Are you hard killing the brokers? And is this issue reproducible?
On Sat, Dec 21, 2013 at 11:39 AM, Drew Goya <d...@gradientx.com> wrote: > Hey guys, another small issue to report for 0.8.1. After a couple days 3 > of my brokers had fallen off the ISR list for a 2-3 of their partitions. > > I didn't see anything unusual in the log and I just restarted one. It came > up fine but as it loaded its logs I these messages showed up: > > [2013-12-21 19:25:19,968] WARN [ReplicaFetcherThread-0-2], Replica 1 for > partition [Events2,58] reset its fetch offset to current leader 2's start > offset 1042738519 (kafka.server.ReplicaFetcherThread) > [2013-12-21 19:25:19,969] WARN [ReplicaFetcherThread-0-14], Replica 1 for > partition [Events2,28] reset its fetch offset to current leader 14's start > offset 1043415514 (kafka.server.ReplicaFetcherThread) > [2013-12-21 19:25:20,012] WARN [ReplicaFetcherThread-0-2], Current offset > 1011209589 for partition [Events2,58] out of range; reset offset to > 1042738519 (kafka.server.ReplicaFetcherThread) > [2013-12-21 19:25:20,013] WARN [ReplicaFetcherThread-0-14], Current offset > 1010086751 for partition [Events2,28] out of range; reset offset to > 1043415514 (kafka.server.ReplicaFetcherThread) > [2013-12-21 19:25:20,036] WARN [ReplicaFetcherThread-0-14], Replica 1 for > partition [Events2,71] reset its fetch offset to current leader 14's start > offset 1026871415 (kafka.server.ReplicaFetcherThread) > [2013-12-21 19:25:20,036] WARN [ReplicaFetcherThread-0-2], Replica 1 for > partition [Events2,44] reset its fetch offset to current leader 2's start > offset 1052372907 (kafka.server.ReplicaFetcherThread) > [2013-12-21 19:25:20,036] WARN [ReplicaFetcherThread-0-14], Current offset > 993879706 for partition [Events2,71] out of range; reset offset to > 1026871415 (kafka.server.ReplicaFetcherThread) > [2013-12-21 19:25:20,036] WARN [ReplicaFetcherThread-0-2], Current offset > 1020715056 for partition [Events2,44] out of range; reset offset to > 1052372907 (kafka.server.ReplicaFetcherThread) > > Judging by the network traffic and disk usage changes after the reboot > (both jumped up) a couple of the partition replicas had fallen behind and > are now catching up. > > > On Thu, Dec 19, 2013 at 4:37 PM, Neha Narkhede <neha.narkh...@gmail.com > >wrote: > > > Hi Drew, > > > > That problem will be fixed by > > https://issues.apache.org/jira/browse/KAFKA-1074. I think we are close > to > > checking that in to trunk. > > > > Thanks, > > Neha > > > > > > On Wed, Dec 18, 2013 at 9:02 AM, Drew Goya <d...@gradientx.com> wrote: > > > > > Thanks Neha, I rolled upgrades and completed a rebalance! > > > > > > I ran into a few small issues I figured I would share. > > > > > > On a few Brokers, there were some log directories left over from some > > > failed rebalances which prevented the 0.8.1 brokers from starting once > I > > > completed the upgrade. These directories contained an index file and a > > > zero size log file, once I cleaned those out the brokers were able to > > start > > > up fine. If anyone else runs into the same problem, and is running > RHEL, > > > this is the bash script I used to clean them out: > > > > > > du --max-depth=1 -h /data/kafka/logs | grep K | sed s/.*K.// | sudo rm > -r > > > > > > > > > On Tue, Dec 17, 2013 at 10:42 AM, Neha Narkhede < > neha.narkh...@gmail.com > > > >wrote: > > > > > > > There are no compatibility issues. You can roll upgrades through the > > > > cluster one node at a time. > > > > > > > > Thanks > > > > Neha > > > > > > > > > > > > On Tue, Dec 17, 2013 at 9:15 AM, Drew Goya <d...@gradientx.com> > wrote: > > > > > > > > > So I'm going to be going through the process of upgrading a cluster > > > from > > > > > 0.8.0 to the trunk (0.8.1). > > > > > > > > > > I'm going to be expanding this cluster several times and the > problems > > > > with > > > > > reassigning partitions in 0.8.0 mean I have to move to trunk(0.8.1) > > > asap. > > > > > > > > > > Will it be safe to roll upgrades through the cluster one by one? > > > > > > > > > > Also are there any client compatibility issues I need to worry > about? > > > > Am I > > > > > going to need to pause/upgrade all my consumers/producers at once > or > > > can > > > > I > > > > > roll upgrades through the cluster and then upgrade my clients one > by > > > one? > > > > > > > > > > Thanks in advance! > > > > > > > > > > > > > > >
