Post

Changed Block Tracking for OpenShift VMs

OpenShift 5 has two Changed Block Tracking previews, CSI and QEMU. Both remain Technology Preview. Incremental backup still needs its own timed restore.

Changed Block Tracking for OpenShift VMs

Personal site note: views expressed here are my own and do not necessarily represent Red Hat or any other organization.

On November 11, 2026, Red Hat’s OpenShift Virtualization 5 post described storage-agnostic incremental backups powered by QEMU Changed Block Tracking (CBT), as a Technology Preview, with Commvault, Cohesity, Dell, and Rubrik. The same release train documents a second CBT feature on the Container Storage Interface (CSI) side.

A design review that says “CBT” without identifying the bitmap risks testing the wrong thing. One bitmap lives in QEMU and libvirt. The other lives in a CSI driver, between two snapshots. Both are previews. Neither feature is a disaster recovery (DR) design.

My OADP for OpenShift VMs: Backup Is Not DR post explains the distinction between restoring a backup and declaring a failover. This post is the incremental layer people will put on top of that split. It is a solutions-architect map as of November 13, 2026, not a backup-product install guide.

Two bitmaps, two owners

 QEMU / KubeVirtCSI
What it diffsDisk blocks changed since a libvirt checkpointBlocks changed between two snapshots of one volume, plus allocated ranges in a single snapshot
Who implements itThe backup product, against the KubeVirt backup APIThe CSI driver and the backup product
StorageStorage-agnostic. The array does not keep the bitmapOnly drivers that implement CSI CBT
What comes backThe changed blocks, copied by the backup productMetadata. The backup product reads the snapshot data itself
Support statement I can cite todayThe November 11 product blog calls it Technology Preview. The OpenShift Virtualization 5.0 release notes do not list itKCS 7146879 and the OKD 5 CSI CBT overview: Technology Preview on OpenShift 5.0

Do not put either path in a production runbook. Technology Preview features are not covered by Red Hat production service level agreements, may be incomplete, and are not intended for production. See the Technology Preview Features Support Scope. For the QEMU path, also wait until the OpenShift Virtualization release notes, or a solution you can cite, describe the build you will run. The product blog is not that note.

QEMU CBT is an upstream alpha API

KubeVirt v1.8 added storage-agnostic incremental backup. The upstream proposal uses QEMU’s changed-block tracking and a libvirt checkpoint as the marker for each backup. The API group in that proposal is backup.kubevirt.io/v1alpha1. The objects are VirtualMachineBackup and VirtualMachineBackupTracker. Creating a tracker requires the IncrementalBackup feature gate.

The tracker stores the latest successful checkpoint for one virtual machine (VM) and one backup product. The next backup that names the tracker copies blocks changed since that checkpoint. A backup with no tracker, or a tracker that has no checkpoint yet, is a full backup. forceFullBackup forces a full copy even when a checkpoint exists.

Two details belong in the vendor conversation:

  • One tracker per backup product. The proposal expects each product to keep its own checkpoint chain. Two products sharing a tracker will not have independent incrementals.
  • A VM restart drops libvirt’s checkpoints. Libvirt is recreated when the VM starts. The proposal’s answer is to redefine checkpoints from the tracker’s stored metadata during startup. If that metadata is missing or stale, the next backup falls back to a full copy. A reboot between backups is a test, not an edge case.

Veeam has written that VEP-25 and VEP-26 are the upstream proposals, and that Kasten can read the QEMU dirty bitmap instead of scanning the whole disk. Red Hat’s November 11 sentence names Commvault, Cohesity, Dell, and Rubrik. Ask which sentence matches the build you will install, on which OpenShift z-stream, and whether that build is the one that creates the tracker.

I do not see OpenShift API for Data Protection (OADP) in that partner sentence, or in the OpenShift Virtualization 5.0 release notes. Keep the OADP design you already have — CSI snapshot, or CSI with DataMover — until a version note says OADP consumes this bitmap.

The practical gain, when the chain is intact, is less data read. Veeam’s illustration is a 2 TB disk with 50 GB of daily change: the backup product reads the changed blocks instead of scanning the disk. That is backup-window math. It is not a recovery time objective (RTO), and it is not proof the guest could recover.

CSI CBT compares two snapshots

OpenShift 5.0 promotes CSI changed block tracking from Developer Preview to Technology Preview. The OKD overview describes what a backup application can ask for:

  • Block ranges that changed between two snapshots of the same CSI block volume
  • Allocated ranges inside one snapshot, so a full backup can skip unallocated space

The call returns metadata. The backup application reads and stores the snapshot data on its own. The driver has to implement the API, and the backup product has to call it. Snapshot support does not imply support for changed-block queries. Confirm the driver on the cluster you will use. The storage shortlist is where that driver was chosen; this preview does not reopen it.

CSI CBT is also the path that can get faster still once arrays implement it. QEMU CBT exists so a backup can proceed when the array has not. Using the QEMU bitmap against a driver that already returns changed blocks is a product choice, not two copies of the same feature.

Incremental backup is still backup

CBT changes how much data the backup job reads. The OADP post still draws the lines around that job.

CBT can shrinkCBT leaves this alone
Bytes read from the diskWhether the guest was quiesced
Data moved to the backup targetRestoring after the original cluster is lost
The backup window, when the checkpoint chain is validThe time you measure for that restore
 Replication, failover, and failback
 Fencing a dead node and restarting the VM

Guest consistency is still a guest-agent problem. A crash-consistent incremental is a crash-consistent copy that happened to be smaller. Document whether the backup is application-consistent or crash-consistent, alongside the recovery point objective (RPO). VolumeGroupSnapshot does not become generally available because a bitmap exists.

Cluster loss is still a restore from wherever the backup product put the blocks. Time it. A faster incremental does not change the download.

Declared DR is still the array or OpenShift Data Foundation Metro-DR and Regional-DR story. Fencing a gray host is still Fence Gray Host Failures on OpenShift Virtualization. Live migration is still a move. None of those is a checkpoint.

What a pilot has to show

Run this on a non-production cluster, with a backup vendor build that names the OpenShift version it supports. The storage setup in the OpenShift PoC docs is the floor: a storage class the VM can use, before anyone talks about incrementals.

For the QEMU path, ask the vendor for five facts before you change a disk:

  1. The OpenShift and OpenShift Virtualization z-stream, and the feature gate their build expects.
  2. Whether their product creates a VirtualMachineBackupTracker per VM, and refuses to share that tracker with another product.
  3. What the next backup does after the VM restarts, when libvirt no longer has the checkpoint.
  4. Bytes read from a disk with a known change set, next to the size of that disk.
  5. A restore into a scratch namespace, with a stopwatch, and a statement of whether the guest was frozen.

For the CSI path, confirm the driver implements changed block tracking on that OpenShift version, then ask for the same restore time. The metadata call is not the backup. The snapshot data still has to move.

Keep the source until the restored application has been verified. An incremental that reports success is not a recovered application. The same rule as storage migration in OpenShift Virtualization 5.0 applies here: verify the copy, then delete the thing you can still roll back to.

The solutions architect takeaway

  1. Name the bitmap. QEMU checkpoints and CSI snapshot diffs are different Technology Previews.
  2. QEMU CBT is upstream alpha. backup.kubevirt.io/v1alpha1, gate IncrementalBackup, one tracker per backup product. The November 11 blog calls the OpenShift feature Technology Preview. The OpenShift Virtualization 5.0 release notes do not list it.
  3. CSI CBT is Technology Preview on OpenShift 5.0. It returns metadata for drivers that implement it. It does not return the backup.
  4. Restart the VM during the test. The checkpoint chain is the feature. A reboot that forces a full copy means the test did not demonstrate incremental continuity, even if the backup job succeeded.
  5. Time the restore. Write the number you measured. A smaller read on the way out does not set the RTO.
  6. Leave DR, fencing, and live migration on their own slides. Incremental backup is still backup.

Want help deciding whether a backup vendor’s CBT claim belongs in this design? Work with your Red Hat account team, and keep the pilot off production until the release notes match the build.

Further reading

This post is licensed under CC BY 4.0 by the author.