Improving the quality of APIs through the analysis of software crash reports

Modern programs depend on apis to implement a significant part of their functionality. Apart from the way developers use apis to build their software, the stability of these programs relies on the apis design and implementation. In this work, we evaluate the reliability of apis, by examining software telemetry data, in the form of stack traces, coming from Android application crashes. We got 4.9 GB worth of crash data that thousands of applications send to a centralized crash report management service. We processed that data to extract approximately a million stack traces, stitching together parts of chained exceptions, and established heuristic rules to draw the border between applications and api calls. We examined 80% of the stack traces to map the space of the most common application failure reasons. Our findings show that the top ones can be attributed to memory exhaustion, race conditions or deadlocks, and missing or corrupt resources. At the same time, a significant number of our stack traces (over 10%) remains unclassified due to generic unchecked exceptions, which do not highlight the problems that lead to crashes. Finally, given the classes of crash causes we found, we argue that api design and implementation improvements, such as specific exceptions, non-blocking algorithms, and default resources, can eliminate common failures.