Defcon: Preventing Overload with Graceful Feature Degradation.

Justin Meza, Thote Gowda, Ahmed Eid, Tomiwa Ijaware, Dmitry Chernyshev, Yi Yu, Md Nazim Uddin, Rohan Das, Chad Nachiappan, Sari Tran, Shuyang Shi, Tina Luo, David Ke Hong,Sankaralingam Panneerselvam, Hans Ragas, Svetlin Manavski, Weidong Wang, Francois Richard


引用 2|浏览1
Every day, billions of people depend on Internet services for communication, commerce, and entertainment. Yet planetary-scale data center infrastructures consisting of millions of servers experience unplanned capacity outages and unexpected demand for resources; how can such infrastructures remain reliable in the face of capacity and workload flux? In this paper, we introduce Defcon, a system for improving the availability of large-scale, globally-distributed Internet services using graceful feature degradation. In response to overload conditions, Defcon enables site operators to gradually disable less-critical features in order to reduce resource demand. Defcon presents a common interface to product developers to define feature knobs that represent degradation capabilities. Defcon automatically tests knobs to understand each knob's product- and infrastructure-level trade-offs. At Meta, we have used Defcon to improve global product availability in the face of worldwide demand-surges in addition to large-scale infrastructure failures.
AI 理解论文