[Pacemaker] Migration atomicity

Vladislav Bogdanov bubble at hoster-ok.com
Fri Mar 16 11:35:46 UTC 2012


16.03.2012 05:46, Andrew Beekhof wrote:
> On Thu, Mar 15, 2012 at 3:22 PM, Vladislav Bogdanov
> <bubble at hoster-ok.com> wrote:
>> 15.03.2012 01:49, Andreas Kurz wrote:
>>> On 03/14/2012 08:40 AM, Vladislav Bogdanov wrote:
>>>> Hi,
>>>>
>>>> I'm observing a little bit unintuitive behavior of migration logic when
>>>> transition is aborted (due to CIB change) in the middle of the resource
>>>> migration.
>>>>
>>>> That is:
>>>> 1. nodea: migrate_to nodeb
>>>> 2. transition abort
>>>> 3. nodeb: stop
>>>> 4. nodea: migrate_to nodec
>>>> 5. nodec: migrate_from nodea
>>>> (note: no stop on nodea)
>>>>
>>>> While I expect migration operation pair to be "more atomic":
>>>> 1. nodea: migrate_to nodeb
>>>> 2. transition abort
>>>> 3. nodeb: migrate_from nodea
>>>> 4. nodea: stop
>>>> 5. nodeb: migrate_to nodec
>>>> 6. nodec: migrate_from nodeb
>>>> 7. nodeb: stop
>>>>
>>>> Is the current behavior intended?
>>>
>>> You mean that a migration is rolled-back due to a transition abort --
>>> depending on its progress? I think that is the defined (and intended)
>>> behavior since quite a long time ... maybe Andrew likes to comment on that?
>>
>> RA is very fast, it mostly operates on a pseudo-resource, so migrate_to
>> should be finished somewhere near transition abort (which is caused by
>> another resource which starts in the same time and modifies CIB).
> 
> It doesn't matter how fast it is, we will wait for it to finish.
> We don't guarantee the _from will get executed after that though.

Is it an issue that could/should be fixed?

4. nodea: migrate_to nodec
should fail if RA cares about actual resource state because resource is
already moved. And this probably has connection to what I've described
in a message with two hb_reports I sent you privately 01.12.2011.

Best,
Vladislav




More information about the Pacemaker mailing list