Skip to content

The chart keeps the namespace it makes and can be told not to make one, and a count the cluster did not take is sent again

Decision 742 · accepted · 2026-10-01 · issue #303

Issue #303, D44 and D40 in docs/developer/defects.md. templates/namespace.yaml rendered the release's namespace as an ordinary resource, so helm uninstall, a reinstall, or a GitOps controller's remediation that uninstalls deleted it with everything else in it: an OpenBao beside the plane, the Secrets the chart expects, volume claims, the WorkerProfile resources. It now carries helm.sh/resource-policy: keep, as the default TroupePolicy already did; an upgrade puts the annotation on the live namespace and in the release's manifest, which is what an uninstall reads. createNamespace (true by default, so an upgrade changes nothing else) leaves the namespace out where something else made it: Helm refuses to take over a namespace it did not make, and the install guide's order, the Secrets first, makes one. Turning it off on an install from an older chart is safe only after one upgrade with it on, because an upgrade deletes what the chart stops rendering unless the live object says keep. Not chosen: dropping the namespace from the chart, which would delete it at the next upgrade of every install that has no keep yet.

In direct mode the scaler wrote the row and then the cluster, and judged the next tick by the row, so a write the cluster refused was never sent again. On a tick that changes no count it now reads the resource's spec.replicas back and sends the row's count where they differ, the comparison gitops mode's pass already makes. Not chosen: writing the row only after the cluster took it, which would let any other write from the row (an edit, a grant, a release) put the old count back in between. Proof: scaling_test.exs ("a count the cluster did not take") and the chart job's "The namespace outlives the chart".

Governs: charts/troupe/templates/namespace.yaml