On 6/17/12 8:21 AM, Sean Fulton wrote: > This was a Linux-HA cluster with a floating IP that the clients would > mount off of whichever server is active. So I set up a two-node > replicated cluster, which the floating IP and heartbeat, and the > client mounted the drive over the floating IP. I'm using the NFS > server built into gluster. So rpcbind and nfslock are running on the > server, but not nfs. The client writes to the one server with the > floating IP, and gluster takes care of keeping the volume in sync > between the two servers. I thought that was the way to do it. It has never been clear to me how well you can fail over a NFS mount using a floating IP address, especially with the Gluster NFS server. I typically install gluster on the client and have it mount localhost:/whatever and all the brick routing is handles locally. Not practical with a lot of clients, but it's a simpler configuration in a smaller environment. If it makes you feel better, I'm in the process of restoring a production SVN repo because Gluster 3.2.5 ate it after a node reboot last night. Not had time to dig through the logs in detail, but it seems like a self-heal from a 4-way replica did something wrong.
