How are you iterating through your RDDs in parallel? In the past (Spark 1.5.2) when I've had actions being performed on multiple RDDs concurrently using futures, I've encountered some pretty bad behavior in Spark, especially during job retries. Very difficult to explain things, like records from one RDD leaking into a totally different (though shared lineage) RDD during job retries. I'm not sure what documentation exists around parallelizing on top of Spark's existing parallelization approach, but I would venture a guess that that could be the source of your concurrent access problems, and potentially other hidden issues. Have you tried a version of your app which doesn't parallelize actions on RDDs, but instead serially processes each RDD? I'm sure it isn't ideal performance-wise, but it seems like a good choice for an A/B test.
The poll.ms issue could very well be settings or capability of your kafka cluster. I think other (non-Spark) approaches may have less consumer churn and be less susceptible to things like GC pauses or coordination latency. It could also be that the number of consumers being simultaneously created on each executor causes a thundering herd problem during initial phases (which then causes job retries, which then causes more consumer churn, etc.). Sean > On Nov 12, 2016, at 11:14 AM, Ivan von Nagy <i...@vadio.com> wrote: > > The code was changed to use a unique group for each KafkaRDD that was created > (see Nov 10 post). There is no KafkaRDD being reused. The basic logic (see > Nov 10 post for example) is get a list of channels, iterate through them in > parallel, load a KafkaRDD using a given topic and a consumer group that is > made from the topic (each RDD uses a different topic and group), process the > data and write to Parquet files. > > Per my Nov 10th post, we still get polling timeouts unless the poll.ms > <http://poll.ms/> is set to something like 10 seconds. We also get concurrent > modification exceptions as well. I believe the key here is the processing of > data in parallel is where we encounter issues so we are looking for some > possible answers surrounding this. > > Thanks, > > Ivan > > > On Fri, Nov 11, 2016 at 12:12 PM, Cody Koeninger <c...@koeninger.org > <mailto:c...@koeninger.org>> wrote: > It is already documented that you must use a different group id, which as far > as I can tell you are still not doing. > > On Nov 10, 2016 7:43 PM, "Shixiong(Ryan) Zhu" <shixi...@databricks.com > <mailto:shixi...@databricks.com>> wrote: > Yeah, the KafkaRDD cannot be reused. It's better to document it. > > On Thu, Nov 10, 2016 at 8:26 AM, Ivan von Nagy <i...@vadio.com > <mailto:i...@vadio.com>> wrote: > Ok, I have split he KafkaRDD logic to each use their own group and bumped the > poll.ms <http://poll.ms/> to 10 seconds. Anything less then 2 seconds on the > poll.ms <http://poll.ms/> ends up with a timeout and exception so I am still > perplexed on that one. The new error I am getting now is a > `ConcurrentModificationException` when Spark is trying to remove the > CachedKafkaConsumer. > > java.util.ConcurrentModificationException: KafkaConsumer is not safe for > multi-threaded access > at > org.apache.kafka.clients.consumer.KafkaConsumer.acquire(KafkaConsumer.java:1431) > at > org.apache.kafka.clients.consumer.KafkaConsumer.close(KafkaConsumer.java:1361) > at > org.apache.spark.streaming.kafka010.CachedKafkaConsumer$$anon$1.removeEldestEntry(CachedKafkaConsumer.scala:128) > at java.util.LinkedHashMap.afterNodeInsertion(LinkedHashMap.java:299) > > > Here is the basic logic: > > Using KafkaRDD - This takes a list of channels and processes them in parallel > using the KafkaRDD directly. They each use a distinct consumer group > (s"$prefix-$topic"), and each has it's own topic and each topic has 4 > partitions. We routinely get timeout errors when polling for data when the > poll.ms <http://poll.ms/> is less then 2 seconds. This occurs whether we > process in parallel. > > Example usage with KafkaRDD: > val channels = Seq("channel1", "channel2") > > channels.toParArray.foreach { channel => > val consumer = new KafkaConsumer[K, V](kafkaParams.asJava) > > // Get offsets for the given topic and the consumer group "$prefix-$topic" > val offsetRanges = getOffsets(s"$prefix-$topic", channel) > > val ds = KafkaUtils.createRDD[K, V](context, > kafkaParams asJava, > offsetRanges, > PreferConsistent).toDS[V] > > // Do some aggregations > ds.agg(...) > // Save the data > ds.write.mode(SaveMode.Append).parquet(somePath) > // Save offsets using a KafkaConsumer > consumer.commitSync(newOffsets.asJava) > consumer.close() > } > > I am not sure why the concurrent issue is there as I have tried to debug and > also looked at the KafkaConsumer code as well, but everything looks like it > should not occur. The things to figure out is why when running in parallel > does this occur and also why the timeouts still occur. > > Thanks, > > Ivan > > On Mon, Nov 7, 2016 at 11:55 AM, Cody Koeninger <c...@koeninger.org > <mailto:c...@koeninger.org>> wrote: > There definitely is Kafka documentation indicating that you should use > a different consumer group for logically different subscribers, this > is really basic to Kafka: > > http://kafka.apache.org/documentation#intro_consumers > <http://kafka.apache.org/documentation#intro_consumers> > > As for your comment that "commit async after each RDD, which is not > really viable also", how is it not viable? Again, committing offsets > to Kafka doesn't give you reliable delivery semantics unless your > downstream data store is idempotent. If your downstream data store is > idempotent, then it shouldn't matter to you when offset commits > happen, as long as they happen within a reasonable time after the data > is written. > > Do you want to keep arguing with me, or follow my advice and proceed > with debugging any remaining issues after you make the changes I > suggested? > > On Mon, Nov 7, 2016 at 1:35 PM, Ivan von Nagy <i...@vadio.com > <mailto:i...@vadio.com>> wrote: > > With our stream version, we update the offsets for only the partition we > > operating on. We even break down the partition into smaller batches and then > > update the offsets after each batch within the partition. With Spark 1.6 and > > Kafka 0.8.x this was not an issue, and as Sean pointed out, this is not > > necessarily a Spark issue since Kafka no longer allows you to simply update > > the offsets for a given consumer group. You have to subscribe or assign > > partitions to even do so. > > > > As for storing the offsets in some other place like a DB, it don't find this > > useful because you then can't use tools like Kafka Manager. In order to do > > so you would have to store in a DB and the circle back and update Kafka > > afterwards. This means you have to keep two sources in sync which is not > > really a good idea. > > > > It is a challenge in Spark to use the Kafka offsets since the drive keeps > > subscribed to the topic(s) and consumer group, while the executors prepend > > "spark-executor-" to the consumer group. The stream (driver) does allow you > > to commit async after each RDD, which is not really viable also. I have not > > of implementing an Akka actor system on the driver and send it messages from > > the executor code to update the offsets, but then that is asynchronous as > > well so not really a good solution. > > > > I have no idea why Kafka made this change and also why in the parallel > > KafkaRDD application we would be advised to use different consumer groups > > for each RDD. That seems strange to me that different consumer groups would > > be required or advised. There is no Kafka documentation that I know if that > > states this. The biggest issue I see with the parallel KafkaRDD is the > > timeouts. I have tried to set poll.ms <http://poll.ms/> to 30 seconds and > > still get the issue. > > Something is not right here and just not seem right. As I mentioned with the > > streaming application, with Spark 1.6 and Kafka 0.8.x we never saw this > > issue. We have been running the same basic logic for over a year now without > > one hitch at all. > > > > Ivan > > > > > > On Mon, Nov 7, 2016 at 11:16 AM, Cody Koeninger <c...@koeninger.org > > <mailto:c...@koeninger.org>> wrote: > >> > >> Someone can correct me, but I'm pretty sure Spark dstreams (in > >> general, not just kafka) have been progressing on to the next batch > >> after a given batch aborts for quite some time now. Yet another > >> reason I put offsets in my database transactionally. My jobs throw > >> exceptions if the offset in the DB isn't what I expected it to be. > >> > >> > >> > >> > >> On Mon, Nov 7, 2016 at 1:08 PM, Sean McKibben <grap...@graphex.com > >> <mailto:grap...@graphex.com>> wrote: > >> > I've been encountering the same kinds of timeout issues as Ivan, using > >> > the "Kafka Stream" approach that he is using, except I'm storing my > >> > offsets > >> > manually from the driver to Zookeeper in the Kafka 8 format. I haven't > >> > yet > >> > implemented the KafkaRDD approach, and therefore don't have the > >> > concurrency > >> > issues, but a very similar use case is coming up for me soon, it's just > >> > been > >> > backburnered until I can get streaming to be more reliable (I will > >> > definitely ensure unique group IDs when I do). Offset commits are > >> > certainly > >> > more painful in Kafka 0.10, and that doesn't have anything to do with > >> > Spark. > >> > > >> > While i may be able to alleviate the timeout by just increasing it, I've > >> > noticed something else that is more worrying: When one task fails 4 > >> > times in > >> > a row (i.e. "Failed to get records for _ after polling for _"), Spark > >> > aborts > >> > the Stage and Job with "Job aborted due to stage failure: Task _ in > >> > stage _ > >> > failed 4 times". That's fine, and it's the behavior I want, but instead > >> > of > >> > stopping the Application there (as previous versions of Spark did) the > >> > next > >> > microbatch marches on and offsets are committed ahead of the failed > >> > microbatch. Suddenly my at-least-once app becomes more > >> > sometimes-at-least-once which is no good. In order for spark to display > >> > that > >> > failure, I must be propagating the errors up to Spark, but the behavior > >> > of > >> > marching forward with the next microbatch seems to be new, and a big > >> > potential for data loss in streaming applications. > >> > > >> > Am I perhaps missing a setting to stop the entire streaming application > >> > once spark.task.maxFailures is reached? Has anyone else seen this > >> > behavior > >> > of a streaming application skipping over failed microbatches? > >> > > >> > Thanks, > >> > Sean > >> > > >> > > >> >> On Nov 4, 2016, at 2:48 PM, Cody Koeninger <c...@koeninger.org > >> >> <mailto:c...@koeninger.org>> wrote: > >> >> > >> >> So basically what I am saying is > >> >> > >> >> - increase poll.ms <http://poll.ms/> > >> >> - use a separate group id everywhere > >> >> - stop committing offsets under the covers > >> >> > >> >> That should eliminate all of those as possible causes, and then we can > >> >> see if there are still issues. > >> >> > >> >> As far as 0.8 vs 0.10, Spark doesn't require you to assign or > >> >> subscribe to a topic in order to update offsets, Kafka does. If you > >> >> don't like the new Kafka consumer api, the existing 0.8 simple > >> >> consumer api should be usable with later brokers. As long as you > >> >> don't need SSL or dynamic subscriptions, and it meets your needs, keep > >> >> using it. > >> >> > >> >> On Fri, Nov 4, 2016 at 3:37 PM, Ivan von Nagy <i...@vadio.com > >> >> <mailto:i...@vadio.com>> wrote: > >> >>> Yes, the parallel KafkaRDD uses the same consumer group, but each RDD > >> >>> uses a > >> >>> single distinct topic. For example, the group would be something like > >> >>> "storage-group", and the topics would be "storage-channel1", and > >> >>> "storage-channel2". In each thread a KafkaConsumer is started, > >> >>> assigned the > >> >>> partitions assigned, and then commit offsets are called after the RDD > >> >>> is > >> >>> processed. This should not interfere with the consumer group used by > >> >>> the > >> >>> executors which would be "spark-executor-storage-group". > >> >>> > >> >>> In the streaming example there is a single topic ("client-events") and > >> >>> group > >> >>> ("processing-group"). A single stream is created and offsets are > >> >>> manually > >> >>> updated from the executor after each partition is handled. This was a > >> >>> challenge since Spark now requires one to assign or subscribe to a > >> >>> topic in > >> >>> order to even update the offsets. In 0.8.2.x you did not have to worry > >> >>> about > >> >>> that. This approach limits your exposure to duplicate data since > >> >>> idempotent > >> >>> records are not entirely possible in our scenario. At least without a > >> >>> lot of > >> >>> re-running of logic to de-dup. > >> >>> > >> >>> Thanks, > >> >>> > >> >>> Ivan > >> >>> > >> >>> On Fri, Nov 4, 2016 at 1:24 PM, Cody Koeninger <c...@koeninger.org > >> >>> <mailto:c...@koeninger.org>> > >> >>> wrote: > >> >>>> > >> >>>> So just to be clear, the answers to my questions are > >> >>>> > >> >>>> - you are not using different group ids, you're using the same group > >> >>>> id everywhere > >> >>>> > >> >>>> - you are committing offsets manually > >> >>>> > >> >>>> Right? > >> >>>> > >> >>>> If you want to eliminate network or kafka misbehavior as a source, > >> >>>> tune poll.ms <http://poll.ms/> upwards even higher. > >> >>>> > >> >>>> You must use different group ids for different rdds or streams. > >> >>>> Kafka consumers won't behave the way you expect if they are all in > >> >>>> the > >> >>>> same group id, and the consumer cache is keyed by group id. Yes, the > >> >>>> executor will tack "spark-executor-" on to the beginning, but if you > >> >>>> give it the same base group id, it will be the same. And the driver > >> >>>> will use the group id you gave it, unmodified. > >> >>>> > >> >>>> Finally, I really can't help you if you're manually writing your own > >> >>>> code to commit offsets directly to Kafka. Trying to minimize > >> >>>> duplicates that way doesn't really make sense, your system must be > >> >>>> able to handle duplicates if you're using kafka as an offsets store, > >> >>>> it can't do transactional exactly once. > >> >>>> > >> >>>> On Fri, Nov 4, 2016 at 1:48 PM, vonnagy <i...@vadio.com > >> >>>> <mailto:i...@vadio.com>> wrote: > >> >>>>> Here are some examples and details of the scenarios. The KafkaRDD is > >> >>>>> the > >> >>>>> most > >> >>>>> error prone to polling > >> >>>>> timeouts and concurrentm modification errors. > >> >>>>> > >> >>>>> *Using KafkaRDD* - This takes a list of channels and processes them > >> >>>>> in > >> >>>>> parallel using the KafkaRDD directly. they all use the same consumer > >> >>>>> group > >> >>>>> ('storage-group'), but each has it's own topic and each topic has 4 > >> >>>>> partitions. We routinely get timeout errors when polling for data. > >> >>>>> This > >> >>>>> occurs whether we process in parallel or sequentially. > >> >>>>> > >> >>>>> *Spark Kafka setting:* > >> >>>>> spark.streaming.kafka.consumer.poll.ms > >> >>>>> <http://spark.streaming.kafka.consumer.poll.ms/>=2000 > >> >>>>> > >> >>>>> *Kafka Consumer Params:* > >> >>>>> metric.reporters = [] > >> >>>>> metadata.max.age.ms <http://metadata.max.age.ms/> = 300000 > >> >>>>> partition.assignment.strategy = > >> >>>>> [org.apache.kafka.clients.consumer.RangeAssignor] > >> >>>>> reconnect.backoff.ms <http://reconnect.backoff.ms/> = 50 > >> >>>>> sasl.kerberos.ticket.renew.window.factor = 0.8 > >> >>>>> max.partition.fetch.bytes = 1048576 > >> >>>>> bootstrap.servers = [somemachine:31000] > >> >>>>> ssl.keystore.type = JKS > >> >>>>> enable.auto.commit = false > >> >>>>> sasl.mechanism = GSSAPI > >> >>>>> interceptor.classes = null > >> >>>>> exclude.internal.topics = true > >> >>>>> ssl.truststore.password = null > >> >>>>> client.id <http://client.id/> = > >> >>>>> ssl.endpoint.identification.al > >> >>>>> <http://ssl.endpoint.identification.al/>gorithm = null > >> >>>>> max.poll.records = 1000 > >> >>>>> check.crcs = true > >> >>>>> request.timeout.ms <http://request.timeout.ms/> = 40000 > >> >>>>> heartbeat.interval.ms <http://heartbeat.interval.ms/> = 3000 > >> >>>>> auto.commit.interval.ms <http://auto.commit.interval.ms/> = 5000 > >> >>>>> receive.buffer.bytes = 65536 > >> >>>>> ssl.truststore.type = JKS > >> >>>>> ssl.truststore.location = null > >> >>>>> ssl.keystore.password = null > >> >>>>> fetch.min.bytes = 1 > >> >>>>> send.buffer.bytes = 131072 > >> >>>>> value.deserializer = class > >> >>>>> com.vadio.analytics.spark.storage.ClientEventJsonOptionDeserializer > >> >>>>> group.id <http://group.id/> = storage-group > >> >>>>> retry.backoff.ms <http://retry.backoff.ms/> = 100 > >> >>>>> sasl.kerberos.kinit.cmd = /usr/bin/kinit > >> >>>>> sasl.kerberos.service.name <http://sasl.kerberos.service.name/> = > >> >>>>> null > >> >>>>> sasl.kerberos.ticket.renew.jitter = 0.05 > >> >>>>> ssl.trustmanager.algorithm = PKIX > >> >>>>> ssl.key.password = null > >> >>>>> fetch.max.wait.ms <http://fetch.max.wait.ms/> = 500 > >> >>>>> sasl.kerberos.min.time.before.relogin = 60000 > >> >>>>> connections.max.idle.ms <http://connections.max.idle.ms/> = 540000 > >> >>>>> session.timeout.ms <http://session.timeout.ms/> = 30000 > >> >>>>> metrics.num.samples = 2 > >> >>>>> key.deserializer = class > >> >>>>> org.apache.kafka.common.serialization.StringDeserializer > >> >>>>> ssl.protocol = TLS > >> >>>>> ssl.provider = null > >> >>>>> ssl.enabled.protocols = [TLSv1.2, TLSv1.1, TLSv1] > >> >>>>> ssl.keystore.location = null > >> >>>>> ssl.cipher.suites = null > >> >>>>> security.protocol = PLAINTEXT > >> >>>>> ssl.keymanager.algorithm = SunX509 > >> >>>>> metrics.sample.window.ms <http://metrics.sample.window.ms/> = 30000 > >> >>>>> auto.offset.reset = earliest > >> >>>>> > >> >>>>> *Example usage with KafkaRDD:* > >> >>>>> val channels = Seq("channel1", "channel2") > >> >>>>> > >> >>>>> channels.toParArray.foreach { channel => > >> >>>>> val consumer = new KafkaConsumer[K, V](kafkaParams.asJava) > >> >>>>> > >> >>>>> // Get offsets for the given topic and the consumer group > >> >>>>> 'storage-group' > >> >>>>> val offsetRanges = getOffsets("storage-group", channel) > >> >>>>> > >> >>>>> val ds = KafkaUtils.createRDD[K, V](context, > >> >>>>> kafkaParams asJava, > >> >>>>> offsetRanges, > >> >>>>> PreferConsistent).toDS[V] > >> >>>>> > >> >>>>> // Do some aggregations > >> >>>>> ds.agg(...) > >> >>>>> // Save the data > >> >>>>> ds.write.mode(SaveMode.Append).parquet(somePath) > >> >>>>> // Save offsets using a KafkaConsumer > >> >>>>> consumer.commitSync(newOffsets.asJava) > >> >>>>> consumer.close() > >> >>>>> } > >> >>>>> > >> >>>>> > >> >>>>> *Example usage with Kafka Stream:* > >> >>>>> This creates a stream and processes events in each partition. At the > >> >>>>> end > >> >>>>> of > >> >>>>> processing for > >> >>>>> each partition, we updated the offsets for each partition. This is > >> >>>>> challenging to do, but is better > >> >>>>> then calling commitAysnc on the stream, because that occurs after > >> >>>>> the > >> >>>>> /entire/ RDD has been > >> >>>>> processed. This method minimizes duplicates in an exactly once > >> >>>>> environment. > >> >>>>> Since the executors > >> >>>>> use their own custom group "spark-executor-processor-group" and the > >> >>>>> commit > >> >>>>> is buried in private > >> >>>>> functions we are unable to use the executors cached consumer to > >> >>>>> update > >> >>>>> the > >> >>>>> offsets. This requires us > >> >>>>> to go through multiple steps to update the Kafka offsets > >> >>>>> accordingly. > >> >>>>> > >> >>>>> val offsetRanges = getOffsets("processor-group", "my-topic") > >> >>>>> > >> >>>>> val stream = KafkaUtils.createDirectStream[K, V](context, > >> >>>>> PreferConsistent, > >> >>>>> Subscribe[K, V](Seq("my-topic") asJavaCollection, > >> >>>>> kafkaParams, > >> >>>>> offsetRanges)) > >> >>>>> > >> >>>>> stream.foreachRDD { rdd => > >> >>>>> val offsetRanges = rdd.asInstanceOf[HasOffsetRanges].offsetRanges > >> >>>>> > >> >>>>> // Transform our data > >> >>>>> rdd.foreachPartition { events => > >> >>>>> // Establish a consumer in the executor so we can update > >> >>>>> offsets > >> >>>>> after each partition. > >> >>>>> // This class is homegrown and uses the KafkaConsumer to help > >> >>>>> get/set > >> >>>>> offsets > >> >>>>> val consumer = new ConsumerUtils(kafkaParams) > >> >>>>> // do something with our data > >> >>>>> > >> >>>>> // Write the offsets that were updated in this partition > >> >>>>> kafkaConsumer.setConsumerOffsets("processor-group", > >> >>>>> Map(TopicAndPartition(tp.topic, tp.partition) -> > >> >>>>> endOffset)) > >> >>>>> } > >> >>>>> } > >> >>>>> > >> >>>>> > >> >>>>> > >> >>>>> -- > >> >>>>> View this message in context: > >> >>>>> > >> >>>>> http://apache-spark-user-list.1001560.n3.nabble.com/Instability-issues-with-Spark-2-0-1-and-Kafka-0-10-tp28017p28020.html > >> >>>>> > >> >>>>> <http://apache-spark-user-list.1001560.n3.nabble.com/Instability-issues-with-Spark-2-0-1-and-Kafka-0-10-tp28017p28020.html> > >> >>>>> Sent from the Apache Spark User List mailing list archive at > >> >>>>> Nabble.com. > >> >>>>> > >> >>>>> > >> >>>>> --------------------------------------------------------------------- > >> >>>>> To unsubscribe e-mail: user-unsubscr...@spark.apache.org > >> >>>>> <mailto:user-unsubscr...@spark.apache.org> > >> >>>>> > >> >>> > >> >>> > >> >> > >> >> --------------------------------------------------------------------- > >> >> To unsubscribe e-mail: user-unsubscr...@spark.apache.org > >> >> <mailto:user-unsubscr...@spark.apache.org> > >> >> > >> > > > > > > > >
