Not sure what could be going on.. What version of the client and the broker are you on?
Can you verify from the state change logs the time it took for leadership to move to the preferred leader? Were there long GCs in your brokers? Can you also look for zookeeper session expirations in your broker logs? It is odd that restarting your consumers appears to have resolved your issues. What config overrides did you use for your consumers? E.g., did you override the max wait time? How many consumers/producers are we talking about here? Thanks, Joel On Thu, Jan 15, 2015 at 11:55:35AM -0800, Allen Wang wrote: > We are using the scala producer. From producer side, we have seen a lot of > error messages in producer during the time of incoming message drop: > > Produce request with correlation id 31616255 failed due to > [trace_annotation,10]: kafka.common.NotLeaderForPartitionException > > And a few (far less than the NotLeaderForPartitionException) of those: > > 2015-01-15 17:57:54,412 WARN KafkaSink-kafka_0_8_logtrace_asg-Sender-2 > DefaultEventHandler - Failed to send producer request with correlation id > 31554484 to broker 20 with > data for partitions [trace_annotation,9] > java.net.SocketTimeoutException > > What's interesting is that the broker cluster recovered from the message > drop only after we restarted consumers. Also during that time, we have > observed is that the garbage collection time for the brokers increased 5 > times. The AllBrokersFetchRequestRateAndTimeMs_9X metric from the consumer > side also increased from a few hundred ms to several seconds. > > What we don't know is whether the garbage collection time increase is the > cause or the effect of the problem. It seems that after the rebalance, some > resources in the brokers was tied up and it was only released after restart > of consumers. > > > On Thu, Jan 15, 2015 at 8:15 AM, Joel Koshy <jjkosh...@gmail.com> wrote: > > > > Is leadership rebalance a safe operation? > > > > Yes - we use it routinely. For any partition, there should only be a > > brief (order of seconds) period of rejected messages as leaders move. > > When that happens the client should refresh metadata and discover the > > new leader. Are you using the Java producer? Do you see any errors in > > the producer logs? > > > > On Wed, Jan 14, 2015 at 06:36:27PM -0800, Allen Wang wrote: > > > Hello, > > > > > > We did a manual leadership rebalance (using > > > PreferredReplicaLeaderElectionCommand) under heavy load and found that > > > there is a significant drop of incoming messages to the broker cluster > > for > > > more than an hour. Looking at broker log, we found a lot of errors like > > > this: > > > > > > 2015-01-15 00:00:03,330 ERROR kafka.utils.Logging$class:103 > > > [kafka-processor-7101-0] [error] Closing socket for /10.213.156.41 > > > because of error > > > java.io.IOException: Connection reset by peer > > > at sun.nio.ch.FileDispatcherImpl.read0(Native Method) > > > at sun.nio.ch.SocketDispatcher.read(SocketDispatcher.java:39) > > > at sun.nio.ch.IOUtil.readIntoNativeBuffer(IOUtil.java:223) > > > at sun.nio.ch.IOUtil.read(IOUtil.java:197) > > > at sun.nio.ch.SocketChannelImpl.read(SocketChannelImpl.java:379) > > > at kafka.utils.Utils$.read(Utils.scala:375) > > > at > > kafka.network.BoundedByteBufferReceive.readFrom(BoundedByteBufferReceive.scala:54) > > > at kafka.network.Processor.read(SocketServer.scala:347) > > > at kafka.network.Processor.run(SocketServer.scala:245) > > > at java.lang.Thread.run(Thread.java:745) > > > > > > > > > Is leadership rebalance a safe operation? > > > > > > Thanks. > > > >