Hi JD, You can enable unclean leader election (disabled by default) to use ex-ISR followers as fallbacks for new leaders - but be mindful of data loss.
Thanks, Alexandre Le mar. 10 nov. 2020 à 06:33, JD Zheng <jdh...@gmail.com> a écrit : > > Hi, everyone, > > Not sure if this is the right place to ask this question. I sent it to the > user mailing list and didn't get any response. Where should I ask such > questions? > We configure the producer write Ack=ALL and broker min.insync.replica = 2. > The replication factor is 3. In this case, if the leader is struggling, and > the two other replicas can not catchup with the leader and are both kicked > out of the isr set, it is still safe to fail over to the most updated > follower. Because the updates which are only available on the failed > leader won't have been acked back to the client as success, no data loss in > terms of client updates in this case. But I didn't find any > information/tool to failover to the newer follower in case the only isr > replica is the failed leader. How can we recover the partition when the > only replica in the isr failed and two other healthy followers can not > catch up? Is there any way to manually bring the two replicas back to isr > without actually catchup with the down leader? Thanks, > > -JD > > ---------- Forwarded message --------- > From: JD Zheng <jdh...@gmail.com> > Date: Sun, Nov 8, 2020 at 7:41 PM > Subject: Recover partitions from a failed broker leader which is the only > replica in isr. > To: <us...@kafka.apache.org> > > > Hi, > > We have a case where one of our kafka brokers's data drive failed which > kicks all the followers of the partitions with this broker as the leader > out of the isr before the broker is completed down. In this case, all the > partitions (this broker was the leader) become offline. What's the best way > to recover these partitions to re-elect a new leader among the non-isr > replicas? Is there any way to elect the replica with the highest ELO? > > Thanks, > > -JD
