Aloke, your findings are actually perfectly consistent with the Google Report. You have to look at what the term “age” means very carefully. For Google’s statistics, they are looking at drives that are always and without exception “on”. A drive’s “age” in the Google Report and its runtime in years are both exactly the same thing. If one were to suppose that the correct age of a drive were its runtime in years only, and not its chronological age, the MAID findings would be perfectly consistent with Google’s data.

Joe Kraska
BAE Systems
San Diego CA
USA

Does anyone know anything further about these drives. The reason I ask is because we have quite a few of these drives currently in production. They cost a little bit more than other models. We imagine that these improvements increase the longevity of these drives. We would be happy to report back in 5 years.

More information on these drives is available from the Seagate product url.

If you are concerned about your most important digital assets after reading this article, please contact us.

All the best form the Lucid Team.

Where can I get one of those long-lived hard drives ??

30,000 hours, maybe ??

Nowadays the servers keep running during those weeks and besides the used power there are no negative side-effects.

That being said, this year papers from CMU and google do have their indisputable values. They show by real system data that disks are not so reliable after all. One of course wouldn’t be so naive to believe his/her disk can last hundreds of years in real systems. ( A sideline, you can find some statistical data on CD and DVD lifetime from a post at Storage Advisor’s blog, another good storage site. )

A more important question an IT manager faces is how much reliable is enough and how to realize this in system set-up, e.g. google guys choose triplication. Disks evolve, in a few years hopefully, we will see reasonable priced holographic HDs with much bigger capacity and lifetime, but as many people here already said, storage systems fail not only due to disks, but too many other factors, thus the architecture issue always will be there. One way to go may be more smartly clusterred systems, a favorite theme here. Again, triplication is not a good solution in long term, even for google, simply due to too much management, space and power costs. RAID-5 can be easily extended to a clustered system. How about RAID-n ( n > 5 )? The truth is so far, technical we ( both academia and industry ) don’t have a sound solution. Current existing RAID-6 or RAID-7 needs too much CPU and thus causes high I/O latency and low thoughput. This is part of the reason array vendors are not pushing you RAID-6 products. Any one here can provide real RAID-6 experiences?

Back to data loss: a piece of data can lose even though the host disk in whole still functions. This is the so-called sector error. The google paper mentioned it a bit by “scan error”. This is much more common and happens much more frequently from whole disk failure. The effect though is the same, if your data happens to be stored on that failed sector. In this sense, the CMU and google papers are just a beginning. If they are good, they should continue to explore the sector error statistics, which certainly will be much harder. This year’s ACM SIGMETRICS 2007 will have a paper from NetApp on this and hopefully this group of people have done a better job than the FAST 2004 group. ( The paper’s title is : An Analysis of Latent Sector Errors in Disk Drives, see http://www.cs.cmu.edu/~sigm07/ )

Oh, well, I will stop here to see if any one from EMC or NetApp or other start-ups can provide their views on systems …

http://slashdot.org/comments.pl?sid=222978&cid=18058500