After all of this, I still don't get the use case for LVM backed VMs. Other than possibly, possibly a super IO heavy database. Even then, it's questionable.
That's roughly it, and yes, it remains questionable at the best of times.
In the cases where you need LVM fat, you almost certainly also need to avoid LVM because that itty bitty overhead is still too much.
Preallocated qcow2 images are 99% as fast as LVM volumes. Even with just preallocating just the metadata I've had almost native disk write speeds. You lose all of the advantages of qcow2 like libguestfs, the qemu agent, internal and external snapshots, etc.
that said, no idea how the eff you do that with ProxMox. That was just KVM.
It's the default actually. We use Qcow2 on LVM-Thin mostly.
I meant the preallocation. I'd be surprised if they expose that because you can either fully preallocate and zero out the blocks, preallocate and just mark the beginning and end, or just preallocate the metadata.