A pilot study of cross-system failures

Abstract

Today’s distributed system infrastructures usually consist of multiple systems that cooperate to deliver computing and storage services together. The reliability of distributed system infrastructures not only depends on the reliability of individual systems, but also on the reliability of the interactions among multiple systems. In fact, failures of system infrastructures are often caused by failures of interactions across multiple systems. We term such failures as cross-system failures.

In this thesis, we conduct a pilot study on cross-system failures on seven real-world systems that are commonly used to build distributed system infrastructures. We analyze the characteristics of real-world cross-system failures in four different dimensions: failure impacts, triggering events, root causes of the failures, and fixes of the root causes. Our goal is to provide insights to guide future research on resolving cross-system failures, and to shed light on the techniques that improve the reliability of cross-system interactions and distributed system infrastructures.