Prevent removal of apiVersions while they are still used

[sbueringer]

What would you like to be added (User Story)?

As a user it would be nice if Cluster API would block removal of apiVersions of CRDs as long as they are still used.

Detailed Description

Today it is possible that during a Cluster API upgrade (e.g. when bumping an infra provider) an apiVersion is removed. If this apiVersion was still referenced in other resources, they won't reconcile anymore.

Affected resources:

• ClusterClass
• Cluster, MD, MS, Machine, ...
  • Refs are probably not affected as they don't actually use the apiVersion from the object spec (they calculate the version they use based on the contract label on the CRD).
    • Would be good to double check this again.
  • OwnerRefs
    • They might not be super relevant as our controllers always keep the ownerRefs up-to-date

A possible solution is to implement a webhook which validates CRD updates. In this webhook we could check if the update on the CRD removes an apiVersion which is still referenced in relevant resources (at least ClusterClass).

Note: This webhook should probably have failurePolicy: Ignore to avoid deadlocks if Cluster API is not working for some reason.

Anything else you would like to add?

Somewhat related: #6539 (as it might lead to us dropping apiVersion from all our refs except the ones in ClusterClass)

Label(s) to be applied

/kind feature
One or more /area label. See https://github.com/kubernetes-sigs/cluster-api/labels?q=area for the list of labels.

[sbueringer]

cc @killianmuldoon @srm09 @fabriziopandini

[killianmuldoon]

/triage accepted

[killianmuldoon]

One clarification - how does this relate to

Ensure clusterctl upgrade warns users if the upgrade would break them (e.g. v0.3 => v1.x (version without v1alpha3 apiVersion))
This should already be covered by our CRD migration code, but let's verify it.
This might only become relevant with v1.6 when we actually remove the apiVersion from the CRDs

From #8038

[sbueringer]

Hm totally independent. The point from #8038 is only that clusterctl should correctly fail the CRD update if we jump from a version which doesn't have v1beta1 to a version which drops v1alpha3 (basically because that's not a valid CRD update in general)

[fabriziopandini]

This is a nice check to have, but it comes with a couple of problems to be addressed, top of mind:

1. Where to do the check (webhook or clusterctl)
2. How to encode the knowledge of where references exist

For 1, it seems to me that the simplest way forward is to rely on clusterctl, because it already has all the discovery logic that is required for performing such a check (I'm assuming the discovery logic that exists in clusterctl move can be re-used for this task).

For 2, we can start thinking of a solution that focuses on references existing in core CAPI types, so we cover 90% of the issue. Then, in a follow-up iteration, we can extend this for references existing in provider types.

/help

[k8s-ci-robot]

@fabriziopandini:
This request has been marked as needing help from a contributor.

Guidelines

Please ensure that the issue body includes answers to the following questions:

• Why are we solving this issue?
• To address this issue, are there any code changes? If there are code changes, what needs to be done in the code and what places can the assignee treat as reference points?
• Does this issue have zero to low barrier of entry?
• How can the assignee reach out to you for help?

For more details on the requirements of such an issue, please see here and ensure that they are met.

If this request no longer meets these requirements, the label can be removed
by commenting with the /remove-help command.

In response to this:

This is a nice check to have, but it comes with a couple of problems to be addressed, top of mind:

1. Where to do the check (webhook or clusterctl)
2. How to encode the knowledge of where references exist

For 1, it seems to me that the simplest way forward is to rely on clusterctl, because it already has all the discovery logic that is required for performing such a check (I'm assuming the discovery logic that exists in clusterctl move can be re-used for this task).

For 2, we can start thinking of a solution that focuses on references existing in core CAPI types, so we cover 90% of the issue. Then, in a follow-up iteration, we can extend this for references existing in provider types.

/help

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes/test-infra repository.

[sbueringer]

This is a nice check to have, but it comes with a couple of problems to be addressed, top of mind:

1. Where to do the check (webhook or clusterctl)
2. How to encode the knowledge of where references exist

For 1, it seems to me that the simplest way forward is to rely on clusterctl, because it already has all the discovery logic that is required for performing such a check (I'm assuming the discovery logic that exists in clusterctl move can be re-used for this task).

For 2, we can start thinking of a solution that focuses on references existing in core CAPI types, so we cover 90% of the issue. Then, in a follow-up iteration, we can extend this for references existing in provider types.

1. I would do it in the webhook so everyone profits not only clusterctl users. (we don't need the discovery logic if we only implement it for core resources)
2. I think we don't have to encode that knowledge if we only implement it for core CAPI resources. Specifically I think only ClusterClass, in all other resources we automatically bump the apiVersion based on contract (which is also what providers should do). ClusterClass should be the only case where we intentionally use specific apiVersions (because we have to).

[ykakarap]

So a validating webhook for kind: CustomResourceDefinition that checks that when a CRD whose .spec.group is one of controlplane.cluster.x-k8s.io or bootstrap.cluster.x-k8s.io or infrastructure.cluster.x-k8s.io is updated does not drop a version that is used by at least one ClusterClass. Is my understanding correct?

[sbueringer]

Pretty much. I think we don't have to filter early by group. We can just take the group+ kind etc and check if it occurs in a ClusterClass

This way we can also support CRDs in other groups. Afaik there is no contract that a provider has to use one of those groups