[Pacemaker] notifications for cloned resources

[Andrew Beekhof]

On 14 Aug 2014, at 12:33 am, Steve Feehan <feehans at ncbi.nlm.nih.gov> wrote:

> On Tue, Aug 12, 2014 at 04:56:06PM +1000, Andrew Beekhof wrote:
>> 
>> What is ganeti doing with the information though?
>> Like GFS2, OCFS2 and the dlm, it might be more appropriate for it to get membership information directly from corosync.
> 
> ganeti wants the ganeti-node-role resource to run on all nodes, but it
> only performs any action on the master node.
> 
> It expects to receive notifications when a node is down and then sets
> its internal state as offline.

Is it a problem that several seconds could go by between the node going offline and the notification arriving?
I would usually expect the answer to be yes.

> The only OCF action that it ipmlements is
> 'notify'. When its invoked in this mode it does this:
> 
> notify_action() {
>  is_master || exit 0
>  [[ -f $NORUNFILE ]] && exit 0
>  # TODO: also implement the "start" operation for readding a node
>  [[ $OCF_RESKEY_CRM_meta_notify_operation == "stop" ]] || exit 0
>  [[ $OCF_RESKEY_CRM_meta_notify_type == "post" ]] || exit 0
>  local -r target=$OCF_RESKEY_CRM_meta_notify_stop_uname
>  local -r node=$(gnt-node list --no-headers -o name $target)
>  # TODO: use drain_node when we can
>  offline_node $node
>  exit 0
> }
> 
> ganeti provides a harep utility that will perform actions to heal a
> cluster when a node is marked offline.
> 
> All I need is some way to mark a node offline when it is down/fenced.

1. Many problems are far easier to express than to solve
2. Normal resources are not supposed to care.
Can you imagine co-ordinating the notification of 400+ resources every time a node changed state?

Those that do care (eg. cluster filesystems) usually have a daemon that a) monitors the corosync membership directly and/or b) subscribes to stonithd fencing notifications.
They do this because they can't wait for resource based notification.

What is the usecase for nesting a ganeti cluster inside a pacemaker one?

> This has to be done from the master. Master failover works, so if the
> master is down pacemaker will promote one of the other nodes to master.
> 
> There are two cases:
> 
> 1. A node that is not the master is down. On the master mark the node
>    as offline and harep will do the rest.

Will ganeti on the master (I'm assuming there is a daemon of some kind) not notice that it can't see the offline node?
If so, it could use the fencing API to request it be shot, wait for the answer, mark it as offline and then use harep.

> 
> 2. A node that is the master is down. Pacemaker will start the master
>    on another node (this works), the new master will mark the old master
>    as offline, and then harep will do the rest.
> 
> -- 
> Steve Feehan [Contractor]
> 
> _______________________________________________
> Pacemaker mailing list: Pacemaker at oss.clusterlabs.org
> http://oss.clusterlabs.org/mailman/listinfo/pacemaker
> 
> Project Home: http://www.clusterlabs.org
> Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf
> Bugs: http://bugs.clusterlabs.org

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 841 bytes
Desc: Message signed with OpenPGP using GPGMail
URL: <https://lists.clusterlabs.org/pipermail/pacemaker/attachments/20140814/db16cd0f/attachment-0003.sig>