Shared provider setup
Before any tool: CSI snapshots
Section titled “Before any tool: CSI snapshots”All three tools capture volumes with CSI snapshots. Check that your cluster supports them:
kubectl get volumesnapshotclasskubectl get storageclassYou need one VolumeSnapshotClass for your CSI driver and a StorageClass that the driver
serves. The examples use the kind dev cluster names csi-hostpath-snapclass and
csi-hostpath-sc. Replace them in the files, or patch them with a kustomize overlay.
Tested with csi-driver-host-path v1.19.0 and external-snapshotter v8.7.0. For production, use
the snapshot controller and driver that your storage vendor documents.
Every tool copies the snapshot data to S3. Use one bucket, or one prefix, per tool. Do not share a repository between tools.
The docs use the endpoint s3.example.internal. Replace it with your own. The example files use
versitygw.storage.svc.cluster.local:7070, the endpoint of the kind dev cluster: Versity Gateway
with the posix backend and the buckets kopiur, velero and volsync. MinIO’s open source
distribution is no longer maintained, so this project does not use it.
Keep S3 keys and repository passwords out of git. Create the Secrets with kubectl, for example:
kubectl -n wordpress create secret generic kopiur-repo-creds \ --from-literal=AWS_ACCESS_KEY_ID=<access key> \ --from-literal=AWS_SECRET_ACCESS_KEY=<secret key> \ --from-literal=KOPIA_PASSWORD=<a strong password you keep safe>Each tool page names the Secret and keys it needs.
Point in time across PVCs
Section titled “Point in time across PVCs”Kopiur and VolSync back up each PVC on its own. The html and mysql backups of one run can be
minutes apart, so the restored pair is not one point in time. A check that compares the two, as
the WordPress check does, can then fail on a write that happened between the backups. Back up
the PVCs close together, or quiesce the app during the backup.
Velero backs up all PVCs of the namespace in one Backup. Its PVCs are one point in time.
Throttle restore throughput
Section titled “Throttle restore throughput”A restore test reads the whole backup from S3 and writes it to a volume. It can saturate the network or the disks. bret has no throttle field. Each tool has its own levers. Checked against the tool docs in November 2026.
Kopiur
Section titled “Kopiur”- Kopia throttle:
Repository.spec.moverDefaults.throttlewithdownloadBytesPerSecond,uploadBytesPerSecond,readOpsPerSecond,writeOpsPerSecond.- Pro: a real bandwidth and ops cap. Restore movers honor it. No CNI support needed.
- Con: set per repository, so it also slows backup, maintenance and replication movers.
- Con: no per-
Restorefield. - Con: cached content bypasses the cap. Small objects stay well below the cap.
- Restore parallelism:
Restore.spec.options.parallel, default8.- Pro: set per restore.
- Con: limits concurrency, not bytes or ops.
- Mover pod annotations:
Repository.spec.moverDefaults.podAnnotations.- Pro: carries the
kubernetes.io/ingress-bandwidthannotation. - Con: set per repository, so it hits every mover. Needs a CNI that enforces ingress limits.
- Pro: carries the
Velero
Section titled “Velero”- No bandwidth or IOPS cap.
- Files in parallel:
velero restore create --parallel-files-download=<n>.- Pro: set per restore.
- Con: limits concurrency only. Only the file-system data mover. No effect on block mode.
- node-agent ConfigMap
loadConcurrency(globalConfig,perNodeConfig).- Pro: limits the data mover pods per node.
- Con: set per install. Also limits backups. Needs a node-agent restart.
- node-agent ConfigMap
podResources(cpuLimit,memoryLimit).- Pro: caps CPU of the restore pods.
- Con: an indirect brake, not a throughput cap. Set per install.
- node-agent ConfigMap
podAnnotations/podLabels(PodVolumeRestoreandDataDownloadpods).- Pro: carries the
kubernetes.io/ingress-bandwidthannotation. - Con: set per install, so it also hits backups. Replaces Velero’s own annotations, such as IAM role annotations, so add those back. Needs a CNI that enforces ingress limits.
- Pro: carries the
VolSync
Section titled “VolSync”- No bandwidth or IOPS cap.
moverResources(CPU, memory) on the mover spec.- Pro: set per
ReplicationDestination, so only the restore. - Con: an indirect brake, not a throughput cap. Tight limits can crash the mover.
- Pro: set per
moverPodLabels.- Pro: set per
ReplicationDestination. - Con: labels only, no annotations, so the bandwidth annotation cannot be set.
- Pro: set per
- Restic env from the repository Secret.
- Con: the docs name no bandwidth variable.
Pod bandwidth annotation
Section titled “Pod bandwidth annotation”kubernetes.io/ingress-bandwidthon the mover pod caps the download.- Pro: the same mechanism for every tool.
- Con: works only if the CNI enforces it, for example the CNI
bandwidthplugin. Cilium ignores the ingress annotation. - Con: no tool can set it per restore. VolSync cannot set it at all. A mutating webhook would be needed to inject it into restore pods only.
- Kubernetes has no per-pod disk IOPS limit. Use storage QoS (StorageClass parameters) instead.
