[Pacemaker] Resources are not restarted on definition change after f59d7460bdde (devel)

Andrew Beekhof andrew at beekhof.net
Thu Jul 28 20:24:34 EDT 2011

On Wed, Jul 27, 2011 at 6:12 PM, Florian Haas <florian.haas at linbit.com> wrote:
> On 2011-07-27 03:46, Andrew Beekhof wrote:
>> On Fri, Jul 1, 2011 at 4:59 PM, Andrew Beekhof <andrew at beekhof.net> wrote:
>>> Hmm.  Interesting. I will investigate.
>> This is an unfortunate side-effect of my history compression patch.
>> Since we only store the last successful and last failed operation, we
>> don't have the md5 of the start operation around to check when a
>> resource's definition is changed.
>> Solutions appear to be either:
>> a) give up the space savings and revert the history compression patch
>> b) always restart a resource if a non-matching md5 is detected - even
>> if the operation was a recurring monitor
>> I'd favor b) along with dropping the per-operation parameters.
>> The only valid use-case I've heard for those is setting OCF_LEVEL or
>> depth or whatever it was called - and I think we're in basic agreement
>> that we need a better solution for that anyway.
> We are, and you know my opinion that OCF_CHECK_LEVEL is hideous
> (although lmb, for one, seems to disagree).

His disagreement (IIUC) is with changing it in the RA API - not how
its exposed to users in the config.

> But dropping it now does
> clearly count as a regression and I'd really hate to see that happen unless
> a) there is a replacement method for tuning the thoroughness of checking
> the resource state during monitor, _and_

That would be the part about promoting it to the 'op' level albeit
with a different name (it would still appear as OCF_CHECK_LEVEL in the
RA's environment).

> b) there is an automated or semi-automated ("cibadmin --upgrade"?) means
> of transitioning off OCF_CHECK_LEVEL and replacing it with its successor
> feature.

Thats what our xslt's do.

In any case, the dropping of per-op params is not a pre-req for the
"restart the resource on any op change" part.
That OCF_CHECK_LEVEL is the only param that is normally set this way
just makes option b) palatable.

More information about the Pacemaker mailing list