""" One could imagine a very imperative “cluster operating system,” like the above, which exposed primitives like “allocate 5 CPUs worth of compute” or “create a new virtual network,” which in turn backed onto configuration changes either in the system’s internal abstractions or into calls into the EC2 API (or other underlying cloud provider). """

There is a name for this: Single System Image. There is no need to write the operating system layer - Linux works well.

I built a SSI. Each grid can host several apps, 'gridapps'. As priorities change during the day, an operator can rebalance resources using a bash-like shell. Each gridapp has (1) core nodes, which need to keep running, and (2) scale nodes which can be started and stopped. Nodes can be configured so they are taskset-bound to an individual processor, or sharing a processor between several nodes. There are equivalent primitives for memory and persistent storage. Commercial, 1000 cores.

The author contrasts this with K8s by saying that K8s is declarative. This is not the important distinction. My system had declarative configuration, and did a fitting at startup.

The key difference is motive. The design above is the natural choice when you build a grid computer from first-principles.

K8s is designed to solve a different problem: scaling up the traditional model of programming. Here, a developer writes an algorithm on their workstation. The deployment circumstances and interactions with other systems are an afterthought. Often this developer does not deploy the code themselves, this is done by another team who don't understand the internals of the code.