Ken Gaillot napsal(a):
On Mon, 2017-08-14 at 12:33 -0500, Ken Gaillot wrote:
On Wed, 2017-08-02 at 09:59 +0000, 井上 和徳 wrote:
Hi,
In Pacemaker-1.1.17, the attribute updated while starting pacemaker is not
displayed in crm_mon.
In Pacemaker-1.1.16, it is displayed and results are different.
https://github.com/ClusterLabs/pacemaker/commit/fe44f400a3116a158ab331a92a49a4ad8937170d
This commit is the cause, but the following result (3.) is expected behavior?
This turned out to be an odd one. The sequence of events is:
1. When the node leaves the cluster, the DC (correctly) wipes all its
transient attributes from attrd and the CIB.
2. Pacemaker is newly started on the node, and a transient attribute is
set before the node joins the cluster.
3. The node joins the cluster, and its transient attributes (including
the new value) are sync'ed with the rest of the cluster, in both attrd
and the CIB. So far, so good.
4. Because this is the node's first join since its crmd started, its
crmd wipes all of its transient attributes again. The idea is that the
node may have restarted so quickly that the DC hasn't yet done it (step
1 here), so clear them now to avoid any problems with old values.
However, the crmd wipes only the CIB -- not attrd (arguably a bug).
Whoops, clarification: the node may have restarted so quickly that
corosync didn't notice it left, so the DC would never have gotten the
Corosync always notice left of node no matter if left is longer or
within token timeout.
"peer lost" message that triggers wiping its transient attributes.
I suspect the crmd wipes only the CIB in this case because we assumed
attrd would be empty at this point -- missing exactly this case where a
value was set between start-up and first join.
5. With the older pacemaker version, both the joining node and the DC
would request a full write-out of all values from attrd. Because step 4
only wiped the CIB, this ends up restoring the new value. With the newer
pacemaker version, this step is no longer done, so the value winds up
staying in attrd but not in CIB (until the next write-out naturally
occurs).
I don't have a solution yet, but step 4 is clearly the problem (rather
than the new code that skips step 5, which is still a good idea
performance-wise). I'll keep working on it.
[test case]
1. Start pacemaker on two nodes at the same time and update the attribute
during startup.
In this case, the attribute is displayed in crm_mon.
[root@node1 ~]# ssh -f node1 'systemctl start pacemaker ; attrd_updater -n
KEY -U V-1' ; \
ssh -f node3 'systemctl start pacemaker ; attrd_updater -n
KEY -U V-3'
[root@node1 ~]# crm_mon -QA1
Stack: corosync
Current DC: node3 (version 1.1.17-1.el7-b36b869) - partition with quorum
2 nodes configured
0 resources configured
Online: [ node1 node3 ]
No active resources
Node Attributes:
* Node node1:
+ KEY : V-1
* Node node3:
+ KEY : V-3
2. Restart pacemaker on node1, and update the attribute during startup.
[root@node1 ~]# systemctl stop pacemaker
[root@node1 ~]# systemctl start pacemaker ; attrd_updater -n KEY -U V-10
3. The attribute is registered in attrd but it is not registered in CIB,
so the updated attribute is not displayed in crm_mon.
[root@node1 ~]# attrd_updater -Q -n KEY -A
name="KEY" host="node3" value="V-3"
name="KEY" host="node1" value="V-10"
[root@node1 ~]# crm_mon -QA1
Stack: corosync
Current DC: node3 (version 1.1.17-1.el7-b36b869) - partition with quorum
2 nodes configured
0 resources configured
Online: [ node1 node3 ]
No active resources
Node Attributes:
* Node node1:
* Node node3:
+ KEY : V-3
Best Regards
_______________________________________________
Users mailing list: Users@clusterlabs.org
http://lists.clusterlabs.org/mailman/listinfo/users
Project Home: http://www.clusterlabs.org
Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf
Bugs: http://bugs.clusterlabs.org
_______________________________________________
Users mailing list: Users@clusterlabs.org
http://lists.clusterlabs.org/mailman/listinfo/users
Project Home: http://www.clusterlabs.org
Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf
Bugs: http://bugs.clusterlabs.org
