[PVE-User] HA scalability and predictability

[Christian Balzer]

On Mon, 10 Dec 2018 07:39:59 +0100 (CET) Alexandre DERUMIER wrote:

> >>I presume a "node1:2,node8:1" and "restricted 1" should do the trick here.  
> 
> restricted is only used, if both both node1 && node8 are down, the vm don't go to another node.
> 
Yes, that's exactly what I want here.
Any other (non-listed) node would be already at capacity and migrating the
VMs there would be potentially fatal.
A double node failure is something that will need to be dealt with
manually.

> But the weight indeed, should do the trick.
> 
> for example, n nodes with :2  , and spare(s) node(s) with :1
> 
> ----- Mail original -----
> De: "Christian Balzer" <chibi at gol.com>
> À: "proxmoxve" <pve-user at pve.proxmox.com>
> Envoyé: Lundi 10 Décembre 2018 04:45:27
> Objet: [PVE-User] HA scalability and predictability
> 
> Hello, 
> 
> still investigating PVE as a large ganeti cluster replacement. 
> 
> Some years ago we did our owh HA VM cluster based on Pacemaker, libvirt 
> (KVM) and DRBD. 
> While this worked well it also showed the limitations in Pacemaker and LRM 
> in particular. Things got pretty sluggish with 60VMs and a total of 120 
> resources. 
> This cluster will have about 800VMs, has anybody done this number of HA 
> VMs with PVE and what's their experience? 
> 
> Secondly, it is an absolute requirement that a node failure will result in 
> a predictable and restricted failover. 
> I.e. the cluster will have a n+1 (or n+2) redundancy with at least one 
> node being essentially a hot spare. 
> Failovers should only go to the spare(s), never another compute node. 
> 
> I presume a "node1:2,node8:1" and "restricted 1" should do the trick here. 
> 
> Regards, 
> 
> Christian 


-- 
Christian Balzer        Network/Systems Engineer                
chibi at gol.com   	Rakuten Communications