EOFException related to memory segments during run

问题:

I'm trying to run an Apache Beam pipeline on Flink on our test cluster. It has been failing with an EOFException at org.apache.flink.runtime.io.disk.SimpleCollectingOutputView:79 during the encoding of an object through serialisation. I haven't been able to reproduce the error locally, yet. You can find the entire job log here. Some values have been replaced with fake data.

The command used to run the pipeline:

bin/flink run \
     -m yarn-cluster                                         \
     --yarncontainer                 1                       \
     --yarnslots                     4                       \
     --yarnjobManagerMemory          2000                    \
     --yarntaskManagerMemory         2000                    \
     --yarnname "EBI"        \
     pipeline.jar               \
     --runner=FlinkRunner \
     --zookeeperQuorum=hdp-master-001.fake.org:2181

While I think it's not related, the object-to-be-serialised is serialisable and has had both an implicit and an explicit coder, but this doesn't affect the situation.

What might be causing this situation and what can I do to address it?

---

For now, increasing the heap memory of the managers to somewhere between 4 and 8GiB seems to prevent the exception. Still unsure if this is supposed to be normal Flink behaviour (shouldn't it spill to disk?). Doesn't seem like a solution that would scale.

回答1:

The EOFException is thrown because Flink ran out of memory buffers. Flink expects an EOFException as a notification to start to write data to disk.

This problem is caused by Beam's SerializableCoder wraps the EOFException in a CoderException. Hence, Flink does not catch the expected EOFException and fails.

The problem can be solved by using a custom coder that does not wrap the EOFException but forwards it.