FIELD NOTE / INSTAGRAM
A namespace is not a tenant security boundary by itself.
The short film, the complete written thought, and the evidence behind it.
A namespace is not a tenant security boundary by itself.
Video caption
A namespace is not a tenant security boundary by itself.
The CNI may not enforce the NetworkPolicy object.
Use disposable identities and harmless denial cases.
Name the enforced boundary and its limits.
#EricFieldNotes
Full written post / accessibility read
Kubernetes namespaces organize API resources. They do not automatically give a tenant least-privilege RBAC, a working deny-by-default network policy or compute isolation. A time-sliced GPU also lacks MIG's memory and fault isolation.
Picture two customer namespaces with a deny policy. If the installed network plugin does not implement it, packets may still flow. A dashboard that lists the policy is evidence of configuration, not of a blocked cross-tenant path.
From Tenant A, attempt a forbidden API read and an ordinary network connection to Tenant B in a test cluster. Verify denial outside the initiating session. Run shared-GPU interference tests for the offered mechanism. Keep the namespace, service account, CNI and device-sharing mode with the result.
Use separate nodes or stronger virtualization when the customer's risk requires it; use shared modes only with explicit terms and tests. Do this because a YAML namespace and a GPU replica are labels unless the data and control paths reject the forbidden action.
#EricFieldNotes
Four-beat scene transcript
1. A namespace is not a tenant security boundary by itself.
Kubernetes namespaces organize API resources. They do not automatically give a tenant least-privilege RBAC, a working deny-by-default network policy or compute isolation. A time-sliced GPU also lacks MIG's memory and fault isolation.
Visual: Sharing GPUs also shares a cluster and its mistakes.
2. A policy file can be present and ineffective.
Picture two customer namespaces with a deny policy. If the installed network plugin does not implement it, packets may still flow. A dashboard that lists the policy is evidence of configuration, not of a blocked cross-tenant path.
Visual: The CNI may not enforce the NetworkPolicy object.
3. Probe each tenant boundary separately.
From Tenant A, attempt a forbidden API read and an ordinary network connection to Tenant B in a test cluster. Verify denial outside the initiating session. Run shared-GPU interference tests for the offered mechanism. Keep the namespace, service account, CNI and device-sharing mode with the result.
Visual: Use disposable identities and harmless denial cases.
4. Sell an isolation class you can demonstrate.
Use separate nodes or stronger virtualization when the customer's risk requires it; use shared modes only with explicit terms and tests. Do this because a YAML namespace and a GPU replica are labels unless the data and control paths reject the forbidden action.
Visual: Name the enforced boundary and its limits.
Research and claim limits
- Kubernetes multi-tenancy guidance (S190)
- NVIDIA GPU time-slicing (S184)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.