Predictive Failure Risk Assessment Using Temporal Dependency Modeling in Distributed Computing Environments

Abstract Predicting adverse operational events in distributed computing environments remains challenging because relevant patterns are often distributed across historical system logs and machine-event records. This study investigates temporal feature-based predictive assessment using three supervised prediction tasks and an additional anomaly-label validation task. The experimental methodology combines chronological event aggregation, historical-window feature construction, block-level sequence representation, and machine-specific event-history analysis. Logistic Regression and Random Forest were evaluated using BGL supercomputer logs, HDFS_v1 block sequences, and Google Cluster machine-event records, while OpenStack logs were examined for injected-anomaly identifier validation. For BGL next-minute alert prediction, Logistic Regression achieved an F1-score of 0.3377 and PR-AUC of 0.2982, outperforming Random Forest, which achieved 0.2887 and 0.1729, respectively. For HDFS prefix-based block-outcome classification, Random Forest achieved an F1-score of 0.4481 and PR-AUC of 0.7325, compared with 0.1983 and 0.6350 for Logistic Regression. For Google subsequent machine-removal prediction, Logistic Regression achieved a higher F1-score of 0.9379, whereas Random Forest obtained a higher PR-AUC of 0.9639 compared with 0.9249. The OpenStack validation identified 120 occurrences of four injected-anomaly VM identifiers in the abnormal log and no occurrences in the two normal logs. These findings demonstrate that the effectiveness of temporal operational features and conventional machine-learning classifiers depends on the prediction target, data characteristics, and evaluation metric. The results provide a dataset-specific experimental baseline for operational event-risk assessment rather than direct evidence of independently verified physical failure prediction.