DB-BERT extracts tuning hints for database systems from text documents such as the manual. It uses extracted hints as a starting point for automated performance tuning.
To analyze text, DB-BERT uses pre-trained language models such as BERT. During tuning, DB-BERT iteratively selects parameter settings and measures performance on a user-defined benchmark. It selects settings to try via reinforcement learning, integrating information extracted from text as well as performance measurements in past iterations. By exploiting text as additional input, DB-BERT tends to find promising configurations faster than baselines.
- SIGMOD 2022 DB-BERT: a database tuning tool that “reads” the manual. Immanuel Trummer.
- SIGMOD Record 2021 Database tuning using natural language processing. Immanuel Trummer.
- VLDB 2021 The case for nlp-enhanced database tuning: towards tuning tools that “read the manual”. Immanuel Trummer.