LAMP: A Dual-Mode Framework for Database Workload Memory Prediction

Precise prediction of working memory consumption for query workloads is crucial for preventing out-of-memory errors and optimizing resource utilization in modern database systems. While existing approaches focus on performance prediction for individual queries, they fail to capture the complex memory dynamics arising from concurrent execution and resource contention. To remedy this, we present Load-Aware Memory Prediction framework (LAMP), a novel framework that addresses workload-level memory prediction through two complementary operational modes: accuracy-oriented prediction for development environments and robustness-oriented prediction for production deployments. LAMP utilizes a neural model that integrates Tree Convolutional Neural Networks for encoding query execution plans alongside DeepSets-based aggregation to effectively manage variable-sized workloads while maintaining permutation invariance. This approach allows the model to capture both the unique characteristics of individual queries and the overarching patterns of collective memory consumption. In production environments, we propose an uncertaintyaware memory allocation strategy grounded in heteroscedastic regression. This strategy quantifies prediction confidence and facilitates risk-calibrated provisioning through the implementation of adaptive safety buffers. Extensive experiments demonstrate that LAMP achieves up to 40.98% reduction in prediction error compared to state-of-the-art methods in accuracy-oriented mode, while in robustness-oriented mode it achieves the optimal tradeoff between system stability and prediction accuracy, providing database administrators with a principled approach to balance resource efficiency and system reliability.