The problem is the lack of controls in the timescales that the CPU scheduler is using, that do not necessarily match the timescales of applications. This is cla

The problem is the lack of controls in the timescales that the CPU scheduler is using, that do not necessarily match the timescales of applications. This is classic statistical multiplexing and burstiness problem often encountered in the network queueing world. I wrote a blog and a couple of synthetic benchmarks that highlight the issues a couple of months ago that you might find interesting: https://medium.com/engineering-at-palo-alto-networks/kubernetes-scheduling-and-timescales-e98d8e31d304 https://medium.com/engineering-at-palo-alto-networks/kuberne...