spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Nicolas Phung <nicolas.ph...@gmail.com>
Subject Re: Resume checkpoint failed with Spark Streaming Kafka via createDirectStream under heavy reprocessing
Date Wed, 29 Jul 2015 12:09:13 GMT
Hello,

I'm using 4Gb for the driver memory. The checkpoint is between 1 Gb and 10
Gb depending if I'm reprocessing all the data from beginning or just
getting the latest offset from the real time processed. Is there any best
practice to be aware of with driver memory relating to checkpoint size ?

Regards,
Nicolas P.

On Tue, Jul 28, 2015 at 4:42 PM, Cody Koeninger <cody@koeninger.org> wrote:

> That stacktrace looks like an out of heap space on the driver while
> writing checkpoint, not on the worker nodes.  How much memory are you
> giving the driver?  How big are your stored checkpoints?
>
> On Tue, Jul 28, 2015 at 9:30 AM, Nicolas Phung <nicolas.phung@gmail.com>
> wrote:
>
>> Hi,
>>
>> After using KafkaUtils.createDirectStream[Object, Object,
>> KafkaAvroDecoder, KafkaAvroDecoder, Option[AnalyticEventEnriched]](ssc,
>> kafkaParams, map, messageHandler), I'm encountering the following issue:
>>
>> 15/07/28 00:29:57 ERROR actor.ActorSystemImpl: Uncaught fatal error from
>> thread [sparkDriver-akka.actor.default-dispatcher-24] shutting down
>> ActorSystem [sparkDriver]
>> java.lang.OutOfMemoryError: Java heap space
>>     at
>> java.io.ObjectOutputStream$HandleTable.growEntries(ObjectOutputStream.java:2351)
>>     at
>> java.io.ObjectOutputStream$HandleTable.assign(ObjectOutputStream.java:2276)
>>     at
>> java.io.ObjectOutputStream.writeOrdinaryObject(ObjectOutputStream.java:1428)
>>     at
>> java.io.ObjectOutputStream.writeObject0(ObjectOutputStream.java:1178)
>>     at
>> java.io.ObjectOutputStream.defaultWriteFields(ObjectOutputStream.java:1548)
>>     at
>> java.io.ObjectOutputStream.writeSerialData(ObjectOutputStream.java:1509)
>>     at
>> java.io.ObjectOutputStream.writeOrdinaryObject(ObjectOutputStream.java:1432)
>>     at
>> java.io.ObjectOutputStream.writeObject0(ObjectOutputStream.java:1178)
>>     at java.io.ObjectOutputStream.writeArray(ObjectOutputStream.java:1378)
>>     at
>> java.io.ObjectOutputStream.writeObject0(ObjectOutputStream.java:1174)
>>     at java.io.ObjectOutputStream.writeObject(ObjectOutputStream.java:348)
>>     at
>> scala.collection.mutable.HashMap$$anonfun$writeObject$1.apply(HashMap.scala:137)
>>     at
>> scala.collection.mutable.HashMap$$anonfun$writeObject$1.apply(HashMap.scala:135)
>>     at
>> scala.collection.mutable.HashTable$class.foreachEntry(HashTable.scala:226)
>>     at scala.collection.mutable.HashMap.foreachEntry(HashMap.scala:39)
>>     at
>> scala.collection.mutable.HashTable$class.serializeTo(HashTable.scala:124)
>>     at scala.collection.mutable.HashMap.serializeTo(HashMap.scala:39)
>>     at scala.collection.mutable.HashMap.writeObject(HashMap.scala:135)
>>     at sun.reflect.GeneratedMethodAccessor22.invoke(Unknown Source)
>>     at
>> sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
>>     at java.lang.reflect.Method.invoke(Method.java:483)
>>     at
>> java.io.ObjectStreamClass.invokeWriteObject(ObjectStreamClass.java:988)
>>     at
>> java.io.ObjectOutputStream.writeSerialData(ObjectOutputStream.java:1496)
>>     at
>> java.io.ObjectOutputStream.writeOrdinaryObject(ObjectOutputStream.java:1432)
>>     at
>> java.io.ObjectOutputStream.writeObject0(ObjectOutputStream.java:1178)
>>     at
>> java.io.ObjectOutputStream.defaultWriteFields(ObjectOutputStream.java:1548)
>>     at
>> java.io.ObjectOutputStream.defaultWriteObject(ObjectOutputStream.java:441)
>>     at
>> org.apache.spark.streaming.dstream.DStreamCheckpointData$$anonfun$writeObject$1.apply$mcV$sp(DStreamCheckpointData.scala:128)
>>     at org.apache.spark.util.Utils$.tryOrIOException(Utils.scala:1137)
>>     at
>> org.apache.spark.streaming.dstream.DStreamCheckpointData.writeObject(DStreamCheckpointData.scala:123)
>>     at sun.reflect.GeneratedMethodAccessor21.invoke(Unknown Source)
>>     at
>> sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
>>
>> I don't know why, after that, it's eating all the CPU on one of the node
>> till the entire job stopped. It tries to resume from checkpoint several
>> times but failed with this error too. I think I have enough spared memory
>> with 4 nodes with 24 Gb per nodes. It has processed successfully around 40
>> gb before that and looking into storage in Spark UI, I don't have a big rdd
>> stored in memory/disk. I notice on this node, there's an increase in
>> connection to kafka that are not closed too.
>>
>> Regards,
>> Nicolas P.
>>
>> On Fri, Jul 24, 2015 at 3:32 PM, Cody Koeninger <cody@koeninger.org>
>> wrote:
>>
>>> It's really a question of whether you need access to the
>>> MessageAndMetadata, or just the key / value from the message.
>>>
>>> If you just need the key/value, dstream map is fine.
>>>
>>> In your case, since you need to be able to control a possible failure
>>> when deserializing the message from the MessageAndMetadata, I'd just go
>>> ahead and do the work in the messageHandler.
>>>
>>> On Fri, Jul 24, 2015 at 2:46 AM, Nicolas Phung <nicolas.phung@gmail.com>
>>> wrote:
>>>
>>>> Hello,
>>>>
>>>> I manage to read all my data back with skipping offset that contains a
>>>> corrupt message. I have one more question regarding messageHandler method
>>>> vs dstream.foreachRDD.map vs dstream.map.foreachRDD best practices. I'm
>>>> using a function to read the serialized message from kafka and convert it
>>>> into my appropriate object with some enrichments and sometimes add filter
>>>> after that. Where's the best spot to put this logic inside messageHandler
>>>> method (convert each message within this handler) or dstream.foreachRDD.map
>>>> (map rdd) or dstream.map.foreachRDD (map dstream) ?
>>>>
>>>> Thank you for your help Cody.
>>>> Regards,
>>>> Nicolas PHUNG
>>>>
>>>> On Tue, Jul 21, 2015 at 4:53 PM, Cody Koeninger <cody@koeninger.org>
>>>> wrote:
>>>>
>>>>> Yeah, I'm referring to that api.
>>>>>
>>>>> If you want to filter messages in addition to catching that exception,
>>>>> have your mesageHandler return an option, so the type R would end up
being
>>>>> Option[WhateverYourClassIs], then filter out None before doing the rest
of
>>>>> your processing.
>>>>>
>>>>> If you aren't already recording offsets somewhere, and need to find
>>>>> the offsets at the beginning of the topic, you can take a look at this
>>>>>
>>>>>
>>>>> https://github.com/apache/spark/blob/branch-1.3/external/kafka/src/main/scala/org/apache/spark/streaming/kafka/KafkaCluster.scala#L143
>>>>>
>>>>> as an example of querying offsets from Kafka.
>>>>>
>>>>> That code is private, but you can either use it as an example, or
>>>>> remove the private[spark] and recompile just the spark-streaming-kafka
>>>>> package.  That artifact is included in your job assembly, so you won't
have
>>>>> to redeploy spark if you go that route.
>>>>>
>>>>>
>>>>> On Tue, Jul 21, 2015 at 6:42 AM, Nicolas Phung <
>>>>> nicolas.phung@gmail.com> wrote:
>>>>>
>>>>>> Hi Cody,
>>>>>>
>>>>>> Thanks for your answer. I'm with Spark 1.3.0. I don't quite
>>>>>> understand how to use the messageHandler parameter/function in the
>>>>>> createDirectStream method. You are referring to this, aren't you
?
>>>>>>
>>>>>> def createDirectStream[ K: ClassTag, V: ClassTag, KD <: Decoder[K]:
>>>>>> ClassTag, VD <: Decoder[V]: ClassTag, R: ClassTag] ( ssc:
>>>>>> StreamingContext, kafkaParams: Map[String, String], fromOffsets:
Map[
>>>>>> TopicAndPartition, Long], messageHandler: MessageAndMetadata[K, V]
=>
>>>>>> R ): InputDStream[R] = { new DirectKafkaInputDStream[K, V, KD, VD,
R
>>>>>> ]( ssc, kafkaParams, fromOffsets, messageHandler) }
>>>>>>
>>>>>> So, I must supply the fromOffsets parameter too, but how do I tell
>>>>>> this method to read from the beginning of my topic ?
>>>>>>
>>>>>> If I have a filter (e.g. a R.date field) on my R class, I can put
a
>>>>>> filter in the messageHandler function too ?
>>>>>>
>>>>>> Regards,
>>>>>> Nicolas P.
>>>>>>
>>>>>> On Mon, Jul 20, 2015 at 7:08 PM, Cody Koeninger <cody@koeninger.org>
>>>>>> wrote:
>>>>>>
>>>>>>> Yeah, in the function you supply for the messageHandler parameter
to
>>>>>>> createDirectStream, catch the exception and do whatever makes
sense for
>>>>>>> your application.
>>>>>>>
>>>>>>> On Mon, Jul 20, 2015 at 11:58 AM, Nicolas Phung <
>>>>>>> nicolas.phung@gmail.com> wrote:
>>>>>>>
>>>>>>>> Hello,
>>>>>>>>
>>>>>>>> Using the old Spark Streaming Kafka API, I got the following
around
>>>>>>>> the same offset:
>>>>>>>>
>>>>>>>> kafka.message.InvalidMessageException: Message is corrupt
(stored
>>>>>>>> crc = 3561357254, computed crc = 171652633)
>>>>>>>>         at kafka.message.Message.ensureValid(Message.scala:166)
>>>>>>>>         at
>>>>>>>> kafka.consumer.ConsumerIterator.makeNext(ConsumerIterator.scala:102)
>>>>>>>>         at
>>>>>>>> kafka.consumer.ConsumerIterator.makeNext(ConsumerIterator.scala:33)
>>>>>>>>         at
>>>>>>>> kafka.utils.IteratorTemplate.maybeComputeNext(IteratorTemplate.scala:66)
>>>>>>>>         at
>>>>>>>> kafka.utils.IteratorTemplate.hasNext(IteratorTemplate.scala:58)
>>>>>>>>         at
>>>>>>>> org.apache.spark.streaming.kafka.ReliableKafkaReceiver$MessageHandler.run(ReliableKafkaReceiver.scala:265)
>>>>>>>>         at
>>>>>>>> java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:511)
>>>>>>>>         at java.util.concurrent.FutureTask.run(FutureTask.java:266)
>>>>>>>>         at
>>>>>>>> java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
>>>>>>>>         at
>>>>>>>> java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
>>>>>>>>         at java.lang.Thread.run(Thread.java:745)
>>>>>>>> 15/07/20 15:56:57 INFO BlockManager: Removing broadcast 4641
>>>>>>>> 15/07/20 15:56:57 ERROR ReliableKafkaReceiver: Error handling
>>>>>>>> message
>>>>>>>> java.lang.IllegalStateException: Iterator is in failed state
>>>>>>>>         at
>>>>>>>> kafka.utils.IteratorTemplate.hasNext(IteratorTemplate.scala:54)
>>>>>>>>         at
>>>>>>>> org.apache.spark.streaming.kafka.ReliableKafkaReceiver$MessageHandler.run(ReliableKafkaReceiver.scala:265)
>>>>>>>>         at
>>>>>>>> java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:511)
>>>>>>>>         at java.util.concurrent.FutureTask.run(FutureTask.java:266)
>>>>>>>>         at
>>>>>>>> java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
>>>>>>>>         at
>>>>>>>> java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
>>>>>>>>         at java.lang.Thread.run(Thread.java:745)
>>>>>>>>
>>>>>>>> I found some old topic about some possible corrupt Kafka
message
>>>>>>>> produced by the new producer API with Snappy compression
on. My question
>>>>>>>> is, is it possible to skip/ignore those offsets when full
processing with
>>>>>>>> KafkaUtils.createStream or KafkaUtils.createDirectStream
?
>>>>>>>>
>>>>>>>> Regards,
>>>>>>>> Nicolas PHUNG
>>>>>>>>
>>>>>>>> On Mon, Jul 20, 2015 at 3:46 PM, Cody Koeninger <cody@koeninger.org
>>>>>>>> > wrote:
>>>>>>>>
>>>>>>>>> I'd try logging the offsets for each message, see where
problems
>>>>>>>>> start, then try using the console consumer starting at
those offsets and
>>>>>>>>> see if you can reproduce the problem.
>>>>>>>>>
>>>>>>>>> On Mon, Jul 20, 2015 at 2:15 AM, Nicolas Phung <
>>>>>>>>> nicolas.phung@gmail.com> wrote:
>>>>>>>>>
>>>>>>>>>> Hi Cody,
>>>>>>>>>>
>>>>>>>>>> Thanks for you help. It seems there's something wrong
with some
>>>>>>>>>> messages within my Kafka topics then. I don't understand
how, I can get
>>>>>>>>>> bigger or incomplete message since I use default
configuration to accept
>>>>>>>>>> only 1Mb message in my Kafka topic. If you have any
others informations or
>>>>>>>>>> suggestions, please tell me.
>>>>>>>>>>
>>>>>>>>>> Regards,
>>>>>>>>>> Nicolas PHUNG
>>>>>>>>>>
>>>>>>>>>> On Thu, Jul 16, 2015 at 7:08 PM, Cody Koeninger <
>>>>>>>>>> cody@koeninger.org> wrote:
>>>>>>>>>>
>>>>>>>>>>> Not exactly the same issue, but possibly related:
>>>>>>>>>>>
>>>>>>>>>>> https://issues.apache.org/jira/browse/KAFKA-1196
>>>>>>>>>>>
>>>>>>>>>>> On Thu, Jul 16, 2015 at 12:03 PM, Cody Koeninger
<
>>>>>>>>>>> cody@koeninger.org> wrote:
>>>>>>>>>>>
>>>>>>>>>>>> Well, working backwards down the stack trace...
>>>>>>>>>>>>
>>>>>>>>>>>> at java.nio.Buffer.limit(Buffer.java:275)
>>>>>>>>>>>>
>>>>>>>>>>>> That exception gets thrown if the limit is
negative or greater than the buffer's capacity
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>> at kafka.message.Message.sliceDelimited(Message.scala:236)
>>>>>>>>>>>>
>>>>>>>>>>>> If size had been negative, it would have
just returned null, so
>>>>>>>>>>>> we know the exception got thrown because
the size was greater than the
>>>>>>>>>>>> buffer's capacity
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>> I haven't seen that before... maybe a corrupted
message of some
>>>>>>>>>>>> kind?
>>>>>>>>>>>>
>>>>>>>>>>>> If that problem is reproducible, try providing
an explicit
>>>>>>>>>>>> argument for messageHandler, with a function
that logs the message offset.
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>> On Thu, Jul 16, 2015 at 11:28 AM, Nicolas
Phung <
>>>>>>>>>>>> nicolas.phung@gmail.com> wrote:
>>>>>>>>>>>>
>>>>>>>>>>>>> Hello,
>>>>>>>>>>>>>
>>>>>>>>>>>>> When I'm reprocessing the data from kafka
(about 40 Gb) with the new Spark Streaming Kafka method createDirectStream, everything is
fine till a driver error happened (driver is killed, connection lost...). When the driver
pops up again, it resumes the processing with the checkpoint in HDFS. Except, I got this:
>>>>>>>>>>>>>
>>>>>>>>>>>>> 15/07/16 15:23:41 ERROR TaskSetManager:
Task 4 in stage 4.0 failed 4 times; aborting job
>>>>>>>>>>>>> 15/07/16 15:23:41 ERROR JobScheduler:
Error running job streaming job 1437032118000 ms.0
>>>>>>>>>>>>> org.apache.spark.SparkException: Job
aborted due to stage failure: Task 4 in stage 4.0 failed 4 times, most recent failure: Lost
task 4.3 in stage 4.0 (TID 16, slave05.local): java.lang.IllegalArgumentException
>>>>>>>>>>>>> 	at java.nio.Buffer.limit(Buffer.java:275)
>>>>>>>>>>>>> 	at kafka.message.Message.sliceDelimited(Message.scala:236)
>>>>>>>>>>>>> 	at kafka.message.Message.payload(Message.scala:218)
>>>>>>>>>>>>> 	at kafka.message.MessageAndMetadata.message(MessageAndMetadata.scala:32)
>>>>>>>>>>>>> 	at org.apache.spark.streaming.kafka.KafkaUtils$$anonfun$6.apply(KafkaUtils.scala:395)
>>>>>>>>>>>>> 	at org.apache.spark.streaming.kafka.KafkaUtils$$anonfun$6.apply(KafkaUtils.scala:395)
>>>>>>>>>>>>> 	at org.apache.spark.streaming.kafka.KafkaRDD$KafkaRDDIterator.getNext(KafkaRDD.scala:176)
>>>>>>>>>>>>> 	at org.apache.spark.util.NextIterator.hasNext(NextIterator.scala:71)
>>>>>>>>>>>>> 	at scala.collection.Iterator$$anon$11.hasNext(Iterator.scala:327)
>>>>>>>>>>>>> 	at scala.collection.Iterator$$anon$11.hasNext(Iterator.scala:327)
>>>>>>>>>>>>> 	at scala.collection.Iterator$$anon$11.hasNext(Iterator.scala:327)
>>>>>>>>>>>>> 	at org.apache.spark.storage.MemoryStore.unrollSafely(MemoryStore.scala:248)
>>>>>>>>>>>>> 	at org.apache.spark.CacheManager.putInBlockManager(CacheManager.scala:172)
>>>>>>>>>>>>> 	at org.apache.spark.CacheManager.getOrCompute(CacheManager.scala:79)
>>>>>>>>>>>>> 	at org.apache.spark.rdd.RDD.iterator(RDD.scala:242)
>>>>>>>>>>>>> 	at org.apache.spark.rdd.MapPartitionsRDD.compute(MapPartitionsRDD.scala:35)
>>>>>>>>>>>>> 	at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:277)
>>>>>>>>>>>>> 	at org.apache.spark.rdd.RDD.iterator(RDD.scala:244)
>>>>>>>>>>>>> 	at org.apache.spark.rdd.CoalescedRDD$$anonfun$compute$1.apply(CoalescedRDD.scala:93)
>>>>>>>>>>>>> 	at org.apache.spark.rdd.CoalescedRDD$$anonfun$compute$1.apply(CoalescedRDD.scala:92)
>>>>>>>>>>>>> 	at scala.collection.Iterator$$anon$13.hasNext(Iterator.scala:371)
>>>>>>>>>>>>> 	at scala.collection.Iterator$$anon$11.hasNext(Iterator.scala:327)
>>>>>>>>>>>>> 	at org.elasticsearch.spark.rdd.EsRDDWriter.write(EsRDDWriter.scala:48)
>>>>>>>>>>>>> 	at org.elasticsearch.spark.rdd.EsSpark$$anonfun$saveToEs$1.apply(EsSpark.scala:67)
>>>>>>>>>>>>> 	at org.elasticsearch.spark.rdd.EsSpark$$anonfun$saveToEs$1.apply(EsSpark.scala:67)
>>>>>>>>>>>>> 	at org.apache.spark.scheduler.ResultTask.runTask(ResultTask.scala:61)
>>>>>>>>>>>>> 	at org.apache.spark.scheduler.Task.run(Task.scala:64)
>>>>>>>>>>>>> 	at org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:203)
>>>>>>>>>>>>> 	at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
>>>>>>>>>>>>> 	at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
>>>>>>>>>>>>> 	at java.lang.Thread.run(Thread.java:745)
>>>>>>>>>>>>>
>>>>>>>>>>>>> This is happening only when I'm doing
a full data processing
>>>>>>>>>>>>> from Kafka. If there's no load, when
you killed the driver and then
>>>>>>>>>>>>> restart, it resumes the checkpoint as
expected without missing data. Did
>>>>>>>>>>>>> someone encounters something similar
? How did you solve this ?
>>>>>>>>>>>>>
>>>>>>>>>>>>> Regards,
>>>>>>>>>>>>>
>>>>>>>>>>>>> Nicolas PHUNG
>>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>

Mime
View raw message