A Node Failure Prediction Methodology for Cloud-Native Clusters

The rapid development of cloud-native technologies has accelerated the transition from traditional monolithic architectures to microservice architectures. While microservices offer greater flexibility and scalability, ensuring Quality of Service (QoS) remains a significant challenge. In particular, node failures can lead to service degradation or even termination, severely affecting the availability of stateful services. This paper investigates a failure prediction method for nodes in Kubernetes clusters. To address the challenges of temporal imbalance in log data and class imbalance in samples, a dual-channel independent embedding method and a dynamically weighted loss optimization strategy are proposed. Furthermore, a failure prediction model named MSDF-TCN, based on Temporal Convolutional Networks (TCN), is introduced. The model incorporates a dual-branch global pooling strategy to capture both burst anomalies and long-term evolving patterns, thereby improving its performance and efficiency. Experimental results on public datasets demonstrate that the proposed approach achieves an F1 score of 51.1%, outperforming existing methods by an average margin of 8.5%, and providing effective support for proactive failure prediction and intervention in cluster nodes.