Hi there sorry for crossposting,

While implementing a fix for Bug #980[1] i came to conclusion, that
Sequoia misses one central concept of high-availability.
The split-brain scenario[2][3]. If the connection between two or more
nodes are lost, the nodes have to stay online working for themselfs.
In  a split-brain scenario the databases on each node are different,
each node behaves like a 1-node cluster serving select and insert
statements without
loadbalancing or distribution to the other node.

The split-brain scenario even carries on, when both controllers see each
other again.

And thats exactly the sequoia problem, the underlying hedera stuff adds
new members to the group automatically without asking the upper sequoia
logic.
I tried to fix this by changing GroupMembershipListener.joinMember()
signature to return a boolean if join is allowed and changed order of
calls (first call upper listeners to get this boolean and add member to
group in hedera after this successful call) But i think thats not the
right way to act.
I would prefer to enhance hedera.... any ideas... any discussion about
split-brain...

I hope nobody will argue with "split-brain" scenario is not a scenario
for sequoia. Network has to be stable ALL THE TIME. That would be an
illusion.

thanx in advance and happy discussion

Stefan

[1] https://forge.continuent.org/jira/browse/SEQUOIA-980
[2] http://de.wikipedia.org/wiki/Split_Brain#Computercluster
[3] http://www.linux-ha.org/SplitBrain

Attachment: smime.p7s
Description: S/MIME cryptographic signature

_______________________________________________
Hedera mailing list
[email protected]
https://forge.continuent.org/mailman/listinfo/hedera

Reply via email to