hello
hello runs podinfo, a small web page that proves your inputs reach the cluster, that a module converges and that turning one off removes it. On by default, on every cloud; turn it off once you trust the pipeline.
Getting started
Section titled “Getting started”It is already on. Change its message to see an input reach the cluster:
kube = { hello = { replicas = 2 message = "hello from acme" }}Then kubectl -n hello port-forward svc/podinfo 9898:9898 and http://localhost:9898 reads hello from acme — <cluster name> on <cloud>.
Settings
Section titled “Settings”| Attribute | Default | |
|---|---|---|
enabled | true | Turns the module on or off. |
replicas | 1 | podinfo’s replica count. |
message | "hello from socle" | The message on podinfo’s page; the cluster name and the cloud are appended. |
hello has no values and no values_secret: it exists to show inputs
flowing, not to be configured.
Every setting
Section titled “Every setting”kube = { hello = { enabled = true # on by default; false once you trust the pipeline replicas = 1 # podinfo's replica count message = "hello from socle" # the cluster name and the cloud are appended }}Good to know
Section titled “Good to know”- The page’s colour is the cloud’s, and so is the cloud named in the
message: both come from the cloud’s own step,
oci/clusters/<cloud>/hello/patch.yaml, which the cloud inserts into the module before its workload. A module shows there what only that cloud has, and nothing else. - Turning it off removes the release and its namespace: the quickest proof that garbage collection works.
- Upgrades: nothing to adapt; the chart moves with
socle_version.
Under the hood
Installed: chart podinfo 6.15.0 from
oci://ghcr.io/stefanprodan/charts/podinfo, in the hello namespace.
What the socle sets: replicaCount from replicas; requests 10m CPU,
32Mi and 64Mi of ephemeral storage (the chart’s 1m / 16Mi is outside GKE
Autopilot’s CPU:memory ratio); from the cloud step’s
hello-cloud-values ConfigMap, ui.color and ui.message — message with
the cluster name and the cloud appended.
Cloud access: none.
Measured on floci 2.1.0, GitHub runner, 2026-09-30: a value patched into the inputs reached podinfo in 13 s.