Gluster client is installed in Dom0 (the client to access data). But Gluster server are in VMs, so you got more flexibility.
If the node with arbiter goes down, yes, you are in RO. But you won't enter a split brain scenario (which is the worst case in 2 nodes thing).
Eg using DRBD, in 2 nodes in multi-master, if you just lose the replication link, and you wrote on both sides, you are basically f***ed (you'll need to discard data on one node).
There is no miracle: play defensive (RO if one node down) or risky (split brain). We chose the "intermediate" way, safe and having 50% of chance to lose the "right" node without being in RO then.
Obviously, 3 nodes is the best spot when you decide to use hyperconvergence at small scale. Because the usual 3rd physical server used previously for storage, can be also now a "compute" node (hypervisor) with storage, and you could lose any host of the 3 without being in read only (disperse 3).
edit: XOSAN allow to go from 2 to 3 nodes while your VM are running, ie without any service interruption. So you can start with 2 and extend later :)
