Every database cluster archives its write-ahead log to Garage continuously and takes base backups on the schedule in tenant-workload/database/backups/, kept for the retentionPolicy in its backup section. Together they let you recover the data as it was at any moment in that window. This guide restores the current primary into a separate, new cluster, so the running database is not touched. The primary is cluster-1 until a failover, then whichever cluster carries bedrockctl.com/db-role: primary:
kubectl get clusters.postgresql.cnpg.io -A -l bedrockctl.com/db-role=primaryBelow, <primary> is its name, which is also its namespace.
#Before you begin
- Know the moment you want, in UTC, for example
2026-10-19 09:30:00+00. - Make sure Longhorn has room for one more volume the size of the primary’s.
#Write the restore cluster
Create restore.yaml. It lives in the primary’s namespace, so it can reach Garage and already has the backup credentials:
apiVersion: postgresql.cnpg.io/v1
kind: Cluster
metadata:
name: <primary>-restore
namespace: <primary>
labels:
bedrockctl.com/db-role: unmanaged
spec:
instances: <instances>
priorityClassName: bedrock-data
imageName: <imageName of the primary>
enableSuperuserAccess: false
postgresql:
pg_hba:
- hostnossl all all all reject
storage:
size: <storage size of the primary>
storageClass: longhorn-encrypted
env:
- name: AWS_REQUEST_CHECKSUM_CALCULATION
value: when_required
- name: AWS_RESPONSE_CHECKSUM_VALIDATION
value: when_required
- name: AWS_DEFAULT_REGION
value: garage
bootstrap:
recovery:
source: primary-backup
recoveryTarget:
targetTime: "2026-10-19 09:30:00+00"
externalClusters:
- name: primary-backup
barmanObjectStore:
serverName: <primary>
destinationPath: s3://<primary>-backups/wals
endpointURL: http://garage.garage.svc:3900
s3Credentials:
accessKeyId:
name: <primary>-backups
key: AWS_ACCESS_KEY_ID
secretAccessKey:
name: <primary>-backups
key: AWS_SECRET_ACCESS_KEY
wal:
compression: gzip
data:
compression: gzipFill in the placeholders. Copy imageName and the storage size from the primary’s cluster (kubectl -n <primary> get cluster <primary> -o yaml), since the restored data must fit and the PostgreSQL major version must match. The unmanaged role label tells the failover controller to leave this cluster alone, so it is neither watched nor reported as a problem. A single instance is enough to inspect the data; use more if the restored cluster will serve traffic. Change targetTime to your moment. The file has no backup section on purpose: the restored cluster must never write into the primary’s bucket.
#Run the restore
kubectl apply -f restore.yaml
kubectl -n <primary> get cluster <primary>-restore -wCloudNativePG fetches the base backup taken before your moment, replays the WAL up to it and starts the instance. When the status reads Cluster in healthy state, the data is ready.
#Check the data
kubectl -n <primary> exec <primary>-restore-1 -c postgres -- psql -U postgres -d app -c "\dt"Query the tables you care about and compare with what you expected at that moment.
#Use or discard it
The restored cluster is a normal CloudNativePG cluster. You can copy rows out of it with pg_dump and load them into the primary, or keep it while you investigate. When you are done, delete it:
kubectl delete -f restore.yamlIts volume is deleted with it.