ZFS on Linux and Amazon EC2

Putting your zpool as files on an existing file system means you're relying on that file system to provide consistency (which sounds dangerous at best) and also that ZFS can't take good advantage of caching. I'm not sure how well ZFS would handle the transfer from files to physical devices; the file system itself probably wouldn't have any real complaints, but you might run into things like it not liking vdevs going onto smaller devices (from what I've read, a number of people have been bit by this having set autoexpand=on, so you might want to be careful with that property and its cousin autoreplace). Alternatively, you'd be running ZFS on top of LVM, which is probably possible but doesn't allow ZFS to handle the devices intelligently since it'll only see one huge device. Remember that ZFS is not just a file system, it's a volume manager as well, so properly replaces both the regular file system and LVM. Many of its features, including metadata placement on multiple disks and multiple copies of data for redundancy within a zpool, work best when ZFS has a good idea of the physical storage layout.

I've been considering migrating to ZFS as well, and the best option I've been able to come up with for migration involves one more hard disk. Install another hard disk that is at least the size of the smallest physical drive you currently have in the array, make a ZFS pool and file system on it (configured for JBOD but with only one device), and move as much onto it as you can. (Since I'm not running LVM, I'd move everything on the smallest drive onto the ZFS fs.) Reduce the LVM array by removing one physical disk from it, expand the ZFS zpool onto that now-free disk, move some more files, rinse and repeat until done. With clever use of symlinks or good handling of exports, you may even be able to keep the process transparent for anyone who might be using files on that NAS box of yours in the meantime.

After much tinkering and experimentation I've found a solution, albeit with a fairly large tradeoff.

First off, the options I had to rule out:

• Having a second offsite ZFS server with a mirrored pool wasn't an option due to cost. Had it been an option this would by far have been the best approach, utilizing ZFS send / receive to ship snapshots to the remote pool.

• Having a second onsite ZFS mirrored pool, which I could remove disks from to take home. This is more feasible than the first option, but I would need the second pool to always have two disks onsite (or to use two data-copies on a single onsite disk). At present I have four disks, and no more space for a fifth in the server. This would be a fair approach but still not ideal.

• Using ZFS attach and detach to rotate the backup disk into and out of the mirrored pool. This works well, but has to perform a full resilver every time the disk is added. This takes unacceptably long, and so I couldn't rely on this.

My solution is similar to using attach and detach, however it uses online and offline. This has the advantage of performing a delta resilvering versus a full resilvering, but the drawback that the pool always reports a DEGRADED state (the pool always has two disks; the rotating offsite disks are marked offline when they are in remote storage and resilver and then come online when they are onsite).

So, a quick recap and overview of my setup:

I have one ZFS server and four identical disks. ZFS is setup to use a mirrored pool. Two of the four disks are permanent members of this pool. The other two disks rotate; one is always in offsite storage, the other is part of the pool to act as a ready-to-go backup.

When it comes time to rotate the backups:

• I wait for a zfs scrub to complete to reasonably assure the backup disk is error free

• I zfs offline the disk which will be taken remote. After its offline'd I hdparm -Y /dev/id to spin it down. After a minute I partially remove the disk sled (just enough to ensure its lost power) and then give it another minute before fully pulling the drive to guarantee it has stopped spinning. The disk goes in a static bag and then a protective case and goes offsite.

• I bring in the other offsite disk. It gets installed in the hotswap tray and spins up. I use zfs online to restore the disk to the pool and kick off a partial resilvering to make it concurrent.

This system guarantees that at any given time I have two ONLINE mirror disks and one OFFLINE remote disk (which has been scrubbed). The fourth disk is either being resilvered or online, which has the benefit that in case a running drive fails it's probably the pool will still consistent of two online disks.

It's worked well for the past couple weeks, but I'd still consider this a hackish approach. I'll follow up if I run into any major issues.

Update: After running with this for a couple months I've found that in my real-world use the resilvering is taking the same time for either detach/attach and offline/online. In my testing I don't think I was running a scrub--my hunch is that if a drive is offline for a scrub then it requires a full resilver.