RAID-5 and database servers

Jefferson Ogata poweredge at
Fri Mar 12 18:18:06 CST 2010

On 2010-03-12 17:45, Craig White wrote:
> On Fri, 2010-03-12 at 15:57 +0000, Jefferson Ogata wrote:
>> On 2010-03-12 15:39, Craig White wrote:
>>> I don't think I understand your 'odds' model. I interpret the first
>>> example as RAID 50 having 5 times more likelihood of loss than RAID 10
>>> and I presume that isn't what you were after
>> Yes, it is 5 times higher. But it is not 100%; it's actually less than
>> 50%. And the probability for RAID 10 is not 50% as you said it was. I
>> was just correcting your analysis. I'm still not sure what RAID
>> structure you had in mind where a second failure on a
>> RAID 10 has a 50% probability of loss.
> ----
> sorry I wasn't clear but I thought you would figure it out.
> Say you have a 4 disk RAID 10 array. If you lose 2 disks, your chances
> are 50% that the RAID 10 array is unrecoverable. If you lose both
> elements of one stripe or both elements of one mirror. That's my
> understanding anyway.

The odds of losing both elements of one mirror on a 4-disk RAID 10 are 
1/3. After the first disk fails, there are three remaining; only one of 
these will kill your RAID if it dies.

Another way to look at it: number the disks 0-3, and say your RAID 10 is 
a stripe of a 0-1 mirror and a 2-3 mirror. Here are all the ways two 
disks can fail:

0 1 *
0 2
0 3
1 2
1 3
2 3 *

The * are the ones that cause loss, 2 out of 6 cases.

> I admit I am far from the most knowledgeable person on this topic and I
> sat on the sidelines for both of the discussions but felt that the
> article from enterprisestorage needed to be linked because clearly there
> are sufficient issues with the typical high density, large SATA drives
> and RAID 5. I have yet to see anything that would change my mind from
> thinking that the only reason to use RAID 5 is to maximize storage per
> dollar which may very well come with performance and reliability issues
> that should not be unspoken.

There's no dispute that RAID 10 is on average more reliable by itself. 
But there's a lot more to keeping a data center running that the odds of 
losing data on a RAID. If having twice as much disk chews up your budget 
so you can't afford that UPS upgrade, then the next power hit may kill 
your whole data center. If the heat from those extra disks reduces your 
runtime without HVAC from 30 minutes to 20 minutes, you might get there 
10 minutes too late to shut things down without taking 1000 hours off 
the life of all of those disks.

People who tell other people to just go out and buy RAID 10 without 
knowing anything about the workload or the impact on the rest of the 
data center are *not* helping them. I'm not even sure why these people 
stop at RAID 10. How about network RAID with geographic redundancy? How 
about RAID 10 with 3 replicas per mirror?

More information about the Linux-PowerEdge mailing list