Perhaps the feature of https://issues.apache.org/jira/browse/MAPREDUCE-5583 is
what you are looking for.

On Fri, 27 May 2016 at 00:04 Per Ullberg <[email protected]> wrote:

> The fair scheduler would solve this issue, but we need the capacity
> scheduler for other reason. Would it be possible to run multiple schedulers
> in parallel?
>
> /Pelle
>
> On Thursday, May 26, 2016, David Morel <[email protected]> wrote:
>
> > Le 26 mai 2016 9:04 AM, "Per Ullberg" <[email protected]
> > <javascript:;>> a écrit :
> > >
> > > The split is skewed. Just running one sqoop action will cause some
> > > containers to finish early and others to finish late. If we run the
> > actions
> > > concurrently, the early finishers will be idle until all containers for
> > > that action is done and the next action can commence. By running the
> > > actions in parallel, we will finish earlier in total and also utilize
> our
> > > cluster resources better.
> >
> > I used the FairScheduler for exactly this scenario at my previous job.
> >
> > David
> >
> > > regards
> > > /Pelle
> > >
> > > On Thu, May 26, 2016 at 3:09 AM, Robert Kanter <[email protected]
> > <javascript:;>>
> > wrote:
> > >
> > > > Hi,
> > > >
> > > > If you want to only run one of the Sqoop Actions at a time, why not
> > simply
> > > > remove the fork and run the Sqoop Actions sequentially?
> > > >
> > > > - Robert
> > > >
> > > > On Tue, May 3, 2016 at 12:15 AM, Per Ullberg <[email protected]
> > <javascript:;>>
> > > > wrote:
> > > >
> > > > > Hi,
> > > > >
> > > > > We have an oozie workflow that imports data table by table from a
> > RDBMS
> > > > > using sqoop. One action per table. The sqoop commands use "split by
> > > > column"
> > > > > and spread out on a number of mappers.
> > > > >
> > > > > We fork all the actions so basically all sqoop jobs are launched at
> > once.
> > > > >
> > > > > The RDBMS can only accept a fixed number of connections and if this
> > is
> > > > > exceeded, the sqoop action will fail and eventually the whole oozie
> > > > > workflow will fail.
> > > > >
> > > > > We use the yarn capacity scheduler (2.6.0) and have set up a
> specific
> > > > queue
> > > > > for this job to throttle the maximum number of concurrent
> containers.
> > > > > However, this setup is hard to manage because all configurations in
> > the
> > > > > capacity scheduler are relative to the max amount of vcores of the
> > > > cluster
> > > > > and as we add machines or otherwise tune the cluster, the actual
> > number
> > > > of
> > > > > containers granted to the oozie job changes and at times we hit the
> > > > > connection roof.
> > > > >
> > > > > So, is there another way to throttle the number of concurrent
> > containers
> > > > > for an oozie job? I guess you would have to be able to throttle
> both
> > > > > launchers and map-reduce containers?
> > > > >
> > > > > best regards
> > > > > /Pelle
> > > > >
> > > > >
> > > > > --
> > > > >
> > > > > *Per Ullberg*
> > > > > Tech Lead
> > > > > Odin - Uppsala
> > > > >
> > > > > Klarna AB
> > > > > Sveavägen 46, 111 34 Stockholm
> > > > > Tel: +46 8 120 120 00
> > > > > Reg no: 556737-0431
> > > > > klarna.com
> > > > >
> > > >
> > >
> > >
> > >
> > > --
> > >
> > > *Per Ullberg*
> > > Tech Lead
> > > Odin - Uppsala
> > >
> > > Klarna AB
> > > Sveavägen 46, 111 34 Stockholm
> > > Tel: +46 8 120 120 00
> > > Reg no: 556737-0431
> > > klarna.com
> >
>
>
> --
>
> *Per Ullberg*
> Tech Lead
> Odin - Uppsala
>
> Klarna AB
> Sveavägen 46, 111 34 Stockholm
> Tel: +46 8 120 120 00
> Reg no: 556737-0431
> klarna.com
>

Reply via email to