On 9/1/2017 8:53 AM, Jim Kusznir wrote:
Huh...Ok., how do I convert the arbitrar to full replica, then? I was misinformed when I created this setup. I thought the arbitrator held enough metadata that it could validate or refudiate any one replica (kinda like the parity drive for a RAID-4 array). I was also under the impression that one replica + Arbitrator is enough to keep the array online and functional.
I can not speak for the Ovirt implementation of Rep2+Arbiter as I've not used it, but on a standalone libvirt VM host cluster, Arb does exactly what you want. You can lose 'one' of the two replicas and stay online. The Arb maintains quorum. Of course if you lose the second Replica before you have repaired the first failure you have completely lost your data as the Arb doesn't have that. So Rep2+Arb is not as SAFE as Rep3, however it can be faster, especially on less than 10G networks. When any node fails, Gluster will pause for 42 seconds or so (its configurable) before marking the bad node as bad. Then normal activity will resume. On most people's systems, the 'pause' (I think its a read-only event), it noticeable, but not enough to cause issue. One person has reported that his VMs went read-only during that period, but other have not reported that. -wk