Mechanisms and Architectures for Tail-Tolerant System Operations in Cloud

Conducting system operations (such as upgrade, reconfiguration, deployment) for large-scale systems in cloud is error prone and complex. These operations rely heavily on unreliable cloud infrastructure APIs to complete. The inherent uncertainties and inevitable errors cause a long-tail in the completion time distribution of operations. In this paper, we propose mechanisms and deployment architecture tactics to tolerate the long-tail. We wrapped cloud provisioning API calls and implemented deployment tactics at the architecture level for system operations. Our initial evaluation shows that the mechanisms and deployment tactics can effectively reduce the long tail.