AW: [LinuxFailSafe] startup of resource groups with one node down
Martin Bene
martin.bene@icomedias.com
Tue, 7 May 2002 21:43:38 +0200
> this is not true. membership can form even if the tiebreaker=20
> node is not available. In a 2 node cluster with network partition,=20
> if the tiebreaker node is not able to reset the non-tiebreaker=20
> node and non-tiebreaker node can successfully reset
> the tiebreaker node, a membership of one node (non-tiebreaker node)
> can be formed.
The above is what I've observed when I start out with the cluster with =
ha services running on both nodes and then partition the network or =
power one of them down.=20
On startup however, it's not what I'm observing. if the tiebreaker node =
is not available (switched off), the available node neither tries to =
reset the tiebreaker, nor does it form a memership, at least not in the =
time I've given it to try. (~ 30 minutes). cmsd exits with "Cmsd is out =
of membership, will restart after notifying clients" after trying to =
reach the offline node for 30 seconds, and gets restarted by cmond. =
repeat indefinitely.
>> That said, I still don't understand where the first case=20
>> differs from starting up with both nodes present and then=20
>> loosing one :-)
>=20
> The key difference is in the second case, you are making an=20
> assumption that there are no HA resources running on the node=20
> (which is hopefully down).
The differnce in the second case is clear, what still makes me wonder is =
that the cluster ends up in different states when I
1) start ha services with two nodes, switch one off after sucessfull =
startup
2) start ha services for whole cluster (not just for available node) =
with just one node available.
Both scenarios end up with one node in UP and one node (after being =
reset) in DOWN status; cluster in "online with error" status. I'd have =
assumed the cluster state of these two cases to be functionaly =
equivalent now, but it isn't: in case 1) resource groups can be brought =
on/offline without problems, in case 2) they can't. Seems rather strange =
to me.
Bye, Martin