[Pacemaker] About behavior in "Action Lost".

[Andrew Beekhof]

Pushed as:
   http://hg.clusterlabs.org/pacemaker/1.1/rev/8433015faf18

Not sure about applying to 1.0 though, its a dramatic change in behavior.

On Wed, Sep 22, 2010 at 11:18 AM,  <renayama19661014 at ybb.ne.jp> wrote:
> Hi Andrew,
>
> Thank you for comment.
>
>> A long time ago in a galaxy far away, some messaging layers used to
>> loose quite a few actions, including stops.
>> About the same time, we decided that fencing because a stop action was
>> lost wasn't a good idea.
>>
>> The rationale was that if the operation eventually completed, it would
>> end up in the CIB anyway.
>> And even if it didn't, the PE would continue to try the operation
>> again until the whole node fell over at which point it would get shot
>> anyway.
>
> Sorry...
> I did not know the fact that there was such an argument in old days.
>
>
>> Now, having said that, things have improved since then and perhaps,
>> the interest of speeding up recovery in these situations, it is time
>> to stop treating stop operations differently.
>> Would you agree?
>
> That means, you change it in the case of "Action Lost" of the stop this time to carry out stonith?
> If my recognition is right, I agree too.
>
> if(timer->action->type != action_type_rsc) {
> send_update = FALSE;
> } else if(safe_str_eq(task, "cancel")) {
> /* we dont need to update the CIB with these */
> send_update = FALSE;
> }
> ---> delete "else if(safe_str_eq(task, "stop")){..}" ?
>
> if(send_update) {
> /* cib_action_update(timer->action, LRM_OP_PENDING, EXECRA_STATUS_UNKNOWN); */
> cib_action_update(timer->action, LRM_OP_TIMEOUT, EXECRA_UNKNOWN_ERROR);
> }
>
> Best Regards,
> Hideo Yamauchi.
>
> --- Andrew Beekhof <andrew at beekhof.net> wrote:
>
>> On Tue, Sep 21, 2010 at 8:59 AM,  <renayama19661014 at ybb.ne.jp> wrote:
>> > Hi,
>> >
>> > Node was in state that the load was very high, and we confirmed monitor movement of Pacemeker.
>> > Action Lost occurred in stop movement after the error of the monitor occurred.
>> >
>> > Sep &#65533;8 20:02:22 cgl54 crmd: [3507]: ERROR: print_elem: Aborting transition, action lost:
>> [Action 9]:
>> > In-flight (id: prmApPostgreSQLDB1_stop_0, loc: cgl49, priority: 0)
>> > Sep &#65533;8 20:02:22 cgl54 crmd: [3507]: info: abort_transition_graph: action_timer_callback:486
> -
>> > Triggered transition abort (complete=0) : Action lost
>> >
>> >
>> > For the load of the node, We think that the stop movement did not go well.
>> > But cannot nodes execute stonith.
>>
>> A long time ago in a galaxy far away, some messaging layers used to
>> loose quite a few actions, including stops.
>> About the same time, we decided that fencing because a stop action was
>> lost wasn't a good idea.
>>
>> The rationale was that if the operation eventually completed, it would
>> end up in the CIB anyway.
>> And even if it didn't, the PE would continue to try the operation
>> again until the whole node fell over at which point it would get shot
>> anyway.
>>
>> Now, having said that, things have improved since then and perhaps,
>> the interest of speeding up recovery in these situations, it is time
>> to stop treating stop operations differently.
>> Would you agree?
>>
>> _______________________________________________
>> Pacemaker mailing list: Pacemaker at oss.clusterlabs.org
>> http://oss.clusterlabs.org/mailman/listinfo/pacemaker
>>
>> Project Home: http://www.clusterlabs.org
>> Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf
>> Bugs: http://developerbugs.linux-foundation.org/enter_bug.cgi?product=Pacemaker
>>
>
>
> _______________________________________________
> Pacemaker mailing list: Pacemaker at oss.clusterlabs.org
> http://oss.clusterlabs.org/mailman/listinfo/pacemaker
>
> Project Home: http://www.clusterlabs.org
> Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf
> Bugs: http://developerbugs.linux-foundation.org/enter_bug.cgi?product=Pacemaker
>