Re: Hardware/OS recommendations for large databases (

William Yu wrote:
>
> Our SCSI drives have failed maybe a little less than our IDE drives.

Microsoft in their database showcase terraserver project has
had the same experience. They studied multiple configurations
including a SCSI/SAN solution as well as a cluster of SATA boxes.

They measured a
6.4% average annual failure rate of their SATA version and a
5.5% average annual failure rate on their SCSI implementation.

ftp://ftp.research.microsoft.com/pub/tr/TR-2004-107.pdf

"We lost 9 drives out of 140 SATA drives on the Web and
Storage Bricks in one year. This is a 6.4% annual failure rate.
In contrast, the Compaq Storageworks SAN and Web servers lost
approximately 32 drives in three years out of a total of 194
drives.13 This is a 5.5% annual failure rate.

The failure rates indicate that SCSI drives are more
reliable than SATA. SATA drives are substantially
cheaper than SCSI drives. Because the SATA failure rate
is so close to the SCSI failure rate gives SATA a
substantial return on investment advantage."

So unless your system is extremely sensitive to single drive
failures, the difference is pretty small. And for the cost
it seems you can buy enough extra spindles of SATA drives to
easily make up for the performance difference.

> Basically, I've found it's cooling that's most important. Packing the
> drives together into really small rackmounts? Good for your density, not
> good for the drives.

Indeed that was their guess for their better-than-expected
life of their SATA drives as well. From the same paper:

"We were careful about disk cooling – SATA
drives are rarely cooled with the same care that a SCSI
array receives."