Cluster Lock is required

[P Arumugavel]

Hi experts,

For a three node SG cluster, is cluster lock required?

[rariasn]

Hi:

Lock Requirements

A one-node cluster does not require a cluster lock.

A two-node cluster

requires a cluster lock.

In clusters larger than three nodes, a cluster lock is strongly recommended.

If you have a cluster with more than four
nodes, use a quorum server; a cluster lock disk is not allowed for clusters of that size.

rgs,

[Jayakrishnan G Naik]

Hi Vel-UX

Cluster lock is recommended mainly for 2 node clusters and I have heard, for 3 node clusters as well. For 3 or more nodes quorum server will be a better solution.

The cluster lock avoids the split brain syndrom, and in a two node cluster, the quorum is lost when a node fails so the to maintain quorum(more than 50%) another vote is held by the quorum disk.
when we discuss about 3 node clusters I have seen people turning to lock disk rather than a quorum server, and the reason given is that, in case of a node failure -say hardware failure or a planned down time on one node, the cluster reduces to a two nodes . The chance of split brain or lose of quorum due to another failure in running nodes remains and that can be avoided if 3 node cluster have lock disk.

Thanks & regards
Jayakrishnan G Naik

[Stephen Doud]

It took me longer to write this response than to find the answer for you.

Use a browser to navigate to http://docs.hp.com/en/ha.html -> HP Serviceguard -> User Guide -> "Managing Serviceguard Eighteenth Edition, September 2010" manual (at http://bizsupport2.austin.hp.com/bc/docs/support/SupportManual/c02437444/c02437444.pdf).

Then search the PDF document with keyword string "cluster lock disk".
The answer is on page 63

Happy Fishing :)

[Emil Velez]

you can get away without a cluster lock for a 3 node cluster if and only if when you loose a node you will not loose another node.

if you go to a 2 node cluster and loose the next node both nodes will panic without quorum.

IF you have a 3 node cluster and if you loose a node and you know that you will fix the node and return it to service before loosing another node then you do not need a cluster lock.

IF you have a LVM volume group that is accessable by all nodes and that a package uses why not use a cluster lock disk. This way no additional disks or systems are necessary. Nothing wrong with cluster lock. If it fails fix it!!!!

[P Arumugavel]

Hi Thanks...

For an example, if I use a three node cluster and the package is running on Node A.
Node B and Node C are adoptive node respectively. Lock disk is made in one of shared VG. Now let me understand that what are all situations that would make the use of Lock disk?

[Torsten.]

All situation where a "split brain" can occur (e.g. if 1 node is down, you have a 2 node cluster; if something bad happens now - lost communication for example - you need the lock).

[P Arumugavel]

Hi Torsten,

OK. In 2 node cluster, if Communication fails, split brain can occur (50-50% available nodes). Here Lock disk will be acquired by one of the node. But in a three node cluster, what is the cahance there split brain can occur? If my first node fails then package will fail over to the next adoptive node...

[melvyn burnard]

and what if you have one of the 3 nodes "out of service" as far as cluster activities are concerned, perhaps fo patching?
If you then get a failure on one of the two remaining nodes, then ou have a 50-50% split again.
Recommendation is to have some form of arbitration, no matter what size cluster you have

[Torsten.]

This exactly what I was talking about. 2 nodes active, 1 intensionally down for whatever reason.



And IMHO in rare cases you can even have a situation of 3 splitted brains.

[P Arumugavel]

Thanks to all.

With three or four nodes, a cluster lock is recommended but not required. For three nodes, there would never be a situation where the nodes are divided 50/50. However, if a node is pulled out of the cluster for maintenance purposes, then the cluster lock would be used should one of the remaining two nodes in the cluster fail....