| The non-recoverable bit error rate spec. NetApp tracks it with their Nearstore product line which used SATA drives in a NAS box (they have been for a while actually, when I left they had data on about 65 million drive hours) and while Seagate quotes it a 1x10^15 bits but its actually closer to 5 in 10^15 bits. A 3TB drive has 3x10^13 bits of data (closer to 3x10^14 when you account for track markers and error recovery bits). If you're bored some time try reading every sector from one of these drives. To maximize your chance of success make sure you operate the drive at a slightly warm temperature (keeps the lubricant from sticking) and isolate it from vibration. Its worse if you read it randomly (you will get some arm servo movement just because the drive will have replaced some blocks from spares, but minimizing it also keeps vibrations down.) Long before it became an issue on single drives, like it is today, it was an issue when trying to reconstruct a RAID4 (or 5) group that was 3.5TB (which at the time was a 7 disk raid group of .5T drives. 14 disk groups (a full shelf) were pretty much guaranteed to see a second error in the shelf during reconstruction. Which was also way RAID6 or dual-parity RAID became a must have enterprise feature back in 2005 or thereabouts. On an interesting side note, because the chance of hitting an unrecoverable read error is evenly distributed through a drive, 3X replication is still recoverable even with intermittent read failures. There isn't really a RAID number for that but it does work reasonably well and avoids a pesky parity calculation if you embed check data in your blocks as they do in GFS. [1] https://www.usenix.org/legacy/publications/library/proceedin... -- Peter Corbett's paper (he is the guy who invented NetApp's dual parity system, and from that paper the following -- "Disks protect against media errors by relocating bad blocks, and by undergoing elaborate retry sequences to try to extract data from a sector that is difficult to read [10]. Despite these precautions, the typical media error rate in disks is specified by the manufacturers as one bit error per 1014 to 1015 bits read, which corresponds approximately to one uncorrectable error per 10TBytes to 100TBytes transferred. The actual rate depends on the disk construction. There is both a static and a dynamic aspect to this rate. It represents the rate at which unreadable sectors might be encountered during normal read activity. Sectors degrade over time, from a writable and readable state to an unreadable state." And in experience from the field put it at about 15TB transferred, so 3TB into 15TB, one in five. |
[1 - (1 - 5 10^-15)^(3 * 10^12)] ~ 0.01488...
If your 20% figure is accurate, the actual uncorrectable bit error rate would need to be something like 7 in 10^14. I am not disputing your empirical information, but your numbers are do not agree with it. The difference in what your numbers say and what you say is only about 1 order of magnitude. Doing statistical calculations with better records could allow the cause of that to be identified.