When a broker shuts down due to hardware failure or otherwise, the leader/master replica removes that broker from the list of in sync replicas for that partition. This is important if you have producers that are using producer.num.acks=-1 that requires the produce request to reach all in sync replicas before the producer gets an acknowledgement for that request. One thing to note about hard crash is that, it will take zookeeper upto session.timeout.ms to discover that the broker is dead, effectively delaying the removal of that broker from the list of in sync replicas.
Thanks, Neha On Fri, Nov 30, 2012 at 12:06 PM, S Ahmed <[email protected]> wrote: > Yes, I'm referring to increase the # of replicas (new broker id) > > Thanks, i'll look into that. > > > On Fri, Nov 30, 2012 at 2:59 PM, Joel Koshy <[email protected]> wrote: > >> Node replacement does not require anything special - i.e., as long as it >> has the same broker id, just bringing it back into the cluster would >> suffice, and it should catch up on partitions that are assigned to it. Not >> sure what you mean by "increase a replica" - or you mean increase the >> number of replicas for a partition. That will require an admin tool - (see >> KAFKA-347). >> >> Thanks, >> >> Joel >> >> >> >> On Fri, Nov 30, 2012 at 11:22 AM, S Ahmed <[email protected]> wrote: >> >> > Hello, >> > >> > I am watching the video of your meetup, where Jun is going over the new >> .8 >> > replica feature. >> > >> > Since the replica's for a master broker are in zookeeper, when there is a >> > hardware crash it will somehow (?) notice it is down and then ignore that >> > node. When it comes back up, it will first get up to speed and then >> become >> > an active replica. >> > >> > If we replace the node, or increase a replica, do we have to restart all >> > brockers or simply update zooker? >> > >>
