High-Accuracy Cloud Job Failure Prediction Using a Two-Layer Voting Architecture

Abstract

Modern cloud computing systems require precise job failure prediction for system dependability, resource efficiency, and ongoing service delivery. This study adds CatBoost, a strong gradient boosting algorithm, to a multilayer voting-based prediction framework to improve accuracy and resilience. CatBoost handles categorical features and complicated data distributions in cloud workloads better than traditional models, and its symmetric tree structure and built-in overfitting control promote generalization across varied and dynamic contexts. The enhanced system trains and evaluates the model using the Google Cluster 2019 trace dataset, outperforming machine learning and ensemble methods. In addition to model improvement, a Flask-based web interface allows users to upload task data and instantaneously receive failure predictions and categorization results. This integration connects theoretical models to practice. Experimental findings reveal that the CatBoost-based extension outperforms competing framework classifiers and delivers dependable results even on unseen data. The suggested addition enhances the cloud job failure prediction system's accuracy, scalability, and usability, making it appropriate for cloud infrastructure management and decision-making.