Multi-disk VM snapshot delete leaves DB inconsistent when one disk's merge times out after another disk's merge already succeeded

Begin with the reproduction on KVM using SharedMountPoint, Filesystem, or NetworkFilesystem primary storage and a multi-disk VM, focusing on the snapshot deletion path and qcow2.delta.merge.timeout behavior. Done means a timed-out disk cannot leave another successfully merged disk pointing to a deleted file, with the operation either remaining atomic or persisting per-disk completion consistently.

problem

Deleting a disk-only VM snapshot on a multi-disk VM can leave CloudStack's database permanently inconsistent with the actual state of primary storage if one disk's merge completes successfully while another disk's merge in the same operation times out. The disk that succeeded has its underlying file correctly committed and removed on the hypervisor, but the corresponding database update is never applied - because the whole operation is treated as a single all-or-nothing unit. The affected volume is left pointing at a file that no longer exists, and the VM cannot subsequently be started.

versions

4.21.0.0+
KVM hypervisor, primary storage of type Filesystem / NetworkFilesystem / SharedMountPoint, VM with 2+ disks, disk-only VM snapshot (no memory).

The steps to reproduce the bug

1. Create a VM with at least two disks on KVM/SharedMountPoint (or Filesystem/NetworkFilesystem) primary storage - one small disk, one large disk with enough real delta data that a commit takes noticeably longer than qcow2.delta.merge.timeout (or 1 hour, if using the running-VM/non-events path).
2. Take a disk-only VM snapshot.
3. Write enough data to the large disk that its subsequent commit will exceed the timeout.
4. Delete the VM snapshot.
5. Observe the small disk's merge completes and its delta file is deleted, while the large disk's commit is killed by the timeout.

What to do about it?

• The operation is fully atomic: no disk's file is deleted/merged unless all disks in the snapshot are confirmed to have completed successfully, and the operation cleanly fails/rolls back to a well-defined error state if any disk times out; or
• Per-disk completion is persisted incrementally as each disk finishes, so a disk that successfully merged is never left with a database record pointing at a deleted file, regardless of what happens to other disks in the same VM snapshot.