These are all great points, but I wanted to share something that I wish I'd known before I spun up my array... The configuration of your array matters a lot. I had originally chosen to use RAIDZ1 as it's the most efficient with capacity while still offering a little fault tolerance. This was a mistake, but in my defense, the hard data on this really wasn't distributed until long after I had moved my large (for me) dataset to the array. I really wish I had gone with a Striped Mirror configuration. The benefits are pretty overwhelming:
Performance is better than even RAIDZ2, especially as individual disk size increases.
Fault tolerance is better as you could have up to 50% of the disks fail, so long as one disk in a mirrored set remains functional.
Fault recovery is better. With traditional arrays with distributed chunks, you have to resilver (rebuild) the entire array, requiring more time, costing performance and shortening the life of the unaffected drives.
You can stripe mismatched sets of mirrored drives, so long as the mirrored set is identical, without having the array default to the size of the smallest member. This allows you to grow your array more organically, rather than having to replace every drive, one at a time, resilvering after each change.
Yes, you pay for these gains with less usable space, but platter drives are getting cheaper and cheaper, the trade seems more worth it than ever. Oh and I realize that it wasn't obvious, but I am still using ZFS to manage the array, just not in a RAIDZn configuration.