[Pacemaker] spilit brain situation

[Andrew Beekhof]

On Feb 6, 2009, at 12:11 PM, Lars Marowsky-Bree wrote:

> On 2009-02-06T10:39:42, Andrew Beekhof <beekhof at gmail.com> wrote:
>
>> which is why two node clusters are good for demonstrations and  
>> thats about
>> it.
>> there are some limited uses for them, but in general you need 3  
>> nodes for a
>> sane cluster.
>
> Andrew, this is not true. The majority of all clusters out ther - 90%+
> would be my guess - are 2 nodes only, and _work perfectly fine_ with
> STONITH.


I'm not going to rehash the entire IRC conversation I just had with  
Lars, suffice to say that there are traps to setting up a two node  
cluster that I consider non-obvious to people new to clustering...  
resulting in clusters that appear to work right up until they implode.

I will concede though that they can be made to work safely :-)

However I maintain that adding an old/underpowered box to act as a tie- 
breaker (it doesn't even need to run resources to be useful) is still  
the saner choice as it means the cluster behaves more predictably and  
allows it to respond better to transient failures (because its no  
longer necessary to prevent stonith'd nodes from rejoining the cluster).