Re: Would ZFS and gmirror work well together in a two-node failover cluster?

On 18 Jul 2008, at 19:16, Maurice Volaski wrote:

> For one thing, how small can small window of data be, something  
> measured in minutes or hours? And this would be a system that could  
> be moving 100 MB+ per second, so the data could get outdated quickly.

As I understand it (although I've not attempted this myself) there's  
no reason not to take snapshots at minute intervals or less -  
snapshots are pretty low cost. How far behind it gets depends on how  
much data you write to the main system and how fast you can sync it to  
the backup.

The rough principal is that you take an initial snapshot and send/ 
receive that to the backup system. Then you take a second snapshot and  
send a "diff" stream to the backup. Then you can drop the initial  
snapshot and take a third and do the same. You drop the second, take a  
fourth, etc., etc. You have to avoid there being too many existent  
snapshots as the utilities become very slow when you have a lot.

If you have a couple of spare machines, I'd give it a whirl and get  
some feeling for how well it works.

> Plus, I'm looking for the failover to be automatic and near  
> instantaneous. That is, if I pull the power cord on the primary,  
> could the secondary go hot in under a minute?

The ZFS filesystem(s) on the backup machine would be live at all  
times. You can even immediately start writing to them - though at that  
point you can't receive any more snapshots to the backup obviously.  
Failover of whatever systems were on top of these would be another  
question. Presumably this is a file server or something? Are you going  
to failover the IP address(es)?

Jonathan