hadoop-common-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Chris Nauroth <cnaur...@hortonworks.com>
Subject Re: How to call Hadoop job from a web service in a non-blocking fashion?
Date Mon, 27 Apr 2015 16:42:19 GMT
Hello Marko,

Job#waitForCompletion is implemented as a polling loop around
Job#isComplete and Job#isSuccessful.  Both of those calls are non-blocking.

http://hadoop.apache.org/docs/r2.7.0/api/org/apache/hadoop/mapreduce/Job.ht
ml#isComplete()


http://hadoop.apache.org/docs/r2.7.0/api/org/apache/hadoop/mapreduce/Job.ht
ml#isSuccessful()


(Technically they do block on I/O due to RPC calls to check job status,
but the idea is that they don't block waiting for the entire job to
complete the way waitForCompletion does.)

Do you think you could build the solution that you need around using these
2 methods?  I believe Oozie implements it in a similar way.  Even though
you aren't using Oozie, you might consider looking at the Oozie codebase
for inspiration.  I think the relevant class in Oozie would be
JavaActionExecutor.  Disclaimer: I really don't know Oozie very well.  :-)

--Chris Nauroth




On 4/27/15, 5:35 AM, "Marko Dinic" <marko.dinic@nissatech.com> wrote:

>Hello,
>
>I have a sequence of jobs that depend on each other, output of one job
>is input for the next one. Also, there is a loop in one part of the
>sequence, containing two jobs executing in a row.
>
>Until now I was able to run this job by simply creating Job objects and
>using waitForCompletition(true). In that way, I was forwarding output of
>one job as input to next one.
>
>The problem is, waitForCompletition(true) will block the web service I'm
>trying to use, so I need a way to run this sequence of dependent jobs,
>but not to get stuck waiting for result of the whole sequence. So, I
>want the next model - user uploads some files, starts the job and gets
>the response that the job has been started. After the sequence has
>finished user should be notified in some way.
>
>I wouldn't like to use Oozie, since this the jobs are more low-level
>(it's actually an algorithm similar to those implemented in Mahout), and
>I don't know if I may use JobControl, since there is a loop, and how to
>do it.
>
>Any help would be highly appreciated.
>
>Regards,
>Marko


Mime
View raw message