[
https://issues.apache.org/jira/browse/KAFKA-702?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13555219#comment-13555219
]
Sriram Subramanian commented on KAFKA-702:
------------------------------------------
I would like to add my thoughts to this.
1. Load shedding arbitrary clients will bound the memory but would essentially
cause the system to fail most of the requests and not recover from it till the
load goes down. We have quite a few inter-dependencies between requests
(producer depends on replica requests, replica depends on produce requests and
consumer requests depend on produce requests) and dropping requests would
essentially cause the requests depending on it to stay longer in the purgatory
and fail.
2. Having client quotas may not work because we do not have one faulty client.
Each client can at most have only one request.
Few improvements might reduce the failure scenarios
1. Currently replica request wait on a hard limit (min bytes). Instead they
could be made to return earlier to free the purgatory and accept more requests
during high load scenarios.
2. Direct consumers to read from other replicas in the isr that have lesser
load. This is going to be harder.
> Deadlock between request handler/processor threads
> --------------------------------------------------
>
> Key: KAFKA-702
> URL: https://issues.apache.org/jira/browse/KAFKA-702
> Project: Kafka
> Issue Type: Bug
> Components: network
> Affects Versions: 0.8
> Reporter: Joel Koshy
> Assignee: Jay Kreps
> Priority: Blocker
> Labels: bugs
> Fix For: 0.8
>
> Attachments: KAFKA-702-v1.patch
>
>
> We have seen this a couple of times in the past few days in a test cluster.
> The request handler and processor threads deadlock on the request/response
> queues bringing the server to a halt
> "kafka-processor-10251-7" prio=10 tid=0x00007f4a0c3c9800 nid=0x4c39 waiting
> on condition [0x00007f46f698e000]
> java.lang.Thread.State: WAITING (parking)
> at sun.misc.Unsafe.park(Native Method)
> - parking to wait for <0x00007f48c9dd2698> (a
> java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject)
> at java.util.concurrent.locks.LockSupport.park(LockSupport.java:158)
> at
> java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.await(AbstractQueuedSynchronizer.java:1987)
> at
> java.util.concurrent.ArrayBlockingQueue.put(ArrayBlockingQueue.java:252)
> at kafka.network.RequestChannel.sendRequest(RequestChannel.scala:107)
> at kafka.network.Processor.read(SocketServer.scala:321)
> at kafka.network.Processor.run(SocketServer.scala:231)
> at java.lang.Thread.run(Thread.java:619)
> "kafka-request-handler-7" daemon prio=10 tid=0x00007f4a0c57f000 nid=0x4c47
> waiting on condition [0x00007f46f5b80000]
> java.lang.Thread.State: WAITING (parking)
> at sun.misc.Unsafe.park(Native Method)
> - parking to wait for <0x00007f48c9dd6348> (a
> java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject)
> at java.util.concurrent.locks.LockSupport.park(LockSupport.java:158)
> at
> java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.await(AbstractQueuedSynchronizer.java:1987)
> at
> java.util.concurrent.ArrayBlockingQueue.put(ArrayBlockingQueue.java:252)
> at kafka.network.RequestChannel.sendResponse(RequestChannel.scala:112)
> at kafka.server.KafkaApis.handleProducerRequest(KafkaApis.scala:198)
> at kafka.server.KafkaApis.handle(KafkaApis.scala:58)
> at kafka.server.KafkaRequestHandler.run(KafkaRequestHandler.scala:41)
> at java.lang.Thread.run(Thread.java:619)
> This is because there is a cycle in the wait-for graph of processor threads
> and request handler threads. If the request handling slows down on a busy
> server, the request queue fills up. All processor threads quickly block on
> adding incoming requests to the request queue. Due to this, those threads do
> not processes responses filling up their response queues. At this moment, the
> request handler threads start blocking on adding responses to the respective
> response queues. This can lead to a deadlock where every thread is holding a
> lock on one queue and asking a lock for the other queue. This brings the
> server to a halt where it accepts connections but every request gets timed
> out.
> One way to resolve this is by breaking the cycle in the wait-for graph of the
> request handler and processor threads. Instead of having the processor
> threads dispatching the responses, we can have one or more dedicated response
> handler threads that dequeue responses from the queue and write those on the
> socket. One downside of this approach is that now access to the selector will
> have to be synchronized.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
