Ready. It tells you where Ankra records what went wrong, which reads it can run on the servers for you, and which remedies exist. You end up knowing why setup stopped and either a cluster that finished on a retry or a clear reason to remove it and create it again.
Before you start
- The cluster’s create has started. A create refused before anything was built, such as “The cluster was not created” with Open Plan and billing, is covered in Create a cluster.
- The
clusters.operatepermission on the cluster, for Try again, retries and the diagnose reads. See Roles and access. - For the CLI steps, the Ankra CLI signed in to the organisation, with the cluster selected (
ankra cluster select <cluster>).
Find where setup stopped
1
Read the progress page
After Create cluster, the progress page follows the setup in stages. Its lines say how far it got:
“The platform has not reported a setup stage yet. Progress updates automatically.” means the operation has not started its first step. Wait for it.Check: if the page shows a setup failure card, carry on with the next step. If it is still moving, leave it; a large cluster takes longer than a small one.
2
Read the failure card
When setup stops, the page shows what stopped and what to do. Where Ankra recognises the failure, the card names it and lists the steps to take; the common ones are under What can go wrong. Otherwise the card reads Setup stopped at this step: followed by the step and the server it ran on, or Setup stopped before it finished, with the provider’s or the installer’s error below it. “The platform recorded no error message for this failure.” means there was none to show.Open Technical details on the card and choose the failed task to read its full log.Check: you can name the step that failed (for example a server create, the Kubernetes install on one control plane, or the agent install) and you have its error.
3
Open the operation
Every create is an operation. Choose View operation on the progress page, or a task under Technical details on the failure card, to see every step, its status and its log. The Operations page explains the list, retries and attention states.From the CLI:Check: the failed step in the operation matches the one the failure card named.
4
Try again
Many failures are temporary: a provider out of capacity for a few minutes, a network connection that timed out. Choose Try again on the failure card. It reads Starting until the new attempt has started. A retry starts a new operation that re-plans only the resources that did not finish.From the CLI:Without the permission, the card says “You do not have permission to run setup again. Ask an organisation owner to run it for you.”Check: the progress page moves past the step that failed. If the same step fails again with the same error, the cause is not temporary: read What can go wrong and Read the servers.
Read the servers
When the operation’s own log does not explain the failure, Ankra can read the servers for you. These reads change nothing. The node and bastion reads run through the bastion with the platform’s own access, so you do not need the cluster’s SSH key;nodes console-log reads the OVHcloud API instead. The platform waits up to two minutes for a diagnose, 90 seconds for nodes cloud-init-log and one minute for nodes console-log; a slower read hands back an operation id to follow with ankra cluster operations list. Running a nodes read again attaches to the one already running; a second bastion diagnose is refused until the first finishes.
Find node ids first:
A node that did not finish its first boot
A server is set up by cloud-init at its first boot. On the providers that offer it (see What each provider offers), read the outcome:cloud-init status --long and the tail of /var/log/cloud-init-output.log.
A node that joined but is NotReady
A node that isNotReady, or whose kubelet stopped posting status, is below what the cluster’s own API shows. The node diagnose read snapshots the node’s OS: the Kubernetes service unit and its journal, disk, memory, uptime, the kernel log and the firewall path to the API server, and names the symptoms it finds.
A node that hung
An OVHcloud node that no longer answers over SSH can still be read from its serial console, through the OVHcloud API. A kernel soft lockup, hung-task warnings, an out-of-memory kill or a root filesystem gone read-only is often the last thing it printed.The bastion
Every read above exceptnodes console-log goes through the bastion. When they fail too, read the bastion itself: sshd configuration, failed logins, disk, failed units, journal errors and listening ports.
What each provider offers
On Proxmox VE and HPE Morpheus, read the failed step’s log in the operation, then the server’s console in Proxmox VE or in your Morpheus appliance.
Verify
- Under Infrastructure → Clusters, the cluster’s status reads Online.
-
ankra cluster operations list --attention openlists no open failure for the create. A failed attempt stays in the history; once a retry finished the resources it failed on, it carries the Resolved by a later run label in the portal. -
Every node is
Ready. Pointkubectlat the cluster through Ankra (this needs a Cluster Access grant):
What can go wrong
The first six rows are failures the progress page names, with the steps it gives. The last two are not on the page: they are what to check when setup looks finished but the cluster is not.
For a cluster that cannot be saved, terminate it and create it again: see Terminate a cluster, and the provider guide for what terminating deletes.
Next steps
- Create a cluster - the create form, and the refusals before anything is built
- Operations - retries, cancels and attention states
- Operate a cluster Ankra built - node groups, the bastion, SSH access and upgrades
- Provider guides: Hetzner, Scaleway, Ankra Cloud, Proxmox VE, HPE Morpheus