Networking
Service exposure, traffic routing, DNS, certificates, network policies, and trust.
Foundation uses several cooperating components to expose services safely and predictably, and to restrict the traffic between them.
Main roles
| Component | Main role |
|---|---|
| Gateway API (Envoy Gateway) | primary ingress controller for all HTTP(S) traffic |
| Internal model gateway | private ClusterIP gateway, kserve-internal, between the BSQAI API and model workloads |
| Infrastructure load balancer | makes the shared gateway reachable from the target network |
| External DNS | publishes DNS records for exposed services when a DNS provider is configured |
| cert-manager | manages certificates and TLS lifecycle |
| trust-manager and Reflector | distribute CA bundles, the wildcard certificate, and other shared secrets across namespaces |
| Kubernetes network policies | admit only the callers each platform workload serves, with optional default-deny for platform namespaces and default-deny for tenant namespaces |
Typical traffic path
A common Foundation request path looks like this:
- a hostname is published in the DNS zone, by External DNS when a DNS provider is configured (Azure DNS, an RFC2136 server or Amazon Route 53), otherwise by the site
- TLS uses the wildcard certificate that cert-manager issues, or custom TLS material from the platform configuration
- traffic reaches the shared gateway through the environment's load-balancer implementation
- Gateway API routes the request based on
HTTPRouterules attached to the sharedGatewayresource - the request is forwarded to the target Foundation or higher-layer service
See DNS and TLS outside Azure for the supported DNS paths.
Gateway API
The platform uses Gateway API (Envoy Gateway) as the primary ingress controller. Each platform component defines its own HTTPRoute resource that registers its endpoints with a shared Gateway.
The routing model has these properties:
- HTTPRoute resources replace
Ingressresources: routing is defined per component, not centrally. - Envoy is the data plane: the
EnvoyProxyconfiguration andGatewayClassare managed by the platform. - Gateway: public traffic enters through one shared
Gatewaywith an HTTPS listener for*.<base-domain>that terminates TLS with the wildcard certificate. Its HTTP listener serves only a redirect: every plain-HTTP request receives301to HTTPS. - Security policies: Envoy Gateway security policies validate a presented Keycloak token on the BSQAI API route and run the OIDC sign-in for some operator UIs. Services still authorize each request.
- Headers: the gateway drops request headers whose name contains an underscore, so they can neither break nor impersonate platform headers.
- Source ranges:
GATEWAY_API_GATEWAY_LOAD_BALANCER_SOURCE_RANGESlimits which source CIDRs reach the public load balancer. Empty leaves it open.
Some routes are created at run time rather than rendered. Where the site enables app previews of the agent sandbox, the BSQAI API creates one HTTPRoute per chat preview on <chat-id>.sandbox.<base-domain>, routed through the preview proxy that checks the preview session. The wildcard certificate then also covers *.sandbox.<base-domain>.
Internal model gateway
Model workloads have no public route. A second Gateway, kserve-internal, uses the same GatewayClass with a ClusterIP Service. It admits routes only from namespaces labeled for model serving: kserve for the Global Organization, and each tenant namespace once inference is enabled for the tenant. The BSQAI API sends every call for an in-cluster model through it. See Inference for the path allow-list and the callers it admits.
Network policies
BullSequana AI renders an allow policy set for its own components by default. Each selected workload accepts traffic only from the callers it serves, such as the gateway or Prometheus, and reaches only the services it uses, cluster DNS, and the Kubernetes API. Argo CD gets its own set: it reaches cluster DNS, the Kubernetes API, and the Git, registry, and proxy ports, and argocd-server accepts traffic only from the gateway.
Two default-deny layers close what the allow policies do not open:
- Platform namespaces: with
default_deny = true, every platform namespace that is not excluded receives adefault-deny-ingresspolicy. It is off by default, and a site turns it on in stages. - Tenant namespaces: the tenant operator gives each tenant namespace its own default-deny policy. It is on by default.
The internal model gateway admits only the BSQAI API workloads that call models. Tenant databases, private MLflow, and Managed MCP Servers each admit only their named callers.
Policies take effect only on a cluster whose network plugin enforces Kubernetes network policies. Calico is an optional bootstrap install for clusters without one. See Network policies for the settings and the staged rollout, and bsqai test-npe to check enforcement with real traffic.
Why this matters
The routing model defines how users, systems, and services reach the platform. It is also where network exposure, DNS, TLS, access control, and network isolation come together.