Workload-driven learned database components leverage historical query workloads to build predictive models. These models typically take queries or query plans as input and generate predictions or decisions to optimize query performance. However, the effectiveness of these models heavily depends on the quality and distribution of the training query workload (training workload). While prior studies focus on scenarios where the test query workload (test workload) distribution matches that of the training workload, real-world workloads often change over time, leading to workload distribution drift. In this paper, we explore the impact of workload drift on workload-driven models across different database components. We generate a variety of workload drift scenarios based on IMDb and STATS datasets, and analyze how different levels and types of drift affect estimation performance of each component and overall query execution performance 1 . Our findings highlight the limitations of current workload-driven models under workload drift and provide insights into designing more robust and adaptive solutions.
Meng et al. (Mon,) studied this question.