Abstract

When a software system crashes, users have the option to report the crash using automated bug tracking systems. These tools capture software crash and failure data (e.g., stack traces, memory dumps, etc.) from end-users. These data are sent in the form of bug (crash) reports to the software development teams to uncover the causes of the crash and provide adequate fixes. The reports are first assessed (usually in a semi-automatic way) by a group of software analysts, known as triagers. Triagers assign priority to the bugs and redirect them to the software development teams in order to provide fixes.
The triaging process, however, is usually very challenging. The problem is that many of these reports are caused by similar faults. Studies have shown that one way to improve the bug triaging process is to detect automatically duplicate (or similar) reports. This way, triagers would not need to spend time on reports caused by faults that have already been handled. Another issue is related to the prioritization of bug reports. Triagers often rely on the information provided by the customers (the report submitters) to prioritize bug reports. However, this task can be quite tedious and requires tool support. Next, triagers route the bug report to the responsible development team based on the subsystem, which caused the crash. Since having knowledge of all the subsystems of an ever-evolving industrial system is impractical, having a tool to automatically identify defective subsystems can significantly reduce the manual bug triaging effort.
The main goal of this research is to investigate techniques and tools to help triagers process bug reports. We start by studying the effect of the presence of stack traces in analyzing bug reports. Next, we present a framework to help triagers in each step of the bug triaging process. We propose a new and scalable method to automatically detect duplicate bug reports using stack traces and bug report categorical features. We then propose a novel approach for predicting bug severity using stack traces and categorical features, and finally, we discuss a new method for predicting faulty product and component fields of bug reports.
We evaluate the effectiveness of our techniques using bug reports from two large open-source systems. Our results show that stack traces and machine learning methods can be used to automate the bug triaging process, and hence increase the productivity of bug triagers, while reducing costs and efforts associated with manual triaging of bug reports.