I just found out the hard way that we have a bug in snapshot-create-as,
when I corrupted a guest.  I was testing with an offline domain:

# virsh dumpxml win64 |grep -C1 driver
    <disk type='block' device='disk'>
      <driver name='qemu' type='raw'/>
      <source dev='/dev/sda4'/>
# virsh snapshot-create-as win64 --disk-only --diskspec
hda,file=/var/lib/libvirt/images/win64.qcow2 --no-metadata
Domain snapshot 1417850911 created
# virsh dumpxml win64 |grep -C1 driver
    <disk type='block' device='disk'>
      <driver name='qemu' type='qcow2'/>
      <source dev='/var/lib/libvirt/images/win64.qcow2'/>
# systemctl restart libvirtd
# virsh dumpxml win64 |grep -C1 driver
    <disk type='block' device='disk'>
      <driver name='qemu' type='raw'/>
      <source dev='/dev/sda4'/>

Even though we don't want to save the snapshot metadata, we DO need to
save the XML change.  Otherwise, the restarted libvirtd sees the wrong
disk as the source; worse, if the user started the guest before
restarting libvirtd, they run the risk of accidentally reverting to the
pre-snapshot state and losing all changes they made in the meantime,
perhaps even getting their guest filesystem into an inconsistent state.

I'm out of time to get to a root cause or fix before next week, but
wanted to report it now.  I'm not sure if --no-metadata is essential to
reproducing the bug, or if it happens in all cases.