[
https://issues.apache.org/jira/browse/CHUKWA-4?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12850369#action_12850369
]
Jerome Boulon commented on CHUKWA-4:
------------------------------------
I'm exploring a couple of options:
1- use the local fileSystem instead of HDFS in the first place and it's working
pretty well for me since my HDFS is S3 so I can not really write directly to it
2- At startup time, any .chukwa files with last access time greater than 2
rotations period cannot be file that collectors are writing to. So we can
2.1 Open the file and read it, if not then you have todo 2.2 so not sure if
always doing 2.2 will not be a better option
2.2 Open the file, Create a new SequenceFile and copy data from one to the
other, then close the file and rename. To avoid being trapped to the same issue
and case of Kill -9/Crash, we need to use another extension like .recover and
at start time we can delete any .recover file.
Also at this point, there's no a valid way to process invalid SequenceFile in
hadoop so paying the price when we know that something could be wrong seems to
be better that crashing a M/R with 300 input files and then having to parse
them all ....
> Collectors don't finish writing .done datasink from last .chukwa datasink
> when stopped using bin/stop-collectors
> ----------------------------------------------------------------------------------------------------------------
>
> Key: CHUKWA-4
> URL: https://issues.apache.org/jira/browse/CHUKWA-4
> Project: Hadoop Chukwa
> Issue Type: Bug
> Components: data collection
> Environment: I am running on our local cluster. This is a linux
> machine that I also run Hadoop cluster from.
> Reporter: Andy Konwinski
> Priority: Minor
>
> When I use start-collectors, it creates the datasink as expected, writes to
> it as per normal, i.e. writes to the .chukwa file, and roll overs work fine
> when it renames the .chukwa file to .done. However, when I use
> bin/stop-collectors to shut down the running collector it leaves a .chukwa
> file in the HDFS file system. Not sure if this is a valid sink or not, but I
> think that the collector should gracefully clean up the datasink and rename
> it .done before exiting.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
