[Linux-cluster] Adding a node back to cluster failing

[Jan Pokorný]

On 07/11/13 13:35 +0800, Zama Ques wrote:
>> A. The host "%s" is already a member of cluster "%s"
>> 
>> or
>> 
>> B. %s is already a member of a cluster named "%s"
>> 
>> or some other message (interpolate "%s" above with your values)?
>> 
>> 
>>> But the node name is not there in cluster.conf file.
>> 
>> you are talking about the original surviving node now, right?
>> 
>> It seems more likely to me that B. from above is right, so then,
>> please make sure there is no /etc/cluster/cluster.conf on the node you
>> are trying to add (may be a leftover from the original setup of this
>> recovered node).
> 
> Thanks Jan, you assumed correctly.  
> 
> Deleting /etc/cluster/cluster.conf actually resolved the issue.
> Node is successfully added back to cluster.

Great to hear that :)


The problem there was that we distinguish two situations in luci:

a. node is associated with the cluster (as per clusternode entry
   in cluster.conf), however not an active member of the cluster,
   (due to cman service not running there as e.g., when the node
   has been a subject of "leave cluster" action in luci)

   - this node is explicitly listed amongst the cluster nodes
     with "not a cluster member" status and from here, it
     can be selected for "join cluster" action, which will
     start cman + rgmanager on that node again, leading
     to a cluster membership (if nothing goes wrong)

   - in this case the node is expected (enforced) to carry
     cluster.conf (which is also being updated throughout the changes
     in cluster configuration as long as ricci is run there
     and cluster.conf is not deleted manually in between)

b. node is not a priori associated with the cluster (it's not mentioned
   in the cluster.conf across the cluster)

   - this node is apparently not (no hint for that) listed amongst
     the cluster nodes and can be added via "add" from that view

   - in this case the node is expected *not* to contain cluster.conf
     upon being added to an existing cluster; simply because by
     mistake, an attempt to add a node being already a member of
     a different cluster could be made, possibly leading to some
     inconsistencies in the luci view of the cluster
     NB: we might tighten this constraint and allow the cluster
         configuration to be already present on the node being
	 added provided that the cluster name matches the destination
	 (which would help in this very case, IMHO)


Apparently this was case of b. and the above description explains
why removing cluster.conf from the node to be added was of help.

It's up to consideration if to provide a solution to such class
of cases as suggested, feel free to comment on:

    https://bugzilla.redhat.com/show_bug.cgi?id=1028092

note: I mentioned some other little discrepancies I've discovered
      there

-- 
Jan