Re: NPE with Flink Streaming from Kafka

Posted by Stephan Ewen on
URL: http://deprecated-apache-flink-user-mailing-list-archive.369.s1.nabble.com/NPE-with-Flink-Streaming-from-Kafka-tp3821p3860.html

Mihail!

The Flink windows are currently in-memory only. There are plans to relax that, but for the time being, having enough memory in the cluster is important.

@Gyula: I think window state is currently also limited when using the SqlStateBackend, by the size of a row in the database (because windows are not key/value state currently)

Here are some simple rules-of-thumb to work with:

1) For windows, the number of expected keys can be without bound. It is important to have a rough upper bound for the number of "active keys at a certain time". For example, if you have your time windows (let's say by 10 minutes or so), it only matters how many keys you have within each 10 minute interval. Those define how much memory you need.

2) If you work with the "OperatorState" abstraction, then you need to think about cleanup a bit. The OperatorState keeps state currently for as long until you set the state for the key to "null". This manual state is explicitly designed to allow you to keep state across windows and across very long time. On the flip side, you need to manage the amount of state you store, by releasing state for keys.

3) If a certain key space grows infinite, you should "scope the state by time". A pragmatic solution for that is to define a session window:

  - The session length defines after what inactivity the state is cleaned (let's say 1h session length or so)

  - The trigger implements this session (there are a few mails on this list already that explain how to do this) and take care of evaluating on every element.

  - A count(1) evictor makes sure only one element is ever stored

Greetings,

Stephan

On Wed, Dec 2, 2015 at 11:37 AM, Gyula Fóra <[hidden email]> wrote:

Hi, 

I am working on a use case that involves storing state for billions of keys. For this we use a MySql state backend that will write each key-value state to MySql server so it will only hold a limited set of key-value pairs on heap while maintaining the processing guarantees.

This will keep our streaming job from running out of memory as most of the state is off heap. I am not sure if this is relevant to your use case but if the state size grows indefinitely you might want to give it a try.

I will write a detailed guide in some days but if you want to get started check this one out:

https://docs.google.com/document/d/1xx3J88ZR_kuYYHAil3HD3YSBLqBTKV4Pu1SLwhiRiGs/edit?usp=sharing

There are some pending improvements that I will commit in the next days that will increase the performance of the MySql adapter

Let me know if you are interested in this!

Cheers,

Gyula

Vieru, Mihail <[hidden email]> ezt írta (időpont: 2015. dec. 2., Sze, 11:26):

Hi Aljoscha,

we have no upper bound for the number of expected keys. The max size for an element is 1 KB.

There are 2 Maps, a KeyBy, a timeWindow, a Reduce and a writeAsText operators in the job. In the first Map we parse the contained JSON object in each element and forward it as a Flink Tuple. In the Reduce we update the state for each key. That's about it.

Best,

Mihail

2015-12-02 11:09 GMT+01:00 Aljoscha Krettek <[hidden email]>:

Hi Mihail,
could you please give some information about the number of keys that you are expecting in the data and how big the elements are that you are processing in the window.

Also, are there any other operations that could be taxing on Memory. I think the different exception you see for 500MB mem size is just because Java notices that it ran out of memory at a different part in the program.

Cheers,
Aljoscha

> On 02 Dec 2015, at 10:57, Vieru, Mihail <[hidden email]> wrote:
>
> Yes, with the "start-cluster-streaming.sh" script.
> If the TaskManager gets 5GB of heap it manages to process ~100 million messages and then throws the above OOM.
> If it gets only 500MB it manages to process ~8 million and a somewhat misleading exception is thrown:
>
> 12/01/2015 19:14:07    Source: Custom Source -> Map -> Map(1/1) switched to FAILED
> java.lang.Exception: Java heap space
>     at org.apache.flink.streaming.connectors.kafka.internals.LegacyFetcher.run(LegacyFetcher.java:242)
>     at org.apache.flink.streaming.connectors.kafka.FlinkKafkaConsumer.run(FlinkKafkaConsumer.java:399)
>     at org.apache.flink.streaming.api.operators.StreamSource.run(StreamSource.java:58)
>     at org.apache.flink.streaming.runtime.tasks.SourceStreamTask.run(SourceStreamTask.java:55)
>     at org.apache.flink.streaming.runtime.tasks.StreamTask.invoke(StreamTask.java:218)
>     at org.apache.flink.runtime.taskmanager.Task.run(Task.java:584)
>     at java.lang.Thread.run(Thread.java:745)
> Caused by: java.lang.OutOfMemoryError: Java heap space
>     at org.json.simple.parser.Yylex.<init>(Yylex.java:231)
>     at org.json.simple.parser.JSONParser.<init>(JSONParser.java:34)
>     at de.zalando.saiki.gerakanflink.ItemPriceAvgPerOrder$1.map(ItemPriceAvgPerOrder.java:70)
>     at de.zalando.saiki.gerakanflink.ItemPriceAvgPerOrder$1.map(ItemPriceAvgPerOrder.java:65)
>     at org.apache.flink.streaming.api.operators.StreamMap.processElement(StreamMap.java:37)
>     at org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.collect(OperatorChain.java:316)
>     at org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.collect(OperatorChain.java:300)
>     at org.apache.flink.streaming.runtime.io.CollectorWrapper.collect(CollectorWrapper.java:48)
>     at org.apache.flink.streaming.runtime.io.CollectorWrapper.collect(CollectorWrapper.java:29)
>     at org.apache.flink.streaming.runtime.tasks.SourceStreamTask$SourceOutput.collect(SourceStreamTask.java:97)
>     at org.apache.flink.streaming.api.operators.StreamSource$NonTimestampContext.collect(StreamSource.java:92)
>     at org.apache.flink.streaming.connectors.kafka.internals.LegacyFetcher$SimpleConsumerThread.run(LegacyFetcher.java:450)
>
>
>
>
> 2015-12-02 10:45 GMT+01:00 Robert Metzger <[hidden email]>:
> Its good news that the issue has been resolved.
>
> Regarding the OOM, did you start Flink in the streaming mode?
>
> On Wed, Dec 2, 2015 at 10:18 AM, Vieru, Mihail <[hidden email]> wrote:
> Thank you, Robert! The issue with Kafka is now solved with the 0.10-SNAPSHOT dependency.
>
> We have run into an OutOfMemory exception though, which appears to be related to the state. As my colleague, Javier Lopez, mentioned in a previous thread, state handling is crucial for our use case. And as the jobs are intended to run for months, stability plays an important role in choosing a stream processing framework.
>
> 12/02/2015 10:03:53    Fast TumblingTimeWindows(5000) of Reduce at main(ItemPriceAvgPerOrder.java:108) -> Sink: Unnamed(1/1) switched to FAILED
> java.lang.OutOfMemoryError: Java heap space
>     at java.util.HashMap.resize(HashMap.java:703)
>     at java.util.HashMap.putVal(HashMap.java:662)
>     at java.util.HashMap.put(HashMap.java:611)
>     at org.apache.flink.runtime.state.AbstractHeapKvState.update(AbstractHeapKvState.java:98)
>     at de.zalando.saiki.gerakanflink.ItemPriceAvgPerOrder$3.reduce(ItemPriceAvgPerOrder.java:121)
>     at de.zalando.saiki.gerakanflink.ItemPriceAvgPerOrder$3.reduce(ItemPriceAvgPerOrder.java:108)
>     at org.apache.flink.streaming.runtime.operators.windowing.KeyMap.putOrAggregate(KeyMap.java:196)
>     at org.apache.flink.streaming.runtime.operators.windowing.AggregatingKeyedTimePanes.addElementToLatestPane(AggregatingKeyedTimePanes.java:50)
>     at org.apache.flink.streaming.runtime.operators.windowing.AbstractAlignedProcessingTimeWindowOperator.processElement(AbstractAlignedProcessingTimeWindowOperator.java:210)
>     at org.apache.flink.streaming.runtime.io.StreamInputProcessor.processInput(StreamInputProcessor.java:166)
>     at org.apache.flink.streaming.runtime.tasks.OneInputStreamTask.run(OneInputStreamTask.java:63)
>     at org.apache.flink.streaming.runtime.tasks.StreamTask.invoke(StreamTask.java:218)
>     at org.apache.flink.runtime.taskmanager.Task.run(Task.java:584)
>     at java.lang.Thread.run(Thread.java:745)
>
>
>
>
> 2015-12-01 17:42 GMT+01:00 Maximilian Michels <[hidden email]>:
> Thanks! I've linked the issue in JIRA.
>
> On Tue, Dec 1, 2015 at 5:39 PM, Robert Metzger <[hidden email]> wrote:
> > I think its this one https://issues.apache.org/jira/browse/KAFKA-824
> >
> > On Tue, Dec 1, 2015 at 5:37 PM, Maximilian Michels <[hidden email]> wrote:
> >>
> >> I know this has been fixed already but, out of curiosity, could you
> >> point me to the Kafka JIRA issue for this
> >> bug? From the Flink issue it looks like this is a Zookeeper version
> >> mismatch.
> >>
> >> On Tue, Dec 1, 2015 at 5:16 PM, Robert Metzger <[hidden email]>
> >> wrote:
> >> > Hi Gyula,
> >> >
> >> > no, I didn't ;) We still deploy 0.10-SNAPSHOT versions from the
> >> > "release-0.10" branch to Apache's maven snapshot repository.
> >> >
> >> >
> >> > I don't think Mihail's code will run when he's compiling it against
> >> > 1.0-SNAPSHOT.
> >> >
> >> >
> >> > On Tue, Dec 1, 2015 at 5:13 PM, Gyula Fóra <[hidden email]> wrote:
> >> >>
> >> >> Hi,
> >> >>
> >> >> I think Robert meant to write setting the connector dependency to
> >> >> 1.0-SNAPSHOT.
> >> >>
> >> >> Cheers,
> >> >> Gyula
> >> >>
> >> >> Robert Metzger <[hidden email]> ezt írta (időpont: 2015. dec. 1.,
> >> >> K,
> >> >> 17:10):
> >> >>>
> >> >>> Hi Mihail,
> >> >>>
> >> >>> the issue is actually a bug in Kafka. We have a JIRA in Flink for this
> >> >>> as
> >> >>> well: https://issues.apache.org/jira/browse/FLINK-3067
> >> >>>
> >> >>> Sadly, we haven't released a fix for it yet. Flink 0.10.2 will contain
> >> >>> a
> >> >>> fix.
> >> >>>
> >> >>> Since the kafka connector is not contained in the flink binary, you
> >> >>> can
> >> >>> just set the version in your maven pom file to 0.10-SNAPSHOT. Maven
> >> >>> will
> >> >>> then download the code planned for the 0.10-SNAPSHOT release.
> >> >>>
> >> >>> On Tue, Dec 1, 2015 at 4:54 PM, Vieru, Mihail
> >> >>> <[hidden email]>
> >> >>> wrote:
> >> >>>>
> >> >>>> Hi,
> >> >>>>
> >> >>>> we get the following NullPointerException after ~50 minutes when
> >> >>>> running
> >> >>>> a streaming job with windowing and state that reads data from Kafka
> >> >>>> and
> >> >>>> writes the result to local FS.
> >> >>>> There are around 170 million messages to be processed, Flink 0.10.1
> >> >>>> stops at ~8 million.
> >> >>>> Flink runs locally, started with the "start-cluster-streaming.sh"
> >> >>>> script.
> >> >>>>
> >> >>>> 12/01/2015 15:06:24    Job execution switched to status RUNNING.
> >> >>>> 12/01/2015 15:06:24    Source: Custom Source -> Map -> Map(1/1)
> >> >>>> switched
> >> >>>> to SCHEDULED
> >> >>>> 12/01/2015 15:06:24    Source: Custom Source -> Map -> Map(1/1)
> >> >>>> switched
> >> >>>> to DEPLOYING
> >> >>>> 12/01/2015 15:06:24    Fast TumblingTimeWindows(5000) of Reduce at
> >> >>>> main(ItemPriceAvgPerOrder.java:108) -> Sink: Unnamed(1/1) switched to
> >> >>>> SCHEDULED
> >> >>>> 12/01/2015 15:06:24    Fast TumblingTimeWindows(5000) of Reduce at
> >> >>>> main(ItemPriceAvgPerOrder.java:108) -> Sink: Unnamed(1/1) switched to
> >> >>>> DEPLOYING
> >> >>>> 12/01/2015 15:06:24    Source: Custom Source -> Map -> Map(1/1)
> >> >>>> switched
> >> >>>> to RUNNING
> >> >>>> 12/01/2015 15:06:24    Fast TumblingTimeWindows(5000) of Reduce at
> >> >>>> main(ItemPriceAvgPerOrder.java:108) -> Sink: Unnamed(1/1) switched to
> >> >>>> RUNNING
> >> >>>> 12/01/2015 15:56:08    Fast TumblingTimeWindows(5000) of Reduce at
> >> >>>> main(ItemPriceAvgPerOrder.java:108) -> Sink: Unnamed(1/1) switched to
> >> >>>> CANCELED
> >> >>>> 12/01/2015 15:56:08    Source: Custom Source -> Map -> Map(1/1)
> >> >>>> switched
> >> >>>> to FAILED
> >> >>>> java.lang.Exception
> >> >>>>     at
> >> >>>>
> >> >>>> org.apache.flink.streaming.connectors.kafka.internals.LegacyFetcher.run(LegacyFetcher.java:242)
> >> >>>>     at
> >> >>>>
> >> >>>> org.apache.flink.streaming.connectors.kafka.FlinkKafkaConsumer.run(FlinkKafkaConsumer.java:397)
> >> >>>>     at
> >> >>>>
> >> >>>> org.apache.flink.streaming.api.operators.StreamSource.run(StreamSource.java:58)
> >> >>>>     at
> >> >>>>
> >> >>>> org.apache.flink.streaming.runtime.tasks.SourceStreamTask.run(SourceStreamTask.java:55)
> >> >>>>     at
> >> >>>>
> >> >>>> org.apache.flink.streaming.runtime.tasks.StreamTask.invoke(StreamTask.java:218)
> >> >>>>     at org.apache.flink.runtime.taskmanager.Task.run(Task.java:584)
> >> >>>>     at java.lang.Thread.run(Thread.java:745)
> >> >>>> Caused by: java.lang.NullPointerException
> >> >>>>     at
> >> >>>>
> >> >>>> org.I0Itec.zkclient.ZkConnection.writeDataReturnStat(ZkConnection.java:115)
> >> >>>>     at org.I0Itec.zkclient.ZkClient$10.call(ZkClient.java:817)
> >> >>>>     at
> >> >>>> org.I0Itec.zkclient.ZkClient.retryUntilConnected(ZkClient.java:675)
> >> >>>>     at
> >> >>>> org.I0Itec.zkclient.ZkClient.writeDataReturnStat(ZkClient.java:813)
> >> >>>>     at org.I0Itec.zkclient.ZkClient.writeData(ZkClient.java:808)
> >> >>>>     at org.I0Itec.zkclient.ZkClient.writeData(ZkClient.java:777)
> >> >>>>     at kafka.utils.ZkUtils$.updatePersistentPath(ZkUtils.scala:332)
> >> >>>>     at kafka.utils.ZkUtils.updatePersistentPath(ZkUtils.scala)
> >> >>>>     at
> >> >>>>
> >> >>>> org.apache.flink.streaming.connectors.kafka.internals.ZookeeperOffsetHandler.setOffsetInZooKeeper(ZookeeperOffsetHandler.java:112)
> >> >>>>     at
> >> >>>>
> >> >>>> org.apache.flink.streaming.connectors.kafka.internals.ZookeeperOffsetHandler.commit(ZookeeperOffsetHandler.java:80)
> >> >>>>     at
> >> >>>>
> >> >>>> org.apache.flink.streaming.connectors.kafka.FlinkKafkaConsumer$PeriodicOffsetCommitter.run(FlinkKafkaConsumer.java:632)
> >> >>>>
> >> >>>>
> >> >>>> Any ideas on what could cause this behaviour?
> >> >>>>
> >> >>>> Best,
> >> >>>> Mihail
> >> >>>
> >> >>>
> >> >
> >
> >
>
>
>