Skip to content

Ethere's choices

Ethere is Kubernetes: the API, kubectl, Helm charts and Flux work as anywhere. It also makes choices another cluster leaves open, each of which changes something for manifests written elsewhere.

Choice Why What it changes
IPv6 only No IPv4 to allocate, every pod addressable Services ipFamilies: [IPv6], servers listening on ::; IPv4-only destinations by name (Network)
Every pod denied by default but DNS An application reaches and is reached by what it declares, nothing more A CiliumNetworkPolicy per flow: the Gateway in, the names and services out (Network)
Gateway API One Gateway, its certificates managed, names checked per namespace HTTPRoute instead of Ingress; no ingress controller to install (Exposing)
Names granted per namespace A route can't take another tenant's name A route's hostnames among the namespace's allowed names, never empty (Exposing)
Flux, locked down per tenant Deployments by commit, no cross-namespace access A HelmRelease names the service account flux; sources and charts in its own namespace (Deploying)
Servers replaced, never updated Every update is a new image on new servers Nothing lasts on a node's disk: data in volumes, hostPath never (Storage)
Volumes in the zone of their pod Created where the pod is scheduled (WaitForFirstConsumer) Nothing to set; a volume follows its pod when a server is replaced (Storage)
The cloud's metadata service closed A pod can't read its server's credentials Nothing reaches 169.254.169.254, no policy reopens it

Coming from manifests written elsewhere

  • An Ingress becomes an HTTPRoute attached to the Gateway public of the namespace ethere-gateway, its hostnames a name of the namespace, plus a CiliumNetworkPolicy letting the Gateway in (Exposing). Annotations of an ingress controller (rewrites, redirects, timeouts) become the route's filters and rules.
  • IPv4: 0.0.0.0 becomes :: (most servers then accept both), an IPv4 literal in a configuration becomes a name, a Service declares ipFamilies: [IPv6].
  • Egress: whatever an application calls (an API, a database outside, an SMTP server) is a toFQDNs rule, the Kubernetes API a toEntities: [kube-apiserver] rule; between the application's own pods, toEndpoints.
  • Cluster-scoped objects (CRDs, ClusterRoles, webhooks, IngressClass, StorageClass) aren't a tenant's: a chart installing them needs them disabled, or the admins.
  • hostNetwork, hostPath, privileged pods: not for applications.