Alexey Filin wrote:
I understand it, I say about configuration when afr master node crashes before close() (it's indifferent is afr master on glfs slave or on glfs server node), in such a situation replicas can be inconsistent but afr version xattr is the same (old) on all replicas. To discover the inconsistency without minor version xattr it is required e.g. to calculate checksum of all replicas or to use workarounds like modification time check misbehaving in some cases.
Oh, I'm all for the AFR minor version. I'm just leery of the benefits of having a separated journal when the AFR minor version suggestion you had seemed so elegant to me. I don't know what a separated journal would give you that the minor version wouldn't and the minor version is more distributed and robust.
-- -Kevan Benson -A-1 Networks
