There's a pretty obscure corner case in the Pacemaker Remote node
fencing code that takes a lot of fairly special setup to occur, but
still seems like a bug worth fixing.
To make a very long story short, if you have a cluster with three or
more nodes and a remote node configured, and the connection resource for
that remote node is running on a node that leaves the cluster, it is
still possible for that node to fence the remote node. This can happen
even if the node is in a partition without quorum, and even if the
remote node's connection resource can be started on the partition with
quorum.
It is highly unlikely that you'll hit this bug. I suspect it has been
present for a very long time, perhaps as far back as when Pacemaker
Remote nodes were introduced. It's also possible that somewhere,
someone is relying on this behavior.
Therefore, we've introduced a fix for this behavior but we've made it
opt-in. If you are seeing a problem like what I am describing, you can
enable the fix with the following:
# pcs property set fence-remote-without-quorum=false
If that property is unset (which is the default for everyone who doesn't
take any action), or is set to true, you will continue to get the
previous fencing behavior.
If all of this is somewhat confusing and you're not sure what to do, my
advice is:
If you're not seeing any odd fencing problems with remote nodes and
partitioned clusters, don't do anything. If you are, try the new
property out.
- Chris
_______________________________________________
Manage your subscription:
https://lists.clusterlabs.org/mailman/listinfo/users
ClusterLabs home: https://www.clusterlabs.org/
