[Openais] cpg behavior on transitional membership change

[Vladislav Bogdanov]

Hi Jiaju,

03.09.2011 19:52, Jiaju Zhang wrote:
> On Fri, Sep 02, 2011 at 10:12:11PM +0300, Vladislav Bogdanov wrote:
>> 02.09.2011 20:55, David Teigland wrote:
>> [snip]
>>>
>>> I really can't make any sense of the report, sorry.  Maybe reproduce it
>>> without pacemaker, and then describe the specific steps to create the
>>> issue and resulting symptoms.  After that we can determine what logs, if
>>> any, would be useful.
>>>
>>
>> I just tried to ask a question about cluster components logic based on
>> information I discovered from both logs and code analysis. I'm sorry if
>> I was unclear in that, probably some language barrier still exists.
>>
>> Please see my previous mail, I tried to add some explanations why I
>> think current logic is not complete.
> 
> Hi Vladislav, I guess I have known the problem what you described;)
> I'd like to give a example to make the things more clear.
> 
> 3-node cluster, for whatever reason, especially on heavy workload,

(my case)

> corosync may detect one node disappear and reappear again. So the

BTW, could this be prevented (at least for majority of cases) by some
corosync timeout params?
Steve?

> membership information changes are as follows:
> membership 1: nodeA, nodeB, nodeC
> membership 2: nodeB, nodeC
> membership 3: nodeA, nodeB, nodeC

Exactly.

> 
> From the membership change 1 -> 2, dlm_controld konws nodeA is down,
> and have many things to do, like check_fs_done, check_fencing_done ...
> The key point here is dlm need to wait the fencing is really done
> before it proceed. If we employ a cluster filesystem here, like ocfs2,
> it also needs the fencing is really done. I believe in the normal
> cases, pacemaker will fence nodeA and then everything should be OK.
> 
> However, there is a possibility here that pacemaker won't fence nodeA.
> Say nodeA is the original DC of the cluster, when nodeA is down, the
> cluster should elect a new DC. But if the time window where membership
> change 2 -> 3 is too small, node A is up again and attend the election
> too, then node A is elected to be the DC again and it won't fence
> itself.

Ahm, I think you are perfectly right in what exactly happens.
This addresses my case I think, just because that node which left
cluster for a short moment is usually
1. Started earlier on a cold boot (it is a VM, others are bare-metal
which boot diskless via PXE, and there is 1 minute timeout before that
bare-metal systems get their boot image, just because of cisco
implementation of ether-channel)
2. Upgraded (and rebooted if needed) first, before bare-metal nodes are
booted with new image.

So that node is usually a DC.

> Andrew, correct me if my understanding on pacemaker is wrong;)
> 
> So I think the membership change should be like a transaction in
> database or filesystem field, that is, for the membership change
> 1 -> 2, every thing should be done (e.g. fencing nodeA), no matter the
> following change 2 -> 3 will happen or not. For the situation where a
> node magically disappear and reappear, and the situation where a node
> normally down and then up, ocfs2 and dlm should not be able to see any
> difference between them, what they can do is just waiting the fencing
> to be done.
> 
> Any comments? thoughts?

I'd have protection at as many layers as possible. So, if you are right
about pacemaker, then it is great if it could be fixed in pacemaker.
But, I'd prefer to be safe and have DLM et all schedule fencing as well
if they notice unrecoverable problem which leads to cluster subsystem
freeze if fencing is not done for whatever reason. I suppose that
current behavior is just a artifact from cluster2, where groupd was on
duty for that event (I may be wrong, because I didn't look at groupd
code closely, but comments in other daemons code make me think so).

Thank you for your comments,
Vladislav