Unit 8.1: Architecture Decision Frameworks
Introduction
Modules 2 to 7 each made one decision at a time. A real design makes them all at once, and the decisions interact: a compute class decides which credential covers a tier, a load balancer decides where filtering can live, and a database engine decides what a failover can lose. This unit collects the decisions of the course into five selection matrices that you can reuse on any IONOS CLOUD engagement.
Each matrix has the same columns. Hard constraints remove an option before any comparison, because no amount of preference can make it fit. Differentiating criteria rank what is left. Derived in points to the unit that explains the mechanics. Read the hard constraints first, every time.
1. How to Use the Matrices
Work through a decision in three passes:
- Eligibility. Strike every option whose hard constraints the requirement breaks: a location it is not offered in, a shape that cannot change, a protocol it does not carry.
- Ranking. Among the remaining options, compare the differentiating criteria against the tier's needs: isolation, operating model, cost, failure behaviour.
- Composition check. Test the choice against its neighbours in the layered architecture (Section 8.1), then run the sovereignty filter over the assembled design (Section 7).
Record the result as an architecture decision record: the option, the constraint that removed each rival, and the unit that proves it. FinCorp's records from Modules 2 to 7 are the worked examples behind every row below.
2. Compute
| Option | Differentiating criteria | Hard constraints | Choose when | Derived in |
|---|---|---|---|---|
| Dedicated Core Server | Exclusive CPU use; CPU family selected at provisioning and changeable with a restart; 99.95% uptime SLA | Up to 62 cores and 230 GB RAM; CPU families vary by location | Regulated or latency-bound tiers; software licensed to a CPU family | 4.1 |
| vCPU Server | Shared CPU use at a lower price | Up to 60 vCPUs and 230 GB RAM; CPU family cannot be chosen or changed | Development, test, tolerant internal tools | 4.1 |
| Cubes | Fixed template with a direct-attached NVMe volume; up to 23 add-on Block Storage devices | Template cannot change after provisioning; 99.9% uptime SLA; outside the IT-Grundschutz scope | Short-lived, non-regulated workloads such as build agents | 4.1 |
| Confidential VM | AMD SEV-SNP memory encryption; your own attestation service | de/fra/2 on AMD EPYC Turin only; no live migration; vCPU and RAM fixed at creation; no cloud-init; no startup-volume snapshots |
Data in use must stay opaque to the operator | 1.4, 4.1 |
| Cloud GPU VM | GPU templates for self-hosted models | de/fra/2 and gb/bhx only; zone always Auto; excluded from the SLA |
Fine-tuning, or a model the AI Model Hub does not serve | 6.5 |
| VM Auto Scaling group | Horizontal replicas from one replica configuration | Early Access; horizontal only; minimum one replica | Stateless tiers outside production-critical use | 4.3 |
| Private Cloud | Dedicated VMware: vSphere Enterprise Plus, vSAN, NSX-T | At least three hosts and at most 24 per cluster; offered in fewer locations than VDC compute | VMware-bound workloads that keep their operating model | 4.4, 7.4 |
As shown above, location and immutability remove more options than price does. A Confidential VM requirement fixes the VDC to de/fra/2 before any sizing, and a Cube is eligible only for a tier that will never need a different shape or the IT-Grundschutz scope.
3. Container Platform
| Option | Differentiating criteria | Hard constraints | Choose when | Derived in |
|---|---|---|---|---|
| Managed Kubernetes, public node pools | Control plane free of charge in the chosen region; node pools under the Compute Engine SLA terms; within the IT-Grundschutz scope | A LoadBalancer Service is a static IP on one node; node pool nodes cannot join NSGs; node pool access cannot switch to private | Internet-facing container services behind the ALB | 6.1, 6.2 |
| Managed Kubernetes, private node pools | Nodes with internal addresses only, behind a NAT Gateway the service deploys | No LoadBalancer Services and no static node IPs; node pools in VDCs of the cluster's region; the API stays internet-reachable, so set Allowed IPs | Nodes that must not be reachable from the internet | 6.3 |
| Red Hat OpenShift | The OpenShift operating model | Not a managed service of IONOS CLOUD; control plane and cluster lifecycle managed by Red Hat CCSP partners | An organisation standardised on OpenShift | 6.4 |
| SUSE Rancher Prime | The Rancher operating model | Self-managed; you install, scale and maintain the management server | An organisation standardised on Rancher | 6.4 |
| Private Container Registry | Tokens scoped per repository | DE/FRA only; no public repositories; vulnerability scanning cannot be disabled once enabled | Every image pipeline | 6.4 |
The second axis is isolation inside the platform. Separate clusters when exposure or data class differs, as FinCorp does with fc-k8s-portal and fc-k8s-ai; share a cluster through namespaces and default-deny network policies when teams differ but exposure does not.
4. Storage
| Option | Differentiating criteria | Hard constraints | Choose when | Derived in |
|---|---|---|---|---|
| Block Storage: Essential, Balanced, Performance | Single-VM volumes; Balanced and Performance scale with size | Class and zone fixed at provisioning; Balanced and Performance need at least 100 GB for full performance; Zones 1, 2, 3 and Auto; no cross-region redundancy | Boot and data disks of one server | 4.2, 5.1, 7.3 |
| Network File Storage | Shared NFS access for many clients | One location and one LAN of one VDC; minimum 2 TiB; not offered in GB/BHX | Shared assets of a server pair | 5.1 |
| IONOS Cloud Object Storage | S3-compatible API; Object Lock in Governance or Compliance mode | Object Lock only at bucket creation; bucket names globally unique; cross-regional replication for user-owned buckets only | Audit archive, dumps, datasets, dead-letter tail | 5.2 |
| Backup Service | Acronis Cyber Protect agent backups of VMs, files and applications | Not a backup for managed databases; outside the IT-Grundschutz scope | Self-managed servers and their data | 5.7 |
| Volume snapshot | Fast rollback point | Bound to its location; not transaction-consistent | Before a risky change, deleted afterwards | 4.2, 5.7 |
5. Database Engine
| Option | Differentiating criteria | Hard constraints | Choose when | Derived in |
|---|---|---|---|---|
| DBaaS for PostgreSQL | Asynchronous (default) or Strictly Synchronous replication; managed connection pooler in transaction or session mode; PITR | Strictly Synchronous needs three instances; up to five instances; standbys serve no reads; no superuser access; private LANs only | Accounts and payments, integrity across many entities | 5.3 |
| DBaaS for MariaDB | MariaDB compatibility | Asynchronous replication only, so a failover can lose committed data | Applications built for MariaDB that accept that loss | 5.3 |
| MongoDB | Document model; Business or Enterprise edition | Business: one or three members; Enterprise: up to seven members, sharding and PITR; seven days of backups | Records read and written as a whole, with an evolving shape | 5.4 |
| DBaaS for In-Memory DB | Valkey-based, compatible with the Redis OSS API | Asynchronous replication; up to five nodes; one network connection per cluster | Read scaling and session state | 5.5 |
| Event Streams for Apache Kafka | Partitioned, replicated event log | Three brokers; default replication factor 3; mTLS authentication | Change events, integration feeds, the AI ingestion spine | 5.6 |
| Self-managed engine on Dedicated Core Servers | Full control of engine and version | You own replication, patching and backup through the Backup Service agent | A licensed engine or feature the managed services do not offer | 4.1, 5.7 |
As shown above, the managed engines differ most in what a failover can lose. Read the replication row of a requirement before its feature list: a zero-loss requirement on relational data selects Strictly Synchronous PostgreSQL with three or more instances and removes the alternatives.
6. Networking Primitive
| Primitive | Differentiating criteria | Hard constraints | Choose when | Derived in |
|---|---|---|---|---|
| Managed Application Load Balancer | Layer 7 routing on header, path, query, method, host, cookie or source IP; TLS offloading | No SNAT; a public listener needs a reserved public IP; IPv4 only; NSGs do not attach to it | HTTP and HTTPS edges, host and path routing | 3.3 |
| Managed Network Load Balancer | Layer 4 TCP proxy; Round Robin, Least Connection, Random or Source IP | No SNAT; targets must be provisioned first; rules and health checks are TCP-based only | Non-HTTP or end-to-end encrypted traffic, the private internal tier | 3.4 |
| Managed NAT Gateway | Outbound internet access for private LANs | SNAT only, no DNAT; reserved IPs; up to six private networks; you set the default route in each VM | The egress path of every private tier | 3.6 |
| VPN Gateway | IPSec with IKEv2 and PSK, or WireGuard | Static routing only, no BGP; up to 1 Gbps per tunnel; cannot connect directly to LANs managed by Managed Kubernetes | On-premises links and VDC-to-VDC links across regions | 3.6 |
| Cross Connect | Private VDC-to-VDC link | Same region and same contract; one IP range; a LAN joins a single Cross Connect | VDCs in one location and contract, such as core and AI | 3.6 |
| Cloud Connect | Dedicated Layer 3 line from the company network | VDC in DE/FRA, DE/TXL or US/LAS; a dedicated line to that data center | Sustained bulk transfer during migration | 3.6 |
| IP failover group | Shared reserved IP across your own HA pair | Public LAN without a load balancer; reserved IPv4 only; the HA software runs in your VMs | Vendor appliances using VRRP | 3.5 |
| Cloud DNS | Records your monitor rewrites through the API | TTL floor 60 seconds, default 3600 seconds; steers new connections only | Cross-site steering | 3.7, 7.1 |
7. The Sovereignty Filter, Applied Last
Unit 1.4 fixed the location before the first workload landed. That first pass is necessary but not sufficient, because every later choice was made on functional grounds and can bring a component, a backup location or a credential gap into the design. Run the filter again over the finished design, component by component:
| Filter | Final-pass question | FinCorp finding |
|---|---|---|
| Location | Does every resource, backup location, snapshot target and bucket sit in an allowed location? | The backup location of fc-pg-portal, chosen outside Frankfurt for site loss, is checked against the allowed locations too |
| Service | Does the credential the requirement cites name this service in scope? | Portal tier on Dedicated Core Servers, not Cubes; Backup Service recorded separately |
| Region detail | Is the regional credential of each endpoint final? | Audit archive in eu-central-4, because the IT-Grundschutz certification for eu-central-3 is pending |
| Facility | Are physical controls evidenced by the right data-center credentials? | Berlin facility credentials cited for the Private Cloud estate |
| Trust model | Must the operator be unable to read data in use? | Fraud scoring on Confidential VM in de/fra/2 |
8. The Two Composition Errors
8.1 Pairing Incompatible Primitives
Each primitive is correct alone; the pair breaks a hard constraint of one of them.
| Pairing | Why it fails | Native pattern |
|---|---|---|
| IP failover group on the ALB's public LAN | IP failover works only in public LANs that do not contain load balancers | A separate public LAN for the appliance pair (Unit 3.5) |
| NSG rules for Managed Kubernetes nodes | Node pool nodes cannot be NSG members | In-cluster network policies, filtering at the ALB's targets (Unit 6.1) |
| LoadBalancer Service in a private node pool | Not supported for private node pools | Inbound from a private LAN the node pool joins (Unit 6.3) |
| VPN Gateway on the cluster's own LAN | Not supported for LANs managed by Managed Kubernetes | An additional LAN attached to the node pool (Unit 3.6) |
| Cross Connect from Frankfurt to Berlin | Cross Connect needs the same region | VPN Gateway between the VDCs (Unit 3.6) |
| Strictly Synchronous PostgreSQL with two instances | The mode needs three instances | Three instances, or accept asynchronous loss (Unit 5.3) |
| Targets expected to reach the internet through the ALB or NLB | Neither balancer supports SNAT | NAT Gateway with a default route (Unit 3.6) |
8.2 A Correct Choice Outside the Required Scope
The design works and fails the audit. The fix is a different component, not a different configuration.
| Choice | Scope problem | Correction |
|---|---|---|
| Cubes for a tier that must be in IT-Grundschutz scope | Cubes are outside the certificate | Dedicated Core Servers (Unit 4.1) |
| Backup Service cited as the in-scope backup control | Backup Service is outside the certificate | Record it separately; add locked Object Storage copies (Unit 5.7) |
| C5 cited for a service without checking the report | C5 scope is per service | Confirm each service in the attestation report (Unit 1.4) |
| OpenShift or Rancher control plane evidenced by the provider's certificate | Its operation is your workload, not the managed service | Evidence its operation yourself, or stay on Managed Kubernetes (Unit 6.4) |
Audit archive in eu-central-3 |
Its IT-Grundschutz certification is pending | eu-central-4 (Unit 2.3) |
Decision Summary
| Step | Use | Rule |
|---|---|---|
| 1 | Hard-constraint columns of Sections 2 to 6 | An option that breaks one is out, whatever its other merits |
| 2 | Differentiating-criteria columns | Rank only what survived step 1 |
| 3 | Section 8.1 | Check every pair of neighbours in the layered path |
| 4 | Section 7 | Re-run the sovereignty filter over the whole design |
| 5 | Section 8.2 | Replace any component that fails the credential the requirement cites |
Summary
The frameworks turn the course into five reusable matrices and one final gate. Hard constraints such as location, immutability, protocol support and replication behaviour decide eligibility; differentiating criteria rank what remains. The finished design then passes the sovereignty filter once more, because most compliance gaps enter through choices that were functionally right.
Key Points:
- Read hard constraints before features; location and immutability remove more options than price.
- Zero-loss relational data means Strictly Synchronous PostgreSQL with at least three instances.
- Check each neighbour pair: IP failover and load balancers, NSGs and Kubernetes nodes, Cross Connect and regions.
- Re-run the sovereignty filter last, service by service and region by region.
- A correct choice outside the cited credential is fixed by changing the component.