Unit 6.4: Container Registry and Platform Selection
Introduction
Every container platform needs a trusted source of images, and every image source is an access-control problem. On IONOS CLOUD that problem is solved with tokens, not with users and roles inside the registry. How you cut those tokens decides how far a leaked pipeline secret reaches.
The second decision is the platform itself. Managed Kubernetes is one option; OpenShift and Rancher Prime run on IONOS CLOUD infrastructure as well, but someone other than IONOS CLOUD runs their control planes. The unit ends by creating FinCorp's registry and its first tokens in the Data Center Designer (DCD).
1. Registry Access Is Token Discipline
1.1 What the Registry Enforces
The Private Container Registry is reachable from the public internet, but authentication is always required. It cannot be configured for unauthenticated access, and repositories cannot be made public. Every client, whether a developer, a build server or a cluster node, connects with an authentication token that carries the right permissions.
Inside the IONOS CLOUD account, managing registries is a privilege: contract administrators, owners and users with Access and manage Private Container Registry can create them, and the same credentials manage every registry and repository in the account. Access to images is then defined per token. A token's scope has three parts:
| Scope field | Values | Design meaning |
|---|---|---|
| Type | Repository (manage the contents of a repository) or Registry (list the repositories) | Most tokens are Repository tokens |
| Path | Repository names, or * for all repositories |
Name the repositories; * is the widest blast radius |
| Actions | Admin (delete artifacts), Push (requires Pull), Pull | Push implies Pull; Admin is for cleanup only |
As shown above, least privilege is written into the token, not into a role. You set one scope when you create a token and can add or remove scopes later.
1.2 Lifecycle Rules That Shape the Design
A token's name cannot be changed after creation. Its state can be set to enabled or disabled, which is your kill switch for an incident. Its expiry date must be at least one hour in the future, or you choose no expiry. When the expiry date is reached, the token is deleted, not disabled, so an expired pipeline credential cannot be revived and must be reissued. The token password is shown only once, at creation.
These rules give the governance pattern:
- One token per pipeline stage and per consumer. The build stage gets Push and Pull on its own repositories; each cluster gets Pull only.
- Named repositories, not
*. A token for the portal pipeline lists the portal repositories. - An expiry on every automation token, with rotation before the date, because expiry deletes the token.
- Disable first, then delete. On suspicion, disable the token; delete it once the incident is closed.
- No Admin in pipelines. Keep the delete action for a separate cleanup token.
1.3 Vulnerability Scanning Is a One-Way Toggle
Vulnerability scanning is a paid add-on. You can turn it on when you create a registry or later, but you cannot turn it off once it is enabled. Images are scanned for Common Vulnerabilities and Exposures (CVEs) when they are pushed and again when CVE databases are updated; images pushed or pulled in the past 30 days are rescanned as new information becomes available. Scanning is non-blocking, so an image can be pulled while it is scanned.
For a regulated estate the decision is simple: enable it at creation, because the report is evidence. Budget for it as a permanent cost, since the toggle only goes one way. Gating deployments on scan results is a pipeline task (Developer course).
1.4 Other Facts That Constrain the Design
The registry is available in Frankfurt (de/fra) only, and its name and location cannot be changed after creation. Registry names must be globally unique across all customers. Data is encrypted at rest with platform-managed keys; you cannot provide your own. Network traffic is not charged. Each client can hold up to 100 concurrent connections and send up to 60 requests per second, beyond which requests receive HTTP 429, so a large node fleet pulling at once needs staggered rollouts. Garbage collection is turned off by default and runs on a weekly schedule once enabled, and the registry is in read-only mode while it runs.
2. Platform Selection
2.1 Who Runs the Control Plane
| Platform | Control plane run by | Licensing | Notes |
|---|---|---|---|
| Managed Kubernetes | IONOS CLOUD; free of charge | Included | Within the IT-Grundschutz certificate scope (Unit 1.4) |
| Red Hat OpenShift | Red Hat Certified Cloud Service Provider (CCSP) partners | Subscriptions from Red Hat or a subscription distributor; IONOS CLOUD does not sell them | Validated version 4.20; Red Hat validation, scheduled through an IONOS CLOUD representative, before production |
| SUSE Rancher Prime | You, self-managed | Bring Your Own Subscription (BYOS) through SUSE or a distributor | You install, scale and maintain the management server |
As shown above, only Managed Kubernetes is a managed service of IONOS CLOUD. IONOS CLOUD does not provide a managed OpenShift, and it does not offer Rancher Prime as a managed service. For both, infrastructure is billed at the regular IONOS CLOUD rates, with no additional handling fee for OpenShift and no hidden integration fees for Rancher Prime.
2.2 Who Carries the Evidence
The compliance consequence follows the control plane. With Managed Kubernetes, the provider's certificate covers the service, and you evidence your workloads (Unit 6.1, Section 3.4). With OpenShift or Rancher Prime, the control plane is a workload on Compute Engine servers. The infrastructure can sit within a credential's scope, but the operation of the control plane, its patching, its access control and its availability, is evidence that the partner or your own team must produce. Choose a customer-deployed platform when you need that platform's own tooling or an existing operating model built on it, and budget for the team that runs it.
2.3 Separate Clusters or Shared Namespaces
Isolation has two levels. Namespaces share one control plane, one CNI and one set of node pools; separation relies on Kubernetes RBAC, network policies and resource quotas, all of which you configure. Separate clusters add a boundary that IONOS CLOUD enforces: their own API endpoint with its own Allowed IPs, their own IAM grants and their own node pools in their own VDC.
Because the control plane is free, a second cluster costs only its nodes, including the warm node of each autoscaled pool. Use separate clusters for different trust zones, regulatory scopes or administrator groups, and namespaces for teams within one zone.
FinCorp applies both levels. The public fc-k8s-portal and the private fc-k8s-ai are separate clusters because their exposure and data differ. Within fc-k8s-portal, the portal and partner API teams share the cluster in separate namespaces with default-deny network policies. FinCorp stays on Managed Kubernetes, so that its evidence pack can cite the provider's certificate for the control plane.
DCD Implementation Walkthrough
You will create FinCorp's registry with scanning enabled, issue a push token for the portal pipeline and a pull token for fc-k8s-portal, and log a client in. This realises the token discipline of Section 1.2.
Prerequisites: an account with the Access and manage Private Container Registry permission, or the Administrator role; a secrets store for the token passwords.
Build goal: Create a registry, issue a scoped token, authenticate a client.
Steps (in the Data Center Designer):
- Go to Menu > Containers > Container Registry and select Add a Registry.
- Enter a globally unique Name, for example
fincorp-prod-images: 3 to 63 characters, alphanumeric characters and dashes, starting with a letter a-z and ending with an alphanumeric character. Choose the Frankfurt Location. Neither can be changed later. - Turn on Vulnerability Scanning. It is a paid add-on and cannot be disabled once enabled.
- Select Add Registry. The registry is ready when its status is Running.
- Select the registry, open the Tokens tab and select Add Token. Enter the Name
fc-ci-portal-push, turn on State, and set Expire on to the end of FinCorp's rotation period. - Under Scope, select Type Repository, enter the portal repository names in Path, and select Push and Pull. Select Add Token.
- Copy the password and the Docker Login command shown now into the secrets store. The password cannot be recovered.
- Repeat steps 5 to 7 for
fc-k8s-portal-pullwith Pull only. Add scopes later through OPTIONS > Manage Scope. - In the Properties tab, expand Garbage Collection, select a Day and Time (UTC) outside the deployment window, and select Update Schedule.
On the build server, the client authenticates with the hostname, username and password from step 7. Reading the password from standard input keeps it out of the shell history:
echo "$FC_CI_PORTAL_PUSH_PASSWORD" | docker login <registry-hostname> \
--username <token-username> --password-stdin
Common mistakes:
- Issuing one
*token with Push for every pipeline and cluster. A leak then reaches every repository. - Relying on expiry as a pause. An expired token is deleted; use State to disable a token you may need again.
- Losing the password at creation. It is shown once; reissue the token instead of searching for it.
- Enabling scanning on a test registry "to try it". It cannot be turned off, and it is billed.
- Choosing a generic registry name. Names are globally unique and fixed; prefix them with your organisation.
- Rolling out hundreds of nodes that pull at the same moment. Per-client rate limits return HTTP 429; stagger the rollout.
Summary
Registry security on IONOS CLOUD is token design: narrow repository paths, the smallest set of actions, an expiry on every automation token, and disabling as the incident switch. Scanning is a one-way, billed decision worth making at creation. Platform choice follows who runs the control plane: IONOS CLOUD for Managed Kubernetes, a CCSP partner for OpenShift, your own team for Rancher Prime, and the compliance evidence goes with it.
Key Points:
- Authentication is always required, and repositories cannot be public.
- Scope each token by type, repository path and actions; Push requires Pull.
- An expired token is deleted, not disabled; a token can be disabled manually.
- Vulnerability scanning is a paid add-on that cannot be disabled once enabled.
- Only Managed Kubernetes is managed by IONOS CLOUD; OpenShift is partner-run and Rancher Prime is self-managed.
- Use separate clusters for trust zones and namespaces for teams within a zone.