Unit 8.2: The Reference Enterprise Architecture
Introduction
Unit 1.2 drew an empty frame: a public Layer 7 edge, a stateless compute tier, a private Layer 4 balancer and a private-only data tier, with sites and endpoints attached. Every module since then decided one part of it for FinCorp. This unit puts the decisions on one page and reads them as one system.
Reading the system as a whole changes the questions. A load balancer is no longer a networking topic; it is where TLS ends, where health decides failover inside a site, and where filtering cannot live. A credential is no longer a sovereignty topic; it is a column in the inventory that every component must pass. The capstone in Unit 8.3 builds the Frankfurt core of this design.
1. The Full System
1.1 The Picture
FinCorp's design spans two contracts (Unit 2.1). The production contract holds the regulated estate:
Clients and partners On-premises VMware site
| | IPSec tunnels: VPN Gateway fc-vpn-onprem
Cloud DNS portal (TTL 60) -> Frankfurt ALB, or Berlin ALB after a declared failover
|
FRANKFURT fc-prod-fra-core (de/fra)
[fc-edge] fc-alb-edge HTTPS listener: /api/ -> fc-tg-api, cluster host -> fc-tg-k8s, default -> fc-tg-portal
[fc-app] fc-app-01 (Zone 1), fc-app-02 (Zone 2) fc-k8s-portal: one node pool per zone
NFS share portal-assets fc-nlb-data listener fc-nat-app -> internet egress
[fc-data] payment switch fc-data-01 (Zone 1), fc-data-02 (Zone 2) behind fc-nlb-data
fc-pg-portal (.2) fc-imdb-portal (.3) fc-mongo-cases (.4-.6) fc-kafka-events (.7-.9)
| Cross Connect
FRANKFURT fc-prod-fra-ai (de/fra): fc-k8s-ai (private node pools), fc-pg-rag -> AI Model Hub
FRANKFURT VDC in de/fra/2: fraud scoring on Confidential VM
BERLIN fc-prod-txl-dr: pilot light (stopped portal servers, ALB, reserved IPs) <- VPN Gateway
BERLIN Private Cloud cluster: core banking on vSphere and vSAN, replicated in by VCDA
OBJECT STORAGE eu-central-4: audit archive under Object Lock, dumps, Kafka archive, AI corpus
The non-production contract holds staging, development and early AI experiments, with vCPU Servers, Cubes for build agents, and the staging portal as a VM Auto Scaling group (Units 4.1 and 4.3).
1.2 Where Each Box Was Decided
| Layer or site | Components | Decided in |
|---|---|---|
| Boundaries | Two contracts, VDC naming, region per VDC | 2.1 |
| Identity | fc-prod-platform-ops, fc-prod-deploy with svc-fc-prod-deploy, pre-created federated users |
2.2 |
| Edge | fc-alb-edge on fc-edge, reserved block fc-prod-fra-edge |
3.1, 3.3 |
| Compute | Dedicated Core Servers per tier and zone; Confidential VM in de/fra/2 |
4.1, 4.2 |
| Internal balancing | fc-nlb-data, listener on fc-app, backend on fc-data |
3.4 |
| Data | fc-pg-portal, fc-imdb-portal, fc-mongo-cases, fc-kafka-events, NFS |
5.1 to 5.6 |
| Containers | fc-k8s-portal (public), fc-k8s-ai (private), registry fincorp-prod-images |
6.1 to 6.4 |
| AI | AI Model Hub, RAG on fc-pg-rag, corpus in Object Storage |
6.5, 6.6 |
| Connectivity | VPN Gateway, NAT Gateway, Cross Connect, Cloud Connect for bulk waves | 3.6 |
| Recovery | Pilot light in fc-prod-txl-dr, monitored portal record |
3.7, 7.1 |
| VMware | Private Cloud in Berlin, VCDA waves | 4.4, 7.4 |
2. Cross-Cutting Concerns
A component list hides the concerns that run through it. The following table reads the same design by concern rather than by box.
| Concern | Edge | Compute and containers | Data | Sites |
|---|---|---|---|---|
| High availability | The ALB's own HA; listener on a reserved IP | Members in explicit Zone 1 and Zone 2, volumes in the same zones; one node pool per zone | PostgreSQL standby promotion; three-instance clusters; Kafka replication factor 3 | VPN Gateway Premium with High Availability; vSAN host redundancy |
| Load balancing | ALB: TLS offloading, host and path rules, HTTP health checks | Ingress controller behind the ALB, not a LoadBalancer Service | NLB with TCP health checks in front of the payment switch | Cloud DNS record across sites |
| Security | No NSG on the ALB; filtering on its targets against the backend address | fc-nsg-app and fc-nsg-data, default NSG narrowed; network policies inside clusters; Allowed IPs on each Kubernetes API |
No internet access element on fc-data; private DBaaS endpoints |
Static routes per host; PSK and key rotation in the runbook |
| Connectivity | Reserved public IPv4; IPv4 only for the managed edge | NAT Gateway for outbound; the private cluster's own NAT Gateway | Cross Connect to the AI VDC | VPN to Berlin and on-premises; Cloud Connect for bulk waves |
| Observability | Flow logs on the egress path | Monitoring pipeline and alert for VMs; traces for the portal and the assistant | Central Logging and Central Monitoring for the databases | Collectors in Private Cloud pushing to the Frankfurt pipelines; SIEM fan-in |
| Cost | Lean ALB rule set | Dedicated cores sized for peak in scope; Savings Plans on the floor | Storage sized for I/O, not only for data | Cost alert at 80% of the monthly production budget |
As shown above, no concern lives in one layer. High availability is a zone decision in compute, a replication decision in data and a steering decision across sites, and the weakest of the three sets the real availability. Security is the same: the ALB, the Kubernetes nodes and the managed databases each take their control in a different place, because NSGs attach to VMs and NICs only.
3. The Native-Substitution Set, End to End
Unit 1.3 named six capabilities the platform composes rather than sells. In the finished design each one is a concrete set of components and a concrete piece of work:
| Capability | Realised in FinCorp's design | What the team owns |
|---|---|---|
| API routing | fc-alb-edge path rule for /api/ and host rule to fc-tg-k8s; ingress controller in fc-k8s-portal |
Two rule sets kept consistent; target list checked after node maintenance |
| Read scaling | Cache-aside on fc-imdb-portal; transaction pooler of fc-pg-portal |
Key deletion in the write path; pool and connection sizing |
| Failover | In site: balancer health checks, standby promotion, VRRP with IP failover for the partner gateway. Across sites: the monitor rewrites portal after the Berlin data is restored |
The monitor, the declaration step, the drill record |
| Change-data-capture | The portal's outbox relay publishes to payments.events, keyed by account ID |
Write-and-publish consistency; consumer group offsets |
| VM import | Image conversion and backup-restore waves to VDC compute; VCDA waves into Private Cloud | Guest preparation, wave gates, rollback |
| Access control | IAM groups per team and environment; tokens per pipeline stage; registry tokens per repository | Joiner, mover and leaver runbook; token rotation |
Each row crosses at least two modules. That is why the substitutions are design decisions rather than workarounds: their glue has an owner, a monitor and a test, and Module 7 put each one into the operating model.
4. Service-by-Service Compliance Mapping
4.1 The Two Service Credentials
Two BSI recognitions apply, with different types and scopes (Unit 1.4):
- IT-Grundschutz: ISO 27001 certificate on the basis of IT-Grundschutz, BSI-IGZ-0730-2025, issued 14 September 2025 and valid until 13 September 2028, subject to annual surveillance audits. Its scope is the provision and operation of IONOS Cloud Services in Germany by IONOS SE: the ISMS and infrastructure, Compute Engine, Managed Kubernetes and S3 Object Storage including provisioning, and the IONOS Backbone.
- C5: the C5:2020 Type 1 attestation of 28 May 2026 by PwC, covering 33 services in 10 service clusters. The report is available on request, and scope is confirmed per service against it.
4.2 The Mapping
| Service in FinCorp's design | IT-Grundschutz (BSI-IGZ-0730-2025) | C5:2020 Type 1 | Evidence action |
|---|---|---|---|
| Compute Engine: portal and payment-switch servers | In scope, Germany | Confirm in the report | Cite the certificate with its scope |
Managed Kubernetes: fc-k8s-portal, fc-k8s-ai |
In scope, Germany | Confirm in the report | Cite for the service; evidence images, RBAC, policies and secrets as FinCorp controls |
IONOS Cloud Object Storage in eu-central-4 |
In scope, Germany | Confirm in the report | Keep regulated buckets out of eu-central-3 while its certification is pending |
| AI Model Hub | Not named in the scope | In scope | Cite C5; record FinCorp's AI Act deployer controls separately (Unit 6.6) |
| Backup Service | Outside the scope | Confirm in the report | Record the gap; locked Object Storage copies as the in-scope layer |
| Cubes (non-production only) | Outside the scope | Confirm in the report | Keep out of every in-scope tier |
| DBaaS, Event Streams for Apache Kafka, load balancers, gateways, Cloud DNS, observability services | Not named in the scope | Confirm each in the report | No IT-Grundschutz citation; map each service individually |
| Private Cloud in Berlin | Not named in the scope | Confirm in the report | Cite the Berlin facility credentials (BSI IT-Grundschutz, ISO/IEC 27001, operated by IONOS CLOUD Group) and record the service-level position |
| OpenShift or Rancher Prime (not chosen) | Infrastructure only | Infrastructure only | Their control-plane operation would be FinCorp's own evidence (Unit 6.4) |
As shown above, only three of FinCorp's services carry the IT-Grundschutz certificate, and that is enough for the tiers the risk office placed in its scope: the portal and payment compute, the clusters and the audit archive. The rest of the design is mapped, not assumed. Two rules close the table. Every credential describes how IONOS CLOUD runs a service, never FinCorp's workloads on it. And IONOS CLOUD is the first German cloud provider to hold both the C5 attestation and the IT-Grundschutz certification of the BSI (as of May 2026), which is a provider-level statement and does not put every service under both.
5. FinCorp's Design Review
FinCorp's architecture board reviews the design against the frameworks of Unit 8.1 before the capstone build. Three pairings get explicit sign-off. The partner file gateway stays on its own public LAN, because IP failover works only in public LANs without a load balancer. The AI consumers read Kafka over the Cross Connect, which therefore joins fc-data; the AI VDC's LAN on that Cross Connect shares one IP range with it, and fc-data joins no other Cross Connect. And the evidence pack lists DBaaS and Kafka under C5 only after the risk office has checked each service in the attestation report.
Decision Summary
| Concern | Rule of the reference design | Where it is enforced |
|---|---|---|
| Exposure | Only the ALB listener sits on a public LAN; data LANs have no internet access element | Topology, Unit 3.1 |
| Availability | Explicit zones inside a site; data restored before DNS moves across sites | Units 7.1, 3.7 |
| Filtering | NSGs on VMs and NICs; network policies in clusters; targets filter balancer backends | Units 3.2, 6.1 |
| Egress | Every private tier leaves through a NAT Gateway | Unit 3.6 |
| Substitutions | Each composed capability has an owner, a monitor and a drill | Units 1.3, 7.1, 7.2 |
| Compliance | Map every service to each credential; cite only what the scope names | Units 1.4, 8.1 |
Summary
The reference architecture is the layered shape of Unit 1.2 with every box decided: a Frankfurt core with its edge, compute, internal balancer and private data tier, an AI VDC on a Cross Connect, a Confidential VM tier, a Berlin pilot light, a Private Cloud for core banking, and Object Storage underneath all of it. Read by concern, high availability, security, connectivity, observability and cost each cross every layer. Read by credential, only Compute Engine, Managed Kubernetes and Object Storage carry the IT-Grundschutz certificate, AI Model Hub is in C5 scope, and every other service is mapped individually.
Key Points:
- The design is one layered path plus attached sites; nothing requires a fifth tier.
- The weakest of zone placement, data replication and cross-site steering sets the real availability.
- Each native substitution has a named owner, a monitor and a drill in the operating model.
- IT-Grundschutz BSI-IGZ-0730-2025 covers Compute Engine, Managed Kubernetes and S3 Object Storage in Germany; Backup Service and Cubes are outside it.
- C5:2020 Type 1 (28 May 2026) is confirmed per service against the report; credentials never cover customer workloads.