Any organization that’s serious about its data integrity uses storage systems based on RAID level 6 (that's block-level striping with double distributed parity) which can tolerate the failure of two simultaneous disk failures.
While RAID 6 is a sound choice it’s also expensive to implement as this scheme requires a minimum of 4 drives to implement and 1-2/n% of the total system capacity is lost to support data integrity (i.e. 50% for 4 drives, 60% for 5 drives, 66% for 6 drives, and so on). But there’s an even greater overhead no matter what kind of RAID is being used: If a drive fails then a technician is needed to replace the failed unit and until the system has been rebuilt (a process that can take days with large drives) there’s the possibility of catastrophic failure from subsequent failures.
To overcome these problems a group of researchers, Jehan-François Pâris of University of Houston; Ahmed Amer Darrell of Santa Clara University; D. E. Long of University of California; and Thomas Schwarz of Universidad Católica del Uruguay, have come up with a better storage resiliency solution that gets around maintenance problems by never requiring disks to be replaced and ensuring "five nines" reliability. The authors point out that with maintenance, in particular, "… the cost of the service call is likely to exceed that of the equipment being replaced. This is even truer for installations that are far away from metropolitan areas."
What the researchers' paper, Self-Repairing Disk Arrays, proposes are:
… disk arrays that contain enough spare disks to free users from all maintenance tasks over the expected lifetime of each array. Human intervention will only be needed if the observed disk failure rates significantly exceed 4 to 5 percent per year. This would be the case if the installed disks happened to belong to a bad batch. The solution would be to replace the defective array with a new one.
The paper explains that to be cost-effective there are three challenges to be met:
… the initial cost of the array, its performance and its long-term reliability. The results we present here show that we can build disk arrays that can achieve 99.999 percent reliability over four years with a space overhead not exceeding that of mirroring. Our model assigns equal failure rates to both active drives and spare drives, and assumes that these rates will follow a bathtub curve with high failure rates for the first eighteen months, much lower failure rates for the next eighteen months and much higher failure rates after that.
The model the researchers came up with was based on data from Backblaze, a leading online backup service, on the disk failure rates of over 25,000 drives. The model’s assumptions were that “disks will fail at a rate of 5.1 percent per year during their first 18 months, then at 1.4 percent per year during the next 18 months, and at 11.8 percent per year after that”, that spare disks that are yet unused fail at the same rate as the other disks (“which is a very pessimistic assumption”), and the disk array will remain operational for four years.
Using a discrete simulation program they tested various configurations of data, parity, and spare disks and concluded that a configuration of 45 data disks with 10 parity disks would achieve “five nines” reliability with the lowest space overhead of all of the analyzed configurations, namely 48.86% (for 33 spares) and 49.44% (34 spares).
Interestingly and somewhat counter-intuitively more drives doesn’t improve reliability: "… an array with 66 data disks and 12 parity disks will never be able to achieve five nines over four years, even when provided with an unlimited supply of spares."
The implications of this storage design for enterprise IT are huge in terms of reducing costs and maximizing data integrity. In fact, the tradeoff between the increased price of arrays built this way compared to the expense of having humans service equipment should actually reduce overall operational costs.
This is, however, a research paper and products based on this technology have yet to be built. That said, given the advantages of this approach, we can expect to see storage systems based on these designs in the near future.