[ClusterLabs] Antw: Re: 2-Node Cluster Pointless?

[Digimer]

On 18/04/17 08:50 PM, Dimitri Maziuk wrote:
> On 04/18/2017 07:05 PM, Digimer wrote:
> 
>> Certainly, the people creating the software have to assume that a
>> split-brain is devastating. Same for people who teach others and people
>> who write documentation.
> 
> <sigh.../> split brain is devastating when you're drbd. When you're a
> mere floating ip, assume makes an ass of you and me.

If you want to talk about a hypothetical cluster where the only thing
hosted is an IP, and nothing else, then yes, a split-brain would mean
that you have a random node accessible. This is not a common
configuration, to be sure, if you don't care that the active node will
waffle back and forth whenever a node says that it has the IP.

The majority of clusters are not so simple. In the *vast* majority of
real-world clusters, a split-brain is devastating. We've seen for years
how people misbelieve that fencing is optional, and we've seen countless
people coming along asking why their cluster "that worked fine for
years!", is not offline or, worse, destroyed.

Given how rare your proposed scenario is, it strikes me that suggesting
fencing is optional is a dis-service with potentially devastating
consequences for those new to the technology.

So, I stand by my statement that all clusters need fencing.

-- 
Digimer
Papers and Projects: https://alteeve.com/w/
"I am, somehow, less interested in the weight and convolutions of
Einstein’s brain than in the near certainty that people of equal talent
have lived and died in cotton fields and sweatshops." - Stephen Jay Gould