Runtime Prediction of Failure Modes from System Error Logs

Predicting potential failure occurrences during runtime is important to achieve system resilience and avoid hazardous consequences of failures. Existing failure prediction techniques in software systems involve forecasting failure counts, effects, and occurrences. Most of these techniques predict failures that may occur in future runtime intervals and only few techniques predict them at runtime. However, they do not estimate the failure modes and they require extensive instrumentation of source code. In this paper, we provide an approach for predicting failure occurrences and modes during system runtime. Our methodology utilizes system error log records to craft runtime error-spread signature. Using system error log history, we determine a predictive function (estimator) for each failure mode based on these signatures. This estimator can be used to predict a failure mode eventuality measure (a probability of failure mode occurrence) from system error log during system runtime. An experimental evaluation using PostgreSQL opensource database is provided. Our results show high accuracy of failure occurrence and mode predictions.