MVGPT: Generative Materialized View Forecasting

Existing materialized view (MV) selection methods rely on historical queries to select MVs. However, queries are dynamically and continually changing, and it is ineffective to generate MVs based only on historical queries. To address this limitation, we propose a novel MV forecasting framework, MVGPT, which utilizes large language models (LLMs) to predict effective MVs for evolving query workloads. First, we finetune an LLM for MV forecasting, retrieve relevant events to incorporate real-time event knowledge to instruct the query workload evolution inference. We then propose a predicate-level MV model based on the Bayesian network to generate valid and effective MVs. Finally, we propose a variable set representation method as the interactive interface between the LLM and the MV model to address the risk that the LLM generates incorrect results. We have implemented our system MVGPT and deployed it into Alibaba's advertising analysis platform, and experimental results on real datasets show that our method outperforms existing state-of-the-art approaches.