We usually package my jobs as a single jar that contains a /lib directory in the jar that contains all other jars that the job code depends on. Hadoop understands this layout when run as 'hadoop jar'. So the jar layout would be something like:
/META-INF/manifest.mf /com/mypackage/MyMapperClass.class /com/mypackage/MyReducerClass.class /lib/dependency1.jar /lib/dependency2.jar etc. If you use Maven or some other build tool with dependency management, you can usually produce this jar as part of your build. We also have Maven write the main class to the manifest, such that there is no need to type it. So for us, submitting a job looks like: hadoop jar jar-with-all-deps-in-lib.jar arg1 arg2 argN Then Hadoop will take care of submitting and distributing, etc. Of course you pay the penalty of always sending all of your dependencies over the wire (the job jar gets replicated to 10 machines by default). Pre-distributing sounds tedious and error prone to me. What if you have different jobs that require different versions of the same dependency? HTH, Friso On 16 nov. 2011, at 15:42, Something Something wrote: Bejoy - Thanks for the reply. The '-libjars' is not working for me with 'hadoop jar'. Also, as per the documentation (http://hadoop.apache.org/common/docs/current/commands_manual.html#jar): Generic Options The following options are supported by dfsadmin<http://hadoop.apache.org/common/docs/current/commands_manual.html#dfsadmin>, fs<http://hadoop.apache.org/common/docs/current/commands_manual.html#fs>, fsck<http://hadoop.apache.org/common/docs/current/commands_manual.html#fsck>, job<http://hadoop.apache.org/common/docs/current/commands_manual.html#job> and fetchdt<http://hadoop.apache.org/common/docs/current/commands_manual.html#fetchdt>. Does it work for you? If it does, please let me know. "Pre-distributing" definitely works, but is that the best way? If you have a big cluster and Jars are changing often it will be time-consuming. Also, how does Pig do it? We update Pig UDFs often and put them only on the 'client' machine (machine that starts the Pig job) and the UDF becomes available to all machines in the cluster - automagically! Is Pig doing the pre-distributing for us? Thanks for your patience & help with our questions. On Wed, Nov 16, 2011 at 6:29 AM, Something Something <mailinglist...@gmail.com<mailto:mailinglist...@gmail.com>> wrote: Hmm... there must be a different way 'cause we don't need to do that to run Pig jobs. On Tue, Nov 15, 2011 at 10:58 PM, Daan Gerits <daan.ger...@gmail.com<mailto:daan.ger...@gmail.com>> wrote: There might be different ways but currently we are storing our jars onto HDFS and register them from there. They will be copied to the machine once the job starts. Is that an option? Daan. On 16 Nov 2011, at 07:24, Something Something wrote: > Until now we were manually copying our Jars to all machines in a Hadoop > cluster. This used to work until our cluster size was small. Now our > cluster is getting bigger. What's the best way to start a Hadoop Job that > automatically distributes the Jar to all machines in a cluster? > > I read the doc at: > http://hadoop.apache.org/common/docs/current/commands_manual.html#jar > > Would -libjars do the trick? But we need to use 'hadoop job' for that, > right? Until now, we were using 'hadoop jar' to start all our jobs. > > Needless to say, we are getting our feet wet with Hadoop, so appreciate > your help with our dumb questions. > > Thanks. > > PS: We use Pig a lot, which automatically does this, so there must be a > clean way to do this.
