Retain PVCs on PostgresCluster delete

[nbrns]

Overview

When deleting a PostgresCluster CRD, the operator also deletes it's PVCs. I'd love to see a configuration option on the operator/cluster CRD that allows us to retain the PVC and the attached PV.

I know there is a guide to retain the PVs by doing some manual patching. However, it is cumbersome to reattach the PVs to a new Cluster, especially in a fully automated scenario. Newly provisioned persistent volumes claim do not work well with label selectors (e.g. dynamic provisioning with selector labels is not supported on AKS and EKS).

Use Case

• Easier recovery by mounting existing PVCs to StatefulSet deployed by the operator instead of having to reattach PVs to PVCs
• Many operators keep the PVCs by default, i.e. they do not prune the data upon CRD deletion. This is more intuitive in my opinion

Desired Behavior

A configuration option on either the operator or the PostgresCluster CRD, that allows to retain PVCs on PostgresCluster deletion. E.g. retainPvcs: true

Environment

Tell us about your environment:

Please provide the following details:

• Platform: AKS
• Platform Version: 1.26.3
• PGO Image Tag: (e.g. ubi8-5.3.1-0)
• Postgres Version (e.g. 15)
• Storage: (e.g. hostpath, nfs, or the name of your storage class)
• Number of Postgres clusters: 2

Additional Information

NA

[tjmoore4]

Hello @nbrns. Thank you for your use-case information. I've added a note about this feature request to our backlog for future development. Regarding your specific scenario, do I understand correctly that you always keep all associated PVCs or do you only need to retain a subset of them?

[nbrns]

Thanks for noting our use case!

I would like to keep all PVCs associated to a Postgres Cluster and its respective StatefulSets

[tjmoore4]

Understood, thanks! I've updated our backlog accordingly.

[szelenka]

This would be a very welcome addition, similar to #3654

[rgherta]

I think we need to understand what layer should handle this logic otherwise the postgres-operator will become more complex than it should and in time difficult to maintain.

When you setup your cluster, you normally set up storage provisioner(s) - lvm, S3, nfs etc... Based on this you define some storage classes that describe your data and refer one of these storage provisioners.

Storage classes have a specific key called

StorageClass.reclaimPolicy

By setting its value to "Retain" you make sure the PVs will be retained for manual reclamation by the cluster admin.

This storage class you will reference in the postgrescluster crd via the "storageClassName" key.

And so the storage provisioner will handle the lifecycle of the PVCs and not the postgres operator. Otherwise conflicts may arise because postgres-operator will manage a resource that is the storage controller's responsibility.

[cawilliams97]

Is this feature request still being tracked? +1 to its utility.

As things currently stand, is there a way purely using declarative YAML to configure that PVs associated with a given PostgresCluster are to be retained rather than deleted? I am seeing tutorials like this: https://www.crunchydata.com/blog/keep-storage-and-backups-when-dropping-cluster-with-pgo

But they require running shell commands after the cluster has already started. I'm working within constraints that disallow such an approach - our cluster needs to be fully configurable purely using YAML that we deploy once.