It's a been a while since this thread has been updated, so I figure I should share what I've learned. First: here's a script which reproduces the bug reliably and includes steps to revert: https://gist.github.com/dmeiselman/ecb35ffcb675c38e6eeb14292948c2e3 Don't run this on your production cluster. Here's my understanding of this bug and it's current state: Some operation sets the snapshot_of metadata field of a valid snapshot to NULL (We can do this intentionally by making a VM, taking a snapshot of it, then deleting the parent disk.) If you perform a revert of this invalid snapshot, revert's "Updating the snapshot_of fields" step looks for VDIs whose snapshot_of matches the snapshot's source disk, so it can relink them to the new disk. If the source is NULL, the search becomes "every VDI with NULL", which matches ISOs, base images, ordinary VM disks. Congrats, every real disk and ISO in every SR now looks like a snapshot. Now these disks have a value set for snapshot_of which means XOA hides the VDI in the web UI. The fix is to delete or clone the super-parent, and all of its snapshots. sr_scan will correct the metadata for all the VMs and ISOs and you're back where you started. This happens in line 401 of https://github.com/xapi-project/xen-api/blob/v26.1.19/ocaml/xapi/xapi_vm_snapshot.ml#L389-L408 PR 7290 https://github.com/xapi-project/xen-api/pull/7290 has a fix, which looks like it's coming in xapi v26.1.21. (It's worth reading the Pull Request, as it has a more graceful description of the issue, it really helped me wrap my head around it.)

Update: switching the remote to “multiple data blocks” fixed the 2 TiB overflow. The full backup now succeeds (≈1.1M blocks, 13 h) and file restore works. However, the next incremental always falls back to a full The delta export request gets no data for exactly 5 minutes, then undici’s body timeout fires. So “can’t compute delta” is actually a timeout. My guess is that the host takes more than 5 min to compute the qcow2 delta of a 3 TB disk before sending the first byte (single SATA SMR disk, local EXT SR). XapiQcow2StreamSource.close() throws because the source was never initialized. In earlier attempts, this may have left tapdisks for the qcow2 snapshots stuck and attached to dom0 (tap-ctl close timing out, SR_BACKEND_FAILURE_1200). I had to kill them manually before the VBDs could be unplugged. CBT is disabled on all VDIs. Is there a way to increase this timeout for qcow2 delta exports, or is a fix planned (e.g. keep-alive while the delta is computed)? Is the close() error on an uninitialized source a known bug? Same environment as above (XO from sources 3923756ff, XCP-ng 8.3.0, sm-3.2.12-25.1).

Sharing some results. I enabled this on my homelab server, a Dell R630 with Intel Xeon E5-2630 v4 (Broadwell). No adverse impact noticed. The CDM results on Windows is higher than previous runs however, I do not go in depth with I/O benchmarking. (I am using the latest win-pv-drivers on Windows and the rust-based guest tools on Linux) Controller is in HBA mode. VM SR is using SMAPIv3 (ZFS, RAID-0, noop scheduler) [image: R630_XCP-ng_APICv.png] CDM results. Windows Server 2019 4 vCPU/8GB RAM/single VHD: [image: 2610_WIN2019_ZFS.png] I installed ScyllaDB on Alma Linux 10.2 (4 vCPU/8GB RAM & two VHDs, ScyllaDB uses a dedicated drive for data) with io_uring enabled and noop scheduler set on VM. Here are the IOTune results from the ScyllaDB installation: Measuring sequential write bandwidth: 826 MiB/s (deviation 13%) Measuring sequential read bandwidth: 1061 MiB/s (deviation 6%) Measuring random write IOPS: 50791 IOPS (deviation 11%) Measuring random read IOPS: 62977 IOPS (deviation 16%)