hadoop-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Bertrand Dechoux <decho...@gmail.com>
Subject Re: Question about MapReduce
Date Wed, 14 Nov 2012 08:51:57 GMT
That would be "mapred.task.timeout".
You should download hadoop source, that way you would have the conf files
with all the default configuration.
(You could also get them online but lots of links are broken.)

And the name of the property is only half the story.
The error says explicitly that no report status was sent for this period
which is the reason for the killing.
If you do send a report status, you don't need to change the
mapred.task.timeout value which is already pretty high (IMHO/YMMV).

The way you can send a status would depend on whether you are using mapred
or mapreduce API.

Regards

Bertrand

PS : I cced this discussion to the non hbase specific user mailing list.
Replies could be done only on this done because it is a non HBase specific
question.

On Wed, Nov 14, 2012 at 3:41 AM, Jean-Marc Spaggiari <
jean-marc@spaggiari.org> wrote:

> One more question about MapReduce.
>
> One of my servers is slower than the others. I don't have any time
> constraint for the job to finish.
>
> But I'm getting this message:
>
> "Task attempt_201211122318_0014_m_000021_0 failed to report status for
> 601 seconds. Killing!"
>
> Where can I chance this timeout to something like 1800 seconds? Is it
> on the mapred-site.xml file? If so, which property should I insert?
>
> Thanks,
>
> JM
>
> 2012/11/2, Jean-Marc Spaggiari <jean-marc@spaggiari.org>:
> > That was my initial plan too, but I was wondering if there was any
> > other best practice about the delete.  So I will go that way.
> >
> > Thanks,
> >
> > JM
> >
> > 2012/11/2, Shrijeet Paliwal <shrijeet@rocketfuel.com>:
> >> Not sure what exactly is happening in your job. But in one of the delete
> >> jobs I wrote I was creating an instance of HTable in setup method of my
> >> mapper
> >>
> >> delTab = new HTable(conf, conf.get(TABLE_NAME));
> >>
> >> And performing delete in map() call using delTab. So no, you do not have
> >> access to table directly *usually*.
> >>
> >>
> >> -Shrijeet
> >>
> >>
> >> On Fri, Nov 2, 2012 at 12:47 PM, Jean-Marc Spaggiari <
> >> jean-marc@spaggiari.org> wrote:
> >>
> >>> Sorry, one last question.
> >>>
> >>> On the map method, I have access to the row using the values
> >>> parameter. Now, based on the value content, I might want to delete it.
> >>> Do I have access to the table directly from one of the parameters? Or
> >>> should I call the delete using an HTableInterface from my pool?
> >>>
> >>> Thanks,
> >>>
> >>> JM
> >>>
> >>> 2012/11/2, Jean-Marc Spaggiari <jean-marc@spaggiari.org>:
> >>> > Yep, you perfectly got my question.
> >>> >
> >>> > I just tried and it's working perfectly!
> >>> >
> >>> > Thanks a lot! I now have a lot to play with.
> >>> >
> >>> > JM
> >>> >
> >>> > 2012/11/2, Shrijeet Paliwal <shrijeet@rocketfuel.com>:
> >>> >> JM,
> >>> >>
> >>> >> I personally would chose to put it neither hadoop libs nor hbase
> >>> >> libs.
> >>> >> Have
> >>> >> them go to your application's own install directory.
> >>> >>
> >>> >> Then you could sent the variable HADOOP_CLASSPATH to have your
jar
> >>> >> (also
> >>> >> include hbase jars, hbase dependencies and dependencies your program
> >>> >> needs)
> >>> >> And to execute fire 'hadoop jar' command.
> >>> >>
> >>> >> An example[1]:
> >>> >>
> >>> >> Set classpath:
> >>> >> export HADOOP_CLASSPATH=`hbase
> >>> classpath`:mycool.jar:mycooldependency.jar
> >>> >>
> >>> >> Fire following to launch your job:
> >>> >> hadoop jar mycool.jar hbase.experiments.MyCoolProgram
> >>> >> -Dmapred.running.map.limit=50
> >>> >> -Dmapred.map.tasks.speculative.execution=false aCommandLineArg
> >>> >>
> >>> >>
> >>> >> Did I get your question right?
> >>> >>
> >>> >> [1] In the example I gave `hbase classpath` gets you set with all
> >>> >> hbase
> >>> >> jars.
> >>> >>
> >>> >>
> >>> >>
> >>> >> On Fri, Nov 2, 2012 at 11:56 AM, Jean-Marc Spaggiari <
> >>> >> jean-marc@spaggiari.org> wrote:
> >>> >>
> >>> >>> Hi Shrijeet,
> >>> >>>
> >>> >>> Helped a lot! Thanks!
> >>> >>>
> >>> >>> Now, the only think I need is to know where's the best place
to put
> >>> >>> my
> >>> >>> JAR on the server. Should I put it on the hadoop lib directory?
Or
> >>> >>> somewhere on the HBase structure?
> >>> >>>
> >>> >>> Thanks,
> >>> >>>
> >>> >>> JM
> >>> >>>
> >>> >>> 2012/10/29, Shrijeet Paliwal <shrijeet@rocketfuel.com>:
> >>> >>> > In line.
> >>> >>> >
> >>> >>> > On Mon, Oct 29, 2012 at 8:11 AM, Jean-Marc Spaggiari <
> >>> >>> > jean-marc@spaggiari.org> wrote:
> >>> >>> >
> >>> >>> >> I'm replying to myself ;)
> >>> >>> >>
> >>> >>> >> I found "cleanup" and "setup" methods from the TableMapper
> table.
> >>> >>> >> So
> >>> >>> >> I
> >>> >>> >> think those are the methods I was looking for. I will
init the
> >>> >>> >> HTablePool there. Please let me know if I'm wrong.
> >>> >>> >>
> >>> >>> >> Now, I still have few other questions.
> >>> >>> >>
> >>> >>> >> 1) context.getCurrentValue() can throw a InterrruptedException,
> >>> >>> >> but
> >>> >>> >> when can this occur? Is there a timeout on the Mapper
side? Of
> >>> >>> >> it's
> >>> >>> >> if
> >>> >>> >> the region is going down while the job is running?
> >>> >>> >>
> >>> >>> >
> >>> >>> > You do not need to call  context.getCurrentValue(). The
'value'
> >>> >>> > argument
> >>> >>> to
> >>> >>> > map method[1] has the information you are looking for.
> >>> >>> >
> >>> >>> >
> >>> >>> >> 2) How can I pass parameters to the Map method? Can
I use
> >>> >>> >> job.getConfiguration().put to add some properties
there, can get
> >>> them
> >>> >>> >> back in context.getConfiguration.get?
> >>> >>> >>
> >>> >>> >
> >>> >>> > Yes, thats how it is done.
> >>> >>> >
> >>> >>> >
> >>> >>> >> 3) What's the best way to log results/exceptions/traces
from the
> >>> >>> >> map
> >>> >>> >> method?
> >>> >>> >>
> >>> >>> >
> >>> >>> > In most cases, you'll have mapper and reducer classes
as nested
> >>> static
> >>> >>> > classes within some enclosing class. You can get handle
to the
> >>> >>> > Logger
> >>> >>> from
> >>> >>> > the enclosing class and do your usual LOG.info, LOG.warn
yada
> >>> >>> > yada.
> >>> >>> >
> >>> >>> > Hope it helps.
> >>> >>> >
> >>> >>> > [1] map(KEYIN key, *VALUEIN value*, Context context)
> >>> >>> >
> >>> >>> >>
> >>> >>> >> I will search on my side, but some help will be welcome
because
> >>> >>> >> it
> >>> >>> >> seems there is not much documentation when we start
to dig a bit
> >>> >>> >> :(
> >>> >>> >>
> >>> >>> >> JM
> >>> >>> >>
> >>> >>> >> 2012/10/27, Jean-Marc Spaggiari <jean-marc@spaggiari.org>:
> >>> >>> >> > Hi,
> >>> >>> >> >
> >>> >>> >> > I'm thinking about my firs MapReduce class and
I have some
> >>> >>> >> > questions.
> >>> >>> >> >
> >>> >>> >> > The goal of it will be to move some rows from
one table to
> >>> >>> >> > another
> >>> >>> >> > one
> >>> >>> >> > based on the timestamp only.
> >>> >>> >> >
> >>> >>> >> > Since this is pretty new for me, I'm starting
from the
> >>> >>> >> > RowCounter
> >>> >>> >> > class to have a baseline.
> >>> >>> >> >
> >>> >>> >> > There are few things I will have to update. First,
the
> >>> >>> >> > createSumittableJob method to get timestamp range
instead of
> >>> >>> >> > key
> >>> >>> >> > range, and "play2 with the parameters. This part
is fine.
> >>> >>> >> >
> >>> >>> >> > Next, I need to update the map method, and this
is where I
> have
> >>> >>> >> > some
> >>> >>> >> > questions.
> >>> >>> >> >
> >>> >>> >> > I'm able to find the timestamp of all the cf:c
from the
> >>> >>> >> > context.getCurrentValue() method, that's fine.
Now, my concern
> >>> >>> >> > is
> >>> >>> >> > on
> >>> >>> >> > the way to get access to the table to store this
field, and
> the
> >>> >>> >> > table
> >>> >>> >> > to delete it. Should I instantiate an HTable
for the source
> >>> >>> >> > table,
> >>> >>> >> > and
> >>> >>> >> > execute and delete on it, then do an insert on
another HTable
> >>> >>> >> > instance?  Should I use an HTablePool? Also,
since I’m already
> >>> >>> >> > on
> >>> >>> >> > the
> >>> >>> >> > row, can’t I just mark it as deleted instead
of calling a new
> >>> >>> >> > HTable?
> >>> >>> >> >
> >>> >>> >> > Also, instead of calling the delete and put one
by one, I
> would
> >>> >>> >> > like
> >>> >>> >> > to put them on a list and execute it only when
it’s over 10
> >>> >>> >> > members.
> >>> >>> >> > How can I make sure that at the end of the job,
this is
> >>> >>> >> > flushed?
> >>> >>> >> > Else,
> >>> >>> >> > I will lose some operations. Is there a kind
of “dispose”
> >>> >>> >> > method
> >>> >>> >> > called on the region when the job is done?
> >>> >>> >> >
> >>> >>> >> > Thanks,
> >>> >>> >> >
> >>> >>> >> > JM
> >>> >>> >> >
> >>> >>> >>
> >>> >>> >
> >>> >>>
> >>> >>
> >>> >
> >>>
> >>
> >
>



-- 
Bertrand Dechoux

Mime
View raw message