With Spark you can indeed run batch jobs on multiple VMs. We are currently working having dynamic allocation and preemption on Kubernetes. This is for us essent

With Spark you can indeed run batch jobs on multiple VMs. We are currently working having dynamic allocation and preemption on Kubernetes. This is for us essential for our on premise cluster, as we want fair sharing. But for our cloud GCP jobs we could indeed save money by having a pool of spot instances. If deadlines are important, you could replace them by normal instances.