A scalable machine learning strategy for resource allocation in database

Modern cloud computing systems require intelligent resource allocation strategies that balance quality-of-service (QoS), operational costs, and energy sustainability. Existing deep Q-learning (DQN) methods suffer from sample inefficiency, centralization bottlenecks, and reactive decision-making during workload spikes. Transformer-based forecasting models such as Temporal Fusion Transformer (TFT) offer improved accuracy but introduce computational overhead, limiting real-time deployment. We propose LSTM-MARL-Ape-X, a novel framework integrating bidirectional Long Short-Term Memory (BiLSTM) for workload forecasting with Multi-Agent Reinforcement Learning (MARL) in a distributed Ape-X architecture. This approach enables proactive, decentralized, and scalable resource management through three innovations: high-accuracy forecasting using BiLSTM with feature-wise attention, variance-regularized credit assignment for stable multi-agent coordination, and faster convergence via adaptive prioritized replay. Experimental validation on real-world traces demonstrates 94.6% SLA compliance, 22% reduction in energy consumption, and linear scalability to over 5,000 nodes with sub-100 ms decision latency. The framework converges 3.2× faster than uniform sampling baselines and outperforms transformer-based models in both accuracy and inference speed. Unlike decoupled prediction-action frameworks, our method provides end-to-end optimization, enabling robust and sustainable cloud orchestration at scale.