On Mon, Nov 07, 2016 at 11:53:20AM +0000, Guy Plunkett wrote:
>> On 7 Nov 2016, at 11:50, Kevin Lemonnier <[email protected]> wrote:
>> 
>>> moving the /etc/pve/nodes/pve11/qemu-server/*.conf files to another node
>>> worked well. I didn't have to restart any services.
>> 
>> Still, that seems like a fairly usual operation that the GUI could
>> very well support, I think.
>> Would it be possible to add that in the future ?
>it would also be nice for the GUI not to lose the details of the VMs if the 
>node is down, and indeed allow the transfer of the VM to an active node.
>
>I know this can be done today with the HA but it would be nice to also do this 
>for non HA VMs

the problem is that PVE has a notion of nodes being owners of their
respective ressources (like containers or VMs). the technical background
is that we only allow changes to their configuration in processes
holding a node-local lock, to prevent race conditions and conflicts.
cluster-wide locking is very expensive, and is thus only used when
absolutely needed (like allocating/freeing disk images on shared
storage).

a consequence of this architectural design is that node A cannot "steal"
a VM/container (or rather, its configuration file) from another node B,
because there is no way for a process on node A to lock this file (only
node B is able and allowed to locally lock and modify it). the only
exception to this is for HA-managed ressources, because our HA stack
has a system of cluster-wide locking and watchdogs to ensure that it
only steals ressources from actually dead nodes (and in a conflict-free
manner).

so there are only three options:
- use HA for automatic migration on node failure
- move the configuration files manually as root when a node has died,
  and take the blame if you moved from a node that was not actually dead
  (in which case you violated one of the basic assumptions of PVE and
  "anything" might happen)
- change this architectural design (not likely, as it is one of the
  basic principles of PVE's clustering ;))

I hope this makes it clear why this case is handled like it is, and that
it's not some arbitrary unwillingness to add this functionality to the
API/GUI ;)

p.s. top-posting makes email threads rather hard to follow, please
consider switching to in-line replies or bottom-posting

_______________________________________________
pve-user mailing list
[email protected]
http://pve.proxmox.com/cgi-bin/mailman/listinfo/pve-user

Reply via email to