Thank you for analyzing the problem Gordon! We can not upgrade Kryo anytime soon because state in old savepoints is still serialized with the current Kryo version. I would propose to add our own JavaThrowableSerializer to Flink and document how users can register that serializer if they run into the error.
Shannon and Bruno can just use the serializer in their current Flink version. On Sat, Mar 11, 2017 at 12:00 PM, Tzu-Li (Gordon) Tai <tzuli...@apache.org> wrote: > FYI: Here’s the JIRA ticket to track this issue - > https://issues.apache.org/jira/browse/FLINK-6025. > > > On March 11, 2017 at 6:27:36 PM, Tzu-Li (Gordon) Tai (tzuli...@apache.org) > wrote: > > Hi Shannon, > > Thanks a lot for providing the example, it was very helpful in reproducing > the problem. > > I think this is actually a Kryo bug, that was just recently fixed: > https://github.com/EsotericSoftware/kryo/pull/483 > <https://github.com/EsotericSoftware/kryo/commit/19a6b5e> > It will be available in Kryo 4.0.1, which unfortunately doesn’t seem to be > released yet. > > The problem is that when Kryo defaults to Java serialization for the > exception instance, the `ObjectInputStream` used to read the object does > not correctly use Kryo’s configured class loader (i.e., the user code class > loader). That’s why it's complaining that the class cannot be found. > > We can “workaround” this by registering our own `JavaSerializer` as the > serializer for Throwables in Kryo, but I’m not sure if we should actually > do this, or just wait for the Kryo fix to be released. > > - Gordon > > > On March 11, 2017 at 9:54:03 AM, Shannon Carey (sca...@expedia.com) wrote: > > Here ya go (see attached). > > > From: Robert Metzger <rmetz...@apache.org> > Date: Friday, March 10, 2017 at 1:18 PM > To: "user@flink.apache.org" <user@flink.apache.org> > Subject: Re: AWS exception serialization problem > > Can one of you guys provide us with a minimal example to reproduce the > issue? (Ideally locally, not using EMR?) > I think once we can reproduce the issue its easy to fix. > > On Thu, Mar 9, 2017 at 1:24 AM, Bruno Aranda <brunoara...@gmail.com> > wrote: > >> Hi Stephan, we are running Flink 1.2.0 on Yarn (AWS EMR cluster) >> >> On Wed, 8 Mar 2017, 21:41 Stephan Ewen, <se...@apache.org> wrote: >> >>> @Bruno: How are you running Flink? On yarn, standalone, mesos, docker? >>> >>> On Wed, Mar 8, 2017 at 2:13 PM, Bruno Aranda <brunoara...@gmail.com> >>> wrote: >>> >>> Hi, >>> >>> We have seen something similar in Flink 1.2. We have an operation that >>> parses some JSON, and when it fails to parse it, we can see the >>> ClassNotFoundException for the relevant exception (in our case >>> JsResultException from the play-json library). The library is indeed in the >>> shaded JAR, otherwise we would not be able to parse the JSON. >>> >>> Cheers, >>> >>> Bruno >>> >>> On Wed, 8 Mar 2017 at 12:57 Tzu-Li (Gordon) Tai <tzuli...@apache.org> >>> wrote: >>> >>> Hi Shannon, >>> >>> Just to clarify: >>> >>> From the error trace, it seems like that the messages fetched from Kafka >>> are serialized `AmazonS3Exception`s, and you’re emitting a stream of >>> `AmazonS3Exception` as records from FlinkKafkaConsumer? >>> Is this correct? If so, I think we should just make sure that the >>> `com.amazonaws.services.s3.model.AmazonS3Exception` class exists in the >>> user fat jar. >>> >>> Also, what is the Flink version you are using? >>> >>> Cheers, >>> Gordon >>> >>> >>> >
