Controling resource deletions

[vavdoshka]

Is your feature request related to a problem? Please describe.

We are at early stage of adopting the messaging-topology-operator and graduating it's usage to production. Currently removal of a k8s resource triggers the removal of that object in RMQ, we don't feel it safe for our production workloads and would like to control the removal behavior.

Describe the solution you'd like

As a flag provided in runtime controlling resource deletion behavior (and/or) on the resource level.

Describe alternatives you've considered

Using admission controller to prevent resource removal

Additional context

topology-operator can reconcile resources which were requested previously, so it seems the missing piece is logic to ignore delete.
relevant request on rmq-server side rabbitmq/rabbitmq-server#12772

[Zerpet]

A flag to skip deletion of topology objects defeats the point of having an Operator in the first place. The Operator is meant to abstract the management and lifecycle of topology objects in RabbitMQ. Skipping deletion removes a fundamental part of the Operator functionality, and would result in a mismatch between the Kubernetes state and RabbitMQ state; in other words, it would result in objects not present in Kubernetes, and present in RabbitMQ.

[vavdoshka]

Introducing a flag to control resource deletion behavior enhances the safety and flexibility of the Messaging Topology Operator without undermining its purpose. Many successful operators, like Crossplane, ArgoCD, Cert-Manager and others provide similar mechanisms to address real-world production needs, ensuring critical resources are not inadvertently deleted. While mismatches between Kubernetes and RabbitMQ states are a concern, they can be mitigated through logging and reconciliation mechanisms. Additionally, we already tested the reconciliation of resources created by the Operator after simulating the orphaned deletion and the existing implementation successfully detected and managed the orphaned resource.

Being used wisely this feature should allow easier and broader adoption of the operator for the both teams being early in their k8s journey and teams managing very sensitive RMQ workloads.

[ansd]

@vavdoshka Why can't you simply disallow deletion of these K8s objects via RBAC (Role Based Access Control)?

[ansd]

Also, please reference where these other projects support this feature.

A quick Google search suggests that this is an alpha feature in Crossplane:
https://docs.crossplane.io/latest/concepts/usages/

It's an open issue in ArgoCD:
argoproj/argo-cd#7301
argoproj/argo-cd#16063

[Zerpet]

Introducing a flag to control resource deletion behavior enhances the safety and flexibility of the Messaging Topology Operator without undermining its purpose.

I disagree. The mission of the Operator is to reconcile Kubernetes state into RabbitMQ. This flag is completely opposed to that mission.

ensuring critical resources are not inadvertently deleted

How can you inadvertently delete a resource when a deletion requires a explicit kubectl delete action, or an API DELETE request? It sounds like we are trying to fix human carelessness, and inherently unfixable problem.

Being used wisely this feature should allow easier and broader adoption of the operator for the both teams being early in their k8s journey and teams managing very sensitive RMQ workloads.

Given that this is the first time this feature has been requested in the ~4 years of existence of this Operator, I'm highly skeptical that it will help to spread adoption.

[vavdoshka]

Thanks for the feedback. As you rightfully mentioned, there is a logical mistake in mocking the delete; however, it is not exactly what I propose. I propose to let users decide whether to optionally retain the resources that might contain a state influenced by factors outside the operator's responsibility or concern, i.e., queues or streams with data. Optional retain behavior, similar to what K8s implements for PVC Reclaim Policy. Because there might be use cases where deletion is appropriate and, in some cases, not so much. In addition one use-case which we did not face yet but fesable is migration (split/merge) of resources from one namespace to another.

In terms of what currently supports these use cases, there are two schools indeed, for example, the Kafka Strimzi operator does not support this, while at the same time:

• The ArgoCD ApplicationSet controller has elaborate support.
• The Zalando Postgres Operator supports this.
• AWS Controllers support this.
• The RabbitMQ operator implicitly implements this. It is not exactly the "retain" feature but it solves both problems mentioned above nevertheless.

We are trying to understand if maintainers would be opened to such contribution at current moment in order to plan accordingly the actions on our side. Thanks for this great Operator and your feedback.