Re: Error while reading from hadoop sequence file
u please clarify? I would expect that it should work once you call
> the method and provide the type info, or else what exactly is the exception
> there.
>
> Best,
> Stefan
>
> On 10. Dec 2018, at 13:35, Akshay Mendole wrote:
>
> Hi,
>I have been facing issues while trying to read from a hdfs sequence
> file.
>
> This is my code snippet
>
> DataSource> input = env
> .createInput(HadoopInputs.readSequenceFile(Text.class, Text.class,
> ravenDataDir),
> TypeInformation.of(new TypeHint>() {
> }));
>
>
> Upon executing this in yarn cluster mode, I am getting following error
> The type returned by the input format could not be automatically
> determined. Please specify the TypeInformation of the produced type
> explicitly by using the 'createInput(InputFormat, TypeInformation)' method
> instead.
>
> org.apache.flink.api.java.ExecutionEnvironment.createInput(ExecutionEnvironment.java:551)
> flipkart.EnrichementFlink.main(EnrichementFlink.java:31)
>
>
> When I add the TypeInformation myself as follows, I run into the same
> issue.
>
> DataSource> input = env
> .createInput(HadoopInputs.readSequenceFile(Text.class, Text.class,
> ravenDataDir));
>
>
>
>
> When I add these libraries in the lib folder,
> flink-hadoop-compatibility_2.11-1.7.0.jar
>
>
> the error changes to this
>
> java.lang.NoClassDefFoundError:
> org/apache/flink/api/common/typeutils/TypeSerializerSnapshot
> at
> org.apache.flink.api.java.typeutils.WritableTypeInfo.createSerializer(WritableTypeInfo.java:111)
> at
> org.apache.flink.api.java.typeutils.TupleTypeInfo.createSerializer(TupleTypeInfo.java:107)
> at
> org.apache.flink.api.java.typeutils.TupleTypeInfo.createSerializer(TupleTypeInfo.java:52)
> at
> org.apache.flink.optimizer.postpass.JavaApiPostPass.createSerializer(JavaApiPostPass.java:283)
> at
> org.apache.flink.optimizer.postpass.JavaApiPostPass.traverseChannel(JavaApiPostPass.java:252)
> at
> org.apache.flink.optimizer.postpass.JavaApiPostPass.traverse(JavaApiPostPass.java:97)
> at
> org.apache.flink.optimizer.postpass.JavaApiPostPass.postPass(JavaApiPostPass.java:81)
> at org.apache.flink.optimizer.Optimizer.compile(Optimizer.java:527)
> at org.apache.flink.optimizer.Optimizer.compile(Optimizer.java:399)
> at
> org.apache.flink.client.program.ClusterClient.getOptimizedPlan(ClusterClient.java:379)
> at
> org.apache.flink.client.program.ClusterClient.getOptimizedPlan(ClusterClient.java:906)
> at
> org.apache.flink.client.program.ClusterClient.run(ClusterClient.java:473)
> at
> org.apache.flink.client.program.ContextEnvironment.execute(ContextEnvironment.java:62)
>
>
> Can someone help me resolve this issue?
>
> Thanks,
> Akshay
>
>
>
>
>
Re: Error while reading from hadoop sequence file
Could anyone please help me with this?
Thanks,
Akshay
On Mon, 10 Dec 2018, 6:05 pm Akshay Mendole Hi,
>I have been facing issues while trying to read from a hdfs sequence
> file.
>
> This is my code snippet
>
> DataSource> input = env
> .createInput(HadoopInputs.readSequenceFile(Text.class, Text.class,
> ravenDataDir),
> TypeInformation.of(new TypeHint>() {
> }));
>
>
> Upon executing this in yarn cluster mode, I am getting following error
> The type returned by the input format could not be automatically
> determined. Please specify the TypeInformation of the produced type
> explicitly by using the 'createInput(InputFormat, TypeInformation)' method
> instead.
>
> org.apache.flink.api.java.ExecutionEnvironment.createInput(ExecutionEnvironment.java:551)
> flipkart.EnrichementFlink.main(EnrichementFlink.java:31)
>
>
> When I add the TypeInformation myself as follows, I run into the same
> issue.
>
> DataSource> input = env
> .createInput(HadoopInputs.readSequenceFile(Text.class, Text.class,
> ravenDataDir));
>
>
>
>
> When I add these libraries in the lib folder,
> flink-hadoop-compatibility_2.11-1.7.0.jar
>
>
> the error changes to this
>
> java.lang.NoClassDefFoundError:
> org/apache/flink/api/common/typeutils/TypeSerializerSnapshot
> at
> org.apache.flink.api.java.typeutils.WritableTypeInfo.createSerializer(WritableTypeInfo.java:111)
> at
> org.apache.flink.api.java.typeutils.TupleTypeInfo.createSerializer(TupleTypeInfo.java:107)
> at
> org.apache.flink.api.java.typeutils.TupleTypeInfo.createSerializer(TupleTypeInfo.java:52)
> at
> org.apache.flink.optimizer.postpass.JavaApiPostPass.createSerializer(JavaApiPostPass.java:283)
> at
> org.apache.flink.optimizer.postpass.JavaApiPostPass.traverseChannel(JavaApiPostPass.java:252)
> at
> org.apache.flink.optimizer.postpass.JavaApiPostPass.traverse(JavaApiPostPass.java:97)
> at
> org.apache.flink.optimizer.postpass.JavaApiPostPass.postPass(JavaApiPostPass.java:81)
> at org.apache.flink.optimizer.Optimizer.compile(Optimizer.java:527)
> at org.apache.flink.optimizer.Optimizer.compile(Optimizer.java:399)
> at
> org.apache.flink.client.program.ClusterClient.getOptimizedPlan(ClusterClient.java:379)
> at
> org.apache.flink.client.program.ClusterClient.getOptimizedPlan(ClusterClient.java:906)
> at
> org.apache.flink.client.program.ClusterClient.run(ClusterClient.java:473)
> at
> org.apache.flink.client.program.ContextEnvironment.execute(ContextEnvironment.java:62)
>
>
> Can someone help me resolve this issue?
>
> Thanks,
> Akshay
>
>
>
>
Error while reading from hadoop sequence file
Hi,
I have been facing issues while trying to read from a hdfs sequence file.
This is my code snippet
DataSource> input = env
.createInput(HadoopInputs.readSequenceFile(Text.class, Text.class,
ravenDataDir),
TypeInformation.of(new TypeHint>() {
}));
Upon executing this in yarn cluster mode, I am getting following error
The type returned by the input format could not be automatically
determined. Please specify the TypeInformation of the produced type
explicitly by using the 'createInput(InputFormat, TypeInformation)' method
instead.
org.apache.flink.api.java.ExecutionEnvironment.createInput(ExecutionEnvironment.java:551)
flipkart.EnrichementFlink.main(EnrichementFlink.java:31)
When I add the TypeInformation myself as follows, I run into the same issue.
DataSource> input = env
.createInput(HadoopInputs.readSequenceFile(Text.class, Text.class,
ravenDataDir));
When I add these libraries in the lib folder,
flink-hadoop-compatibility_2.11-1.7.0.jar
the error changes to this
java.lang.NoClassDefFoundError:
org/apache/flink/api/common/typeutils/TypeSerializerSnapshot
at
org.apache.flink.api.java.typeutils.WritableTypeInfo.createSerializer(WritableTypeInfo.java:111)
at
org.apache.flink.api.java.typeutils.TupleTypeInfo.createSerializer(TupleTypeInfo.java:107)
at
org.apache.flink.api.java.typeutils.TupleTypeInfo.createSerializer(TupleTypeInfo.java:52)
at
org.apache.flink.optimizer.postpass.JavaApiPostPass.createSerializer(JavaApiPostPass.java:283)
at
org.apache.flink.optimizer.postpass.JavaApiPostPass.traverseChannel(JavaApiPostPass.java:252)
at
org.apache.flink.optimizer.postpass.JavaApiPostPass.traverse(JavaApiPostPass.java:97)
at
org.apache.flink.optimizer.postpass.JavaApiPostPass.postPass(JavaApiPostPass.java:81)
at org.apache.flink.optimizer.Optimizer.compile(Optimizer.java:527)
at org.apache.flink.optimizer.Optimizer.compile(Optimizer.java:399)
at
org.apache.flink.client.program.ClusterClient.getOptimizedPlan(ClusterClient.java:379)
at
org.apache.flink.client.program.ClusterClient.getOptimizedPlan(ClusterClient.java:906)
at org.apache.flink.client.program.ClusterClient.run(ClusterClient.java:473)
at
org.apache.flink.client.program.ContextEnvironment.execute(ContextEnvironment.java:62)
Can someone help me resolve this issue?
Thanks,
Akshay
Re: OutOfMemoryError while doing join operation in flink
Hi Zhijiang, Thanks for the explanation and the workaround suggested. While this can work for the example stated above, we have more complex use cases where we would have to re-tune the above parameters. FYI, we ran into same problems when we did a simple groupBy on the skewed dataset. Thanks, Akshay On Fri, Nov 23, 2018 at 8:29 AM zhijiang wrote: > Hi Akshay, > > Sorrry I have not thought of a proper way to handle single large record in > distributed task managers in flink. But I can give some hints for adjusting > the related memories for work around OOM issue. > Large fraction of memories in task manager are managed by flink for > efficiency, and these memories are long live persistent in JVM not recycled > by gc. You can check the parameter "taskmanager.memory.fraction" for this > and the default value is 0.7 if you have not changed, that means 7GB * 0.7 > are used by framework. > > I am not sure what is the flink version you used. If I rememberd > correctly, before release-1.5 the network buffers also uses heap memories > by default, so you should also minus this part of memory from total task > manager memory. > > If not considering network buffer used by framework, you only leave 7GB * > 0.3 temporaray memories for other parts. The temporaray memories in > serializer will exceed twice as current size every time if not covering the > record size, that means one serializer may need 2GB overhead memories for > your 1GB record. You have 2 slots per task manager for running two tasks, > so the total overhead memories may need 4GB almost. So you can decrease > the "taskmanager.memory.fraction" in low fraction or increase the total > task manager to cover this overhead memories, or set one slot for each task > manager. > > Best, > Zhijiang > > -- > 发件人：Akshay Mendole > 发送时间：2018年11月23日(星期五) 02:54 > 收件人：trohrmann > 抄 送：zhijiang ; user ; > Shreesha Madogaran > 主 题：Re: OutOfMemoryError while doing join operation in flink > > Hi, > Thanks for your reply. I tried running a simple "group by" on just one > dataset where few keys are repeatedly occurring (in order of millions) and > did not include any joins. I wanted to see if this issue is specific to > join. But as I was expecting, I ran into the same issue. I am giving 7GBs > to each task manager with 2 slots per task manager. From what I understood > so far, such cases where individual records somewhere in the pipeline > become so large that they should be handled in distributed manner instead > of handling them by a simple data structure in single JVM. I am guessing > there is no way to do this in Flink today. > Could you please confirm this? > Thanks, > Akshay > > > On Thu, Nov 22, 2018 at 9:28 PM Till Rohrmann > wrote: > Hi Akshay, > > Flink currently does not support to automatically distribute hot keys > across different JVMs. What you can do is to adapt the parallelism/number > of partitions manually if you encounter that one partition contains a lot > of hot keys. This might mitigate the problem by partitioning the hot keys > into different partitions. > > Apart from that, the problem seems to be as Zhijiang indicated that your > join result is quite large. One record is 1 GB large. Try to decrease it or > give more memory to your TMs. > > Cheers, > Till > > On Thu, Nov 22, 2018 at 1:08 PM Akshay Mendole > wrote: > Hi Zhijiang, > Thanks for the quick reply. My concern is more towards > how flink perform joins of two *skewed *datasets. Pig > <https://wiki.apache.org/pig/PigSkewedJoinSpec> and spark > <https://wiki.apache.org/pig/PigSkewedJoinSpec> seems to support the join > of skewed datasets. The record size that you are mentioning about in your > reply is after join operation takes place which is definitely going to be > huge enough not to fit in jvm task manager task slot in my use case. We > want to know if there is a way in flink to handle such skewed keys by > distributing their values across different jvms. Let me know if you need > more clarity on the issue. > Thanks, > Akshay > > On Thu, Nov 22, 2018 at 2:38 PM zhijiang > wrote: > Hi Akshay, > > You encountered an existing issue for serializing large records to cause > OOM. > > Every subpartition would create a separate serializer before, and each > serializer would maintain an internal bytes array for storing intermediate > serialization results. The key point is that these overhead internal bytes > array are not managed by framework, and their size would exceed with the > record size dynamically. If your job has many subpartitions with large > re
Re: OutOfMemoryError while doing join operation in flink
Hi, Thanks for your reply. I tried running a simple "group by" on just one dataset where few keys are repeatedly occurring (in order of millions) and did not include any joins. I wanted to see if this issue is specific to join. But as I was expecting, I ran into the same issue. I am giving 7GBs to each task manager with 2 slots per task manager. From what I understood so far, such cases where individual records somewhere in the pipeline become so large that they should be handled in distributed manner instead of handling them by a simple data structure in single JVM. I am guessing there is no way to do this in Flink today. Could you please confirm this? Thanks, Akshay On Thu, Nov 22, 2018 at 9:28 PM Till Rohrmann wrote: > Hi Akshay, > > Flink currently does not support to automatically distribute hot keys > across different JVMs. What you can do is to adapt the parallelism/number > of partitions manually if you encounter that one partition contains a lot > of hot keys. This might mitigate the problem by partitioning the hot keys > into different partitions. > > Apart from that, the problem seems to be as Zhijiang indicated that your > join result is quite large. One record is 1 GB large. Try to decrease it or > give more memory to your TMs. > > Cheers, > Till > > On Thu, Nov 22, 2018 at 1:08 PM Akshay Mendole > wrote: > >> Hi Zhijiang, >> Thanks for the quick reply. My concern is more towards >> how flink perform joins of two *skewed *datasets. Pig >> <https://wiki.apache.org/pig/PigSkewedJoinSpec> and spark >> <https://wiki.apache.org/pig/PigSkewedJoinSpec> seems to support the >> join of skewed datasets. The record size that you are mentioning about in >> your reply is after join operation takes place which is definitely going to >> be huge enough not to fit in jvm task manager task slot in my use case. We >> want to know if there is a way in flink to handle such skewed keys by >> distributing their values across different jvms. Let me know if you need >> more clarity on the issue. >> Thanks, >> Akshay >> >> On Thu, Nov 22, 2018 at 2:38 PM zhijiang >> wrote: >> >>> Hi Akshay, >>> >>> You encountered an existing issue for serializing large records to cause >>> OOM. >>> >>> Every subpartition would create a separate serializer before, and each >>> serializer would maintain an internal bytes array for storing intermediate >>> serialization results. The key point is that these overhead internal bytes >>> array are not managed by framework, and their size would exceed with the >>> record size dynamically. If your job has many subpartitions with large >>> records, it may probably cause OOM issue. >>> >>> I already improved this issue to some extent by sharing only one >>> serializer for all subpartitions [1], that means we only have one bytes >>> array overhead at most. This issue is covered in release-1.7. >>> Currently the best option may reduce your record size if possible or you >>> can increase the heap size of task manager container. >>> >>> [1] https://issues.apache.org/jira/browse/FLINK-9913 >>> >>> Best, >>> Zhijiang >>> >>> -- >>> 发件人：Akshay Mendole >>> 发送时间：2018年11月22日(星期四) 13:43 >>> 收件人：user >>> 主 题：OutOfMemoryError while doing join operation in flink >>> >>> Hi, >>> We are converting one of our pig pipelines to flink using apache >>> beam. The pig pipeline reads two different data sets (R1 & R2) from hdfs, >>> enriches them, joins them and dumps back to hdfs. The data set R1 is >>> skewed. In a sense, it has few keys with lot of records. When we converted >>> the pig pipeline to apache beam and ran it using flink on a production yarn >>> cluster, we got the following error >>> >>> 2018-11-21 16:52:25,307 ERROR >>> org.apache.flink.runtime.operators.BatchTask - Error in >>> task code: GroupReduce (GroupReduce at CoGBK/GBK) (25/100) >>> java.lang.RuntimeException: Emitting the record caused an I/O exception: >>> Failed to serialize element. Serialized size (> 1136656562 bytes) exceeds >>> JVM heap space >>> at >>> org.apache.flink.runtime.operators.shipping.OutputCollector.collect(OutputCollector.java:69) >>> at >>> org.apache.flink.runtime.operators.util.metrics.CountingCollector.collect(CountingCollector.java:35) >>> at >>> org.ap
OutOfMemoryError while doing join operation in flink
Hi, We are converting one of our pig pipelines to flink using apache beam. The pig pipeline reads two different data sets (R1 & R2) from hdfs, enriches them, joins them and dumps back to hdfs. The data set R1 is skewed. In a sense, it has few keys with lot of records. When we converted the pig pipeline to apache beam and ran it using flink on a production yarn cluster, we got the following error 2018-11-21 16:52:25,307 ERROR org.apache.flink.runtime.operators.BatchTask - Error in task code: GroupReduce (GroupReduce at CoGBK/GBK) (25/100) java.lang.RuntimeException: Emitting the record caused an I/O exception: Failed to serialize element. Serialized size (> 1136656562 bytes) exceeds JVM heap space at org.apache.flink.runtime.operators.shipping.OutputCollector.collect(OutputCollector.java:69) at org.apache.flink.runtime.operators.util.metrics.CountingCollector.collect(CountingCollector.java:35) at org.apache.beam.runners.flink.translation.functions.SortingFlinkCombineRunner.combine(SortingFlinkCombineRunner.java:140) at org.apache.beam.runners.flink.translation.functions.FlinkReduceFunction.reduce(FlinkReduceFunction.java:85) at org.apache.flink.api.java.operators.translation.PlanUnwrappingReduceGroupOperator$TupleUnwrappingNonCombinableGroupReducer.reduce(PlanUnwrappingReduceGroupOperator.java:111) at org.apache.flink.runtime.operators.GroupReduceDriver.run(GroupReduceDriver.java:131) at org.apache.flink.runtime.operators.BatchTask.run(BatchTask.java:503) at org.apache.flink.runtime.operators.BatchTask.invoke(BatchTask.java:368) at org.apache.flink.runtime.taskmanager.Task.run(Task.java:711) at java.lang.Thread.run(Thread.java:745) Caused by: java.io.IOException: Failed to serialize element. Serialized size (> 1136656562 bytes) exceeds JVM heap space at org.apache.flink.core.memory.DataOutputSerializer.resize(DataOutputSerializer.java:323) at org.apache.flink.core.memory.DataOutputSerializer.write(DataOutputSerializer.java:149) at org.apache.beam.runners.flink.translation.wrappers.DataOutputViewWrapper.write(DataOutputViewWrapper.java:48) at java.io.DataOutputStream.write(DataOutputStream.java:107) at java.io.ObjectOutputStream$BlockDataOutputStream.drain(ObjectOutputStream.java:1877) at java.io.ObjectOutputStream$BlockDataOutputStream.setBlockDataMode(ObjectOutputStream.java:1786) at java.io.ObjectOutputStream.writeNonProxyDesc(ObjectOutputStream.java:1286) at java.io.ObjectOutputStream.writeClassDesc(ObjectOutputStream.java:1231) at java.io.ObjectOutputStream.writeOrdinaryObject(ObjectOutputStream.java:1427) at java.io.ObjectOutputStream.writeObject0(ObjectOutputStream.java:1178) at java.io.ObjectOutputStream.writeFatalException(ObjectOutputStream.java:1577) at java.io.ObjectOutputStream.writeObject(ObjectOutputStream.java:351) at org.apache.beam.sdk.coders.SerializableCoder.encode(SerializableCoder.java:170) at org.apache.beam.sdk.coders.SerializableCoder.encode(SerializableCoder.java:50) at org.apache.beam.sdk.coders.Coder.encode(Coder.java:136) at org.apache.beam.sdk.transforms.join.UnionCoder.encode(UnionCoder.java:71) at org.apache.beam.sdk.transforms.join.UnionCoder.encode(UnionCoder.java:58) at org.apache.beam.sdk.transforms.join.UnionCoder.encode(UnionCoder.java:32) at org.apache.beam.sdk.coders.IterableLikeCoder.encode(IterableLikeCoder.java:98) at org.apache.beam.sdk.coders.IterableLikeCoder.encode(IterableLikeCoder.java:60) at org.apache.beam.sdk.coders.Coder.encode(Coder.java:136) at org.apache.beam.sdk.coders.KvCoder.encode(KvCoder.java:71) at org.apache.beam.sdk.coders.KvCoder.encode(KvCoder.java:36) at org.apache.beam.sdk.util.WindowedValue$FullWindowedValueCoder.encode(WindowedValue.java:529) at org.apache.beam.sdk.util.WindowedValue$FullWindowedValueCoder.encode(WindowedValue.java:520) at org.apache.beam.sdk.util.WindowedValue$FullWindowedValueCoder.encode(WindowedValue.java:480) at org.apache.beam.runners.flink.translation.types.CoderTypeSerializer.serialize(CoderTypeSerializer.java:83) at org.apache.flink.runtime.plugable.SerializationDelegate.write(SerializationDelegate.java:54) at org.apache.flink.runtime.io.network.api.serialization.SpanningRecordSerializer.addRecord(SpanningRecordSerializer.java:88) at org.apache.flink.runtime.io.network.api.writer.RecordWriter.sendToTarget(RecordWriter.java:131) at org.apache.flink.runtime.io.network.api.writer.RecordWriter.emit(RecordWriter.java:107) at org.apache.flink.runtime.operators.shipping.OutputCollector.collect(OutputCollector.java:65) ... 9 more Caused by: java.lang.OutOfMemoryError: Java heap space at
