I say that because while the *average* number of reads between failures is 12TB, that doesn't mean that the likelihood of failure is identical from Read #1 to Read *#n*. The MTBF only means that after reading 12TB of data, you have a 50% chance of having suffered one URE; the actual likelihood of failure at any given point in time is described by a hyperbolic curve that starts at 0% at Read #1, crosses the 50% mark at Read #12billion, and trends towards 100% as it approaches Read #Infinity.

Why does that matter? Because if the likelihood of failure per disk starts out infinitesimal and increases as time and wear increase, then that means that 6 disks reading 2TB of data each will not progress as far along the failure-probability curve as a single disk reading 12TB of data would -- the wear is split across multiple disks, so each disk suffers substantially less wear during the rebuild than a single disk would if it had read the same amount of data all by itself. Likewise, the failure-probabilities for all of the drives added together will not equal the failure-probability of a single drive that had read the same amount of data all by itself.

Last edited: Oct 23, 2008