Skip to content

A profile whose workers are machines has nothing in the cluster: no WorkerProfile in direct mode, and a count of none on the one a repository holds

Decision 738 · accepted · 2026-09-30 · issue #290

Issue #290. An ssh profile's workers are machines somebody registers, and the worker on each dials the plane; but admin.profile.put, a grant, a bundle's projection and a release all wrote it a WorkerProfile, and the operator, which reads no provisioner, made a StatefulSet of pods for it. Provision.apply/2 now asks in_cluster?/1, which is the Kubernetes provisioner and nothing else. In direct mode a profile that is not in the cluster is written nothing, and a resource left from before (a profile that was Kubernetes's, or one saved before this was asked) is deleted, so the row stays the whole of what is wanted; the answer is :not_in_cluster. In gitops mode a repository holds every profile as a WorkerProfile, so the resource is there, and the plane writes spec.replicas: 0 onto it rather than the scaler's count, which for such a profile is a number of machines. Between Flux's apply and the plane's first write the CRD's default of one replica still stands; teaching the operator the troupe.dev/provisioner annotation would close that, and would make it read an annotation it deliberately reads none of. ReleaseImage passes such a profile by: whoever installs the worker on a machine installs the release. Proof: admin_test.exs, provision_test.exs, release_image_test.exs.

Governs: apps/troupe_plane/lib/troupe/plane/provision.ex, apps/troupe_plane/lib/troupe/plane/reach.ex