Auto-Restoring deleted ServiceImports and EndPointSlices

[davidohana]

Not sure whether to call this a bug or an enhancement.

What happened:

• Deleting a service import or an endpoint slice locally - it stays deleted locally and deletion is synced to the broker.
• Deleting a service import or an endpoint slice from the broker, nothing happen

What you expected to happen:

The agent shall recreate the service import / endpoint slice after some time as the "source of truth" is the export and it still exists. Currently this only happens after lighthouse agent restart.

Additional notes:

To compare, if I delete a k8s endpoint slice, it is recreated immediately.
I realize that those edge cases are only possible if someone with sufficient permissions delete those objects manually.

Slack discussion link: https://kubernetes.slack.com/archives/C010RJV694M/p1645966829687039

[vthapar]

@davidohana Could you clarify what you mean by deleting locally? I expect you mean deleting locally on the source cluster where export exists, but just wanted to be sure it is so.

[davidohana]

Yes, that what I mean - deleting on the origin cluster.

[dfarrell07]

There are likely other objects that would fail in similar ways if a user deletes them, because they are internal objects that we don't expect users to mess with. To handle this properly we would need to make a more systematic review of the code base, which would likely be benefited by an EP and a broader discussion about what should change and how.

[stale]

This issue has been automatically marked as stale because it has not had activity for 60 days. It will be closed if no further activity occurs. Please make a comment if this issue/pr is still valid. Thank you for your contributions.

[dfarrell07]

@vthapar pointed out that finalizers may be a good, low-overhead way of doing this.

[stale]

This issue has been automatically marked as stale because it has not had activity for 60 days. It will be closed if no further activity occurs. Please make a comment if this issue/pr is still valid. Thank you for your contributions.

[dfarrell07]

Discussing this again, it's important to note that monitoring resources cross-many-clusters for deletion comes with overhead that doesn't exist for the parallel of an endpoint slices in one K8s cluster. If there's a usecase for this, we'd be happy to discuss that and prioritize this work. Again, this likely applies to other resources and finalizers might enable this with reasonable overhead.

This might be kinda-related to submariner-io/enhancements#161, or at least a good time to think about finalizers for various resources.

[stale]

This issue has been automatically marked as stale because it has not had activity for 60 days. It will be closed if no further activity occurs. Please make a comment if this issue/pr is still valid. Thank you for your contributions.

[tpantelis]

While a finalizer would prevent out-of-band deletion, we wouldn't be able to update the resource since it's in the process of being deleted.

A simple solution would be to specify a resync period with the informers.

[tpantelis]

I tested finalizer behavior and and after deletion with a finalizer still present, the resource can still be updated. You just can't modify the DeletionTimestamp nor add another finalizer. However using a finalizer merely to prevent out-of-band deletion can be problematic if the resource isn't deleted by the controller that added the finalizer. So if Submariner isn't properly uninstalled, finalizers would remain and block namespace deletion requiring manual removal of the finalizers.

A safer solution is to recreate a resource on out-of-band deletion. We can do this fairly simply via the resource syncer. On deletion of a resource, check if the originating resource still exists and, if so, re-queue it so the deleted resource is re-created. If not then process the deleted resource normally. Eg, ServiceImports are created from ServiceExports, so on deletion of a ServiceImport check if the originating ServiceExport still exists in the resource syncer cache.

[t0lya]

My team runs a multi-cluster setup with Submariner's service discovery setup (v0.15.2), where we might be seeing a case of this issue.

Our workloads can be placed on particular cluster based on nodeSelectors in the Deployment spec. When a Deployment's nodeSelectors are updated in a way that triggers a cluster placement change, we have a controller that deletes Service/ServiceExport/Deployment from the old source cluster and immediately re-creates these 3 objects in new source cluster.

We are seeing that this results in the deletion of aggregated ServiceImport by submariner-lighthouse-agent running on old source cluster. submariner-lighthouse-agent running on new source cluster never re-creates the aggregated ServiceImport.

To mitigate the issue, we had to manually re-create Service/ServiceExport on the new cluster to trigger ServiceImport re-creation on the broker cluster.

@tpantelis Will your solution handle above scenario? And is there a timeline for when a solution will be rolled out?

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further
activity occurs. Thank you for your contributions.