An operations team plans a driver upgrade on a shared cluster. What is the best approach?

1. Upgrade every node simultaneously without draining workloads
2. Validate compatibility and use a controlled rollout with recovery plans
3. Remove all monitoring during the upgrade indefinitely
4. Assume every older CUDA application becomes incompatible automatically