Efficient and Scalable Self-Healing Databases Using Meta-Learning and Dependency-Driven Recovery

TL;DR

This paper presents a self-healing database framework that combines meta-learning, graph neural networks, and reinforcement learning to enable real-time anomaly detection, failure prediction, and autonomous recovery with minimal data.

Contribution

It introduces a novel integrated approach for end-to-end self-healing in databases, leveraging dynamic dependency modeling and multi-objective RL for improved performance and transparency.

Findings

Achieved 90.5% anomaly detection F1-score with 5-shot adaptation.

Attained 90.1% accuracy in cascade failure prediction.

Reduced recovery latency by 85.1% in evaluations.

Abstract

Modern database management systems (DBMS) face significant challenges in maintaining performance and availability under dynamic workloads. This paper proposes a novel self-healing framework that integrates Model-Agnostic Meta-Learning (MAML) for few-shot anomaly detection, Graph Neural Networks (GNNs) for dependency-driven cascading failure prediction, and multi-objective Reinforcement Learning (RL) for autonomous recovery. Unlike existing database tuning systems that focus primarily on offline configuration optimization, our framework enables real-time, end-to-end self-healing by rapidly adapting to unseen workload patterns with minimal labeled data. We introduce dynamic GNN-based dependency modeling that captures workload-dependent relationships between database components, enabling proactive cascade prevention. A scalarized multi-objective RL formulation balances latency, resource utilization, and cost during recovery, while SHAP-based explainability ensures operational transparency. Evaluations on Google Cluster Data and TPC benchmarks demonstrate 90.5\% anomaly detection F1-score with 5-shot adaptation, 90.1\% cascade prediction accuracy, and 85.1\% latency reduction in recovery actions, outperforming strong baselines including Isolation Forest, LSTM autoencoders, static GCN, and standard RL methods.