[jira] [Commented] (KAFKA-6491) Large uncompacted __consumer_offsets files should not make broker restarts slow

["Ewen Cheslack-Postava (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/KAFKA-6491?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16375304#comment-16375304 ] 

Ewen Cheslack-Postava commented on KAFKA-6491:
----------------------------------------------

There might be performance improvements, but you can't really avoid having to scan the entire log to get into a functional state just by the nature of how these topics work and that's why you'd encounter this. The default was changed in 0.9.0.1 over two years ago. I'd be inclined to close this as WONTFIX.

Are you manually overriding the value of `log.cleaner.enable` in your broker configs? In 0.10.2.1 the default was definitely to have it enabled, so you'd only be getting this behavior if it was explicitly overridden in the broker config. What distribution are you using? It is possible they have a bad set of default configs in their broker configuration.

> Large uncompacted __consumer_offsets files should not make broker restarts slow
> -------------------------------------------------------------------------------
>
>                 Key: KAFKA-6491
>                 URL: https://issues.apache.org/jira/browse/KAFKA-6491
>             Project: Kafka
>          Issue Type: Bug
>          Components: offset manager
>    Affects Versions: 0.10.2.1
>            Reporter: Mike Mintz
>            Priority: Major
>
> Before discovering the {{log.cleaner.enable}} option, we had several Kafka brokers running 0.10.2.1 with ~500GB of __consumer_offsets files. When restarting the kafka process on these brokers, clients were able to successfully produce and consume messages, but clients failed to commit offsets to __consumer_offsets partitions which the restarted broker was leader for. Running jstack indicated the broker was spending its time in GroupMetadataManager.loadGroupsAndOffsets (example stack trace below).
> Ideally Kafka brokers would be able to start up much faster in the presence of large __consumer_offsets files.
> {noformat}
> "group-metadata-manager-0" #57 daemon prio=5 os_prio=0 tid=0x00007ffb54ec3800 nid=0xb9d runnable [0x00007ffa81139000]
>    java.lang.Thread.State: RUNNABLE
>   at sun.nio.ch.FileDispatcherImpl.pread0(Native Method)
>   at sun.nio.ch.FileDispatcherImpl.pread(FileDispatcherImpl.java:52)
>   at sun.nio.ch.IOUtil.readIntoNativeBuffer(IOUtil.java:220)
>   at sun.nio.ch.IOUtil.read(IOUtil.java:197)
>   at sun.nio.ch.FileChannelImpl.readInternal(FileChannelImpl.java:741)
>   at sun.nio.ch.FileChannelImpl.read(FileChannelImpl.java:727)
>   at org.apache.kafka.common.utils.Utils.readFully(Utils.java:854)
>   at org.apache.kafka.common.record.FileRecords.readInto(FileRecords.java:121)
>   at kafka.coordinator.GroupMetadataManager.loadGroupsAndOffsets(GroupMetadataManager.scala:427)
>   at kafka.coordinator.GroupMetadataManager.kafka$coordinator$GroupMetadataManager$$doLoadGroupsAndOffsets$1(GroupMetadataManager.scala:392)
>   at kafka.coordinator.GroupMetadataManager$$anonfun$loadGroupsForPartition$1.apply$mcV$sp(GroupMetadataManager.scala:403)
>   at kafka.utils.KafkaScheduler$$anonfun$1.apply$mcV$sp(KafkaScheduler.scala:110)
>   at kafka.utils.CoreUtils$$anon$1.run(CoreUtils.scala:57)
>   at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:511)
> {noformat}



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)