You are designing network segmentation for an environment that runs mixed workloads on the same fabric: east-west AI training traffic, storage replication, cluster management, and latency-sensitive inference services. You want isolation strong enough to prevent noisy-neighbor effects and blast-radius issues, but you cannot materially reduce throughput or create an operations model that is too complex to run.
How would you segment traffic for mixed workloads without compromising throughput?