[Pacemaker] RFC: How do you upgrade?

[Andrew Beekhof]

On May 8, 2008, at 8:43 AM, Keisuke MORI wrote:

> Hi,
>
> Andrew Beekhof <abeekhof at suse.de> writes:
>> Me again,
>>
>> I'd like to get a sense of how people upgrade their clusters...
>>
>> * rolling upgrade (node-by-node)
>> * big-bang (all nodes at once with all resources stopped)
>> * detach and reattach (all nodes at once with resources still  
>> running)

actually, i guess a rolling version of "detach and reattach" is also  
possible.

>>
>> * other?
>>
>> I'm interested to know how popular the various methods are.
>
> Our preference is:
>  detach/reattach  > rolling upgrade  > big-bang
> because the shorter service outage the better, of course.
>
> But in the real world, we haven't used the detach/reattach way yet.
>
> In the case of our existing systems that use another commercial
> cluster software, it just had no capability to detach/reattach and
> we've used rolling upgrade as long as possible.
>
> In the case of heartbeat/pacemaker, we have used big-bang only because
> we have too few experiences about the operation to accomplish
> detach/reattach in the production environment.
>
>
> Besides of this, there is another scenario that our customers
> ask us sometime:
> - the application upgrade without stopping the cluster.
> i.e. suspend to control a particular resource from heatbeat
> (no monitor, no failover action for the resource)
> without stopping other resources and the cluster itself.

I think setting is-managed-default=false and disabling the monitor  
actions would achieve this.