[Pacemaker] Odd cluster constraint behaviour

[Oliver Hookins]

On Mon Jul 06, 2009 at 11:11:14 +0200, Dejan Muhamedagic wrote:
>Hi,
>
>On Mon, Jul 06, 2009 at 03:33:13PM +1000, Oliver Hookins wrote:
>> I'm not sure if I've configured things correctly, as the last time I did
>> this was on Heartbeat 2.0.7 or so. It's either that as a bug (and it's far
>> more likely I've stuffed something up):
>>  * there are two resource groups, one with higher priority (master) and one with
>>    lower priority (slave) - note that I'm not actually configuring them as master/slave resources
>>  * two nodes in the cluster (A and B)
>>  * they are constrained to run:
>>   - only if a pingd instance is successfully running
>>   - colocated with -INFINITY to each other (i.e. cannot run together)
>>   - in the order of the "master" resource group and then the "slave" resource group
>>   - preferring resource group "master" running on node A
>> 
>> With the cluster in its default running state of "master" on A, and "slave" on B,
>> everything seems fine. When I "fail" node A, "slave" stops on node B but then no
>> resources are started on node B. It's clearly a conflict of constraints, or
>> the resource group priorities being ignored but I can pick it.
>> 
>> Here are (hopefully) the relevant snippets of CIB:
>> 
>>     <resources>
>>       <group id="master">
>>         <meta_attributes id="master-meta_attributes">
>>           <nvpair id="master-meta_attributes-priority" name="priority" value="1000"/>
>>         </meta_attributes>
>> 
>>       <group id="slave">
>>         <meta_attributes id="slave-meta_attributes">
>>           <nvpair id="slave-meta_attributes-priority" name="priority" value="0"/>
>>         </meta_attributes>
>> 
>>       <clone id="pingdclone">
>>         <meta_attributes id="pingdclone-meta_attributes">
>>           <nvpair id="pingdclone-meta_attributes-globally-unique" name="globally-unique" value="false"/>
>>         </meta_attributes>
>>         <primitive class="ocf" id="pingd" provider="pacemaker" type="pingd">
>>           <instance_attributes id="pingd-instance_attributes">
>>             <nvpair id="pingd-instance_attributes-host_list" name="host_list" value="X.X.X.X"/>
>>             <nvpair id="pingd-instance_attributes-multiplier" name="multiplier" value="100"/>
>>           </instance_attributes>
>>           <operations>
>>             <op id="pingd-monitor-15s" interval="15s" name="monitor" timeout="5s"/>
>>           </operations>
>>         </primitive>
>>       </clone>
>>     </resources>
>>     <constraints>
>>       <rsc_location id="cli-prefer-master0" node="nodeA" rsc="master" score="1000"/>
>>       <rsc_location id="cli-prefer-master1" node="nodeB" rsc="master" score="0"/>
>>       <rsc_colocation id="separate_master_and_slave" rsc="master" score="-INFINITY" with-rsc="slave"/>
>
>This colocation should be the other way around, i.e. exchange rsc
>and with-rsc attributes. If that doesn't help, please post the
>logs.

Yes, that fixed the problem. Can you point me in the direction of the
documentation that explains why this is the case? I've been reading Andrew's
"Configuration Explained 1.0" document but didn't pick up on this.

-- 
Regards,
Oliver Hookins
Anchor Systems