Cluster settings to review
Orchard hardens tenants by default, but a few protections depend on how your cluster is built, so they stay off or only warn until you turn them on. Turning one on without checking first can refuse a tenant’s existing workload, which is why each entry below says what to check first, how to set it, and how to undo it. Start with orchard doctor, which flags the settings that apply to your install.
Block tenant NodePorts
Section titled “Block tenant NodePorts”Value: tenant.nodePortExposure: false.
A NodePort opens a port on every node. On a shared cluster that is a port on machines other tenants’ pods run on. With this set, tenants get raw TCP ports through your public IP (tenant.publicIp) instead.
- Check first.
kubectl get svc -A | grep NodePort. Services that already exist keep working, and so do redeploys of them. Only creating a new NodePort, or switching a port from internal to public, is refused. - Also do. Firewall the node port range (30000-32767) and the kubelet, SSH and API ports at your cloud or host. Orchard cannot do that for you.
- Undo. Set it back to
true.
Check issuers and storage classes
Section titled “Check issuers and storage classes”Values: admissionPolicies.extraIssuers, admissionPolicies.storageClasses.
The admission guard can check which certificate issuer a tenant Ingress names and which StorageClass a tenant PVC names. The issuer rule always allows tenant.certClusterIssuer. Add any other issuer you hand out, such as a DNS-01 wildcard issuer, to extraIssuers. Set storageClasses to the classes tenants may use, for example [longhorn]. An empty list skips that rule.
- Check first.
kubectl get ingress -A -o jsonpath='{range .items[*]}{.metadata.annotations.cert-manager\.io/cluster-issuer}{"\n"}{end}' | sort | uniq -candkubectl get pvc -A -o custom-columns=SC:.spec.storageClassName | sort | uniq -clist what is in use. Every value in those lists must appear in your setting. - Mode. The guard refuses by default, so a missing issuer or class refuses that tenant’s Ingress or PVC. To try a value without that risk, set
admissionPolicies.validationActions: [Warn, Audit]first, read the warnings, then set it back to[Deny]. - Undo. Empty the list.
Move the hardening rules to Deny
Section titled “Move the hardening rules to Deny”Value: admissionPolicies.hardeningActions: [Deny].
The rest of the guard refuses by default (admissionPolicies.validationActions: [Deny]). The hardening rules (drop ALL capabilities, tenant RuntimeClass) only warn, because pods created before them can fail and a refusal would stop them restarting.
- Check first. Look for hardening warnings (events on the namespace, API server audit log). Redeploy the apps that report one.
- Known warning. The Tailscale proxy pod that Orchard does not build itself may fail the “drop ALL capabilities” rule.
- Safe by design. Every policy has
failurePolicy: Ignoreand every binding tolerates a missing parameter, so a fault in the guard cannot block the cluster. cert-manager’s solver objects and the Orchard server are exempt from the object rules. - Undo. Set it back to
[Warn, Audit]. The change applies in seconds. The same goes forvalidationActionsif the guard ever refuses something it should not.
Sandbox databases and disable the superuser login
Section titled “Sandbox databases and disable the superuser login”Value: databases.runtimeClassName (for example gvisor), plus the cnpg-disable-superuser.sh script in the Orchard repository.
Tenant databases are reachable only over the Postgres protocol: there is no shell in the pod, exec is refused, the network policy isolates them, and new clusters are created with the postgres superuser login off. Without a runtime class, database pods run on the node’s kernel, so a Postgres or extension vulnerability would reach the host. gVisor closes that, at a cost in IO and latency.
- Check first. The runtime must exist on every node (
kubectl get runtimeclass). Test with a database you can afford to lose. The setting affects only pods created afterwards, so roll each cluster yourself: switch over, then delete the old primary. - Failure policy.
databases.runtimeClassFailurePolicy: Ignore(default) never blocks pod creation.Failguarantees that no unsandboxed database pod is admitted, and blocks creation if the policy is broken. - Superuser login. A database cluster created by an Orchard version that left the login on keeps it until you run the script. With no flags, the script is a dry run that lists the clusters it would change. With
--apply, it patches clusters one at a time and stops if any instance pod restarts. Start with--limit 1on a database that is not critical. - Undo. Clear the value. Existing pods stay sandboxed until recreated.
Sandbox tenant pods and builds
Section titled “Sandbox tenant pods and builds”Values: tenantSandbox.enabled (gVisor), builderSandbox.enabled (Kata).
These are the strongest isolation Orchard offers and they depend on your nodes. Install the runtime on every node first. A tenant pod that names a missing runtime does not schedule. For what each sandbox protects against, refer to Security and isolation.
Per-database certificates
Section titled “Per-database certificates”Value: publicTls.db.mode: perDatabase.
Gives each public database its own certificate so the shared wildcard key never reaches a tenant. It needs a DNS-01 issuer and every certificate counts against your CA’s rate limits (Let’s Encrypt allows 50 per registered domain per week). Leave it off on a busy public instance until you have a higher limit.
Lock the origin to your edge
Section titled “Lock the origin to your edge”If a CDN or proxy such as Cloudflare fronts your ingress, allow the ingress controller to accept connections only from the proxy’s address ranges. Otherwise anyone who finds the origin address bypasses the proxy’s protections. Do this at your cloud firewall or in the controller’s allow list, and keep your own monitoring addresses in it. Orchard does not manage it, and getting it wrong takes the site down, so test from outside after changing it.
Plain-HTTP registry hosts
Section titled “Plain-HTTP registry hosts”Setting: REGISTRY_HTTP_HOSTS, a comma-separated list of registry hosts Orchard may reach over HTTP, for example an in-cluster service name. An IP address is allowed. Everything in the list is also reserved, so a tenant cannot claim it as a hostname. List only registries on your cluster network.
Ingress controller
Section titled “Ingress controller”Value: ingress.enabled.
The chart’s own IngressRoute is a Traefik resource for the Orchard dashboard. If you use another controller, set ingress.enabled: false and create an Ingress yourself. Tenant ingress follows tenant.ingressClass and is separate.
Defaults for new organizations
Section titled “Defaults for new organizations”New organizations start with no CPU, memory or count caps. On an instance open to public sign-up, set defaults in Instance admin > Configuration > New organization limits, or set a limit on an organization afterwards. For the values and the matching environment variables, refer to Organization limits.