Skip to content

Claudie v0.10

Upgrade to this release from previous version requires manual intervention (due to the Longhorn version upgrade).

Several major internal changes have been made to Claudie. While the v0.10.0 version should be backward compatible with v0.9.16, it was not possible to test all possible cluster configuration scenarios. We advise creating backups before upgrading to v0.10.0 and later versions.

Only v0.10.0 should be backwards compatible with the v0.9.16 version. Any other v0.10.x version will not have this guarantee.

Most notable changes (TL;DR)

  • This version introduces a regular loop that will periodically ensure that the created infrastructure always matches the specs from the InputManifest and it is aligned and corrected if it drifts. This mechanism will be applied on any newly created clusters. Clusters imported from older versions of Claudie will become regularly reconciled after their first modification in InputManifest.

  • Longhorn v1.9.2 will now be deployed for clusters built with Claudie. For existing clusters built with v0.9.16 manual steps need to be done before deploying Claudie v0.10.0:

    • Please read about the manual steps here

  • The Builder service has been completely removed from Claudie. It is also recommended that you delete the Builder deployment after deploying the v0.10.x versions of Claudie. Claudie now uses NATS instead of the builder to dispatch tasks among the other services.

  • The BuilderTTL field, which was internal to Claudie's task dispatching process, was completely removed in favor of a work queue. Previously, when the BuilderTTL reached 0, a new diff with the current desired state was made, even if the scheduled task did not finish. Thus, it was possible for another task to be dispatched. This is no longer possible, as the move to NATS requires an explicit acknowledgment of the task to progress the building of the cluster.

  • The identification and scheduling of tasks has been overhauled. Claudie now has an initial version of a reconciliation loop. In the v0.9.x versions of Claudie, whenever a change was detected after running kubectl apply -f <your-input-manifest> Claudie stopped and did not continue to health check or fix the error, even if the error was simply a network inconvenience, upon either a failure or success of building that change. As of now, with the reconciliation loop, every kubectl apply -f <your-input-manifest> will explicitly state the desired state of your clusters, and Claudie will try endlessly to reach that desired state. This means that, in the event of any errors, changes will be reverted and then reapplied, along with health checking, which helps identify potential misconfigurations or infrastructure issues. Claudie will then try to auto-repair these issues, if possible. The goal is to further improve the reconciliation loop with each release.

  • DynamoDB was removed in favor of native locking supported by newer versions of OpenTofu which ship with Claudie v0.10.x

  • Support for Exoscale

Deployment

To deploy Claudie v0.10.X, please:

  1. Download Claudie.yaml from release page

  2. Verify the checksum with sha256 (optional)

We provide checksums in claudie_checksum.txt you can verify the downloaded yaml files against the provided checksums.

  1. Install Claudie using kubectl

We strongly recommend changing the default credentials for MongoDB, MinIO before you deploy it.

kubectl apply -f https://github.com/berops/claudie/releases/latest/download/Claudie.yaml

To further harden Claudie, you may want to deploy our pre-defined network policies: 

# for clusters using cilium as their CNI
kubectl apply -f https://github.com/berops/claudie/releases/latest/download/network-policy-cilium.yaml

# other
kubectl apply -f https://github.com/berops/claudie/releases/latest/download/network-policy.yaml

v0.10.0

What's Changed

  • Use native state locking provided by OpenTofu instead of relying on DynamoDB #1906

  • Upgrade kubeone to v1.12.1. Claudie now supports building the following Kubernetes versions: v1.32, v1.33, v1.34 #1913

  • Making use of a provider cache in the Terraformer, essentially removing the time spent downloading the provider on a cache hit #1907

  • Preventing kubeone from overriding config.toml which would collide with NvidiaGPU operator overrides #1916

  • Longhorn will now be deployed with the best-effort data-locality setting #1933

  • The Ansibler stage has been tweaked to take less time overall #1917

  • Genesis Cloud provider support dropped #1941

  • The zone field is now optional for dynamic nodepools defined in the Input Manifest. If omitted, Claudie will automatically distribute the nodes across zones #1947

  • Claudie will now deploy Longhorn with version 1.9.2 #1956. Manual steps need to be done before upgrading to Claudie v0.10.0 for Longhorn.

  • Claudie will now support GPU guest accelerator for GCP nodepools #1952 Previously, it was not possible to communicate this information to the templates used to spawn the infrastructure. With the new changes, the GPU type and count will now be passed to the templates, correctly spawning a VM with the requested GPU.

 nodePools:
   dynamic:
     - name: gpu-workers
       providerSpec:
         name: gcp-provider
         region: europe-west1
         zone: europe-west1-b
       count: 1
       serverType: n1-standard-4
       image: ubuntu-2204-lts
       machineSpec:
         nvidiaGpuCount: 1              # <-- specify number of gpus.
         nvidiaGpuType: nvidia-tesla-t4 # <-- specify gpu type

  • Initial version of the reconciliation loop was added to Claudie #1951 Claudie will now endlessly healthcheck and try to fix errors on identified tasks. While currently this only resolves basic scenarios, such as unreachable nodes, the aim is to broaden this with every release.

  • Claudie will no longer expect NGINX to be installed on existing clusters #1980

  • Part of the reconciliation loop is to refresh the current state infrastructure periodically after no tasks have been identified #1979

  • Added support for a new provider Exoscale

Bug fixes

  • Deletion process was fixed for newer versions of Kubernetes #1919

  • Deploy kubelet-csr-approver to approve kubelet server CSRs #1934