This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
This paper proposes E3-Rewrite, a novel framework leveraging a large-scale language model (LLM) to overcome the limitations of existing rule-based SQL query rewriting methods. Existing methods rely on a fixed set of rules, making it difficult to generalize to new query patterns or complex queries and failing to fully capture effective rewriting strategies. E3-Rewrite leverages execution plans and retrieved examples to build context, and designs a reward function targeting feasibility, equivalence, and efficiency to perform optimal query rewriting through reinforcement learning. Through a step-by-step training process, it achieves stable multi-objective learning and demonstrates up to 25.6% reduction in query execution time and up to 24.4% improvement in rewriting success rates compared to state-of-the-art methods on various SQL benchmarks.
Takeaways, Limitations
•
Takeaways:
◦
We demonstrate that LLM can be used to overcome the limitations of existing rule-based approaches and solve complex SQL query rewriting problems.
◦
We demonstrate that feasible, equivalent, and efficient queries can be generated by building context using execution plans and examples and designing a reward function based on reinforcement learning.
◦
In various SQL benchmarks, we achieved shorter query execution times and higher rewrite success rates than existing best-performing models.
•
Limitations:
◦
It depends on the performance of LLM, and the limitations of LLM may also affect the performance of E3-Rewrite.
◦
Further research may be required as the design of the reward function and optimization of the reinforcement learning process have a significant impact on performance.
◦
Generalization performance for certain types of complex queries may not yet be sufficiently validated.
◦
Further evaluation of scalability and stability in real-world operating environments is required.
