GeassCluster

A GeassCluster stands for the K3s cluster that Geass manages. It holds the switches for the platform add-ons (cert-manager, monitoring, Longhorn storage, log collection and sleep-when-idle), the network mode, the default region for projects, and what the operator needs to join new workers: the cluster’s API address and the Secret with the join token. The installer creates one, named default, in geass-system. You change it with kubectl. The dashboard writes only one field of it, the default region.

For the task-oriented view, see How Geass works, Add servers and Storage.

Summary

Property	Value
API version	geass.geass.dev/v1alpha1
Kind	GeassCluster
Plural	geassclusters
Scope	Namespaced. The installer creates it in geass-system.
Subresources	status
Printer column Ready	.status.conditions[?(@.type=="Ready")].status
Printer column Phase	.status.phase
Finalizer	None

Geass expects one cluster. When a project is created and more than one GeassCluster exists, the dashboard uses the one named in GeassPlatformConfig spec.defaultClusterRef, or the first one.

Examples

What the installer applies

default-cluster.yaml

apiVersion: geass.geass.dev/v1alpha1
kind: GeassCluster
metadata:
  name: default
  namespace: geass-system
spec:
  version: v1
  serverURL: https://10.0.0.5:6443
  tokenSecretRef:
    name: geass-token
    namespace: geass-system
  addons:
    certManager:
      acmeEmail: [email protected]
    longhorn:
      enabled: true
    logging:
      enabled: true

The installer writes the certManager block only when GEASS_ACME_EMAIL is set. Let’s Encrypt refuses placeholder domains such as example.com, so use a real address. The installer leaves the other add-ons at their defaults, so cert-manager and monitoring are on.

A multi-site cluster with the full monitoring profile

multi-site-cluster.yaml

apiVersion: geass.geass.dev/v1alpha1
kind: GeassCluster
metadata:
  name: default
  namespace: geass-system
spec:
  version: v1
  serverURL: https://203.0.113.10:6443
  tokenSecretRef:
    name: geass-token
    namespace: geass-system
  network:
    multiSite: true
  defaultRegion: eu
  addons:
    certManager:
      enabled: true
      acmeEmail: [email protected]
    monitoring:
      enabled: true
      profile: full
    longhorn:
      enabled: true
    logging:
      enabled: true
      retentionDays: 14
    serverless:
      enabled: true

Spec fields

Field	Type	Required or default	Validation	Meaning
spec.version	string	Required	None	A label for the Geass cluster. The installer writes v1. It is not a K3s version and Geass does not use it to install anything. The one exception: when a worker joins and no control plane node reports a K3s version, Geass uses this value, but only if it looks like a K3s release such as v1.31.4+k3s1.
spec.serverURL	string	Required	None	The address workers join, for example https://203.0.113.10:6443. The installer writes the control plane’s own address on its default route (https://10.0.0.5:6443, a private address on most cloud servers), and a multi-site install writes the public address instead. A loopback address (127.0.0.1, localhost, ::1) is never used to join a worker: Geass uses the address the control plane node registered under. Set serverURL only to use a different address: see Add servers.
spec.tokenSecretRef	SecretReference	Required	None	The Secret that holds the K3s join token. A standard Kubernetes name and namespace pair. The installer writes geass-token in geass-system. The token is read from the key token, then value, then k3s-token.
spec.tokenSecretRef.name	string	Optional	None	Name of the Secret.
spec.tokenSecretRef.namespace	string	Optional. Defaults to the cluster’s namespace.	None	Namespace of the Secret.
spec.addons.certManager.enabled	boolean	CRD default true	None	Installs cert-manager. Turning it off removes the HelmChart.
spec.addons.certManager.acmeEmail	string	Optional. Empty by default.	Empty, or an address that matches ^[^@\s]+@[^@\s]+\.[^@\s]+$. At most 254 characters	The contact address of the Let’s Encrypt account, which also receives expiry notices. When it is set (and cert-manager is on), the operator creates and keeps the letsencrypt-prod ClusterIssuer that every ingress asks for. When it is empty, no issuer is created and no HTTPS certificate can be issued, which the ClusterIssuerReady condition reports. The installer sets it from GEASS_ACME_EMAIL. Let’s Encrypt refuses placeholder domains such as example.com.
spec.addons.monitoring.enabled	boolean	CRD default true	None	Installs kube-prometheus-stack. Turning it off removes the HelmChart.
spec.addons.monitoring.profile	string	CRD default lite	Enum lite, full	lite turns off Grafana, Alertmanager, the default rules and dashboards. full leaves those at the chart’s defaults.
spec.addons.longhorn.enabled	boolean	CRD default false	None	Installs Longhorn and the storage classes Geass uses. It needs iSCSI on every node. The installer sets it to true. Turning it off never uninstalls Longhorn.
spec.addons.logging.enabled	boolean	CRD default false	None	Installs log collection: a Fluent Bit pod per node and a VictoriaLogs server. The installer sets it to true. Turning it off removes both HelmCharts, and the logs they kept with them.
spec.addons.logging.retentionDays	integer (int32)	CRD default 7	Minimum 1, maximum 90	How long stored logs are kept.
spec.addons.serverless.enabled	boolean	CRD default false	None	Installs KEDA and its HTTP add-on, so that services can sleep when idle. It is opt-in: the installer does not set it. It also lets the K3s Traefik route to a Service that is only a name (a HelmChartConfig named traefik, created only when none exists). Turning it off stops the operator managing it and never uninstalls KEDA.
spec.network.multiSite	boolean	false	None	The nodes are at different providers and reach each other over the internet. See Network mode.
spec.defaultRegion	string	Optional	Pattern ^[a-z0-9]([-a-z0-9]*[a-z0-9])?$, max length 40	The region of projects that do not choose one. Empty means such projects may run on any node.

Note

The Go types and the generated CRD agree for every field on this page.

Network mode

With spec.network.multiSite: true the overlay network between nodes is encrypted (the K3s WireGuard backend) and every node advertises its public address. Set it when you install, with the installer variables GEASS_MULTI_SITE and GEASS_NODE_EXTERNAL_IP. Changing it later means restarting every node, and Geass does not do that. The operator reads the field only in two places: when it joins a worker (the worker then installs WireGuard support, opens its ports if it can, and advertises its public address) and when it renders the monitoring chart (Prometheus then scrapes kubelets at their external addresses). Do not flip it on a running cluster.

Status

Field	Type	Meaning
status.phase	string	Ready when the cluster is ready. Empty otherwise.
status.reconcileRetry	object	Failure counters (reconcileFailures, reconcileFailureStarted, reconcileFailureGeneration).
status.conditions	list	Ready, AddonsReady and, while cert-manager is on, ClusterIssuerReady.

Condition	Reason	Status	Meaning
AddonsReady	AddonsReady	True	Every enabled add-on that gates readiness is installed and running. The message is AddonsReady is ready.
AddonsReady	AddonsPending	False	At least one add-on is still installing. The message is AddonsReady is not ready.
AddonsReady	ReconcileError	False	Geass could not apply an add-on. The message is the error.
Ready	ClusterReady	True	The message is Geass cluster is ready. status.phase is Ready.
Ready	ClusterPending	False	The add-ons are not ready. The message is Geass cluster is not ready. Geass checks again every 30 seconds.
Ready	ReconcileError	False	An error while applying add-ons. The message is the error.
Ready	ProvisionFailed	False	Geass gave up after repeated errors. The message is Provisioning failed after repeated errors. Update settings or platform configuration, then save or redeploy to try again. The controller then stops until you change the spec.

ClusterIssuerReady says whether HTTPS certificates can be issued. It is informational: it never changes Ready, and a cluster without an issuer still runs workloads. The operator removes the condition while spec.addons.certManager.enabled is false.

Status	Reason	Meaning
True	IssuerReady	The letsencrypt-prod ClusterIssuer exists and cert-manager reports it ready. The message is The letsencrypt-prod ClusterIssuer is ready to issue certificates.
False	WaitingForCertManager	cert-manager is not ready yet, or its kinds are not served yet. Checked again after 30 seconds.
False	ACMEEmailMissing	No issuer exists and spec.addons.certManager.acmeEmail is empty. The message names the field and the installer variable.
False	IssuerRegistering	The issuer exists and cert-manager has not yet reported on its ACME account.
False	IssuerNotReady	cert-manager reports the issuer is not ready. The message carries cert-manager’s reason, for example Let’s Encrypt refusing the email.
False	IssuerApplyFailed	The operator could not create or update the issuer. It retries, and its log says why.
False	IssuerAccessDenied	The operator’s ClusterRole may not manage clusterissuers.cert-manager.io.
True	ExternallyManaged	An issuer named letsencrypt-prod exists that Geass did not create (it lacks the label geass.dev/managed-by=geass), and cert-manager reports it ready. Geass leaves it as it is. If cert-manager does not report it ready, the reason is IssuerRegistering or IssuerNotReady instead.
True	ManagedByAnotherCluster	The issuer belongs to another GeassCluster that still exists (label geass.dev/cluster), and cert-manager reports it ready. Geass leaves it as it is.

The retry rules are the same as for GeassDatabase. A cluster has no Retry button: to resume after ProvisionFailed, change something in the spec.

What gates Ready:

• cert-manager and monitoring, when enabled: the HelmChart’s install Job must have succeeded and a pod of the release must be ready.
• Longhorn, when enabled: its HelmChart must be installed, the CSI driver driver.longhorn.io must be registered, and every schedulable node must have a ready longhorn-manager pod. Only then does Geass write the storage classes.
• Logging never gates readiness. It is installed on the side, and a problem is only logged.
• The serverless add-on never gates readiness either. It comes up on the side, and the operator looks again every 20 seconds until it has.

Caution

With Longhorn on, a node that has joined but has no ready Longhorn manager yet keeps AddonsReady and Ready at False until the manager starts. While that lasts, resources that reconcile report project "<project>" cluster "<cluster>" is not ready. This clears by itself, but it can take a few minutes while a new node pulls images. Once any node is a compliance node and every other node is tagged for general volumes, only those other nodes need a ready manager for this check. This comes from reading the controller, not from timing it on a real cluster.

What the dashboard writes

Where	What it does
Settings > Usage > Regions > Default region	Writes spec.defaultRegion of the first cluster. The choices are the regions your nodes carry, and No default (any node). It is only shown once at least one node belongs to a region.
Settings > Usage > Nodes	Shows each node’s role, region, tags, CPU, memory, pressure and status. Edit changes the node: see GeassNode. It does not write the GeassCluster.
Settings > General	Shows the number of clusters.

Everything else is kubectl. The dashboard does not toggle add-ons or edit serverURL.

Editing by hand

It is safe to edit the add-on toggles, defaultRegion and serverURL with kubectl. The operator applies the change on its next pass.

Set the address workers join

kubectl -n geass-system patch geasscluster default --type merge \
  -p '{"spec":{"serverURL":"https://203.0.113.10:6443"}}'

• Do not delete the cluster. It has no finalizer and no owner references, so deleting it leaves the add-on HelmCharts and everything else running. Projects stop validating, because they need their cluster to exist and be Ready.
• Do not change spec.network.multiSite on a running cluster.
• Turning an add-on off removes its HelmChart, which uninstalls the release. For logging that deletes the stored logs. For Longhorn and the serverless add-on it does nothing: take them out by hand if you really want them gone. For Longhorn, move every volume first. For serverless, switch each service’s own sleep setting off first: the operator decides whether a service can sleep by whether KEDA’s HTTPScaledObject kind exists in the cluster, so services keep sleeping after the add-on is switched off.
• The controller does not rewrite the GeassCluster spec. It writes status.

What happens under the hood

Each add-on is a HelmChart object in kube-system. K3s’s Helm controller runs a Job helm-install-<name> for it and installs the release.

Add-on	HelmChart	Namespace	Notes
cert-manager	geass-cert-manager	cert-manager	Chart v1.17.2 with installCRDs: true. The operator then creates the letsencrypt-prod ClusterIssuer when acmeEmail is set, see below.
Monitoring	geass-kube-prometheus-stack	monitoring	Chart 69.8.2. Prometheus keeps 15 days or 4GB on a 5Gi volume. The names are pinned so the dashboard can find the Prometheus Service.
Longhorn	geass-longhorn	longhorn-system	Chart 1.12.1. The chart’s own storage class and the Longhorn UI are switched off.
Logging	geass-victoria-logs and geass-fluent-bit	geass-logging	The store is bounded to 4GiB on a 6Gi volume and your retention. The shippers skip system namespaces.

Beyond the charts the cluster controller keeps these in place:

• Storage classes, once Longhorn is ready: geass-replicated (1, 2 or 3 copies, by the number of schedulable nodes), geass-single (one copy kept on the node the pod runs on, for databases that replicate themselves), for each region geass-replicated-<region> restricted to that region’s nodes and, for each set of node tags a project requires, geass-tagged-<hash> restricted to the nodes that carry them. Existing volumes are raised, never lowered, to the replica count as nodes join.
• Longhorn node tags: each Longhorn node gets geass-region-<region> and geass-tag-<tag> tags that mirror its Kubernetes node, which is how a storage class can name a region or tags. Once any node is a compliance node, every other node also gets geass-general, and the global and regional classes name it, so other projects’ volume copies stay off compliance nodes.
• The ClusterIssuer letsencrypt-prod, once cert-manager is ready and spec.addons.certManager.acmeEmail is set: ACME against Let’s Encrypt’s production server, solved over HTTP-01 through the traefik ingress class, with the account key in the Secret letsencrypt-prod-account-key in the cert-manager namespace. It carries the labels geass.dev/managed-by=geass and geass.dev/cluster=<cluster>. The operator watches it and puts it back when it is deleted or edited. An issuer without the managed-by label is never touched. If you later empty acmeEmail, the issuer Geass made stays in place so that existing certificates can renew.
• Network policies geass-restrict-metrics-store in monitoring and geass-restrict-log-store in geass-logging, so only the platform can read metrics and logs.
• A PodMonitor geass-ingress in monitoring, so Prometheus scrapes Traefik and the request charts fill in.
• A ServiceMonitor and a PrometheusRule named geass-operator in monitoring, while monitoring is on, so Prometheus scrapes the operator’s own metrics and holds six alerts on them. See Logs, metrics and terminal.
• The serverless add-on, while spec.addons.serverless.enabled is true: the HelmCharts geass-keda and geass-keda-http (KEDA 2.21.0 and the HTTP add-on 0.16.0, installed into keda) and, when none exists, a HelmChartConfig traefik in kube-system that sets providers.kubernetesIngress.allowExternalNameServices: true. They install in that order, and the operator looks again every 20 seconds until both are up. A traefik configuration that already exists is left alone, and the operator logs what to add to it.
• The admission policy geass-required-node-tags, a ValidatingAdmissionPolicy and its binding, created the first time a project requires node tags and never removed. See Compliance node tags.

Inspect the cluster

kubectl -n geass-system get geassclusters
kubectl -n geass-system get geasscluster default -o jsonpath='{range .status.conditions[*]}{.type}={.status} {.reason}: {.message}{"\n"}{end}'

# The add-ons
kubectl -n kube-system get helmcharts
kubectl -n kube-system get jobs | grep helm-install
kubectl get storageclass
kubectl -n longhorn-system get pods

# The nodes and their regions
kubectl get nodes -L topology.kubernetes.io/region,topology.kubernetes.io/zone

Troubleshooting

Ready is False with ClusterPending

An add-on is still installing. List the HelmCharts and their install Jobs (the commands above). A Job that keeps failing or a pod in ImagePullBackOff is the usual cause. Geass only logs Waiting for cert-manager add-on, Waiting for monitoring add-on or Waiting for Longhorn add-on; the condition does not name the add-on.

Longhorn never becomes ready

Longhorn needs the iSCSI initiator on every node. The installer and the worker join step install it, but the worker step is optional, so a worker where it failed has no Longhorn manager. Check kubectl -n longhorn-system get pods -o wide for a node without a ready longhorn-manager. Also check kubectl get csidriver driver.longhorn.io. Until every schedulable node has a ready manager, the storage classes are not created and databases report Storage is starting.

The reason is ProvisionFailed

Read the add-on that failed: AddonsReady still carries the last error. Fix the cause, then change a field in the spec so that metadata.generation changes. The controller resumes when the generation changes. Saving the same values again does not change it.

A worker cannot join: cannot reach the cluster API at <url>

The rest of the message says why: a refused connection or a timeout means 6443/tcp is closed between the worker and the control plane, and a loopback address means spec.serverURL is 127.0.0.1 and no control plane node has an address to use instead. See Add servers.

Projects say project "<project>" cluster "<cluster>" is not ready

The GeassCluster is not Ready. See the cases above.

HTTPS certificates are not issued

Read the ClusterIssuerReady condition: kubectl -n geass-system get geasscluster default -o jsonpath='{.status.conditions[?(@.type=="ClusterIssuerReady")]}'. ACMEEmailMissing means no contact address is set: patch spec.addons.certManager.acmeEmail with a real address. IssuerNotReady carries cert-manager’s reason, most often Let’s Encrypt refusing the address (it rejects placeholder domains such as example.com). Then check kubectl get clusterissuer letsencrypt-prod. See Domains and TLS.

A service cannot sleep and says AddonNotInstalled

The serverless add-on is off or not finished. Check kubectl -n kube-system get helmcharts geass-keda geass-keda-http and kubectl -n keda get pods. See Sleep when idle.

Metrics or logs screens are empty

Check that monitoring or logging is enabled and that the pods in monitoring or geass-logging run. After you turn logging off, the screens that read history have nothing to show.

Related

• How Geass works
• Add servers
• Storage
• Observability
• Scaling and placement
• Sleep when idle
• Compliance node tags
• GeassNode
• GeassPlatformConfig
• GeassHAReadiness
• Configuration