Rok v1.1 (unreleased)

This guide describes the necessary steps to upgrade an existing Rok cluster on Kubernetes from version v1.0 to the latest pre-release version, 1.1-rc5.

Upgrade your management environment

We assume that you have followed the Deploy Rok guide, and have successfully set up a full-fledged rok-tools management environment either in local Docker or in Kubernetes.

Before proceeding with the core upgrade steps you need to first upgrade your management environment, in order to use CLI tools and utilities, such as rok-deploy that are compatible with the Rok version you are upgrading to.

Important

When you upgrade your management environment all previous data (GitOps repository, files, user settings, etc.) are preserved either in a Docker volume or Kubernetes PVC, depending on your environment. This volume or PVC is mounted in the new rok-tools container so that old data is adopted.

For Kubernetes simply apply the latest rok-tools manifests:

$ kubectl apply -f <download_root>/rok-tools.yaml

Note

In case you see the following error:

The StatefulSet "rok-tools" is invalid: spec: Forbidden: updates to
statefulset spec for fields other than 'replicas', 'template', and
'updateStrategy' are forbidden

make sure your first delete the existing rok-tools StatefulSet with:

$ kubectl delete sts rok-tools

and then re-apply.

For Docker first delete the old container:

$ docker stop <OLD_ROK_TOOLS_CONTAINER_ID>
$ docker rm <OLD_ROK_TOOLS_CONTAINER_ID>

and then create a new one with previous data and the new image:

$ docker run -ti \
>     -p 8080:8080 \
>     --entrypoint /bin/bash \
>     -v $(pwd)/rok-tools-data:/root \
>     gcr.io/arrikto/rok-tools:release-1.1-l0-release-1.1-rc5

Upgrade manifests

We assume that you have followed the Deploy Rok guide, and have a local GitOps repo with Arrikto-provided manifests. Once Arrikto releases a new Rok version and pushes updated deployment manifests, you have to follow the standard GitOps workflow:

  1. Fetch latest upstream changes, pushed by Arrikto.
  2. Rebase local changes on top of the latest upstream ones and resolve conflicts, if any.
  3. Tweak manifests based on Arrikto-provided instructions, if necessary.
  4. Commit everything.
  5. Re-apply manifests.

When one initially deploys Rok on Kubernetes, either automatically using rok-deploy or manually, they end up with a deploy overlay in each Rok component or external service that is to be applied to Kubernetes. In the GitOps deployment repository, Arrikto provides manifests that include the deploy overlay in each Kustomize app/package as scaffold, so that users can quickly start and set their preferences.

As a result, fetch/rebase might lead to conflicts since both Arrikto and the end-user might modify the same files that are tracked by Git. In this scenario, the most common and obvious solution is to keep the user’s changes since they are the ones that reflect the existing deployment.

In case of breaking changes, e.g., parts of YAML documents that are absolutely necessary to perform the upgrade, or others that might be deprecated, Arrikto will inform users via version-specific upgrade nodes for all actions that need to be taken.

Note

It is the user’s responsibility to apply valid manifests and kustomizations after a rebase. In case of uncertainty do not hesitate to coordinate with Arrikto’s Tech Team for support.

We will use git to update local manifests. You are about to rebase your work on top of latest pre-release branch. To favor local changes upon conflicts, we will use the corresponding merge strategy option.

Important

During rebase, the sides are swapped, i.e., ours is the so-far rebased series, and theirs is the working branch. For more information on the merge strategy read the official git-scm docs.

To upgrade the manifests:

  1. Fetch latest upstream changes:

    $ git fetch --all -p
    
  2. Retrieve the release channel your are currently following. This is actually the upstream branch of your GitOps repo. To retrieve it run:

    $ git rev-parse --abbrev-ref --symbolic-full-name @{u}
    origin/develop
    

    Important

    If you are currently on the release-1.0 channel (i.e., your branch is based on origin/release-1.0) and want to update to develop, please follow the Switch release channel section before proceeding.

  3. Rebase on top of the latest pre-release version:

    $ git rebase -Xtheirs
    

    Rebasing your work, may cause conflicts, e.g., when a file was modified locally but removed from upstream, for example:

    CONFLICT (modify/delete): kubeflow/kfctl_config.yaml deleted in origin/develop and modified in HEAD~61. Version HEAD~61 of kubeflow/kfctl_config.yaml left in tree.
    

    We suggest you go ahead and delete those files, e.g.:

    $ git status --porcelain | awk '{if ($1=="DU") print $2}' | xargs git rm
    

    And proceed with the rebase:

    $ git rebase --continue
    
  4. (Optional) Edit deploy overlays based on version-specific upgrade notes.

  5. Commit changes, if any.

Important

Make sure you mirror the GitOps repo to a private remote to be able to recover it in any case.

Drain rok-csi nodes

To ensure minimal disruption of Rok services, please follow the following instructions to drain Rok CSI nodes, and wait for any pending Rok CSI operations to complete, before performing the upgrade.

During the upgrade, any pending Rok tasks will be canceled, so it is advisable to run the following steps in a period of inactivity, e.g., when no pipelines or snapshot policies run. Since pausing/queuing everything is currently not an option, one can monitor Rok logs and wait until nothing has been logged for, let’s say, 30 secs:

$ kubectl -n rok logs -l app=rok-csi-controller -c csi-controller -f --tail=100

Note

Finding a period of inactivity is an ideal scenario, that depending on the deployment may not be feasible, e.g., when having tens of recurring pipelines running. In such a case the end-user will simply see some of them fail.

  1. Scale down the rok-operator StatefulSet:

    $ kubectl -n rok-system scale sts rok-operator --replicas=0
    
  2. Ensure rok-operator scaled down to zero:

    $ kubectl get sts rok-operator -n rok-system
    
  3. Scale down the rok-csi-controller StatefulSet:

    $ kubectl -n rok scale sts rok-csi-controller --replicas=0
    
  4. Ensure rok-csi-controller scaled down to zero:

    $ kubectl get sts rok-csi-controller -n rok
    
  5. Watch the rok-csi-node logs and ensure that all pending operations have finished, i.e., nothing has been logged for the last 30 secs:

    $ kubectl -n rok logs -l app=rok-csi-node -c csi-node -f --tail=100
    
  6. Continue with the Upgrade components section.

Upgrade components

We assume that you are already running a v1.0 Rok cluster on Kubernetes and that you also have access to the 1.1-rc5 kustomization tree you are upgrading to.

Since a Rok cluster on Kubernetes consists of multiple components, you need to upgrade each one of them. Throughout the guide, we will keep track of these components, as listed in the table below:

Component old new
RokCluster CR  
RokCluster CRD  
Rok Operator  
Rok Disk Manager  
Rok kmod  

During the upgrade, Rok Operator will remove all members from the cluster and add a dedicated one to perform the upgrade. The cluster will be scaled down to zero and a Kubernetes Job will run to upgrade the cluster config on etcd and run any needed migrations. Finally, the cluster will be scaled back up to its initial size.

1. Increase observability (optional)

To gain insight into the status of the cluster upgrade execute the following commands in a separate window:

  • For live cluster status:

    $ watch kubectl get rokcluster -n rok
    
  • For live cluster events:

    $ watch 'kubectl describe rokcluster -n rok rok | tail -n 20'
    

2. Inspect current version (optional)

Get current images and version from the RokCluster CR:

$ kubectl describe rokcluster rok -n rok
...
Spec:
  Images:
    Rok:      gcr.io/arrikto-deploy/roke:l0-release-v1.0
    Rok CSI:  gcr.io/arrikto-deploy/rok-csi:l0-release-v1.0
Status:
  Version:        l0-release-v1.0

3. Upgrade Rok Disk Manager

Apply the latest Rok Disk Manager manifests:

$ kubectl apply -k rok/rok-disk-manager/overlays/deploy
Component old new
RokCluster CR  
RokCluster CRD  
Rok Operator  
Rok Disk Manager  
Rok kmod  

4. Upgrade Rok kmod

Apply the latest Rok kmod manifests:

$ kubectl apply -k rok/rok-kmod/overlays/deploy
Component old new
RokCluster CR  
RokCluster CRD  
Rok Operator  
Rok Disk Manager  
Rok kmod  

Important

Rok 1.1-rc5 ships with a new version of the dm-era module. As rok-kmod tries to load this new version, it is expected to fail (enter in a CrashLoopBackOff) since the old version of the module is in use and cannot be unloaded. For that to be fixed, you will have to perform a rolling reboot of all the nodes in the Kubernetes cluster after the upgrade has finished.

5. Upgrade Rok cluster

Apply the latest Rok cluster manifests:

$ kubectl apply -k rok/rok-cluster/overlays/deploy
Component old new
RokCluster CR  
RokCluster CRD  
Rok Operator  
Rok Disk Manager  
Rok kmod  

6. Upgrade Rok Operator

Apply the latest Operator manifests:

$ kubectl apply -k rok/rok-operator/overlays/deploy

Note

The above command also updates the RokCluster CRD

After the manifests have been applied, ensure Rok Operator has become ready by running the following command:

$ watch kubectl get pods -n rok-system -l app=rok-operator
Component old new
RokCluster CR  
RokCluster CRD  
Rok Operator  
Rok Disk Manager  
Rok kmod  

7. Verify successful upgrade

  1. Check the status of the cluster upgrade Job:

    $ kubectl get job -n rok rok-upgrade-release-1.1-l0-release-1.1-rc5
    
  2. Ensure that Rok is up and running after the upgrade Job finishes:

    $ kubectl get rokcluster -n rok rok
    
  3. Ensure all pods in the rok-system namespace are up-and-running:

    $ kubectl get pods -n rok-system
    
  4. Ensure all pods in the rok namespace are up-and-running:

    $ kubectl get pods -n rok
    

Upgrade Kubeflow manifests

This section describes how to upgrade Kubeflow . If you have not deployed Kubeflow in your cluster, you can safely skip this section.

Kubeflow manifests are upgraded in a fetch/rebase/apply manner. Assuming that you have already gone through Upgrade manifests to upgrade Kubeflow all you need it to re-apply them:

$ rok-deploy --apply install/kubeflow

Afterwards, make sure to validate the updated Kubeflow deployment by following the Verify section.

Rolling reboot of K8s cluster

Rok 1.1-rc5 ships with a new version of the dm-era module. In order for this module to be loaded, you will have to reboot all nodes in the Kubernetes cluster. Use the rok-k8s-reboot tool to perform a rolling reboot of all the nodes on the Kubernetes cluster, by spawning a Job on each node and reboot them:

$ rok-k8s-reboot

Reset the CBT data of all Rok PVCs

Due to an upstream kernel bug in the dm-era module, the CBT (Changed Block Tracking) data that Rok maintains for each PVC might be incomplete, leading to corrupted snapshots.

To work around this you need to reset the CBT data of all Rok PVCs:

$ rok-reset-cbt

This ensures that the next snapshot of all Rok PVCs will copy the whole volume, thus avoiding data loss, and will reset the CBT data.

Upgrade Kubeflow resources

This section describes how to upgrade Kubeflow resources (Notebooks, Pipelines) so they are compatible with Rok 1.1-rc5. If you have not deployed Kubeflow in your cluster, you can safely skip this section.

Rok 1.1-rc5 introduces a number of authentication and authorization improvements across its components that break backwards compatibility with the v1.0 Rok client. As a result, several Kubeflow workloads that rely on the Rok client such as notebooks and pipelines are expected to fail after the upgrade. To restore their functionality, you need to upgrade them to use the 1.1-rc5 Rok client, using the instructions in this section.

Upgrade notebooks

Jupyter notebooks with Kale support use the Rok client to create snapshots of their volumes when submitting pipelines. This feature is expected to break after upgrading to Rok 1.1-rc5. You can verify that this is the case by enabling the Kale panel from withing the notebook, and checking that the Use this notebook’s volumes and Take Rok snapshots during each step switches are disabled. Note that depending on which version of Kale you are using, these switches may be under the Advanced Settings section. Clicking on More info... in the switches’ tooltip should reveal that the rok.check_rok_availability() method failed with either a 401 (Unauthorized) HTTP error, or a ValueError about not being able to retrieve the Rok API URL, depending on whether the notebook’s pods have been restarted after the upgrade.

To fix the above errors, the Docker image of these notebooks needs to be updated to gcr.io/arrikto/jupyter-kale:v1.1-pre-207-g6c6282e72, which includes an up-to-date version of the Rok client. You can achieve this for all notebooks by executing the rok-notebook-upgrade script, which is available within the rok-tools management environment, and following the on-screen instructions:

$ rok-notebook-upgrade --image gcr.io/arrikto/jupyter-kale:v1.1-pre-207-g6c6282e72

Alternatively, you can update the image of each notebook manually by snapshotting it and restoring it with the updated image. You can achieve this via the following steps:

  1. Visit the Snapshots tab of the Kubeflow dashboard.
  2. Create a new bucket or choose one of the existing buckets to store the created snapshot.
  3. Click the Snapshot button from within the bucket.
  4. Select the JupyterLab service.
  5. Provide the namespace, the name of the notebook and the name of the snapshot to create.
  6. Click Snapshot. Once the created task completes, a new snapshot will appear in the bucket with the name you selected.
  7. Visit the Notebooks tab of the Kubeflow dashboard.
  8. Delete the old notebook, if you intend to recreate it with the same name.
  9. Click on + NEW SERVER.
  10. Click on the file chooser in the Rok URL field and select the snapshot you just created.
  11. Provide a name for the notebook.
  12. Make sure Custom image is ticked, and set the image name to gcr.io/arrikto/jupyter-kale:v1.1-pre-207-g6c6282e72.
  13. Click LAUNCH to create the updated notebook.

Upgrade pipelines

Kubeflow pipelines that were created using Kale use the Rok client to create snapshots of their steps. These pipelines are also expected to stop working after the upgrade, and need to be resubmitted. You can ensure that this is the case by selecting the failed step of a failed pipeline run, visiting the Logs tab, and verifying that the step failed with the ValueError: A Rok API URL must be provided either via the `url' argument or the `ROK_GW_URL' environment variable error, due to not being able to retrieve the Rok API URL from the pod’s environment.

You can fix this by re-creating the pipelines using the updated notebooks, via the following steps:

  1. Restore the notebook that was used to create the pipeline, if it no longer exists, changing its image to gcr.io/arrikto/jupyter-kale:v1.1-pre-207-g6c6282e72, by following the manual instructions in the Upgrade notebooks section. Note that alternatively, if the notebook no longer exists but you still have the snapshots created by one of the pipeline’s runs, you can also restore the notebook by using the Rok URL of one of these snapshots.
  2. Submit a new version of the pipeline using Kale. In order for the new pipeline to use the updated image, you need to change it manually from the Advanced Settings section of the Kale panel before submitting it. Note that leaving the image empty will also make Kale fall back to using the current notebook’s image, which should suffice in this case.

Upgrade recurring runs

Any existing recurring runs on Kubeflow that use one of the failing pipelines described in section Upgrade pipelines need to also be recreated. You can check for the existence of such a run via the following steps:

  1. Visit the Jobs tab of the Kubeflow dashboard to see a summary of existing recurring runs.
  2. Click on the experiment each job belongs to and see if it contains any failed runs that belong to that job. Note that you can see the recurring run a run belongs to (if any) in the Recurring run column of this list.
  3. Click on the failed run to view its steps, then click on a step that failed, and verify that the error that caused the failure is the one mentioned in the Upgrade pipelines section.

To update such a failing recurring run, you can follow these steps:

  1. Update the pipeline the recurring run uses via the instructions of section Upgrade pipelines.
  2. Visit the Jobs tab of the Kubeflow dashboard and click on the recurring run you want to recreate.
  3. Click on Clone recurring run.
  4. Select the updated pipeline in the Pipeline field. Note that this should also select the latest version of that pipeline in the Pipeline Version field.
  5. Click Start to create the updated recurring run.
  6. Return to the Jobs tab of the Kubeflow dashboard, select the old failing recurring run and click Delete to delete it.

Upgrade ALB Ingress controller

Until now we were using ALB Ingress Controller v1.1.4. In v2.0 this got renamed to AWS Load Balancer Controller.

To upgrade it one should follow the official migration guide.

This boils down to:

  1. Install the new controller from scratch. To do so please follow our guide.

  2. Purge old controller:

    $ kubectl delete -f rok/alb-ingress-controller/deployment.yaml
    $ kubectl delete -f rok/alb-ingress-controller/rbac.yaml
    
  3. The existing Ingress resources will continue to work as expected.

Upgrade NGINX Ingress Controller manifests

The rok/nginx-ingress-controller/service.yaml and rok/nginx-ingress-controller/ingress.yaml manifests are now deprecated and have been converted into a kustomization package. Please follow the installation guide for NGINX Ingress Controller to configure and re-apply the necessary resources.

Upgrade Istio manifests

The rok/istio/istio-ingress-nginx.yaml manifests are now deprecated and have been converted into a kustomization package. Please follow the installation guide for exposing Istio from scratch to configure and re-apply the necessary resources.