Skip to content

Stage 18: F5 XC Site Networking on GCP

Key takeaways: A site is the unit that connects your infrastructure to XC. On GCP, the standard shape is a two-interface CE: SLO on an outside VPC for control plane and edge tunnels, SLI on an inside VPC for origins. GCP's one-NIC-per-VPC rule is why the design uses two VPCs instead of two subnets.

What a "site" is

In XC, a site is a logical location — a name that groups one or three CE nodes, their interfaces, and their site-local networks. Origin pools, LBs, and discovery all reference the site, not the VM — so the CE can be replaced or scaled without touching application config.

Site types

Type NICs Use when
Ingress gateway 1 Only inbound delivery is needed and inside/outside share one subnet (small sites, App Stack)
Ingress/egress gateway 2 The classic dual-homed design — inbound delivery and outbound controls (this track's pattern)
Secure mesh node 1–2 Site-to-site connectivity between your sites over the XC backbone, no client-facing LB
App Stack (voltstack) cluster 1+ CE also runs a managed K8s and deploys workloads directly

The two-interface ingress/egress mode is what gives the SLO/SLI split its meaning: SLO is a zone, not just an interface — it is where control-plane and RE-facing traffic is allowed to exist.

SLO vs SLI in GCP terms

SLO — Site Local Outside SLI — Site Local Inside
Purpose Control plane, RE tunnels, internet egress Reach origins and site-local services
GCP VPC xc-outside-vpc (slo-subnet) xc-inside-vpc (sli-subnet)
External IP Yes — ephemeral is fine Never
Firewall posture Outbound-friendly; inbound managed by CE itself Tight: only what origins need
Lab addressing 10.190.0.0/24 10.200.0.0/24

SLO/SLI two-VPC Customer Edge design on GCP

Why two VPCs — the GCP constraint

GCP assigns at most one network interface per VPC to a VM. A dual-homed appliance therefore cannot be "two subnets in one VPC"; it needs two VPCs. This has two useful consequences:

  1. The inside VPC stays a clean security boundary — no route, no external IP, no internet-facing listener exists on the workload side by default.
  2. GCP firewall rules can treat the two zones differently: the inside rule in the labs allows tcp:80 from 10.200.0.0/24 to tag xc-origin, which covers the CE's SLI address and nothing else.

Compare with earlier stages

The CE is structurally the dual-homed router of Stage 4 — it forwards across a boundary that has no other path — except the "routing" is policy-driven tunneling to REs rather than destination-prefix forwarding, and the management plane lives in SaaS.

The registration trust model

A CE is not trusted because it has an IP — it is trusted through a staged handshake:

  1. Site token — generated in the Console (or as a Terraform object), delivered to the VM through instance metadata.
  2. Registration request — the CE presents token + hardware identity over SLO to the control plane.
  3. Approval — an admin accepts the pending registration. A stolen token alone cannot silently join a site.
  4. Issued identity — the node gets a certificate and becomes a known, revocable endpoint; all later tunnels use node certs, not the token.

Deployment paths — three ways to the same node

Path How the CE VM is created When to use
Console manual (Lab 02) You launch the F5 marketplace image with 2 NICs + token metadata Learning, demos, one-offs
XC-orchestrated Terraform (Lab 04) volterra_gcp_vpc_site + volterra_tf_params_action — XC runs its own apply using your volterra_cloud_credentials GCP service account Repeatable, auditable infra
Custom automation BYO Compute Engine with the certified image + registration metadata Integrating with existing instance pipelines

Notice that in the Terraform path, your google provider never creates the CE VM — GCP shows the instance, but the resource lives in XC's plane. That distinction decides where you look when the node fails to come up.

IAM and credentials map

Credential Held by Grants
.p12 API certificate You / CI Manage XC objects for the tenant
GCP SA key in volterra_cloud_credentials XC control plane Create/modify Compute Engine VMs in your project
CE's own GCP service account The CE VM GCP API identity + firewall matching by SA
Site token CE VM at boot One-time registration into a named site

Failure modes worth memorizing

Symptom Most likely layer
Node never appears in Registrations SLO egress blocked, bad metadata/token
Registration pending forever Human approval step — by design
Site UP but pools unreachable Origin pool pointed at outside network, or inside VPC firewall missing
CE can't reach origins SLI NIC on wrong VPC, or firewall tag mismatch
CE created but site shows errors XC's SA lacks Compute permissions in your project

Map it to the labs

Lab 01 builds this exact topology by hand; Lab 02 walks the registration path in the Console; Lab 04 turns the whole site into volterra_gcp_vpc_site code.