[Pacemaker] advisory colocation constraint group->service

[Andreas Kurz]

On Wednesday 29 October 2008 15:56:13 Andrew Beekhof wrote:
.....
>> --- cib_group-res-constr.orig.xml    2008-10-29 11:09:47.000000000 +0100
>> +++ cib_group-res-constr.xml 2008-10-29 11:26:13.000000000 +0100
>> @@ -265,7 +265,7 @@
>>        <rsc_colocation id="bgp-upc-colocate-with-ip_upc"
>> from="service_bgp-upc" to="service_ip_upc" score="INFINITY"/>
>>        <rsc_order id="bgp-telekom-after-ip_telekom"
>> from="service_bgp-telekom" type="after" to="service_ip_telekom"
>> score="0"/>
>>        <rsc_colocation id="bgp-telekom-colocated-with-ip_telekom"
>> from="service_bgp-telekom" to="service_ip_telekom" score="INFINITY"/>
>> -       <rsc_colocation id="bgp-upc-colocated-with-bgp_telekom"
>> from="service_bgp-upc" to="service_bgp-telekom" score="1"
>> symmetrical="true"/>
>    ^^^^^^^^^^^
> > My intention for this (^) was to loosely couple them together so
> > they run
> > together on the same host. If I decide to stop one of them  and
> > restart it
> > after a while it should be colocated again with its "partner" ... if
> > i decide
> > to migrate only one of the  two resources to another node the other
> > resource
> > should stay where it is running.
>
> This is not at all related what symmetrical does.
> Symmetrical was a work-around for a problem a long time ago and no-one
> should be setting it to true.

Aaah ... good to know ;-) ... so explicitely creating an extra advisory 
colocation constraint in the opposite direction is the way to go? 

...
> > Please correct me if I'am wrong but for me it looks like the score
> > of the
> > colocation is ignored and always INFINITY.
>
> not quite
>
> > I expected the advisory colocation
> > constraint to not stop an already runing resource to fullfill the
> > constraint ... is this a misinterpretion?
>
> no
>
> the problem was how non-matches were handled.
> matches gained constraint.score (correct), but non-matches always got
> set to -INFINITY (wrong).
>
> The following patch makes your scenario behave, I just need to run it
> through the regression tests.

Great, thanks Andrew! ... will this patch be available in Heartbeat 2.1.4?

Regards,
Andreas

>
> diff -r 61b457707d67 pengine/native.c
> --- a/pengine/native.c	Thu Oct 23 17:05:06 2008 +0200
> +++ b/pengine/native.c	Wed Oct 29 15:53:02 2008 +0100
> @@ -762,9 +762,9 @@ colocation_match(
>   		    }
>
>   		} else if(do_check == FALSE || constraint->score >= INFINITY) {
> -			crm_debug_2("%s: %s.%s = -INFINITY (%s)", constraint->id, rsc_lh-
>
>  >id,
>
> -				  node->details->uname, do_check?"failed":"unallocated");
> -			node->weight = -INFINITY;
> +			crm_debug_2("%s: %s.%s -= %d (%s)", constraint->id, rsc_lh->id,
> +				    node->details->uname, constraint->score,
> do_check?"failed":"unallocated");
> +			node->weight = merge_weights(-constraint->score, node->weight);
>   		}
>
>   		);
.....