[Pacemaker] Question about the behavior when a pacemaker's process crashed

Kazunori INOUE inouekazu at intellilink.co.jp
Thu Jul 25 05:03:21 EDT 2013 

(13.07.25 11:00), Andrew Beekhof wrote:
>
> On 24/07/2013, at 7:40 PM, Kazunori INOUE <inouekazu at intellilink.co.jp> wrote:
>
>> (13.07.18 19:23), Andrew Beekhof wrote:
>>>
>>> On 17/07/2013, at 6:53 PM, Kazunori INOUE <inouekazu at intellilink.co.jp> wrote:
>>>
>>>> (13.07.16 21:18), Andrew Beekhof wrote:
>>>>>
>>>>> On 16/07/2013, at 7:04 PM, Kazunori INOUE <inouekazu at intellilink.co.jp> wrote:
>>>>>
>>>>>> (13.07.15 11:00), Andrew Beekhof wrote:
>>>>>>>
>>>>>>> On 12/07/2013, at 6:28 PM, Kazunori INOUE <inouekazu at intellilink.co.jp> wrote:
>>>>>>>
>>>>>>>> Hi,
>>>>>>>>
>>>>>>>> I'm using pacemaker-1.1.10.
>>>>>>>> When a pacemaker's process crashed, the node is sometimes fenced or is not sometimes fenced.
>>>>>>>> Is this the assumed behavior?
>>>>>>>
>>>>>>> Yes.
>>>>>>>
>>>>>>> Sometimes the dev1 respawns the processes fast enough that dev2 gets the "hey, i'm back" notification before the PE gets run and fencing can be initiated.
>>>>>>> In such cases, there is nothing to be gained from fencing - dev1 is reachable and responding.
>>>>>>
>>>>>> OK... but I want pacemaker to certainly perform either behavior (fence is performed or fence is not performed), since operation is troublesome.
>>>>>> I think that it is better if user can specify behavior as an option.
>>>>>
>>>>> This makes no sense. Sorry.
>>>>> It is wrong to induce more downtime than absolutely necessary just to make a test pass.
>>>>
>>>> If careful of the increase in downtime, isn't it better to prevent fencing, in this case?
>>>
>>> With hindsight, yes.
>>> But we have no way of knowing at the time.
>>> If you want pacemaker to wait some time for it to come back, you can set crmd-transition-delay which will achieve the same thing it does for attrd.
>>
>> I think that only a little is suitable for my demand because crmd-transition-delay is delay.
>
> The only alternative to a delay, either by crmd-transition-delay or some other means, is that the crmd predicts the future.
>
>>
>>>
>>>> Because pacemakerd respawns a broken child process, so the cluster will return to a online state.
>>>> If so, does subsequent fencing not increase a downtime?
>>>
>>> Yes, but only we know that because we have more knowledge than the cluster.
>>
>> Is it because stack is corosync?
>
> No.
>
>> In pacemaker-1.0 with heartbeat, behavior when a child process crashed can be specified by ha.cf.
>> - when specified 'pacemaker respawn', the cluster will recover to online.
>
> The node may still end up being fenced even with "pacemaker respawn".
>
> If the node does not recover fast enough, relative to the "some process died" notification, then the node will get fenced.
> If the "hey the process is back again" notification gets held up due to network congestion, then the node will get fenced.
> Like most things in clustering, timing is hugely significant - consider a resource that fails just before vs. just after a monitor action is run
>
> Now it could be that heartbeat is consistently slow sending out the "some process died" notification (I recall it does not send them at all sometimes), but that would be a bug not a feature.

Sorry, I mistook it.
You're right.
>
>
>> - when specified 'pacemaker on', the node will reboot by oneself.
>
> "by oneself"?  Not because the other side fences it?

Yes, "by oneself".

[14:34:25 root at vm3 ~]$ gdb /usr/lib64/heartbeat/heartbeat 9876
  :
[14:35:33 root at vm3 ~]$ pkill -9 crmd
  :
(gdb) b cl_reboot
Breakpoint 2 at 0x7f0e433bdcf8
(gdb) c
Continuing.

Breakpoint 2, 0x00007f0e433bdcf8 in cl_reboot () from /usr/lib64/libplumb.so.2
(gdb) bt
#0  0x00007f0e433bdcf8 in cl_reboot () from /usr/lib64/libplumb.so.2
#1  0x000000000040d8e4 in ManagedChildDied (p=0x117f6e0, status=<value optimized out>, signo=9,
     exitcode=0, waslogged=1) at heartbeat.c:3906
#2  0x00007f0e433c8fcf in ReportProcHasDied () from /usr/lib64/libplumb.so.2
#3  0x00007f0e433c140c in ?? () from /usr/lib64/libplumb.so.2
#4  0x00007f0e433c0fe0 in ?? () from /usr/lib64/libplumb.so.2
#5  0x0000003240c38f0e in g_main_context_dispatch () from /lib64/libglib-2.0.so.0
#6  0x0000003240c3c938 in ?? () from /lib64/libglib-2.0.so.0
#7  0x0000003240c3cd55 in g_main_loop_run () from /lib64/libglib-2.0.so.0
#8  0x000000000040e8b8 in master_control_process () at heartbeat.c:1650
#9  initialize_heartbeat () at heartbeat.c:1041
#10 0x000000000040f38d in main (argc=<value optimized out>, argv=<value optimized out>, envp=
     0x7fffe0ba9bd8) at heartbeat.c:5133
(gdb) n

Message from syslogd at vm3 at Jul 25 14:36:57 ...
  heartbeat: [9876]: EMERG: Rebooting system.  Reason: /usr/lib64/heartbeat/crmd
>
>> I want to perform a setup and operation (established practice) equivalent to it.

This is a patch to add the option which can choose to reboot a machine at the time of child process failure.
https://github.com/inouekazu/pacemaker/commit/c1ac1048d8
What do you think?

>>
>>>
>>>>
>>>> Best regards.
>>>>
>>>>>
>>>>>>>
>>>>>>> It makes writing CTS tests hard, but it is not incorrect.
>>>>>>>
>>>>>>>>
>>>>>>>> procedure:
>>>>>>>> $ systemctl start pacemaker
>>>>>>>> $ crm configure load update test.cli
>>>>>>>> $ pkill -9 lrmd
>>>>>>>>
>>>>>>>> attachment:
>>>>>>>> STONITH.tar.bz2 : it's crm_report when fenced
>>>>>>>> notSTONITH.tar.bz2 : it's crm_report when not fenced
>>>>>>>>
>>>>>>>> Best regards.
>>>>>>>> <notSTONITH.tar.bz2><STONITH.tar.bz2>_______________________________________________
>>>>>>>> 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://bugs.clusterlabs.org
>>>>>>>
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> 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://bugs.clusterlabs.org

More information about the Pacemaker mailing list