[Linux-HA] Announcement: New book about Linux-HA Version 2

[David Lang]

On Wed, 4 Jun 2008, Dejan Muhamedagic wrote:

> Hi,
>
> On Wed, Jun 04, 2008 at 01:16:31AM +0200, Lars Marowsky-Bree wrote:
>> On 2008-06-03T16:35:22, "Hildebrand, Nils, 232" <Nils.Hildebrand at bamf.bund.de> wrote:
>>
>>>> - [Quality not important for hardware?]
>>> I think the truth is somewhere in between. If you have a cluster it is
>>> not that important if a node goes down (depending on how available your
>>> services have to be).
>>
>> Right, but telling customers to save on hardware because they are
>> getting clustering is, uhm, not something I'd advocate ;-)
>
> Indeed. Clusters are not a replacement for good hardware. A
> perfect cluster is one which is never exercised (unless when
> testing) ;-)

I will disagree with this. I've seed companies spend insane amounts of money on 
hardware that is internally redundant (to the point of having nultiple 
motherboards and cpus running in lock-step and periodicly cross-checking their 
results) when a cluster on much cheaper hardware would have been the right thing 
to do.

and ignoring this extreme case it can still be better (both in reliability and 
in cost) to buy 2 single-drive/single-power boxes and set them up in a cluster 
then to buy 1 fully redundant box, and with a good cluster implementation (like 
heartbeat) the difference in total reliability between a cluster build from two 
fo the cheap boxes and one built with two of the expensive boxes can be very 
small.

I'm asuming for the example above that both the 'cheap' and 'expensive' boxes 
are purchased from the same company.

a perfect cluster is not one that is never excercised, it's one that provides 
uninterrupted service to it's clients. A cluster can do this in the face of many 
upgrades (including hardware forklift upgrades, OS upgrades, kernel upgrades, 
reboots, etc) that no single box solution can match, no matter how much more you 
spend on the hardware.

David Lang