Traffic flow¶
Generative¶
Public ingress (cloud LB or MetalLB on-prem)
→ shaide server
→ per-model Istio Gateway
→ GAIE endpoint picker (InferencePool)
→ ModelService decode pod
The endpoint picker selects a replica per request, accounting for KV-cache locality and load rather than round-robin.
Embedding¶
Embedding traffic bypasses the Gateway and GAIE - the inference gateway understands chat-completion requests only:
shaide server → ms-<slug>-embeddings ClusterIP → ModelService decode pod
Transparent to clients: both paths are reached through the same OpenAI-compatible endpoint.
Ingress by target¶
| Target | Ingress |
|---|---|
| EKS | AWS Load Balancer Controller |
| GKE | Gateway API with managed certificates |
| AKS | Application Gateway for Containers (AGC) |
| RKE2 | MetalLB |
TLS¶
Terminated at ingress. Certificates are issued by cert-manager or supplied by you. See TLS certificates.