Accommodate Apache YARN to long-lived services

Large-scale cluster scheduling and management systems are becoming a dominant platform for a variety of applications and services. These complex ?cloud? systems often run on clusters of thousands of unreliable commodity machines and mush handle all kinds of failures, schedule and deployment requirements while achieving the maximum resource utilization. We take YARN to a full analysis and argue that YARN currently excels in handling batch jobs without interacting with outside system or clients. In this paper, we propose a solution to the service registration, log management, resource reallocation problem of YARN and accommodate YARN to long-lived services.