Service, Meet Network

The situation

Support has two complaints about Code Nexus Assist. First: a creator swears the assistant quoted last quarter's payout policy, and Leo is already worried the model is confused. Second: screenshots creators attach to support tickets sometimes vanish — there one minute, gone the next. Priya and Lydia have a rule before anyone touches the model: check the network and the storage first, because both symptoms look exactly like what a flaky AI would do, and neither one is that.

What you'll learn

• Kubernetes networking and storage: Service selectors, DNS-style addresses, PersistentVolumeClaims and access modes; test the specific ingress rules in the fixture without claiming complete tenant or geographical isolation.
• ClusterIP, NodePort, LoadBalancer and ExternalName Services; Ingress and the Gateway API; CoreDNS names; CNI plugins; service meshes with sidecars, mTLS, Istio and Linkerd; PersistentVolume versus PersistentVolumeClaim, StorageClass, access modes and CSI.

Who you work with

• 

  Priya Naidoo

  Cloud Architect

• 

  Lydia Roe

  Security Engineer

• 8-BIT

  Code Nexus Internal AI Assistant

Scenes

1. 1.It's probably not the model
2. 2.Routing, storage, access
3. 3.How does traffic get in?
4. 4.The assistant with two colleagues
5. 5.Tighten the selector
6. 6.Where the screenshot actually lives
7. 7.Two regions, and a test that tries to fail
8. 8.Write it up
9. 9.Networking and storage check

What you leave with

The fixture's labels, backend counts and requests identified a legacy build still selected by the Service. Ephemeral uploads required the modelled shared claim. Repairs were repeated in emea, then one cross-namespace request was blocked while an intended local request succeeded. Those checks demonstrate the tested ingress behaviour, not complete tenant data isolation. PVC sharing needs supported access modes and a suitable driver; persistence needs backup, and namespace names do not establish residency.