We just need a unique identifier for every server. If such an identifier "magically" appears somehow, then I believe our protocols will be equally happy. Now, a mechanism to assign ids would also have to take into consideration the group scheme we have for hierarchical quorums. To assign servers to groups, we currently use the identifiers we assign manually to servers. If we don't have such identifiers, then we need a different way of configuring groups.


On Sep 29, 2009, at 7:01 PM, Patrick Hunt wrote:

Jason Venner wrote:
I do find having to have a custom file in each zk root somewhat awkward, as I like to rsync my configuration files around. I also would prefer not to have to have all of my zk nodes listed in the configuration file by id.

I think I would prefer it if there was a mechanism for each log directory to come up with a cluster unique id, and then the jvm running on that log dir
would advertise that ID.

You are saying to list the server's ip/port in each of the server config files, but not the id's, correct? The server id is actually used as part of the quorum (zab) protocol - servers with lower id's don't attempt to
connect to servers with higher id when forming quorum (to avoid 2
servers negotiating on 2 channels/threads). Perhaps this can be worked
around though. Flavio?

I'll enter a jira for tracking this feature as soon as apache issue
tracking comes back (seems to be down right now).

If there was a control that would allow runtime changing of the required
quorum, then I could dynamically add and remove nodes.



2009/9/29 Ørjan Horpestad <orj...@gmail.com>

Thanks for all of your answers. I can see more clearly why using an IP
as id could be a bad idea in a ZK setup.

I will indeed try out your Zkconf tool, thanks.

Regards, Orjan

Reply via email to