Home | About | Sematext search-lucene.com search-hadoop.com
 Search Hadoop and all its subprojects:

Switch to Plain View
Flume >> mail # user >> At flume shutdown only one HDFSSink is closed properly

Bhaskar V. Karambelkar 2012-08-23, 22:24
Mike Percy 2012-08-24, 01:35
Bhaskar V. Karambelkar 2012-08-24, 19:13
Mike Percy 2012-08-24, 19:40
Copy link to this message
Re: At flume shutdown only one HDFSSink is closed properly
Hi Mike,
There are indeed Jira issues for this.
A flume jira issue , which was fixed in 1.2.0.  FLUME-1163.

But that is only half the story, the hadoop jira issue for supporting the
'fs.automatic.shutdown' flag
is logged at

There is even a mail thread about it on this mailing list

Hadoop 1.x don't have the 4829 patch, so we can't use Hadoop 1.x with Flume
if we want proper
renaming of .tmp files at flume shutdown.

CDH4, as well as CDH3u5 do have that patch pulled in, so they are fine.


On Fri, Aug 24, 2012 at 3:40 PM, Mike Percy <[EMAIL PROTECTED]> wrote:

> Thanks for the additional info Bhaskar. So is this a known issue in
> vanilla Hadoop 1.0.3? If so do you have a JIRA number?
> Regards,
> Mike
> On Fri, Aug 24, 2012 at 12:13 PM, Bhaskar V. Karambelkar <
>> oops, this is just the same Hadoop's FileSystem.close() shutdown hook
>> issue. I was getting the exception no matter whether I had 1 HDFS sink or
>> more.
>> I was using hadoop vanilla 1.0.3, and looks like that one doesn't respect
>> the fs.automatic.close option.
>> Switched to CDH3u5, and no more problems, all the HDFS sinks correctly
>> rename the file on shutdown.
>> In conclusion, the vanilla hadoop 1.x series is not an option for flume.
>> Go with Hadoop 2.x or CDH3u5, CDH4
>> thanks
>> Bhaskar
>> On Thu, Aug 23, 2012 at 9:35 PM, Mike Percy <[EMAIL PROTECTED]> wrote:
>>> Hmm... this likely happens because Hadoop statically caches the
>>> FileSystem object, so as it turns out, the multiple Sinks are sharing the
>>> same FileSystem objects.
>>> I think the only reason we need to explicitly close the FileSystem
>>> objects is to support the deleteOnExit feature. We are explicitly closing
>>> them because we removed the automatic shutdown hook typically installed by
>>> Hadoop to invoke FileSystem.close(), since it was interfering with the .tmp
>>> rolling. I wonder if we can get away with never closing them in our
>>> case... I'm not sure if we need the deleteOnExit() functionality implicitly
>>> for any reason, or if there are other more important reasons behind why the
>>> FileSystem objects should be closed.
>>> Regards,
>>> Mike
>>> On Thu, Aug 23, 2012 at 3:24 PM, Bhaskar V. Karambelkar <
>>> [EMAIL PROTECTED]> wrote:
>>>> I have 3 HDFS sinks all writing to the same namenode, but different
>>>> paths.
>>>> e.g. sink1 = hdfs://namenode/path1
>>>> sink2 = hdfs://namenode/path2
>>>> etc.
>>>> When flume is shutdown (kill <flume-pid>), the file for the first sink
>>>> is closed correctly and renamed to remove the .tmp extension
>>>> but the second file's closing throws the following exception and the
>>>> file's .tmp extension is also not removed.
>>>> I see this happening very consistently, for 1+ HDFS sinks, only the
>>>> first one is closed properly and renamed, the rest all throw exception
>>>> when being closed, and are not renamed to remove the .tmp extension.
>>>> 2012-08-23 19:51:39,837 WARN hdfs.BucketWriter: failed to close()
>>>> HDFSWriter for file
>>>> (hdfs://hadoop-namnode:9000/flume/avro/2012/08/23/l/avro_.1345751470750.tmp).
>>>> Exception follows.
>>>> java.io.IOException: Filesystem closed
>>>>     at org.apache.hadoop.hdfs.DFSClient.checkOpen(DFSClient.java:264)
>>>>     at org.apache.hadoop.hdfs.DFSClient.access$1100(DFSClient.java:74)
>>>>     at
>>>> org.apache.hadoop.hdfs.DFSClient$DFSOutputStream.sync(DFSClient.java:3667)
>>>>     at
>>>> org.apache.hadoop.fs.FSDataOutputStream.sync(FSDataOutputStream.java:97)
>>>>     at
>>>> org.apache.flume.sink.hdfs.HDFSDataStream.close(HDFSDataStream.java:103)
>>>>     at
>>>> org.apache.flume.sink.hdfs.BucketWriter.doClose(BucketWriter.java:250)
>>>>     at
>>>> org.apache.flume.sink.hdfs.BucketWriter.access$400(BucketWriter.java:48)
Mike Percy 2012-08-29, 08:30