CAREER: DrCloud: Drill-Ready Cloud Computing

云计算无处不在，但云服务中断仍在发生。本提案介绍了如何确保故障恢复在许多部署场景中正常工作。为了解决这一重要问题，该项目提出了可钻探云计算(DrCloud)，这是一种新的可靠性范例，倡导云系统在实际部署中定期执行“故障演练”(即故意安排实际故障，而不是等待意外的实际故障发生)。这种做法可以发现生产中的恢复问题，防止真正的停机。该项目将创建五个可供演练的云计算构建块：方法论、安全性、效率、可用性和通用性。具体地说，这五个子项目将通过对数百个生产中恢复问题的正式研究来充实新的研究方法，设计类似于适当的消防演习准备的保证安全(没有数据丢失和性能中断)的机制，开发最大限度地提高钻井执行的资源和资金效率的技术，设计简化钻井可用性的规范语言及其运行时，以及最终提高钻井的通用性(例如，支持软件升级和配置更改演习)。DrCloud项目将丰富数十年的容错计算研究和文献。该项目还将为社会带来许多直接好处；来自许多地区的用户越来越多地使用大规模存储和计算服务，这取决于可钻探云计算将促进的高可用性和可预测性。该项目还将涉及最先进的横向扩展云系统(Hadoop、Cassandra、HBase等)。为这些系统增加演练准备状态将提供下一代可靠系统的原型。

Cloud computing is pervasive, but cloud service outages still take place. This proposal addresses how to ensure that failure recovery will work robustly in many deployment scenarios. To address this important question, this project proposes drill-ready cloud computing (DrCloud), a new dependability paradigm that advocates cloud systems to routinely perform "failure drills" in real deployments (i.e., deliberately schedule real failures rather than waiting for unexpected real failures to happen). This practice can unearth in-production recovery issues and prevent real outages. This project will create five building blocks of drill-ready cloud computing: methodology, safety, efficiency, usability, and generality. Specifically, these five sub-projects will substantiate a new research methodology via a formal study of hundreds of in-production recovery issues, devise mechanisms that guarantee safety (no data loss and performance disruptions) analogous to a proper fire drill preparation, develop techniques that maximize resource and monetary efficiency of drill execution, design a specification language and its runtime that simplifies drill usability, and finally boost drill generality beyond failure drills (e.g., supporting software upgrade and configuration change drills). The DrCloud project will enrich decades of research and literature in fault-tolerant computing. The project will also bring many direct benefits to the society; users from many areas increasingly use large-scale storage and computation services, depending on high availability and predictability that drill-ready cloud computing will facilitate. The project will also involve state-of-the-art scale-out cloud systems (Hadoop, Cassandra, HBase, etc.). Adding drill-readiness to these systems will provide prototypes of next-generation reliable systems.