3 ms·

What you need to know about disk latent sector errors

7 points by andrewvc 17y ago

Related: Does RAID 6 stop working in 2019? http://storagemojo.com/2010/02/27/does-raid-6-stops-working-in-2019/

andrewvc 17y ago
Related: Does RAID 6 stop working in 2019? http://storagemojo.com/2010/02/27/does-raid-6-stops-working-in-2019/ http://storagemojo.com/2010/02/27/does-raid-6-stops-working-...
- lutorm 17y ago
  That seems a little overly alarmist since it assumes that there is no background check-and-scrub going on.
  - hga 17y ago
    Read the essay, he is considering it. He's been worried about this issue of latent errors for years; I think it was largely his Storage Mojo postings and the papers he brought to my attention in them that got me so interested in ZFS.
- chaosmachine 17y ago
  He's talking about 7200RPM HDDs, but nine years down the road, shouldn't we expect SSDs or whatever comes next to replace moving-parts technology? Nine years is a long time. Back in 2001, most drives will still sitting around 40gb.
  - wmf 17y ago
    The price/GB gap between flash and disk is ~30x and it's shrinking slowly, if at all. In 2019 I expect disk to still be cheaper than flash, although flash may be cheap enough for many applications.
    - andrewvc 17y ago
      Agreed, hell, tape is still economical for some circumstances, disks will be with us for a long time.
hga 17y ago
Hmmm, continues a theme of his that there is often much less difference between consumer and enterprise drives that one would like or expect. Quoting the paper he's discussing, after a bit how power-law distributions model this much better than the more commonly used geometric and Poisson, he bolded: "We find no significant difference in the statistical properties of LSEs in nearline drives versus enterprise class drives." But he then noted that they do have 4X the error rate. As I recall, it may have been the predecessor paper based on the same data set that pointed out that today's disks do not follow the bathtub curve WRT to failure. (Instead they have very few infant failures, after one year or a little more in service they start failing, and the failure rate is then well behaved (at least through year 5).)

7 points by andrewvc 17y ago

Related: Does RAID 6 stop working in 2019? http://storagemojo.com/2010/02/27/does-raid-6-stops-working-in-2019/

Related: Does RAID 6 stop working in 2019? http://storagemojo.com/2010/02/27/does-raid-6-stops-working-in-2019/ http://storagemojo.com/2010/02/27/does-raid-6-stops-working-...

lutorm 17y ago
That seems a little overly alarmist since it assumes that there is no background check-and-scrub going on.
- hga 17y ago
  Read the essay, he is considering it. He's been worried about this issue of latent errors for years; I think it was largely his Storage Mojo postings and the papers he brought to my attention in them that got me so interested in ZFS.
chaosmachine 17y ago
He's talking about 7200RPM HDDs, but nine years down the road, shouldn't we expect SSDs or whatever comes next to replace moving-parts technology? Nine years is a long time. Back in 2001, most drives will still sitting around 40gb.
- wmf 17y ago
  The price/GB gap between flash and disk is ~30x and it's shrinking slowly, if at all. In 2019 I expect disk to still be cheaper than flash, although flash may be cheap enough for many applications.
  - andrewvc 17y ago
    Agreed, hell, tape is still economical for some circumstances, disks will be with us for a long time.

hga 17y ago

Hmmm, continues a theme of his that there is often much less difference between consumer and enterprise drives that one would like or expect. Quoting the paper he's discussing, after a bit how power-law distributions model this much better than the more commonly used geometric and Poisson, he bolded: "We find no significant difference in the statistical properties of LSEs in nearline drives versus enterprise class drives." But he then noted that they do have 4X the error rate. As I recall, it may have been the predecessor paper based on the same data set that pointed out that today's disks do not follow the bathtub curve WRT to failure. (Instead they have very few infant failures, after one year or a little more in service they start failing, and the failure rate is then well behaved (at least through year 5).)