Understanding Online Failure Prediction in Linux Through Complementary Multi-View Explainability

This paper presents an explainable online failure prediction pipeline for Linux, combining consensus-based feature selection, temporal onset analysis, subsystem-level causal analysis, and complementary diagnostic mechanisms. According to the supplied abstract, frozen training artifacts achieved 91–94% detection on unseen workloads with false-alarm rates below 1%. Diagnostic generalization was substantially weaker: warning lead time varied from 38 to 215 seconds by failure mode, while Leave-One-Mode-Out evaluation produced 0% accuracy for unseen failure modes. The metadata is future-dated and therefore requires verification.