06.06.2013 09:02, Andrew Beekhof wrote:
> 
> On 06/06/2013, at 3:45 PM, Vladislav Bogdanov <[email protected]> wrote:
> 
>> 06.06.2013 08:14, Andrew Beekhof wrote:
>>>
>>> On 06/06/2013, at 2:50 PM, Vladislav Bogdanov <[email protected]> wrote:
>>>
>>>> 06.06.2013 07:31, Andrew Beekhof wrote:
>>>>>
>>>>> On 06/06/2013, at 2:27 PM, Vladislav Bogdanov <[email protected]> 
>>>>> wrote:
>>>>>
>>>>>> 05.06.2013 02:04, Andrew Beekhof wrote:
>>>>>>>
>>>>>>> On 05/06/2013, at 5:08 AM, Ferenc Wagner <[email protected]> wrote:
>>>>>>>
>>>>>>>> Dejan Muhamedagic <[email protected]> writes:
>>>>>>>>
>>>>>>>>> On Mon, Jun 03, 2013 at 06:19:06PM +0200, Ferenc Wagner wrote:
>>>>>>>>>
>>>>>>>>>> I've got a script for resource creation, which puts the new resource 
>>>>>>>>>> in
>>>>>>>>>> a shadow CIB together with the necessary constraints, runs a 
>>>>>>>>>> simulation
>>>>>>>>>> and finally offers to commit the shadow CIB into the live config (by
>>>>>>>>>> invoking an interactive crm).  This works well.  My concern is that 
>>>>>>>>>> if
>>>>>>>>>> somebody else (another cluster administrator) changes anything in the
>>>>>>>>>> cluster configuration between creation of the shadow copy and the
>>>>>>>>>> commit, those changes will be silently reverted (lost) by the commit.
>>>>>>>>>> Is there any way to avoid the possibility of this?  According to
>>>>>>>>>> http://article.gmane.org/gmane.linux.highavailability.pacemaker/11021,
>>>>>>>>>> crm provides this functionality for its configure sessions [*], but 
>>>>>>>>>> the
>>>>>>>>>> shadow CIB route has good points as well (easier to script via 
>>>>>>>>>> cibadmin,
>>>>>>>>>> simulation), which I'd like to use.  Any ideas?
>>>>>>>>>
>>>>>>>>> Record the two epoch attributes of the cib tag at the beginning
>>>>>>>>> and check if they changed just before applying the changes.
>>>>>>>>
>>>>>>>> Maybe I don't understand you right, but isn't this just narrowing the
>>>>>>>> time window of the race?  After all, that concurrent change can happen
>>>>>>>> between the epoch check and the commit, can't it?
>>>>>>>
>>>>>>> The CIB will refuse to accept any update with a "lower" version:
>>>>>>>
>>>>>>> http://clusterlabs.org/doc/en-US/Pacemaker/1.1-pcs/html/Pacemaker_Explained/_configuration_version.html
>>>>>>
>>>>>> I recall that LDAP has similar problem, which is easily worked around
>>>>>> with specifying two values, one is original, second is new.
>>>>>> That way you tell LDAP server:
>>>>>> Replace value Y in attribute X to value Z. And if value is not Y at the
>>>>>> moment of modification request, then command fails.
>>>>>
>>>>> "cibadmin --patch" works this way
>>>>
>>>> Who is baking new CIB in that case, cibadmin or cib?
>>>
>>> The patch is applied on the server - so "cib"
>>
>> Ah, one more question. The whole modification request is rejected if any
>> of patch hunks fail, correct?
> 
> Correct (and yes everything is serialized _unless_ you start using the -l 
> cibadmin option)

Great. Thanks.

_______________________________________________
Linux-HA mailing list
[email protected]
http://lists.linux-ha.org/mailman/listinfo/linux-ha
See also: http://linux-ha.org/ReportingProblems

Reply via email to