Lock Prediction for Zero-Downtime Database Encryption
Modern enterprise database systems face significant challenges in balancing data security and performance. Ensuring robust encryption for sensitive information is critical for systems' compliance with security standards. Although holistic database encryption provides strong protection, existing database systems often require a complete backup and restore cycle, resulting in prolonged downtime and increased storage usage. This makes it difficult to implement online encryption techniques in high-throughput environments without disrupting critical operations. To address this challenge, we envision a solution that enables online database encryption aligned with system activity, eliminating the need for downtime, storage overhead, or full-database reprocessing. Central to this vision is the ability to predict which parts of the database will be accessed next, allowing encryption to be applied online. As a step towards this solution, this study proposes a predictive approach that leverages deep learning models to forecast database lock sequences, using IBM Db2 as the database system under study. In this study, we collected a specialized dataset from TPC-C benchmark workloads, leveraging lock event logs for model training and evaluation. We applied deep learning architectures, such as Transformer and LSTM, to evaluate models for various table-level and page-level lock predictions. We benchmark the accuracy of the trained models versus a Naive Baseline across different prediction horizons and timelines. The study experiments demonstrate that the proposed deep learning-based models achieve up to 49% average accuracy for table-level and 66% for page-level predictions, outperforming a Naive Baseline. By anticipating which tables and pages will be locked next, the proposed approach is a step toward online encryption, offering a practical path toward secure, low-overhead database systems.