Hi Jason, A rolling bounce will create an imbalance in the number of leaders distribution between the brokers and this is not ideal. We do plan to have the preferred leader election tool integrated into kafka that periodically balances the leader count across the brokers in the cluster. For now you are left with the tool and you should be able to script it to be invoked whenever you need it.
On 6/23/13 2:10 AM, "Jason Rosenberg" <j...@squareup.com> wrote: >So, I'm running into the case where after issuing a rolling restart, with >controlled shutdown enabled, the last server restarted ends up without any >partitions that it's the leader of. This is more pronounced of course if >I >have only 2 servers in the cluster (during testing). I presume it's kind >of a problem to have imbalance in the number of partitions each server is >leader of? (Or maybe not such a big deal, since the servers still need to >forward messages between themselves anyway?). > >So, I was looking at the 'preferred leader election' tool here: > >https://cwiki.apache.org/confluence/display/KAFKA/Replication+tools > >Is there a way to programmatically invoke this. I was thinking it make >sense to do periodically (e.g. once an hour), and also after a restart. > But I'd like it to happen from within my app (which contains kafka), >rather than invoking an external tool. > >Thanks, > >Jason
