Yes. Semantically all outputs from a ParDo are equivalent, so the watermark
should traverse them all. The only reason a "default" output exists is for
convenience so we don't force users to always specify an output tag.
On Sun, Jul 9, 2017 at 12:03 AM, Thomas Weise <t...@apache.org> wrote:
> This error turns out to be deterministic and debug friendly :) I enabled
> trace and found that the watermark "disappears" between the following two
> operators:
>
> [8/WriteCounts/WriteFiles/WriteBundles:ApexParDoOperator]
>
> [10/WriteCounts/WriteFiles/GroupUnwritten:ApexGroupByKeyOperator]
> GroupUnwritten takes input from additonal outputs, but the watermark is
> only emitted on the main output. When I modify ApexParDoOperator to emit
> the watermark also on additionalOutput1, it traverses the pipeline and the
> test passes.
>
> Are watermarks supposed to be written to additional outputs?
>
> Thanks,
> Thomas
>
>
> On Thu, Jul 6, 2017 at 1:35 PM, Reuven Lax <re...@google.com.invalid>
> wrote:
>
> > Thomas, any suggestions on what we should do? Do you have an idea what's
> > going on, or should we exclude this test for now until you have time to
> > look at it?
> >
> > Reuven
> >
> > On Wed, Jul 5, 2017 at 3:36 PM, Reuven Lax <re...@google.com> wrote:
> >
> > > I wonder if the watermark is accidentally advancing too early, causing
> > > Apex to shut down the pipeline before the final finalize DoFn executes?
> > >
> > > On Wed, Jul 5, 2017 at 1:45 PM, Thomas Weise <t...@apache.org> wrote:
> > >
> > >> I don't think this is a problem with the test and if anything this
> > problem
> > >> to me shows the test is useful in catching similar issues during unit
> > test
> > >> runs.
> > >>
> > >> Is there any form of asynchronous/trigger based processing in this
> > >> pipeline
> > >> that could cause this?
> > >>
> > >> The Apex runner will shutdown the pipeline after the final watermark,
> > the
> > >> shutdown signal traverses the pipeline just like a watermark, but it
> is
> > >> not
> > >> seen by user code.
> > >>
> > >> Thomas
> > >>
> > >> --
> > >> sent from mobile
> > >> On Jul 5, 2017 1:19 PM, "Kenneth Knowles" <k...@google.com.invalid>
> > wrote:
> > >>
> > >> > Upon further investigation, this tests always writes to
> > >> > ./target/wordcountresult-00000-of-00002 and
> > >> > ./target/wordcountresult-00001-of-00002. So after a successful test
> > >> run,
> > >> > any further run without a `clean` will spuriously succeed. I was
> > running
> > >> > via IntelliJ so did not do the ritual `mvn clean` workaround. So
> > >> > reproduction appears to be easy and we could fix the test (if we
> don't
> > >> > remove it) to use a fresh temp dir.
> > >> >
> > >> > This seems to point to a bug in waitUntilFinish() and/or Apex if the
> > >> > topology is shut down before this ParDo is run. This is a ParDo with
> > >> > trivial bounded input but with side inputs. So I would guess the bug
> > is
> > >> > either in watermark tracking / readiness of the side input or just
> how
> > >> > PushbackSideInputDoFnRunner is used.
> > >> >
> > >> > On Wed, Jul 5, 2017 at 12:23 PM, Reuven Lax
> <re...@google.com.invalid
> > >
> > >> > wrote:
> > >> >
> > >> > > I've done a bit more debugging with logging. It appears that the
> > >> finalize
> > >> > > ParDo is never being invoked in this Apex test (or at least the
> > >> LOG.info
> > >> > in
> > >> > > that ParDo never runs). This ParDo is run on a constant element
> > (code
> > >> > > snippet below), so it should always run.
> > >> > >
> > >> > > PCollection<Void> singletonCollection = p.apply(Create.of((Void)
> > >> null));
> > >> > > singletonCollection
> > >> > > .apply("Finalize", ParDo.of(new DoFn<Void, Integer>() {
> > >> > > @ProcessElement
> > >> > > public void processElement(ProcessContext c) throws
> Exception
> > {
> > >> > > LOG.info("Finalizing write operation {}.",
> writeOperation);
> > >> > >
> > >> > >
> > >> > > On Wed, Jul 5, 2017 at 11:22 AM, Kenneth Knowles
> > >> <k...@google.com.invalid
> > >> > >
> > >> > > wrote:
> > >> > >
> > >> > > > Data-dependent file destinations is a pretty great feature. We
> > also
> > >> > have
> > >> > > > another change to make to this @Experimental feature, and it
> would
> > >> be
> > >> > > nice
> > >> > > > to get them both into 2.1.0 if we can unblock this quickly.
> > >> > > >
> > >> > > > I just tried this too, and failed to reproduce it. But Jenkins
> and
> > >> > Reuven
> > >> > > > both have a reliable repro.
> > >> > > >
> > >> > > > Questionss:
> > >> > > >
> > >> > > > - Any ideas about how these configurations differ?
> > >> > > > - Does this actually affect users?
> > >> > > > - Once we have another test that catches this issue, can we
> > delete
> > >> > this
> > >> > > > test?
> > >> > > >
> > >> > > > Every other test passes, including the actual example
> WordCountIT.
> > >> > Since
> > >> > > > the PR doesn't change primitives, it also seems like it is an
> > >> existing
> > >> > > > issue. And the test seems redundant with our other testing but
> > won't
> > >> > get
> > >> > > as
> > >> > > > much maintenance attention. I don't want to stop catching
> whatever
> > >> this
> > >> > > > issue is, though.
> > >> > > >
> > >> > > > Kenn
> > >> > > >
> > >> > > > On Wed, Jul 5, 2017 at 10:31 AM, Reuven Lax
> > >> <re...@google.com.invalid>
> > >> > > > wrote:
> > >> > > >
> > >> > > > > Hi Thomas,
> > >> > > > >
> > >> > > > > This only happens with https://github.com/apache/
> beam/pull/3356
> > .
> > >> > > > >
> > >> > > > > Reuven
> > >> > > > >
> > >> > > > > On Mon, Jul 3, 2017 at 6:11 AM, Thomas Weise <t...@apache.org>
> > >> wrote:
> > >> > > > >
> > >> > > > > > Hi Reuven,
> > >> > > > > >
> > >> > > > > > I'm not able to reproduce the issue locally. I was hoping to
> > see
> > >> > > which
> > >> > > > > > thread is attempting to emit the results. In Apex, only the
> > >> > operator
> > >> > > > > thread
> > >> > > > > > can emit the results, any other thread that is launched by
> the
> > >> > > operator
> > >> > > > > > cannot. I'm not aware of ParDo managing separate threads
> > though
> > >> and
> > >> > > > > assume
> > >> > > > > > this must be a race. If you still have the log, can you send
> > it
> > >> to
> > >> > > me?
> > >> > > > > >
> > >> > > > > > Thanks,
> > >> > > > > > Thomas
> > >> > > > > >
> > >> > > > > >
> > >> > > > > >
> > >> > > > > > On Sat, Jul 1, 2017 at 5:51 AM, Reuven Lax
> > >> > <re...@google.com.invalid
> > >> > > >
> > >> > > > > > wrote:
> > >> > > > > >
> > >> > > > > > > pr/3356 fails in the Apex WordCountTest. The failed test
> is
> > >> here
> > >> > > > > > > <https://builds.apache.org/job/beam_PreCommit_Java_
> > >> > > > > > > MavenInstall/12829/org.apache.beam$beam-runners-apex/
> > >> > > > > > > testReport/org.apache.beam.runners.apex.examples/
> > WordCountTe
> > >> st/
> > >> > > > > > > testWordCountExample/>
> > >> > > > > > > :
> > >> > > > > > >
> > >> > > > > > > Upon debugging, it looks like this is likely a problem in
> > the
> > >> > Apex
> > >> > > > > runner
> > >> > > > > > > itself. A ParDo calls output(), and that triggers an
> > exception
> > >> > > thrown
> > >> > > > > > from
> > >> > > > > > > inside the Apex runner. The Apex runner calls emit on a
> > >> > > > > DefaultOutputPort
> > >> > > > > > > (ApexParDoOperator.java:275), and that throws an exception
> > >> inside
> > >> > > of
> > >> > > > > > > verifyOperatorThread().
> > >> > > > > > >
> > >> > > > > > > I'm going to ignore this failure for now as it seems
> > >> unrelated to
> > >> > > my
> > >> > > > > PR,
> > >> > > > > > > but does someone want to take a look?
> > >> > > > > > >
> > >> > > > > > > Reuven
> > >> > > > > > >
> > >> > > > > >
> > >> > > > >
> > >> > > >
> > >> > >
> > >> >
> > >>
> > >
> > >
> >
>
