osdir.com

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: org.apache.beam.runners.flink.PortableTimersExecutionTest is very flakey


I believe that the ParDos are being invoked in parallel. I'm not sure on the exact semantics, but I believe that beam will execute separate keys on separate threads, when it processes different bundles for different those keys.
I logged the thread IDs in this test, to verify that different threads are invoking this code.
Applying my fix, I was able to pass the test 400/400 runs.

I talked to Luke, and he suggested using PAssert, which is the most thread safe/standard way to verify pipeline results It also simplifies the code a little bit, removing the last unnecessary DoFn.

PTAL at this PR, I recommend committing this in to remove the concurrency issue collecting test results and remove flakeyness in this test.
https://github.com/apache/beam/pull/7214/files

Please also try to follow this pattern on any other tests which you may have.

FWIW, Here is the logged thread ids that I saw, I appended logs to a ConcurrentLinkedQueue and printed them at the end of the test, so this shows the separate threads and the interleaving of them.
processElement collectResults 26 : 5000 results: 1601591000 threadId: pool-32-thread-15
processElement collectResults 34 : 4093 results: 1360449464 threadId: pool-32-thread-14
processElement collectResults 47 : 4093 results: 323962224 threadId: pool-32-thread-19
processElement collectResults 19 : 4093 results: 323962224 threadId: pool-32-thread-19
processElement collectResults 45 : 4093 results: 167183883 threadId: pool-32-thread-18
processElement collectResults 0 : 4093 results: 167183883 threadId: pool-32-thread-18
processElement collectResults 2 : 4093 results: 167183883 threadId: pool-32-thread-18
processElement collectResults 30 : 4093 results: 865903006 threadId: pool-32-thread-21
processElement collectResults 11 : 4093 results: 865903006 threadId: pool-32-thread-21
processElement collectResults 1 : 4093 results: 865903006 threadId: pool-32-thread-21
processElement collectResults 41 : 4093 results: 1183940089 threadId: pool-32-thread-23
processElement collectResults 7 : 4093 results: 1183940089 threadId: pool-32-thread-23
processElement collectResults 13 : 4093 results: 1183940089 threadId: pool-32-thread-23
processElement collectResults 36 : 4093 results: 1183940089 threadId: pool-32-thread-23
processElement collectResults 21 : 4093 results: 907415986 threadId: pool-32-thread-17
processElement collectResults 32 : 4093 results: 907415986 threadId: pool-32-thread-17
processElement collectResults 10 : 4093 results: 907415986 threadId: pool-32-thread-17
processElement collectResults 20 : 4093 results: 907415986 threadId: pool-32-thread-17
processElement collectResults 14 : 4093 results: 907415986 threadId: pool-32-thread-17
processElement collectResults 24 : 4093 results: 1391785351 threadId: pool-32-thread-15
processElement collectResults 46 : 4093 results: 1391785351 threadId: pool-32-thread-15
processElement collectResults 17 : 4093 results: 1391785351 threadId: pool-32-thread-15



On Wed, Dec 5, 2018 at 6:49 AM Maximilian Michels <mxm@xxxxxxxxxx> wrote:
Thank you for looking into the test. For me the flakiness was solely caused by
the non thread-safe GrpcStateService. I have since closed the JIRA issue because
I didn't see another failure since the fix.

Your fixes are valid, but they won't fix flakiness (if present) in the current
testing pipeline. Why? The results are only ever written by 1 worker because the
testing source uses Impulse which generates a signal only received by a single
worker. So the shared results list is not a problem for this test.

Let me quickly comment on the changes you mentioned:

1) Yes, if we had a parallel source, the List should be a concurrent or
synchronized list.

2) Using a static list should be fine for testing purposes. There are no other
tests accessing this map. Tests are not run parallel on a class level. Besides,
there is only one test in this class.

3) If you make the results object transient, then it won't be serialized, so it
will be 'null' after being deserialized.


Thank you for spending the time to look into the test. Do you think it would
make sense to address changes to the test separately of your PR?

I believe the test can be further improved, e.g. to run more parallel. Also, if
you see any flakiness after the merged fix, please post the build log in the
JIRA issue.

Thanks,
Max

On 05.12.18 03:18, Alex Amato wrote:
> Well, here is my hacky solution.
> You can see the changes I make to PortableTimersExecutionTest
> https://github.com/apache/beam/pull/6786/files
>
> I don't really understand why the pipeline never starts running when I make the
> results object transient in PortableTiemrsExecutionTest.
>
> So I instead continue to access a static object, but key it with the test
> parameter, to prevent tests from interfering with each other.
>
> I am not too sure how to proceed. I don't really want to check in this hacky
> solution. But I am not too sure of what else to do with solved the problems.
> Please let me know if you have any suggestions.
>
> On Tue, Dec 4, 2018 at 5:26 PM Alex Amato <ajamato@xxxxxxxxxx
> <mailto:ajamato@xxxxxxxxxx>> wrote:
>
>     Thanks for letting me know Maximillian,
>
>     Btw, I've been looking a this test the last few days as well. I have found a
>     few other concurrency issues. That I hope to send a PR out for.
>
>       * The PortableTimersExecutionTest result variable is using a static
>         ArrayList, but can be writen to concurrently (by multiple thread AND
>         multiple parameterized test instnace) which causing flakeyness.
>       * But just using a ConcurrentLinkedQueue and a non static variable isn't
>         sufficient as that will cause a copy of the results object to be copied
>         during doFn serialization. So that makes all the assertions fail, since
>         nothing get written to the same result object the test is using/
>           o So it should be made private transient final. However, after trying
>             this I am seeing the test timeout, and I am not sure why. Continuing
>             to debug this.
>
>
>     I think that my PR was increasing flakeyness, which is why I saw more of
>     these issues.
>     Just wanted to point these out in the meantime, hopefull it helps with
>     debugging for you too.
>
>     On Fri, Nov 30, 2018 at 7:49 AM Maximilian Michels <mxm@xxxxxxxxxx
>     <mailto:mxm@xxxxxxxxxx>> wrote:
>
>         This turned out to be a tricky bug. Robert and me had a joined debugging
>         session and managed to find the culprit.
>
>         PR pending: https://github.com/apache/beam/pull/7171
>
>         On 27.11.18 19:35, Kenneth Knowles wrote:
>          > I actually didn't look at this one. I filed a bunch more adjacent flake
>          > bugs. I didn't find your bug but I do see that test flaking at the same
>          > time as the others. FWIW here is the list of flakes and sickbayed tests:
>          > https://issues.apache.org/jira/issues/?filter=12343195
>          >
>          > Kenn
>          >
>          > On Tue, Nov 27, 2018 at 10:25 AM Alex Amato <ajamato@xxxxxxxxxx
>         <mailto:ajamato@xxxxxxxxxx>
>          > <mailto:ajamato@xxxxxxxxxx <mailto:ajamato@xxxxxxxxxx>>> wrote:
>          >
>          >     +Ken,
>          >
>          >     Did you happen to look into this test? I heard that you may have
>          >     been looking into this.
>          >
>          >     On Mon, Nov 26, 2018 at 3:36 PM Maximilian Michels
>         <mxm@xxxxxxxxxx <mailto:mxm@xxxxxxxxxx>
>          >     <mailto:mxm@xxxxxxxxxx <mailto:mxm@xxxxxxxxxx>>> wrote:
>          >
>          >         Hi Alex,
>          >
>          >         Thanks for your help! I'm quite used to debugging
>          >         concurrent/distributed
>          >         problems. But this one is quite tricky, especially with regards
>          >         to GRPC
>          >         threads. I try to provide more information in the following.
>          >
>          >         There are two observations:
>          >
>          >         1) The problem is specifically related to how the cleanup is
>          >         performed
>          >         for the EmbeddedEnvironmentFactory. The environment is shutdown
>          >         when the
>          >         SDK Harness exists but the GRPC threads continue to linger for
>          >         some time
>          >         and may stall state processing on the next test.
>          >
>          >         If you do _not_ close DefaultJobBundleFactory, which happens
>         during
>          >         close() or dispose() in the FlinkExecutableStageFunction or
>          >         ExecutableStageDoFnOperator respectively, the tests run just
>          >         fine. I ran
>          >         1000 test runs without a single failure.
>          >
>          >         The EmbeddedEnvironment uses direct channels which are marked
>          >         experimental in GRPC. We may have to convert them to regular
>         socket
>          >         communication.
>          >
>          >         2) Try setting a conditional breakpoint in GrpcStateService
>          >         which will
>          >         never break, e.g. "false". Set it here:
>          >
>         https://github.com/apache/beam/blob/6da9aa5594f96c0201d497f6dce4797c4984a2fd/runners/java-fn-execution/src/main/java/org/apache/beam/runners/fnexecution/state/GrpcStateService.java#L134
>          >
>          >         The tests will never fail. The SDK harness is always shutdown
>          >         correctly
>          >         at the end of the test.
>          >
>          >         Thanks,
>          >         Max
>          >
>          >         On 26.11.18 19:15, Alex Amato wrote:
>          >          > Thanks Maximilian, let me know if you need any help. Usually
>          >         I debug
>          >          > this sort of thing by pausing the IntelliJ debugger to see
>          >         all the
>          >          > different threads which are waiting on various conditions. If
>          >         you find
>          >          > any insights from that, please post them here and we can try
>          >         to figure
>          >          > out the source of the stuckness. Perhaps it may be some
>          >         concurrency
>          >          > issue leading to deadlock?
>          >          >
>          >          > On Thu, Nov 22, 2018 at 12:57 PM Maximilian Michels
>          >         <mxm@xxxxxxxxxx <mailto:mxm@xxxxxxxxxx>
>         <mailto:mxm@xxxxxxxxxx <mailto:mxm@xxxxxxxxxx>>
>          >          > <mailto:mxm@xxxxxxxxxx <mailto:mxm@xxxxxxxxxx>
>         <mailto:mxm@xxxxxxxxxx <mailto:mxm@xxxxxxxxxx>>>> wrote:
>          >          >
>          >          >     I couldn't fix it thus far. The issue does not seem to be
>          >         in the Flink
>          >          >     Runner but in the way the tests utilizes the EMBEDDED
>          >         environment to
>          >          >     run
>          >          >     multiple portable jobs in a row.
>          >          >
>          >          >     When it gets stuck it is in RemoteBundle#close and it is
>          >         independent of
>          >          >     the test type (batch and streaming have different
>          >         implementations).
>          >          >
>          >          >     Will give it another look tomorrow.
>          >          >
>          >          >     Thanks,
>          >          >     Max
>          >          >
>          >          >     On 22.11.18 13:07, Maximilian Michels wrote:
>          >          >      > Hi Alex,
>          >          >      >
>          >          >      > The test seems to have gotten flaky after we merged
>          >         support for
>          >          >     portable
>          >          >      > timers in Flink's batch mode.
>          >          >      >
>          >          >      > Looking into this now.
>          >          >      >
>          >          >      > Thanks,
>          >          >      > Max
>          >          >      >
>          >          >      > On 21.11.18 23:56, Alex Amato wrote:
>          >          >      >> Hello, I have noticed
>          >          >      >>
>          >         that org.apache.beam.runners.flink.PortableTimersExecutionTest
>          >          >     is very
>          >          >      >> flakey, and repro'd this test timeout on the master
>          >         branch in
>          >          >     40/50 runs.
>          >          >      >>
>          >          >      >> I filed a JIRA issue: BEAM-6111
>          >          >      >> <https://issues.apache.org/jira/browse/BEAM-6111>. I
>          >         was just
>          >          >      >> wondering if anyone knew why this may be occurring,
>          >         and to check if
>          >          >      >> anyone else has been experiencing this.
>          >          >      >>
>          >          >      >> Thanks,
>          >          >      >> Alex
>          >          >
>          >
>