[Pacemaker] Ordering constraints between two clone and stateful resources

[King, Christopher]

Hi all,

 

We have a 2-node cluster running on the SLES 11 HAE SP1, plus the latest
supported patches.  (So Pacemaker version 1.1.5, etc).  In the cluster,
we have a clone resource and a stateful resource that have an mandatory
ordering constraint between the two, such that the stateful resource
will start after the clone.

 

>From the output of "crm configure show".  We reproduced the problem with
the stock "Dummy" and "Stateful" RAs.

 

primitive testStateful ocf:heartbeat:Stateful \

        op start interval="0" timeout="1800s" \

        op stop interval="0" timeout="45s" \

        op monitor interval="10s"

 

primitive testDummy ocf:heartbeat:Dummy \

        op monitor interval="20" timeout="10"

 

ms testStateful-ms testStateful \

        meta target-role="Started" master-max="1" master-node-max="1"
clone-max="2" clone-node-max="1" notify="true" ordered="false"
globally-unique="false" is-managed="true"

 

clone testDummy-clone testDummy \

        meta target-role="Started"

 

order testDummy-testStateful-order inf: testDummy-clone
testStateful-ms:start

 

I then perform some experiments:

1)      If the testDummy instance on the same node as the
testStateful:slave instance fails, both the testStateful:slave and
testStateful:master instances are stopped, and restarted when the
testDummy instance is restarted.

2)      If the testDummy instance on the same node as the
testStateful:master instance fails, both the testStateful:slave and
testStateful:master instances are stopped, and restarted when the
testDummy instance is restarted.

The desired behaviour is for the instance of testStateful on the same
node as the failed instance of the testDummy to be stopped, and
restarted when the testDummy instance is restarted.  In other words, we
want to be able to order an instance of a clone/stateful resource on the
instance of a clone/stateful resource on the same node.  For our
purposes, it is actually very harmful if the dependent resource depends
on BOTH instances of the clone, as my experiments show that it does.

 

So questions:

1)      Is it possible to express with a valid configuration a "nodal"
ordering constaint between instances of different clone resources?
I.e., an ordering constraint in which the instance of the stateful
resource depends on the instance of the clone on the same node, but not
the instance of the clone on the other node.  I've read the on-line
documentation which highly discourages referencing a clone's child in an
ordering constraint (and by "clone's child", I am assuming you mean the
clone's primitive, not a specific instance of the primitive) so I don't
want to do that.  But, if I manually edit the cluster configuration to
reference the primitives in the ordering constraint between testDummy
and testStateful, we seem to get the desired behaviour, but the cluster
configuration is technically invalid.  ("crm configure verify" returns
errors.)  So we don't want to do that, but is there a valid way of
configuring for the desired behaviour?

2)      Again, looking at the online documentation, I found a reference
to a meta attribute of clones called "interleave".  The description is a
little open to interpretation; does it refer to changing the behaviour
of ordering constraints between the instances of a clone, or between two
different cloned resources?  We think the former interpretation is most
likely, but if I add a "interleave="true" attribute to the
testDummy-clone and testStateful-ms configurations above, I get the
desired behaviour.  Is that what you would expect, or is this a
side-effect that may not deterministically occur?

 

Thanks very much.

Chris

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.clusterlabs.org/pipermail/pacemaker/attachments/20111102/ba16e7de/attachment-0002.html>