flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Fabian Hueske <fhue...@gmail.com>
Subject Re: How to force the parallelism on small streams?
Date Thu, 03 Sep 2015 14:06:54 GMT
The purpose of rebalance() should be to rebalance the partitions of a data
streams as evenly as possible, right?
If all senders start sending data to the same receiver and there is less
data in each partition than receivers, partitions are not evenly rebalanced.
That is exactly the problem Arnaud ran into.

IMO, that's a bug and should be fixed.

2015-09-03 15:53 GMT+02:00 Matthias J. Sax <mjsax@apache.org>:

> For rebalance() this makes sense. I don't think anything must be
> changed. For regular data, there is no such issues as for this very
> small data set.
>
> However for shuffle() I would expect that each source task uses a
> different shuffle pattern...
>
> -Matthias
>
> On 09/03/2015 03:28 PM, Fabian Hueske wrote:
> > In case of rebalance(), all sources start the round-robin partitioning at
> > index 0. Since each source emits only very few elements, only the first
> 15
> > mappers receive any input.
> > It would be better to let each source start the round-robin partitioning
> at
> > a different index, something like startIdx = (numReceivers / numSenders)
> *
> > myIdx.
> >
> > In case of shuffle(), the ShufflePartitioner initializes Random()
> without a
> > seed (the current time is taken).
> > However, the ShufflePartitioner is only initialized once at the client
> side
> > (if I see that correctly) and then the same instance is deserialized by
> all
> > operators, i.e., all use random number generators with the same seed.
> >
> > I think, the StreamPartitioner class should be extended with a
> > configuration / initialize method which is called on each parallel
> operator.
> >
> > Cheers, Fabian
> >
> > 2015-09-03 15:04 GMT+02:00 Aljoscha Krettek <aljoscha@apache.org>:
> >
> >> Hi,
> >> I don't think it's a bug. If there are 100 sources that each emit only
> 14
> >> elements then only the first 14 mappers will ever receive data. The
> >> round-robin distribution is not global, since the sources operate
> >> independently from each other.
> >>
> >> Cheers,
> >> Aljoscha
> >>
> >> On Wed, 2 Sep 2015 at 20:00 Matthias J. Sax <mjsax@apache.org> wrote:
> >>
> >>> Thanks for clarifying. shuffle() is similar to rebalance() -- however,
> >>> it redistributes randomly and not in round robin fashion.
> >>>
> >>> However, the problem you describe sounds like a bug to me. I included
> >>> dev list. Maybe anyone else can step in so we can identify it there is
> a
> >>> bug or not.
> >>>
> >>> -Matthias
> >>>
> >>>
> >>> On 09/02/2015 06:19 PM, LINZ, Arnaud wrote:
> >>>> Hi,
> >>>>
> >>>> You are right, but in fact it does not solve my problem, since I have
> >>> 100 parallelism everywhere. Each of my 100 sources gives only a few
> lines
> >>> (say 14 max), and only the first 14 next nodes will receive data.
> >>>> Same problem by replacing rebalance() with shuffle().
> >>>>
> >>>> But I found a workaround: setting parallelism to 1 for the source (I
> >>> don't need a 100 directory scanners anyway), it forces the rebalancing
> >>> evenly between the mappers.
> >>>>
> >>>> Greetings,
> >>>> Arnaud
> >>>>
> >>>>
> >>>> -----Message d'origine-----
> >>>> De : Matthias J. Sax [mailto:mjsax@apache.org]
> >>>> Envoyé : mercredi 2 septembre 2015 17:56
> >>>> À : user@flink.apache.org
> >>>> Objet : Re: How to force the parallelism on small streams?
> >>>>
> >>>> Hi,
> >>>>
> >>>> If I understand you correctly, you want to have 100 mappers. Thus you
> >>> need to apply the .setParallelism() after .map()
> >>>>
> >>>>>
> addSource(myFileSource).rebalance().map(myFileMapper).setParallelism(1
> >>>>> 00)
> >>>>
> >>>> The order of commands you used, set the dop for the source to 100
> >> (which
> >>> might be ignored, if the provided source function "myFileSource" does
> not
> >>> implements "ParallelSourceFunction" interface). The dop for the mapper
> >>> should be the default value.
> >>>>
> >>>> Using .rebalance() is absolutely correct. It distributes the emitted
> >>> tuples in a round robin fashion to all consumer tasks.
> >>>>
> >>>> -Matthias
> >>>>
> >>>> On 09/02/2015 05:41 PM, LINZ, Arnaud wrote:
> >>>>> Hi,
> >>>>>
> >>>>>
> >>>>>
> >>>>> I have a source that provides few items since it gives file names
to
> >>>>> the mappers. The mapper opens the file and process records. As the
> >>>>> files are huge, one input line (a filename) gives a consequent work
> to
> >>> the next stage.
> >>>>>
> >>>>> My topology looks like :
> >>>>>
> >>>>>
> addSource(myFileSource).rebalance().setParallelism(100).map(myFileMapp
> >>>>> er)
> >>>>>
> >>>>> If 100 mappers are created, about 85 end immediately and only a
few
> >>>>> process the files (for hours). I suspect an optimization making
that
> >>>>> there is a minimum number of lines to pass to the next node or it
is
> >>>>> “shutdown” ; but in my case I do want the lines to be evenly
> >>>>> distributed to each mapper.
> >>>>>
> >>>>> How to enforce that ?
> >>>>>
> >>>>>
> >>>>>
> >>>>> Greetings,
> >>>>>
> >>>>> Arnaud
> >>>>>
> >>>>>
> >>>>>
> ----------------------------------------------------------------------
> >>>>> --
> >>>>>
> >>>>> L'intégrité de ce message n'étant pas assurée sur internet,
la
> société
> >>>>> expéditrice ne peut être tenue responsable de son contenu ni de
ses
> >>>>> pièces jointes. Toute utilisation ou diffusion non autorisée est
> >>>>> interdite. Si vous n'êtes pas destinataire de ce message, merci
de le
> >>>>> détruire et d'avertir l'expéditeur.
> >>>>>
> >>>>> The integrity of this message cannot be guaranteed on the Internet.
> >>>>> The company that sent this message cannot therefore be held liable
> for
> >>>>> its content nor attachments. Any unauthorized use or dissemination
is
> >>>>> prohibited. If you are not the intended recipient of this message,
> >>>>> then please delete it and notify the sender.
> >>>>
> >>>
> >>>
> >>
> >
>
>

Mime
View raw message