[plug] RAID-1 vs rsync vs ? for high availability

[Denis Brown]

At 23:18 6/12/2002 +0800, levsky at iinet.net.au wrote:
>On Fri, Dec 06, 2002 at 10:56:21PM +0800, Denis Brown wrote:
> > The RAID HOWTO points me in the direction of RAID-1 and by using

>...snip...

>RAID-1 is cool, but you take quite a performance hit on writes.
>Plus, it's *really* wasteful on disk resources.  I'm fond of RAID-5
>personally, if implemented with a decent hardware raid controller.
>Just depends on what you're doing I suppose - If you're just mirroring
>your system drive, then RAID-1 is likely to be more than adequate.
>If you're trying to store lots of data, it's probably not.

Data volume is not going to be a big issue here (famous last words!) at 
least initially, so RAID-1 seemed to be the way to go.  Agree about the 
wasteful nature of the concept re disk resources but I'd rather not have to 
explain why my users' precious appointments have just gone to /dev/null 
:-)   When funds permit (aka when the toe-in-the-door has become a 
full-fledged foot) time to ask for more money to support RAID-5 and a 
hardware controller.

Okay, to a degree I am being a bit pessimistic -- I have one system that 
had a 400+ days uptime and only went down due to major electrical 
switchboard work (12+ hours) where no UPS could have coped, and another 
that had 280+ days continuous so the track record has been pretty good even 
on hardware to which no special attentions in the line of disks, have been 
paid.

>If you're planning to have more than about half a gig of ram in your
>machine, some form of ECC is a must.  With standard bit error rates
>what they are, your probability of having some ram bit errors over
>a couple of months is very high.

Thanks, very good thought.   Which begs the question.. automated RAM 
monitoring?   Maybe the H-A HowTo will have something to say and I have yet 
to hunt the web.   The thought I have is to have a daemon running at a low 
priority which scans memory x-hundred-K at a time.   When a free segment is 
"approved okay" and is contiguously large enough to contain untested but 
currently-occupied memory, the process employing the to-be-tested memory 
segment is stopped, moved to the tested segment, restarted and the 
previously-busy memory is tested.   On a round-Robin basis so that over the 
course of, let's say a day, all memory is tested.   Might get tricky for 
kernel memory occupancy though, eh?

>Really - as you've already intimated, the best solution is to have
>redundant boxes.

Yes, pipe-dream territory now but worthwhile over the long term, especially 
when I can distribute the hardware across physical sites and give users a 
more responsive system into the bargain.

Sincere thanks for your feedback,
Denis