In this talk we will focus on effective Mesos cluster utilization for non-production jobs, which are commonly used for development and testing. While many of such jobs are idle, they consume cluster resources, leaving a huge gap between allocated and actually utilized resources. We will present our approach for detecting idle resources, offering them to other non-production jobs using Mesos oversubscription model, and isolating non-production jobs from production jobs. We will also highlight issues in Mesos’ approach to oversubscription and discuss possible solutions and workarounds.
