Playbooks compose roles against host groups. One playbook per
operation class, not per feature — the roles are idempotent, so a
convergence run touches only what drifted, and a single feature is
scoped at run time with --tags / --limit, never with a new
playbook.
Run from the ansible/ directory:
ansible-playbook playbooks/site.yml --check --diffsite.yml— converge the non-cluster managed hosts.site-k8s.yml— converge the k8s clusters in place (serial: 1, no drain/reboot). The cluster counterpart tosite.yml.site-openbao.yml— converge the OpenBao service VMs (srvvault1..3). The OpenBao counterpart tosite-k8s.yml;site.ymlexcludes theopenbaogroup.site-ceph.yml— converge the microceph storage layer. Today onlyceph_dev(srvk8sdev, whose baseline + microk8s converge throughsite-k8s.yml); the prd Ceph fleet is not yet Ansible-managed.update-k8s.yml— OS patching for k8s nodes (drain → reboot).rebalance-k8s.yml— two descheduler passes on the prd cluster, which evict the pods a drain left on the wrong node so the scheduler places them again;IaC/Scheduled Updateruns it after the roll. The policies are PlatformAddOnsDeploy'sdescheduler/.rebuild-k8s.yml/evict-k8s.yml— full VM rebuild, and the pre-rebuild drain run against the old node.refresh-k8s-addons.yml— post-snap-upgrade microk8s addon refresh.refresh-calico-token.yml— rolling restart ofcalico-node, capping pod uptime below Calico's token-refresh stall window;IaC/Scheduled Calico Rolloutruns it weekly.renew-host-certs.yml/renew-internal-tls.yml— scheduled certificate renewal (SSH host certs;internal_tlsX.509 leaves). Both are threshold-gated no-ops outside the renewal window, andIaC/Scheduled Certsruns both weekly.openbao-backup-secret-id.yml— delivers abackupAppRole secret_id that its caller minted to every srvvault, each proving it with a login first, and does nothing else. SecretRotator runs it when it rotatesbackup(theopenbaorole README, §Backup pipeline).rotate-ansible-key.yml— authorises one public half of theansibleuser's SSH key on every host of theansible_keygroup and logs in to each with its private half, or takes exactly that public half off every one. The always-up hosts are reached over SSH; the VMs that may be off (srvk8sdev, the scratch VMs) through the guest agent of their PVE node, where the caller says they run. SecretRotator'sssh-keykind runs it twice per rotation, adding the new key, then removing the old one; the playbook's header lists its extra vars. Not part ofsite.yml.step-ca.yml— writes step-ca's five Secrets on the prd cluster from thestep_carole's files, and restarts step-ca when their data changed. The single entry point for step-ca's material: cluster bring-up runs it before step-ca is first applied, and every procedure that changes the material runs it. No job runs it.reissue-host-cert.yml— recovery for a host whose SSH host certificate already lapsed; connects over a bootstrap channel that does not depend on it.adopt.yml,grow-disks.yml— one-off operations.