Empirical Characterization of User Reports about Cloud Failures

Abstract

Cloud services are important for healthcare, banking, communication, and other purposes. Inevitably, such services fail, harming the processes and disturbing the people that depend on them. Understanding failure in cloud services is challenging, but important to help preventing them. Much work has studied failure logs and reports provided by infrastructure operators. However, there is a paucity of information about how users perceive the failures of cloud services. In this work, we collect user-reported failures and characterize them empirically. We collect failures reported by users to the trusted aggregator Outage Report for 12 cloud services over 16 months spread across 2019 and 2020. We show evidence that user-reported failures not only capture major failures also self-reported by cloud operators, but also provide information about additional failures. We count and analyze time patterns in these reports. We make 6 main observations about how users perceive failure in cloud services. We find over 10x differences in request failure rates across microservice structures when using user reported traces compared to using a constant failure distribution. Overall, our study provides the first long-term characterization of user-reported cloud failures.