Don't worry, we're all constantly learning. :-)

On Wed, 30 Nov 2016 at 12:00 Ismaël Mejía <[email protected]> wrote:

> Hello, thanks for the clarifying Aljoscha, you are absolutely right, sorry
> for my imprecision. I should do my homework and work more with flink :).
>
> For ref, the iteration API:
>
>
> https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/datastream_api.html#iterations
>
> And the implementation of the benchmark that effectively does not use
> iterations:
>
>
> https://github.com/yahoo/streaming-benchmarks/blob/master/flink-benchmarks/src/main/java/flink/benchmark/AdvertisingTopologyNative.java
>
> My excuses again,
> Ismael
>
> On Wed, Nov 30, 2016 at 11:30 AM, Aljoscha Krettek <[email protected]>
> wrote:
>
> > Hi,
> > there is support for cycles in Flink but the Yahoo benchmark is not
> making
> > use of that feature, if I'm not completely mistaken.
> >
> > Cheers,
> > Aljoscha
> >
> > On Wed, 30 Nov 2016 at 09:57 Ismaël Mejía <[email protected]> wrote:
> >
> > > Hello,
> > >
> > > Shen you should probably first check the benchmark implementation at
> > > github, I am not sure you need cycles to implement the yahoo benchmark.
> > > Notice for example that there is a flink based implementation and AFAIK
> > > there is no support for cycles in Flink (or at least there wasn't at
> the
> > > moment they published the benchmarks).
> > >
> > > However if you are implementing the yahoo benchmark on Beam, that would
> > be
> > > a nice scenario to test the runner performance (vs the native
> > > implementations), so it would be nice if you can share this.
> > >
> > > Regards,
> > > Ismael
> > >
> > > On Tue, Nov 29, 2016 at 7:32 PM, Shen LI <[email protected]> wrote:
> > >
> > > > Hi Maria, Bobby,
> > > >
> > > > Thanks for the explanation.
> > > >
> > > > Regards,
> > > >
> > > > Shen
> > > >
> > > > On Tue, Nov 29, 2016 at 12:37 PM, Bobby Evans
> > > <[email protected]
> > > > >
> > > > wrote:
> > > >
> > > > > In my experience almost all of the time cycles are bad and cause a
> > lot
> > > of
> > > > > debugging problems. Most of the time you can implement what you
> want
> > by
> > > > > using a windowed join or group by instead.
> > > > > - Bobby
> > > > >
> > > > > On Tuesday, November 29, 2016, 11:06:44 AM CST, María García
> Herrero
> > > > > <[email protected]> wrote:Hi Shen,
> > > > >
> > > > > No. Beam pipelines are DAGs:
> > > > > http://beam.incubator.apache.org/documentation/sdks/
> > > > > javadoc/0.3.0-incubating/org/apache/beam/sdk/Pipeline.html
> > > > > Best,
> > > > >
> > > > > María
> > > > >
> > > > > On Tue, Nov 29, 2016 at 7:44 AM, Shen LI <[email protected]>
> > wrote:
> > > > >
> > > > > > Hi,
> > > > > >
> > > > > > Can I use Beam to create a pipeline with cycles? For example, to
> > > > > implement
> > > > > > the Yahoo! Streaming benchmark(
> > > > > >
> > > https://yahooeng.tumblr.com/post/135321837876/benchmarking-streaming-
> > > > > > computation-engines-at),
> > > > > > can a ParDo transform consume a downstream output as a side
> input?
> > > > > >
> > > > > > Thanks,
> > > > > >
> > > > > > Shen
> > > > > >
> > > > >
> > > >
> > >
> >
>

Reply via email to