[funsec] Database design.

[Drsolly]

On Tue, 6 Jun 2006 Valdis.Kletnieks at vt.edu wrote:

> On Tue, 06 Jun 2006 21:49:18 BST, Drsolly said:
> 
> > 300 gb drives cost about $100. You can put four of those in one computer, 
> > total cost maybe $600. For 460 terabytes, you'd need some 400 of those, 
> > which would cost $240,000.
> 
> Yeah, and that DMX3 quote would have been a lot cheaper if they'd just used
> cheap $100 ATA drives, instead of the drives they actually use.  However,
> when you're looking at 300 terabytes of disk, if you're using drives with
> a 3-year MTBF like they use in the average consumer desktop, you're looking

I'm not sure that the MTBF is 3 years. The guarantee on these things is 3 
years. If they're expecting half of the drives they sell, to become 
warranty replacements, then they did their sums wrong. I suspect that the 
MTBF is a lot more than 3 years (and that's borne out by my experience 
with them).

I checked. The mtbf for these drives (according to Maxtor), is 1.2 million 
hours = 137 years.

> at a failure *EVERY DAY* on the average.  Now the fun - even if you have a

So if you have 1000 drives, the mtbf would be around 1200 hours. One drive 
failure every seven weeks.

> hot spare that the array can rebuild onto, what is the probability that
> (a) you'll have another drive fail before the rebuild finishes and (b) that
> the 2nd failure is inside the rebuilding RAID set?

That's not how I'd do it - I'd have three servers running in parallel. If 
one fails, you replace it while the other two are still running.

Also, for a national database system like this, you'd have the database 
replicated in more than one site, to allow for the possibility of an 
outage at one site.
 
> Hmmm.. let's see.  A petabyte would be about 3,000 drives, and a failure
> every 8 hours.  Assuming you create 100 RAID sets of 30 drives each, and
> you can rebuild a 30-drive RAID in 24 hours, you're looking at permanently
> trashing a RAID set on the order of once or twice a year....  (You can
> improve your chances by creating more RAID sets - but then you need to
> allocate more hot spares....)
 
I suggest you redo your calculation with the correct mtbf of 1.2 million 
hours.