How would Ceph/Proxmox handle a switching failure

[chrispage1]

We've been looking at a new setup that has three nodes. They will be running Proxmox and Ceph in a hyperconverged config. Each node has theoretical switch redundancy using LACP.

However, not for the first time our providers switching has had some kind of failure with a switch and LACP standby hasn't taken effect, meaning all three nodes for a period of ~20 mins are on their own. Of course this scenario is less than ideal and I'm hoping it won't happen, but I want to make sure if it did we could easily recover.

Ceph will be configured to write across all three nodes with a minimum of two.

Does anyone have experience / know how Proxmox & Ceph would handle this when losing quorum in such a dramatic way?

I'm hoping that each respective node will continue running the VMs and no migration will be attempted due to the loss of quorum and still be able to continue reading & writing to it's local Ceph instance. Once the node comes back online the degraded cluster backfills to recover everything to 100% healthy.

Thanks,
Chris.

 

[Lorenz.S]

Hi,

both PVE and Ceph may not function without a quorum. Ceph does not allow any disk writes from VMs without a quorum – disk reads might work, but I am not sure. For the PVE it is not as severe, the VMs will keep running, but any configuration-related changes may not work, and starting/stopping/... VMs is not possible. Furthermore, if HA is active on one of these nodes, this node will fence itself and reboot.

Loosing connections between the nodes should be avoided at all costs.

 

[chrispage1]

Hi,

both PVE and Ceph may not function without a quorum. Ceph does not allow any disk writes from VMs without a quorum – disk reads might work, but I am not sure. For the PVE it is not as severe, the VMs will keep running, but any configuration-related changes may not work, and starting/stopping/... VMs is not possible. Furthermore, if HA is active on one of these nodes, this node will fence itself and reboot.

Loosing connections between the nodes should be avoided at all costs.

Thanks very much for your response. I'll take that on board! Last thing I'd want is everything to be knocked offline by a temperamental switch.

Chris.

 

[chrispage1]

Hi Lorenz.S, I wonder if you could take a look at my new topic - https://forum.proxmox.com/threads/open-vswitch-across-physical-servers.95989/

Is it possible to use Open vSwitch as a cross communication between servers, and run heartbeat through there, eliminating the need for the physical switch?

Chris.