Kevan Benson wrote:
Gordan Bobic wrote:
I suspect this isn't a problem that can be solved without having a
proper journal of metadata per directory, so that upon connection, the
whole journal can be replayed.
You could sort of bodge it and use timestamps as the primary version
and the xattr version as secondary, bit that is no less dangerous - it
only takes one machine to be out of sync, and we are again looking at
massive scope for data loss.
You could bodge the bodge further to work around this by ensuring that
the nodes are heartbeating current times to sync between them and
without the sync no data exchange takes place. But that then
complicates things because what do you do when a node connects and is
out of sync, but in the future? Who wins on time sync? Who has the
latest authoritative copy?
I think the most sane way of addressing this is to have a fully logged
directory metadata journal. But then we are back to the journalling
for fast updates issue with a journal shadow volume, which is
non-trivial to implement.
Unless there is some kind of a major mitigating circumstance, it seems
that between this and the race condition that Martin is talking about
on the other thread, GlusterFS in it's current is just too dangerous
to use in most environments that I can think of. And unlike Gareth a
few days ago, I'm not talking about performance issues - I'm talking
about scope for data loss in very valid and very common use cases. :'(
Hmm, what about trusted.glusterfs.createtime (epoch time) as a major
version number, and trusted.glusterfs.version as the minor version
number. Couple that with a glusterfs master time node (defaults to lock
node) and you should have a fairly consistent cluster, right?
There are several problems with this:
1) The concept of the "lock node" is limiting. The locking should be
distributed.
2) Using creation/modification time as the major number is problematic
due to time syncing. What happens when the master node goes offline? If
the nodes are in not in perfect time sync, you've still got the same
problem.
3) "fairly consistent" is _really_ not good enough when we are talking
about a file system.
IMO, it would be better to come up with a design that solves the problem
once and for all. The order of priorities really has to be: consistency,
reliability, performance.
If that isn't the case, you might as well be using a distributed hash
table and hope that you'll get most of the data back most of the time.
Gordan
