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

Joydeep Chandra, Prabal Manhas · 2025

Read Efficient and Scalable Self-Healing Databases Using Meta-Learning and Dependency-Driven Recovery by Joydeep Chandra, Prabal Manhas (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-13757.txt.

Introduction

Databases must stay fast and up while work loads shift through the day. Joydeep Chandra and Prabal Manhas propose a self-healing frame in 2025 that spots faults, warns of spread, and acts on its own. Their plan joins three tools that learn in distinct ways.

One tool learns to spot odd patterns from just a few examples, so it adapts fast to new loads. A graph-based tool maps how parts of the system lean on each other and flags chains of failure before they grow. A third tool picks repair steps that balance speed, resource use, and cost. The frame also adds plain-language reasons for its choices, so staff can check its work. Tests draw on cluster data from Google.

Catalog record

CREATORJoydeep Chandra, Prabal Manhas
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE22 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-13757

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Efficient and Scalable Self-Healing Databases Using Meta-Learning and Dependency-Driven Recovery Joydeep Chandra, Prabal Manhas Abstract This study explored the development of a novel self-healing framework for databases using meta-learning and reinforcement learning techniques. The primary objective was to address the challenges of real-time adaptability and minimal retraining in dynamic workload environments. The proposed approach integrated Model-Agnostic Meta-Learning (MAML) with reinforcement learning to enable anomaly detection and corrective actions that adapted swiftly to evolving database conditions. Multi-objective optimization was employed to balance performance, resource utilization, and cost efficiency during the healing process. Graph Neural Networks (GNNs) were incorporated to model interdependencies within database components, ensuring holistic recovery strategies. Data efficiency was enhanced through synthetic task augmentation and self-supervised learning, enabling effective training in sparse data regimes. To promote trust and transparency, explainable AI techniques were integrated to provide interpretable insights into anomaly detection and healing actions. Federated meta-learning further enabled privacy-preserving adaptability in distributed database environments. The framework demonstrated significant improvements in adaptability, efficiency, and reliability, contributing to advancements in database management and self-healing systems. I Introduction The self-healing database has received considerable interest as a way of improving the dependability and performance of the DBMS that runs under fluctuating workloads. Previous works have highlighted the need to incorporate anomaly detection and reconfiguration in management of databases. Some of the existing techniques for handling such challenges include: Generative Adversarial Networks for Anomaly detection [1], and reinforcement learning for self- adaptive systems [2]. But historical models commonly take considerable retraining and do not retain their relevance in handling new workloads promptly [3]. Model-Agnostic Meta-Learning (MAML) is a meta-learning model that has recently been explored to train a model that could help achieve fast task adaptation with minimal training data [4]. It has been used successfully in anomaly detection for dynamic systems [5] as well as in few shot learning cases [6] thereby making it a useful starting point for addressing the adaptability gap for self healing systems. Also, it has been applied to improve learning and developing of the recovery strategy in the distributed system [7] and the multi-objective optimization method has been adopted to minimize the trade-off between performance, cost and resource utilization [8]. Despite these advancements, challenges persist in handling sparse data environments and modeling interdependencies within database components. Approaches using graph neural networks (GNNs) have been proposed to capture cascading failures and dependencies [9], while federated meta-learning frameworks have been introduced to enhance privacy-preserving adaptability in distributed systems [10]. Self-supervised learning methods have also emerged as a promising avenue for leveraging unlabeled data in anomaly detection tasks [11]. This work further expands upon these bases by incorporating MAML to the existing reinforcement learning methods in order to create an adaptation for a self-healing database framework set that does not need much retraining when subjected to new workloads. Some of the proposed innovations are multi-objective optimization applied to general target system recovery, concrete dependency graphical model, as well as synthetic task aggrandizement for superior training in the least dense data situations. The features presented in this paper are the key steps and mechanisms that could become a foundation for further development of the proposed approach, thus overcoming the existing systems’ drawbacks and introducing novel concepts for database management and self-healing[1, 2]. II Related Works Self-healing systems have been extensively studied to enhance system resilience through advanced anomaly detection and repair mechanisms. Generative Adversarial Networks (GANs) have been explored as a powerful tool for anomaly detection in high-speed systems, as demonstrated by Cao et al., who used synthetic anomalous data to detect rare failure modes [1]. While GAN-based methods achieved high accuracy in anomaly prediction, their inefficiency in managing large datasets and inability to adapt to dynamic workload patterns limited their applicability in real-time database systems. Furthermore, these approaches lacked sensitivity to evolving workloads, which is critical in unstable and dynamic environments. Fig. 1: Framework of the Self Healing Database using MAML, GNN and RL Reinforcement Learning (RL) has been widely employed in self-adaptive systems for its ability to refine task-specific strategies. Studies have highlighted its potential to optimise system performance by dynamically adjusting recovery actions [2]. However, traditional RL approaches exhibited significant limitations, including lengthy convergence cycles and reliance on static learning models, which restricted adaptability to dynamic database environments. These shortcomings emphasised the necessity for more flexible and efficient methods capable of responding to evolving system states. Automated Machine Learning (AutoML) techniques were introduced to improve anomaly detection by automating model selection and hyperparameter tuning [3]. Bahri et al. concluded that AutoML demonstrated strong performance in static workload scenarios but struggled in dynamic environments where workloads frequently shifted. This inability to adapt to changing patterns further reinforced the need for advanced solutions capable of real-time anomaly detection and adaptation. Meta-learning has emerged as a revolutionary paradigm for task adaptation, with Model-Agnostic Meta-Learning (MAML) being a significant advancement in this field. Finn et al. demonstrated that MAML allows models to rapidly adapt to new tasks with minimal retraining, making it highly suitable for dynamic environments [4]. Chalvidal et al. extended this concept by incorporating self-modifying networks, which enhanced adaptation efficiency and reduced computational overhead [4]. In addition, meta-learning applications for anomaly detection, as discussed by Ding et al., employed few-shot learning to generalise across related tasks, effectively addressing the problem of limited labelled data [6]. Despite these advancements, existing meta-learning frameworks have been unable to adequately capture the complex interdependencies required for cascade failure detection in distributed database systems. Graph Neural Networks (GNNs) have shown significant promise in modelling interdependencies within distributed systems. Studies, including [9], have demonstrated their ability to predict cascading failures and model relationships between database components such as queries, tables, and indexes. However, traditional GNN-based frameworks have primarily focused on reactive strategies, lacking the integration of proactive anomaly detection techniques. This reactive approach often limits their ability to prevent failures before they propagate across the system. The incorporation of federated meta-learning frameworks, as outlined in [10], has been instrumental in addressing privacy concerns in distributed systems by enabling localised learning without centralised data aggregation. However, these frameworks face challenges such as node heterogeneity and inconsistent model updates, which can hinder their scalability and reliability. Multi-objective optimisation techniques for resource allocation and recovery, as discussed in [8], have provided structured methodologies for balancing performance, resource consumption, and cost. Nonetheless, fixed optimisation goals in these approaches restrict their applicability to dynamic database environments where system priorities frequently shift. Self-supervised learning (SSL) has also been explored for training anomaly detection models with unlabelled data, as proposed by Kumagai et al. [11]. SSL improves data efficiency and enables the development of models in sparse data regimes. However, its effectiveness diminishes in low-density datasets and scenarios where synthetic task augmentation, combined with meta-learning techniques, can provide superior outcomes. The proposed framework builds upon these prior advancements by integrating multiple state-of-the-art techniques to address the limitations of existing approaches. Synthetic task augmentation leverages generative models to generate diverse workload patterns and unique failure modes, addressing knowledge gaps highlighted in [1, 11]. Dependency modelling with GNNs extends prior work in [9], enabling precise identification and prevention of cascading failures in real-time. Multi-objective meta-reinforcement learning introduces adaptive trade-off schemes for optimising performance, cost, and

Read the complete text