On Wed, Dec 30, 2020, 7:39 AM Jesper Dybdal <jd-debian-u...@dybdal.dk> wrote:
> I would hope that the most recently modified half of the array would be > the one to overwrite the least recently modified one, so that a > temporary absence of one disk which later comes back unmodified, will > not destroy data. > > Is that how it works? > You would hope that it works that way. Trouble is, you can't necesssrily tell how far the failed drive got in the last write to it. And cause failure occurred, mdadm can't always tell either. In addition, unfixable corruption may have taken place in the filesystem above that, such that even a succesful mdadm recovery still doesn't get your data back. So the backup on separate media is crucial. Some places I've worked use mdadm only for RAID performance reasons. But any failure recovery is a restore from off site backup. They don't even try RAID recovery. Or as I have done in the past, the other half of the RAID 1 is put directly into use, the failed drive is simply removed and destroyed, then replaced with a brand new drive and re-mirrored. If you have hundreds or thousands of servers, weelll..... there's another good argument for virtualisation. Said your boss :-) -- > Jesper Dybdal > https://www.dybdal.dk > >
