DRBD

[Alessandro 123]

Anyone using DRBD in production?
As per my other thread, i'm planning a new environment with HA and Live Migration.

Gluster or Ceph are too complicated and too slow in small clusters. We need at least 8 nodes to get decent performances.

DRBD is way faster and requires only 2 nodes.
I'm thinking about creating a DRBD Active/Passive cluster with volumes exported thourgh NFS.

Should I use NFS or iSCSI ?
Any suggestion to get 100% availability and absolutely no data-loss? Split brains are scaring me.

 

[manu]

Hi Alessandro
You need three nodes for HA and DRDB9 is not production ready.
Read well the requirements of:
https://pve.proxmox.com/wiki/High_Availability_Cluster_4.x
Lots of people are using NFS or iSCSI with success, both of them are valuable options.

 

[Alessandro 123]

Why 3 nodes? DRBD 8 works with only 2 nodes.

 

[Alessandro 123]

Keep in mind that i was talking about DRBD on 2 nodes.
PVE would be on 4-5 or 6 nodes (currently I have 8 XenServer nodes that I would like to move to PVE)

 

[manu]

> I'm thinking about creating a DRBD Active/Passive cluster with volumes exported thourgh NFS.

In that case PVE is only concerned about the NFS server, what you do in the NAS side is transparent for the hypervisor.

 

[Alessandro 123]

I know but I'm just searching for some real usage information

Maybe some users are using drbd in production and are willing to share their experience

 

[mir]

NFS on top of DRBD with a flooting IP is not a good idea since the NFS service run in kernel mode - file locks are acquired in kernel mode and exported to the client. So any operation which relies on persistent connections (VM's) to the NFS server will hang in case of fail-over since locks cannot be transferred between hosts. In your case I would opt for exposing the storage through iSCSI (backed by DRBD) and then implement floating IP using pacemaker or whatever fits your bill. https://pve.proxmox.com/wiki/Storage_Model#LVM_Groups_with_Network_Backing

 

[Alessandro 123]

Thanks for this advice
What about ganesha nfs? It runs in user space

 

[mir]

NFS was never designed with HA in mind so this it is not advised to use NFS for HA. Commercial storage providers (netApp, Nexenta, EMC etc) has this feature but they have made an add-on to NFS to be able to do that. This is therefore something you pay for.