Here are some common examples of bad designs:

🟠 Containers on shared hardware as the only level of isolation between tenants

🟠 VMs on shared hardware as the only level of isolation between tenants (less serious than containers, but still risky)

🟠 Lack of VXLANs or improper configuration, and no concept of a per-tenant VPC on the frontend network, relying on firewalls

🟠 Multi-tenant Kubernetes control planes, where components such as kube-apiserver, the scheduler, helm charts controlling cluster-wide services such as a shared GPUOperator or NetworkOperator, and etcd are shared across tenants

🟠 No hardware, firmware, and OS provisioning automation, leading to a practice where machines are recycled between tenants instead of provisioning things from scratch

🟠 Backend storage on a shared network with no isolation (the VPC comment again)

🟠 Giving any tenant access to the BMC network (IPMI, Redfish) or to any management port on Bluefield DPUs or any SmartNICs

🟠 Leaving Bluefield DPUs in their default host-trusted mode, where anyone with root on the host can reach the DPU’s Arm cores over RShim

🟠 Giving any tenant access to login to backend or frontend switches on a shared network

🟠 Incorrect configuration of InfiniBand security keys such as PKey, MKey and SAKey

🟠 Multi-tenant dashboards where infrastructure is shared between internal-facing logs and customer-facing logs

🟠 Controls where provider employees can receive privileged access to tenant logs and monitoring dashboards without tenant permission. (4/5)