The scale-up vs. scale-out argument is very old. I don't see Kubernetes providing a new answer either way. But it does not have to. Let's stipulate that most

The scale-up vs. scale-out argument is very old. I don't see Kubernetes providing a new answer either way. But it does not have to. Let's stipulate that most analytic applications can run on single servers. You still need a way to schedule work efficiently when the applications in aggregate require more than a single host. Kubernetes does that very well. That's leaving out considerations like HA. The paper is notably silent on this topic. To keep systems up across failures as well as maintenance you need replicas in different locations. There's an implication in the paper that everyone is just going to run off S3, which is assumed to be globally available. Not everyone makes that choice for a variety of reasons. So we will continue to see shared nothing designs that require placement across hosts, racks, AZs, etc. Kubernetes is efficient at deploying these.