Brian Grant

The definition of "scheduling" is to plan an event to happen at a particular time. OS thread schedulers primarily chose which thread to run in the pre-multi-core era. Batch queuing systems perform a similar function: pick tasks from queues. Is that cohyponymic transfer?

Borg took that further, perhaps due to its workqueue heritage. Babysitter had a machine assigner. Like Borg, the "scheduler" in Kubernetes principally does placement.

"Scheduler" in the context of systems like Kubernetes, Mesos, and Peloton has apparently now come to imply the entire computational resource management ecosystem, including queuing, coscheduling, quota and sharing management, autoscaling, oversubscription, placement, etc

And preemption, descheduling, and so on. It's important to understand how all these control loops interact with each other and with their surroundings, such as load balancers and infrastructure autoprovisioners. But in Kubernetes we've tried to make these functions composable

As I mentioned in my podcast, the importance of some of these is greatly dependent on the variety of workloads run, the elasticity of the underlying infrastructure, and priorities. For instance, many use Kubernetes to accelerate deployment more than to increase utilization

If the infrastructure is not very elastic, as in bare metal datacenter environments, then utilization can be a primary concern, and that's where techniques such as vertical scaling, oversubscription, binpacking, mixing workloads, QoS, share management, etc. become important