Inferring Faults from Persistent State Interactions

Опубликовано 7 сентября 2016, 16:24

Misconfiguration of persistent state such as registries or files is a major cause of software faults. An individual machine with a misbehaving application has a large amount of persistent state, so automatic troubleshooting methods are essential for narrowing down the state which could have caused the fault. In this talk, we describe how we apply machine learning techniques to explore the relationship between the state that an application consumes and the exit codes that it produces. Using the temporal ordering of state consumption in combination with decision trees and other forms of mutual information analysis, we describe mechanisms for nominating the important state that influences program behaviour.