PVE 5 : Three Node HA Cluster (Down to one Node)

[Rais Ahmed]

Hi all,
I have setup 3 nodes HA cluster using PVE 5, and trying to achieve that if 2 nodes down, quorum do not break. and all the vms moved to last standing node. i believe it is possible but my understanding is not clear as i am new to PVE. i have found there are few feature available but not tested it personally. how this can be achieve, need input from community regarding below features to achieve above scenario.

1. two_node
2. quorum votes:2 (for every node)
3. auto_tie_breaker
4. last_man_standing & last_man_standing_window
5. auto_tie_breaker (If you have a cluster with an odd number of nodes and auto_tie_breaker enabled then it is compulsory to also specify wait_for_all) unable to understand what will be the behavior of this option.

check below PDF for refference
http://people.redhat.com/ccaulfie/docs/Votequorum_Intro.pdf

if anyone implemented the above scenario or have info please share
Thanks.

 

[dcsapak]

the last_man_standing and auto_tie_breaker config can work theoretically, but with both you are risking a split brain scenario
true quorum can only be achieved when having more than 50% of the votes
so everything apart from the default quorum configuration is not recommended

 

[Rais Ahmed]

After added auto_tie_breaker, last_man_standing & last_man_standing_window in /etc/pve/corosync.conf. i have corrupted my cluster environment. node1 shows cannot initialize cmap service error. tried everything to resolve it but failed, node2 is active after pvecm expected 1 & holds all vms OK, meanwhile node3 shows quorum_initialize failed: 2 . tried to restart every services. after wasting few hours but failed.i have reinstalled all servers again & now creating HA environment again.
so i believe do not play with corosync.conf??

 

[Xahid]

Yes, you are not suppose to play with corosync.conf, if you made a cluster using pve-command tools.

 

[Rais Ahmed]

Thankyou

 

[RolandK]

i can confirm that auto_tie_breaker is dangerous. it did not work for me, too and the cluster service failed with that.