I am wondering if anyone else has noticed the kind of scenario I have been noticing when a Corosync cluster that gets partitioned is repaired ?
We have a Corosync cluster that consists of three nodes - n1, n2 and n3. n1 has the lowest IP addr. and always appears as the ring leader in the full cluster. When n1 is disconnected from the network by physically pulling out a network cable, n2 and n3 form their own ring and n1 is in a ring all by itself. When the n1 is reconnected by plugging the network cable back, the original cluster is restored and things are back to normal, most of the time. Every now then the newly restored cluster splits again
without external intervention. n2 and n3 form a new ring with n2 as their ring leader and new configutation ID , which is numerically 4 higher than the previous configuartion ID of n1, n2 and n3. Both n2 and n3 are delivered this configuration change, so they notice that n1 is missing.
However n1 is not made aware of this configuration change. It does however receive a configuration whose ID is numerically 8 higher than that of the original full cluster, but the membership list remains unchanged. This configuration is delivered to n2 and n3 too. In other words for n1, it appears as though a new configuration got installed
with no change in membership. n2 and n3 in the meanwhile notice that their ring leader n1 is missing first. When they receive a new configuration that includes n1, they wait for an action to be initiated by n1, except that n1 does not initiate that action as it does not know that n2 and n3 left the configuration.
Thanks,
Sathya
_______________________________________________ discuss mailing list discuss@xxxxxxxxxxxx http://lists.corosync.org/mailman/listinfo/discuss
