[
https://issues.apache.org/jira/browse/KAFKA-10134?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17148190#comment-17148190
]
Guozhang Wang commented on KAFKA-10134:
---------------------------------------
Here's my reasoning: if there are still some owned partitions that are
fetchable even during a rebalance, then within the while-loop even if we exit
the updateAssignmentMetadataIfNeeded we can still try to fetch from then and
then break out of the while loop. However, if we do not have fetchable
partitions then we may be blocked in the while-loop for a long time until the
poll timeout expires. During this period of time we would then see high CPU
usage indeed. I think [~neowu0]'s idea is better: we should also use long poll
if there are fetchable partitions but not data fetched.
What's puzzling me however, is that even in that case we should still be able
to exit the busy while-loop eventually since the `timer.update` should still be
called in pollForFetches within that while loop and hence advance the timer,
why with cooperative rebalance we would block inside the while loop forever is
not clear to me.
> High CPU issue during rebalance in Kafka consumer after upgrading to 2.5
> ------------------------------------------------------------------------
>
> Key: KAFKA-10134
> URL: https://issues.apache.org/jira/browse/KAFKA-10134
> Project: Kafka
> Issue Type: Bug
> Components: clients
> Affects Versions: 2.5.0
> Reporter: Sean Guo
> Assignee: Guozhang Wang
> Priority: Blocker
> Fix For: 2.6.0, 2.5.1
>
>
> We want to utilize the new rebalance protocol to mitigate the stop-the-world
> effect during the rebalance as our tasks are long running task.
> But after the upgrade when we try to kill an instance to let rebalance happen
> when there is some load(some are long running tasks >30S) there, the CPU will
> go sky-high. It reads ~700% in our metrics so there should be several threads
> are in a tight loop. We have several consumer threads consuming from
> different partitions during the rebalance. This is reproducible in both the
> new CooperativeStickyAssignor and old eager rebalance rebalance protocol. The
> difference is that with old eager rebalance rebalance protocol used the high
> CPU usage will dropped after the rebalance done. But when using cooperative
> one, it seems the consumers threads are stuck on something and couldn't
> finish the rebalance so the high CPU usage won't drop until we stopped our
> load. Also a small load without long running task also won't cause continuous
> high CPU usage as the rebalance can finish in that case.
>
> "executor.kafka-consumer-executor-4" #124 daemon prio=5 os_prio=0
> cpu=76853.07ms elapsed=841.16s tid=0x00007fe11f044000 nid=0x1f4 runnable
> [0x00007fe119aab000]"executor.kafka-consumer-executor-4" #124 daemon prio=5
> os_prio=0 cpu=76853.07ms elapsed=841.16s tid=0x00007fe11f044000 nid=0x1f4
> runnable [0x00007fe119aab000] java.lang.Thread.State: RUNNABLE at
> org.apache.kafka.clients.consumer.internals.ConsumerCoordinator.poll(ConsumerCoordinator.java:467)
> at
> org.apache.kafka.clients.consumer.KafkaConsumer.updateAssignmentMetadataIfNeeded(KafkaConsumer.java:1275)
> at
> org.apache.kafka.clients.consumer.KafkaConsumer.poll(KafkaConsumer.java:1241)
> at
> org.apache.kafka.clients.consumer.KafkaConsumer.poll(KafkaConsumer.java:1216)
> at
>
> By debugging into the code we found it looks like the clients are in a loop
> on finding the coordinator.
> I also tried the old rebalance protocol for the new version the issue still
> exists but the CPU will be back to normal when the rebalance is done.
> Also tried the same on the 2.4.1 which seems don't have this issue. So it
> seems related something changed between 2.4.1 and 2.5.0.
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
