, semantically meaningful workload units. This design substantially improves model interpretability and controllability, enabling DBAs and system components to directly understand, inspect, and regulate predicted workload patterns at the query-template level.
Li et al. (Sat,) studied this question.