Data scientists occasionally need GPU nodes on a GKE cluster for long-running, non-restartable jobs. Minimize cost while meeting this need. What should you do?

Node auto-provisioning is the most cost-effective solution here. It automatically creates and deletes node pools based on pending pods, including those requiring GPUs. This ensures GPU nodes are only provisioned when needed and deprovisioned when idle, minimizing costs. Creating a Vertical Pod Autoscaler adjusts resource requests/limits for pods but doesn't provision new GPU nodes. A node pool with preemptible VMs and GPUs is cheaper, but preemptible VMs are not suitable for long-running, non-restartable jobs due to their potential for preemption. Creating a node pool with GPUs and autoscaling with a minimum size of 1 means you'll always have at least one GPU node running, incurring unnecessary costs when no GPU jobs are active.