[
https://issues.apache.org/jira/browse/HDFS-7314?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14211584#comment-14211584
]
Ming Ma commented on HDFS-7314:
-------------------------------
Thanks Colin for the good point. I also noticed that during the analysis; but
assumed that is part of the original design.
1. The issue you described is in trunk. It can happen when LeaseRenewer goes
away, due to SocketTimeoutException or RenewerExpired.
2. In your above steps, LeaseRenewer object is added to Factory.INSTANCE in #3,
not in step #4 and #5. But that doesn't change the issue. What will happen is
when the first thread calls endFileLease, it will get hold of LR2. So LR1 will
keep renewing the lease even after all files have been closed.
It appears we have discovered a bunch of race conditions regardless whether the
original issue is addressed or not. Given that, we can consider fixing the
original issue and open another jira to address these race conditions.
As you mentioned, the issues come from the fact that LeaseRenewer tries to
clean up the object and the thread when they are no longer used. IMHO, that is
not necessary; we can just keep LeaseRenewer objects and their threads around
once they are created, the idea in the original patch. LeaseRenewer objects are
keyed by NN address and ugi. In the normal set up, with HDFS federation you
can have several NN addresses, but # of ugis should be limited. So it isn't
expensive to keep these objects and their threads around.
If the long term fix is to keep the LeaseRenewer object and thread around, we
can start with fix for SocketTimeoutException in this patch and open another
patch to address the RenewerExpired scenario later.
> Aborted DFSClient's impact on long running service like YARN
> ------------------------------------------------------------
>
> Key: HDFS-7314
> URL: https://issues.apache.org/jira/browse/HDFS-7314
> Project: Hadoop HDFS
> Issue Type: Improvement
> Reporter: Ming Ma
> Assignee: Ming Ma
> Attachments: HDFS-7314-2.patch, HDFS-7314-3.patch, HDFS-7314-4.patch,
> HDFS-7314-5.patch, HDFS-7314-6.patch, HDFS-7314-7.patch, HDFS-7314.patch
>
>
> It happened in YARN nodemanger scenario. But it could happen to any long
> running service that use cached instance of DistrbutedFileSystem.
> 1. Active NN is under heavy load. So it became unavailable for 10 minutes;
> any DFSClient request will get ConnectTimeoutException.
> 2. YARN nodemanager use DFSClient for certain write operation such as log
> aggregator or shared cache in YARN-1492. DFSClient used by YARN NM's
> renewLease RPC got ConnectTimeoutException.
> {noformat}
> 2014-10-29 01:36:19,559 WARN org.apache.hadoop.hdfs.LeaseRenewer: Failed to
> renew lease for [DFSClient_NONMAPREDUCE_-550838118_1] for 372 seconds.
> Aborting ...
> {noformat}
> 3. After DFSClient is in Aborted state, YARN NM can't use that cached
> instance of DistributedFileSystem.
> {noformat}
> 2014-10-29 20:26:23,991 INFO
> org.apache.hadoop.yarn.server.nodemanager.containermanager.localizer.ResourceLocalizationService:
> Failed to download rsrc...
> java.io.IOException: Filesystem closed
> at org.apache.hadoop.hdfs.DFSClient.checkOpen(DFSClient.java:727)
> at org.apache.hadoop.hdfs.DFSClient.getFileInfo(DFSClient.java:1780)
> at
> org.apache.hadoop.hdfs.DistributedFileSystem$17.doCall(DistributedFileSystem.java:1124)
> at
> org.apache.hadoop.hdfs.DistributedFileSystem$17.doCall(DistributedFileSystem.java:1120)
> at
> org.apache.hadoop.fs.FileSystemLinkResolver.resolve(FileSystemLinkResolver.java:81)
> at
> org.apache.hadoop.hdfs.DistributedFileSystem.getFileStatus(DistributedFileSystem.java:1120)
> at org.apache.hadoop.yarn.util.FSDownload.copy(FSDownload.java:237)
> at org.apache.hadoop.yarn.util.FSDownload.call(FSDownload.java:340)
> at org.apache.hadoop.yarn.util.FSDownload.call(FSDownload.java:57)
> at java.util.concurrent.FutureTask.run(FutureTask.java:262)
> at
> java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:471)
> at java.util.concurrent.FutureTask.run(FutureTask.java:262)
> at
> java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1145)
> at
> java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:615)
> at java.lang.Thread.run(Thread.java:745)
> {noformat}
> We can make YARN or DFSClient more tolerant to temporary NN unavailability.
> Given the callstack is YARN -> DistributedFileSystem -> DFSClient, this can
> be addressed at different layers.
> * YARN closes the DistributedFileSystem object when it receives some well
> defined exception. Then the next HDFS call will create a new instance of
> DistributedFileSystem. We have to fix all the places in YARN. Plus other HDFS
> applications need to address this as well.
> * DistributedFileSystem detects Aborted DFSClient and create a new instance
> of DFSClient. We will need to fix all the places DistributedFileSystem calls
> DFSClient.
> * After DFSClient gets into Aborted state, it doesn't have to reject all
> requests , instead it can retry. If NN is available again it can transition
> to healthy state.
> Comments?
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
