Skip to content

Requirements

Opsta AI Gateway runs entirely on your own Kubernetes cluster. Nothing is sent to a managed cloud, and the whole platform is reproducible from the Helm chart. This page lists what the cluster needs before you install.

Who this is for

Platform engineers deploying and operating the gateway. If you only want to use the gateway, start with Get access instead.

Kubernetes

  • Any conformant Kubernetes distribution ≥ 1.28. The platform is distro-agnostic — it runs on any CNCF-conformant cluster: on-prem (RKE2/Rancher, OpenShift, VMware Tanzu, k3s), managed (GKE, AKS, EKS, OKE, CCE), or local (k3d/kind). Only the cluster provisioning is platform-specific; the install steps on this page are identical everywhere. (The platform installs Gateway API CRDs and needs a recent control plane.)
  • Gateway API v1.4.x — provides the v1 and v1alpha2 route types the gateway needs. The chart can install these for you, or reuse what's already on the cluster.
  • A working default StorageClass (or set global.storageClass) — PostgreSQL, Redis, and the observability stack all use persistent volumes.
  • A CNI that enforces NetworkPolicy if you want network isolation between components (recommended for production). Lightweight CNIs such as k3d/flannel ignore NetworkPolicy — the platform still runs, but the isolation guards become no-ops.

Compute

The tables below are measured on a real Standalone cluster (k3s v1.36, product v1.11.1). Load model: PoC/small-team — ~5 RPS sustained chat-completions, 1 org, 1 project, ~10 users.

Standalone — measured idle and under load (PoC load model)

ComponentIdle CPUIdle RAMUnder load CPUUnder load RAMPVC
Higress gateway7m1263 Mi9m1277 Mi
Higress controller3m83 Mi3m111 Mi
Control plane1m14 Mi1m16 Mi
Console (Next.js)2m72 Mi8m73 Mi
Console oauth2-proxy1m11 Mi1m11 Mi
PostgreSQL (CNPG, 1 instance)11m73 Mi8m104 Mi5 Gi
Redis (standalone)6m4 Mi6m4 Mi1 Gi
Keycloak3m646 Mi3m647 Mi
Keycloak PostgreSQL8m83 Mi7m115 Mi5 Gi
Mimir (all-in-one)10m160 Mi14m131 Mi5 Gi
Loki (single-binary)8m135 Mi7m137 Mi5 Gi
Tempo (single-binary)5m37 Mi4m37 Mi5 Gi
Grafana11m210 Mi13m213 Mi2 Gi
Alloy (collector)15m129 Mi16m128 Mi
cert-manager (3 pods)4m76 Mi4m76 Mi
CNPG operator5m42 Mi4m43 Mi
Redis operator7m30 Mi5m27 Mi
Qdrant (optional — semantic)1m28 Mi10 Gi
Ollama (optional — semantic/local model)1m138 Mi5 Gi
Total (core, no semantic)~108m (0.1 vCPU)~3.8 GiB~113m~3.9 GiB~28 Gi
Total (with semantic)~110m~4.0 GiB~43 Gi

Node total at idle (both k3s nodes including system pods): 385m CPU / 5.1 GiB RAM on a 2-node cluster (16 vCPU / 30 GiB each).

Standalone node sizing recommendation

For a Standalone / PoC deployment at the load model above:

ResourceMinimumComfortable
vCPU48
RAM8 GiB16 GiB
Disk (PVCs)30 Gi RWO50 Gi RWO

Higress gateway is the RAM anchor

The Higress gateway pod holds ~1.3 GiB RAM at idle due to Envoy's filter + Wasm plugin chains. This is the largest single memory consumer — plan for it when sizing the node.

High-availability sizing (estimated)

HA sizing requires a cluster measurement at production load (RA.1 cluster window — pending). The following is derived from the standalone measurements plus expected replica overhead:

ComponentHA replicasEst. CPUEst. RAMStorage
Higress gateway≥2~20m~2.6 GiB
Control plane2~4m~32 Mi
Console2~20m~150 Mi
PostgreSQL (CNPG, 3 instances)3~30m~300 Mi15 Gi
Redis (Replication + Sentinel)3+3~40m~30 Mi3 Gi
Keycloak2~6m~1.3 GiB
LGTM stack (distributed)varies~200m~2 GiBobject store
Operators + cert-managersame~20m~200 Mi
Total (estimated)~340m~7 GiB≥20 Gi + object store

HA node sizing recommendation (per worker node, ≥3 nodes):

ResourceMinimum per nodeComfortable per node
vCPU48
RAM8 GiB16 GiB
Disk20 Gi RWO50 Gi RWO

Enabling the semantic features (Qdrant + Ollama) adds ~170 Mi RAM and ~15 Gi disk per node hosting those pods. See High availability for replica counts.

Networking & DNS

  • A base domain you control (e.g. ai-gateway.example.com). The platform serves several subdomains under it — api., console., grafana., auth., and (if enabled) mcp.. See TLS & domains.
  • The ability to issue a wildcard TLS certificate for that domain — via Let's Encrypt (DNS-01), your own certificate, or a self-signed issuer for air-gapped sites.
  • Inbound reachability to the gateway. You can expose it through your own ingress/load balancer, or use the built-in Cloudflare Tunnel option for environments without inbound public IPs.

Identity

  • An OIDC identity provider for console and dashboard sign-in. The platform ships Keycloak as an embedded broker, so organizations can connect their own corporate IdP — see SSO & IdP brokering.
  • At least one bootstrap administrator email, set at install time, so the first person can sign in and configure everything else.

Tooling

  • Helm 3 and kubectl, configured against the target cluster.
  • Access to the container images — either pull from the public registry, or mirror them into your own registry for air-gapped installs.

Next steps

  • Install — deploy the chart.
  • Configuration — the one config surface and how it maps to environments.

Enterprise AI governance, on infrastructure you own.