Diagnosing mobile applications in the wild

摘要: There are a lot of applications that run on modern mobile operating systems. Inevitably, some these fail in the hands users. Diagnosing failure to identify culprit, or merely reproducing lab is difficult. To get insight into this problem, we interviewed developers five and analyzed hundreds trouble tickets. We find support for diagnosing unexpected application behavior lacking across major platforms. Even when implement heavy-weight logging during controlled trials, they do not discover many dependencies then stressed wild. They also well-equipped understand how monitor large number without impacting phone's limited resources such as CPU battery. Based findings, argue three fundamental changes reporting phones. The first spatial spreading, which exploits phones field by spreading monitoring work them. second statistical inference, builds conditional distribution model between its presence partial information. third adaptive sampling, dynamically varies what each phone monitors, adapt both varying population being learned about failure. propose system called MobiBug combines techniques simplify task applications.