NVIDIA's Central Identity Gateway Cuts Login Events by 55% Across Federated Kubernetes

NVIDIA published a pattern for carrying user identity across federated Kubernetes and AI platforms, separating session ownership from request enforcement. A central identity gateway validates sessions via a shared store, while regional gateways stay stateless and enforce access. The company reports that this approach cut repeated login events by 55% across its internal developer platforms running in AWS and OCI.

The pattern relies on standard OpenID Connect and a minimal /gateway/userinfo validation endpoint. Regional gateways remain stateless, with session data stored in a Redis-backed store that has explicit time-to-live (TTL) values. Each session is keyed by an opaque session ID and associated with an HTTP-only cookie scoped to the platform domain. On every request, a regional gateway calls /gateway/userinfo, receives claims such as user ID, email, groups, roles, and session metadata, then injects standardized identity headers downstream. Applications no longer parse tokens or perform direct OIDC exchanges; they consume identity from the gateway's header contract.

Centralized session ownership

Centralized session ownership handles token refresh and platform-wide logout, keeping refresh cycles coordinated. When an access token nears expiry, the central gateway refreshes it using the stored refresh token and updates the Redis session record. Logout deletes the session record, so subsequent regional gateway calls see an invalid session and deny access. This design means upstream identity-provider load scales with active users instead of the product of users, tools, and clusters. Compared to distributed session ownership, centralized ownership provides single-login, platform-wide logout, coordinated refresh, lower IdP load, and standardized downstream headers.

Security guardrails include mutual TLS or workload identity between gateways, stripping inbound identity headers before injecting trusted ones, and defined failure behavior when the session store is unavailable. The approach applies to federated Kubernetes environments across multiple clouds, on-premises, and specialized execution planes.

Next steps for implementers

NVIDIA suggests experimenting with OAuth2 Proxy locally, following the Istio external authorization sample, adding Rego policy via OPA Envoy Istio, and reviewing Authorino for JWT/API-key validation and trusted upstream headers. For inference engineers serving models across federated clusters, this pattern reduces the overhead of repeated authentication and standardizes how identity reaches model-serving workloads, without requiring changes to the applications themselves.