✨ WIP: Cluster provider and cluster-aware controllers

[embik]

This is a continuation of #2726 and #2207, a prototype for a multi-cluster support in controller-runtime. A design proposal is at #2746 which will need updating depending on the feedback for this PR. I've reorganized code changes into different commits than the previous attempts, but tried to keep authorship of changes faithful to the commits this work is based on.

The main change from the previous PR (#2726) is the update to typed generics support and adjustment to have a BYO request type. The implications of that is a) that you will need to bring your own typed EventHandler and b) that a wrapper function is needed to inject information (mostly the cluster name) when establishing a handler.

Apart from that, the core design of this is the cluster provider that can be plugged into a manager. This unlocks different ways to discover the Kubernetes "fleet" that is being reconciled.

[sprsquish]

We're internally working on a mechanism for dealing with multiple clusters, too. It's been helpful for the cluster identifier to be more flexible than a simple string. We've made it comparable.

Something like this:

type Cluster[ID comparable] interface {
  Identifier() ID
  ...
}

I understand this suggestion would cause a much larger change requiring type parameters in many places. Future work could be done to migrate.

Mostly, I wanted to raise this as something to consider and that calling the method Identifier would provide room to migrate in the future.

[embik]

I'm definitely open to that, I think renaming it to Identifier() should be easy enough! Maybe some maintainers can leave a note if they consider this a future endeavour and this would be a worthwhile change to prepare that.

[sttts]

We were there before, had a logical cluster abstraction, and then went back to string as it just simplifies everything.

@sprsquish Am curious why a string as identifier and some lookup table on the cluster provider side does not work?

[sttts]

Also, adding Name() string to the interface has no impact on the consumers of the Cluster type. But adding a type parameter forces them to pass e.g. string or whatever everywhere including the manager. I don't think that's a good idea. IMO, we should not do it.

[sprsquish]

@sttts

Am curious why a string as identifier and some lookup table on the cluster provider side does not work?

Strings work fine.. We like to work with more structure in our identifiers. For instance, working with multiple clusters across cloud providers it's nice to have a structure like this:

type ClusterID struct {
  Provider string
  Region string
  Account string
  Name string
}

It can be marshaled into a string, of course, it's just nice not to have to.

I understand the need for a simpler approach and appreciate that this has been discussed already. I think there's a way to migrate over time that follows what's been happening with handlers by introducing TypedCluster[ID comparable] and an alias for the common use case type Cluster = TypedCluster[string] (Manager would need the same treatment).

The only change to the current proposal would be to use the more generic Identifier() rather than Name(). That leaves the door open for future discussions regarding generic identifiers without having to do all the work up front.

[sttts]

If I understand your use-case correctly, this pattern should help you elegantly: https://goplay.space/#wRWyxv-GUrj

Or in other words: the type only matters in code that is ClusterID aware. Turning the string back into the rich format is one type conversion away.

[sttts]

is the default cluster a thing here already? I.e. can I always pass empty string?

[sttts]

Should we specify what is returned for an unknown cluster?

[embik]

The default cluster is not a thing here, that is in Manager.GetCluster. The cluster provider doesn't have a concept of a default cluster.

Should we specify what is returned for an unknown cluster?

Probably. I can add a sentence to clarify this.

[embik]

Added a note on what should happen if the cluster is unknown.

[sttts]

"defines methods" is redundant for an interface.

Suggested change

	// Provider defines methods to retrieve clusters by name. The provider is
	// responsible for discovering and managing the lifecycle of each cluster.
	//
	// Provider allows to retrieve clusters by name. The provider is
	// responsible for discovering and managing the lifecycle of each cluster.
	//

It's kind of unclear what managing means in a getter interface.

[sttts]

Maybe be concrete saying:

Suggested change

	// Provider defines methods to retrieve clusters by name. The provider is
	// responsible for discovering and managing the lifecycle of each cluster.
	//
	// Provider allows to retrieve clusters by name, e.g. to reconcilers.
	// The provider is responsible for discovering and managing the lifecycle
	// of each cluster, calling `Engage` and `Disengage` on the manager
	// it is run against, whenever a new cluster is discovered or a cluster
	// is unregistered.

[embik]

+1, I've added that.

[k8s-ci-robot]

[APPROVALNOTIFIER] This PR is NOT APPROVED

This pull-request has been approved by: embik
Once this PR has been reviewed and has the lgtm label, please assign joelanford for approval. For more information see the Code Review Process.

The full list of commands accepted by this bot can be found here.

Details

Needs approval from an approver in each of these files:

• OWNERS

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

[k8s-ci-robot]

@embik: The following test failed, say /retest to rerun all failed tests or /retest-required to rerun all mandatory failed tests:

Test name	Commit	Details	Required	Rerun command
pull-controller-runtime-apidiff	e93979e	link	false	/test pull-controller-runtime-apidiff

Full PR test history. Your PR dashboard. Please help us cut down on flakes by linking to an open issue when you hit one in your PR.

Details

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-sigs/prow repository. I understand the commands that are listed here.

[varshaprasad96]

In the process of reviewing the PR in detail, for now most of my comments are either nits or questions.

This is a really helpful feature in controller-runtime, that can benefit other internal kubernetes-sigs projects.
cc: @sbueringer @joelanford Could we have some eyes on this and the proposal please. Thank you!

[varshaprasad96]

I'm not sure if it matters, I'll leave it up to you - the former is comparing parameter type at runtime and the other is checking for its type at during compile time. Isn't it better to be consistent?

Suggested change

	case reflect.TypeOf(reconcile.Request{}):
	case reflect.TypeFor[reconcile.Request]():

[embik]

Good point, but I tried to just adapt the previous if condition. I'm not opposed to changing it, I just wanted to not fiddle with the condition checked.

[varshaprasad96]

Just curious on why we need to use reflect and set the value instead of doing a direct assignment? If we are using reflect.TypeOf previously, aren't we sure of the type at runtime?
Just to make reading of the code easier.

[embik]

I've asked myself the same question when I stumbled over this part of the code (see below, it's how the code today sets the handler implementation as well). When I try to change it to something like:

case reflect.TypeOf(reconcile.ClusterAwareRequest{}):
    hdler = handler.WithLowPriorityWhenUnchanged(&handler.EnqueueClusterAwareRequestForObject{ClusterName: cl.Name()})

it fails with:

../../pkg/builder/controller.go:335:12: cannot use handler.WithLowPriorityWhenUnchanged(&handler.EnqueueClusterAwareRequestForObject{…}) (value of type handler.TypedEventHandler[client.Object, reconcile.ClusterAwareRequest]) as handler.TypedEventHandler[client.Object, request] value in assignment: handler.TypedEventHandler[client.Object, reconcile.ClusterAwareRequest] does not implement handler.TypedEventHandler[client.Object, request] (wrong type for method Create)
		have Create(context.Context, event.TypedCreateEvent[client.Object], workqueue.TypedRateLimitingInterface[reconcile.ClusterAwareRequest])
		want Create(context.Context, event.TypedCreateEvent[client.Object], workqueue.TypedRateLimitingInterface[request])

My understanding of generics here is a bit limited, but I think this is because we expect TypedRateLimitingInterface to take whatever generic request we give into the function, but we are setting it to a concrete type. I suppose this only works with the reflect magic we currently have.

[varshaprasad96]

Ah, thanks for explaining @embik!

[embik]

No problem! This is at least how I understand this line. It's a bit unclear to me too.

[varshaprasad96]

It would be super helpful if we can expose the set of engagedClusters from the manager. In Kueue for a multi cluster implementation, we have a list of active and missing clusters based on available quota for admitting workloads.

[embik]

For my understanding, you'd be looking for a list of the keys for that map, is that correct (i.e. the cluster identifiers, not the cluster.Cluster objects directly)?

[varshaprasad96]

Yes, a unique identifier for the cluster, so that we can fetch the relevant client and submit workloads in that cluster.

[varshaprasad96]

Please correct me if I'm wrong - While calling Engage we are starting a SyncingSource from multicluster cache (

controller-runtime/pkg/builder/controller.go

Line 397 in e93979e

src := &ctxBoundedSyncingSource[request]{ctx: ctx, src: source.TypedKind(cl.GetCache(), projected, handler.WithLowPriorityWhenUnchanged(h), allPredicates...)}

), but in disengage we aren't stopping the context anywhere in here:

controller-runtime/pkg/controller/multicluster.go

Line 126 in e93979e

func (c *typedMultiClusterController[request]) Disengage(ctx context.Context, cl cluster.Cluster) error {

. Is this expected?

[embik]

The way this works (but I wonder if it's confusing and should be changed to be honest) is that the implementation expects the cluster provider to cancel the clusterCtx used for Engage (and subsequently, the ctxBoundedSyncingSource) when shutting down. Compare the fleet example here canceling the stored context when figuring out that a cluster has to be removed: https://github.com/embik/controller-runtime/blob/e93979ec5cdd656809f5fbb91b858665979cc040/examples/fleet/main.go#L280

// disengage manager
if mgr != nil {
	if err := mgr.Disengage(ctx, k.clusters[name]); err != nil {
		k.log.Error(err, "failed to disengage manager")
        }
}

// stop and forget
k.lock.Lock()
k.cancelFns[name]() // <- HERE
delete(k.clusters, name)
delete(k.cancelFns, name)
k.lock.Unlock()

The ctx passed to Disengage is a different context, one that is currently unused but could be used when disengaging should be cancelable.

I'm open to input here though. I was wondering if the manager should even have a Disengage (and instead clean up its internal cluster list when the respective context gets canceled), but then the manager itself wouldn't be a cluster.Aware.

[sttts]

Suggested change

	// EngageWithProvidedClusters indicates whether the controller should engage
	// EngageWithProviderClusters indicates whether the controller should engage

[embik]

As per #2746 (comment), we're focussing on getting the design presented in this PR up and running in https://github.com/multicluster-runtime/multicluster-runtime. Feel free to follow along on the new project if you are interested.