[
https://issues.apache.org/jira/browse/HDFS-6617?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14050577#comment-14050577
]
Chris Nauroth commented on HDFS-6617:
-------------------------------------
Hi, [~xieliang007]. Thanks for contributing this. I'm curious if putting the
below code snippet into {{TestDFSZKFailoverController}} would fix the problem
for you. This should avoid the disk latency for edit logging during test runs,
so I'm curious if this is another way to fix the problem in your environment.
This also might make the test run faster. If not, then your current patch is
fine too.
{code}
static {
EditLogFileOutputStream.setShouldSkipFsyncForTesting(true);
}
{code}
> Flake TestDFSZKFailoverController.testManualFailoverWithDFSHAAdmin due to a
> long edit log sync op
> -------------------------------------------------------------------------------------------------
>
> Key: HDFS-6617
> URL: https://issues.apache.org/jira/browse/HDFS-6617
> Project: Hadoop HDFS
> Issue Type: Test
> Components: auto-failover, test
> Affects Versions: 3.0.0
> Reporter: Liang Xie
> Assignee: Liang Xie
> Priority: Minor
> Attachments: HDFS-6617.txt
>
>
> Just Hit a false alarm testing while working at HDFS-6614, see
> https://builds.apache.org/job/PreCommit-HDFS-Build/7259//testReport/org.apache.hadoop.hdfs.server.namenode.ha/TestDFSZKFailoverController/testManualFailoverWithDFSHAAdmin/
> After a looking at the log, shows the failure came from a timeout at
> ZKFailoverController.doCedeActive():
> localTarget.getProxy(conf, timeout).transitionToStandby(createReqInfo());
> While stopping active service, see FSNamesystem.stopActiveServices():
> void stopActiveServices() {
> LOG.info("Stopping services started for active state");
> ....
> this corelates with the log:
> "2014-07-01 08:12:50,615 INFO namenode.FSNamesystem
> (FSNamesystem.java:stopActiveServices(1167)) - Stopping services started for
> active state"
> then stopActiveServices will call editLog.close(), which goes to
> endCurrentLogSegment(), see log:
> 2014-07-01 08:12:50,616 INFO namenode.FSEditLog
> (FSEditLog.java:endCurrentLogSegment(1216)) - Ending log segment 1
> but this operation did not finish in 5 seconds, then triggered the timeout:
> 2014-07-01 08:12:55,624 WARN ha.ZKFailoverController
> (ZKFailoverController.java:doCedeActive(577)) - Unable to transition local
> node to standby: Call From asf001.sp2.ygridcore.net/67.195.138.31 to
> localhost:10021 failed on socket timeout exception:
> java.net.SocketTimeoutException: 5000 millis timeout while waiting for
> channel to be ready for read. ch : java.nio.channels.SocketChannel[connected
> local=/127.0.0.1:53965 remote=localhost/127.0.0.1:10021]; For more details
> see: http://wiki.apache.org/hadoop/SocketTimeout
> the logEdit/logSync finally done followed with printStatistics(true):
> 2014-07-01 08:13:05,243 INFO namenode.FSEditLog
> (FSEditLog.java:printStatistics(675)) - Number of transactions: 2 Total time
> for transactions(ms): 0 Number of transactions batched in Syncs: 0 Number of
> syncs: 3 SyncTimes(ms): 14667 74 105
> so obviously, this long sync contributed the timeout, maybe the QA box is
> very slow at that moment, so one possible fix here is setting the default
> fence timeout to a bigger one.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
