What’s new in 0.7.0
0.7.0 is the release where updates become routine. A cluster moves to a new version with one command, one canary node goes first, and any node that does not come back healthy returns to the generation it ran before. Around that, the other parts of the platform (dashboard, Terraform, MCP, the AI agent, Ceph, HA, Vault, Provisioning) are documented as one product for the first time.
Updates
- Install by tag.
kctl update version set v0.7.0resolves a release tag, checks that the controller can read your database, and rolls the cluster. It also downgrades.--dry-runshows what would change. - Canary. The first node updates and soaks while the controller compares it with the rest: version, heartbeats, guest start failures, agent call failures, Ceph OSD latency. A bad canary rolls itself back and the rollout stops. Default soak is one hour, or 24 hours across a NixOS release change.
- Revert on failure. Every node update is a transaction with a guard timer for in-place switches and a one-shot boot entry for kernel changes. A failed health check puts the previous generation back, with no controller needed.
- Terraform.
kcore_cluster_versiondoes the same from a plan. - Dashboard. The Release page sets the version, promotes a canary, and shows each node.
Read Cluster & node upgrades.
Operating the platform
- Multi-document apply.
kctl apply -ftakes a file with several documents, including volumes, snapshot policies, and operators. - Clients match the controller.
kctl, the dashboard, kcore MCP, and the Terraform provider cover the same resource kinds. See Dashboard, Terraform, and MCP and the AI agent. - Certificate revocation for rotated nodes. A node certificate that a rotation replaced is revoked, and nodes without a sub-CA chain can still verify the revocation list.
- Web sessions on nested calls. Workload actions started from the dashboard keep the operator’s session through every call they make.
Hardware
- vGPU slices. Enable a host once and give a VM a slice:
A2-4Con NVIDIA,intel-2vforradeon-4vfon Intel and AMD. NVIDIA A2 slices boot under Cloud Hypervisor. Intel and AMD cards are split into SR-IOV virtual functions, with a count per card. That path is built but has not been run on lab hardware. - Slurm clusters run as kcoreOS host services and are managed with
kind: SlurmCluster. - Cold volume copies read the source with
O_DIRECT.
Provisioning
- A multi-network provisioner that drives Kea (DHCP) and PowerDNS: networks, reservations, leases, and records, with a RouterOS import. The dashboard lists the subnets, leases, and reservations Kea holds.
- Machine profiles, disk write-verify with an optional wipe, and virtual media over Redfish. See kcore Provisioning.
Already in the product, now documented
- High availability for VMs: detection, fencing, placement gates, restart priority, anti-affinity.
- Live and cold migration.
- kcore SAN on Ceph: three hosts, replicated pool, rbd-nbd, fencing before failover.
- kcore Vault and ransomware resistance.
Upgrade notes
- From 0.6.x. A 0.6.x controller has no
update versioncommand and refusescanary. Move the cluster once with a hand-writtenClusterUpdatethat names the commit of thev0.7.0tag. From 0.7.0 on, usekctl update version set. - Tags before this release have no declaration of version and database schema, and are refused by
update version set. - Terraform. Use provider 0.8.0 with a 0.7.0 controller.
strategy = "canary"onkcore_cluster_updateis refused by controllers older than 0.7.0. - Vault and Provisioning are separate products with their own version numbers. They need no change to move a cluster to 0.7.0.
- Back up first. Take a Vault backup of anything you cannot rebuild.
Known limits
- Update bundles for air-gapped sites are designed but not shipped. An air-gapped site needs a tag mirror and a reachable binary cache.
- Live migration does not move VFIO devices. Cold migration remaps named GPUs to a free one of the same family.
- The dashboard has password and TOTP sign-in. SSO is not available.
- Removing a Ceph member is refused when it would leave fewer hosts than the pool’s replica count, and it is one host per apply.