I have a pretty unique case on how I want to increase/decrease the amount of VM’s running in my EKS cluster.

Currently I have RabbitMq setup with jobs to run that are all network intensive(large uploads and downloads from 1GB-1TB). I was thinking it would be interesting to launch a VM per job to run and have a custom k8 scheduler that will place the pod on the appropriate VM. Once the job completes I destroy the pod and the VM as well. With Auto-scaling Groups I don’t see how this mapping can work. Is this feasible with EKS or do I have to use something like Karpenter to control cluster up/down scaling?

I am looking for input on if something like this has been done, and what people would think of scaling EKS cluster VM count to the amount of physical work to process that is all network IO bound.

Thank you.