Backup on volume snapshots

CloudNativePG is one of the first known cases of database operators that directly leverages the Kubernetes native Volume Snapshot API for both backup and recovery operations, in an entirely declarative way.

About standard Volume Snapshots

Volume snapshotting was first introduced in Kubernetes 1.12 (2018) as alpha, promoted to beta in 1.17 (2019), and moved to GA in 1.20 (2020). It’s now stable, widely available, and standard, providing 3 custom resource definitions: VolumeSnapshot, VolumeSnapshotContent and VolumeSnapshotClass.

This Kubernetes feature defines a generic interface for:

• the creation of a new volume snapshot, starting from a PVC
• the deletion of an existing snapshot
• the creation of a new volume from a snapshot

Kubernetes delegates the actual implementation to the underlying CSI drivers (not all of them support volume snapshots). Normally, storage classes that provide volume snapshotting support incremental and differential block level backup in a transparent way for the application, which can delegate the complexity and the independent management down the stack, including cross-cluster availability of the snapshots.

Requirements

For Volume Snapshots to work with a CloudNativePG cluster, you need to ensure that each storage class used to dynamically provision the PostgreSQL volumes (namely, storage and walStorage sections) support volume snapshots.

Given that instructions vary from storage class to storage class, please refer to the documentation of the specific storage class and related CSI drivers you have deployed in your Kubernetes system.

Normally, it is the VolumeSnapshotClass that is responsible to ensure that snapshots can be taken from persistent volumes of a given storage class, and managed as VolumeSnapshot and VolumeSnapshotContent resources.

How to configure Volume Snapshot backups

CloudNativePG allows you to configure a given Postgres cluster for Volume Snapshot backups through the backup.volumeSnapshot stanza.

A generic example with volume snapshots (assuming that PGDATA and WALs share the same storage class) is the following:

apiVersion: postgresql.cnpg.io/v1
kind: Cluster
metadata:
  name: snapshot-cluster
spec:
  instances: 3

  storage:
    storageClass: @STORAGE_CLASS@
    size: 10Gi
  walStorage:
    storageClass: @STORAGE_CLASS@
    size: 10Gi

  backup:
    # Volume snapshot backups
    volumeSnapshot:
       className: @VOLUME_SNAPSHOT_CLASS_NAME@
    # WAL archive
    barmanObjectStore:
       # ...

As you can see, the backup section contains both the volumeSnapshot stanza (controlling physical base backups on volume snapshots) and the barmanObjectStore one (controlling the WAL archive).

The volumeSnapshot.className option allows you to reference the default VolumeSnapshotClass object used for all the storage volumes you have defined in your PostgreSQL cluster.

Once a cluster is defined for volume snapshot backups, you need to define a ScheduledBackup resource that requests such backups on a periodic basis.

Hot and cold backups

By default, CloudNativePG requests an online/hot backup on volume snapshots, using the PostgreSQL defaults of the low-level API for base backups:

• it doesn't request an immediate checkpoint when starting the backup procedure
• it waits for the WAL archiver to archive the last segment of the backup when terminating the backup procedure

You can explicitly change the default behavior through the following options in the .spec.backup.volumeSnapshot stanza of the Cluster resource:

• online: accepting true (default) or false as a value
• onlineConfiguration.immediateCheckpoint: whether you want to request an immediate checkpoint before you start the backup procedure or not; technically, it corresponds to the fast argument you pass to the pg_backup_start/pg_start_backup() function in PostgreSQL, accepting true (default) or false
• onlineConfiguration.waitForArchive: whether you want to wait for the archiver to process the last segment of the backup or not; technically, it corresponds to the wait_for_archive argument you pass to the pg_backup_stop/pg_stop_backup() function in PostgreSQL, accepting true (default) or false

If you want to change the default behavior of your Postgres cluster to take cold backups by default, all you need to do is add the online: false option to your manifest, as follows:

  # ...
  backup:
    volumeSnapshot:
       online: false
       # ...

If you are instead requesting an immediate checkpoint as the default behavior, you can add this section:

  # ...
  backup:
    volumeSnapshot:
       online: true
       onlineConfiguration:
         immediateCheckpoint: true
       # ...

Overriding the default behavior

You can change the default behavior defined in the cluster resource by setting different values for online and, if needed, onlineConfiguration in the Backup or ScheduledBackup objects.

For example, in case you want to issue an on-demand cold backup, you can create a Backup object with .spec.online: false:

apiVersion: postgresql.cnpg.io/v1
kind: Backup
metadata:
  name: snapshot-cluster-cold-backup-example
spec:
  cluster:
    name: snapshot-cluster
  method: volumeSnapshot
  online: false

Similarly, for the ScheduledBackup:

apiVersion: postgresql.cnpg.io/v1
kind: ScheduledBackup
metadata:
  name: snapshot-cluster-cold-backup-example
spec:
  schedule: "0 0 0 * * *"
  backupOwnerReference: self
  cluster:
    name: snapshot-cluster
  method: volumeSnapshot
  online: false

Persistence of volume snapshot objects

By default, VolumeSnapshot objects created by CloudNativePG are retained after deleting the Backup object that originated them, or the Cluster they refer to. Such behavior is controlled by the .spec.backup.volumeSnapshot.snapshotOwnerReference option which accepts the following values:

• none: no ownership is set, meaning that VolumeSnapshot objects persist after the Backup and/or the Cluster resources are removed
• backup: the VolumeSnapshot object is owned by the Backup resource that originated it, and when the backup object is removed, the volume snapshot is also removed
• cluster: the VolumeSnapshot object is owned by the Cluster resource that is backed up, and when the Postgres cluster is removed, the volume snapshot is also removed

In case a VolumeSnapshot is deleted, the deletionPolicy specified in the VolumeSnapshotContent is evaluated:

• if set to Retain, the VolumeSnapshotContent object is kept
• if set to Delete, the VolumeSnapshotContent object is removed as well

The value in VolumeSnapshotContent is determined by the deletionPolicy set in the corresponding VolumeSnapshotClass definition, which is referenced in the .spec.backup.volumeSnapshot.className option.

Please refer to the Kubernetes documentation on Volume Snapshot Classes for details on this standard behavior.

Example

The following example shows how to configure volume snapshot base backups on an EKS cluster on AWS using the ebs-sc storage class and the csi-aws-vsc volume snapshot class.

The following manifest creates a Cluster that is ready to be used for volume snapshots and that stores the WAL archive in a S3 bucket via IAM role for the Service Account (IRSA, see AWS S3):

apiVersion: postgresql.cnpg.io/v1
kind: Cluster
metadata:
  name: hendrix
spec:
  instances: 3

  storage:
    storageClass: ebs-sc
    size: 10Gi
  walStorage:
    storageClass: ebs-sc
    size: 10Gi

  backup:
    volumeSnapshot:
       className: csi-aws-vsc
    barmanObjectStore:
      destinationPath: s3://@BUCKET_NAME@/
      s3Credentials:
        inheritFromIAMRole: true
      wal:
        compression: gzip
        maxParallel: 2

  serviceAccountTemplate:
    metadata:
      annotations:
        eks.amazonaws.com/role-arn: "@ARN@"
---
apiVersion: postgresql.cnpg.io/v1
kind: ScheduledBackup
metadata:
  name: hendrix-vs-backup
spec:
  cluster:
    name: hendrix
  method: volumeSnapshot
  schedule: '0 0 0 * * *'
  backupOwnerReference: cluster
  immediate: true

The last resource defines daily volume snapshot backups at midnight, requesting one immediately after the cluster is created.