So, our Streaming Job fails with the following errors. If you see the errors(highlighted in blue below), they are all related to Kafka losing offsets and OffsetOutOfRangeException.
What are the options we have other than fixing Kafka? We would like to do something like the following. How can we achieve 1 and 2 with Spark Kafka Direct? 1.Need to see a way to skip some offsets if they are not available after the max retries are reached..in that case there might be data loss. 2.Catch that exception and somehow force things to "reset" for that partition And how would it handle the offsets already calculated in the backlog (if there is one)? 3.Track the offsets separately, restart the job by providing the offsets. 4.Or a straightforward approach would be to monitor the log for this error, and if it occurs more than X times, kill the job, remove the checkpoint directory, and restart. ERROR DirectKafkaInputDStream: ArrayBuffer(kafka.common.UnknownException, org.apache.spark.SparkException: Couldn't find leader offsets for Set([test_stream,5])) java.lang.ClassNotFoundException: kafka.common.NotLeaderForPartitionException at org.apache.spark.util.Utils$.logUncaughtExceptions(Utils.scala:1699) java.util.concurrent.RejectedExecutionException: Task org.apache.spark.streaming.CheckpointWriter$CheckpointWriteHandler@a48c5a8 rejected from java.util.concurrent.ThreadPoolExecutor@543258e0[Terminated, pool size = 0, active threads = 0, queued tasks = 0, completed tasks = 12112] org.apache.spark.SparkException: Job aborted due to stage failure: Task 10 in stage 52.0 failed 4 times, most recent failure: Lost task 10.3 in stage 52.0 (TID 255, 172.16.97.97): UnknownReason Exception in thread "streaming-job-executor-0" java.lang.Error: java.lang.InterruptedException Caused by: java.lang.InterruptedException java.lang.ClassNotFoundException: kafka.common.OffsetOutOfRangeException at org.apache.spark.util.Utils$.logUncaughtExceptions(Utils.scala:1699) org.apache.spark.SparkException: Job aborted due to stage failure: Task 7 in stage 33.0 failed 4 times, most recent failure: Lost task 7.3 in stage 33.0 (TID 283, 172.16.97.103): UnknownReason java.lang.ClassNotFoundException: kafka.common.OffsetOutOfRangeException at org.apache.spark.util.Utils$.logUncaughtExceptions(Utils.scala:1699) java.lang.ClassNotFoundException: kafka.common.OffsetOutOfRangeException at org.apache.spark.util.Utils$.logUncaughtExceptions(Utils.scala:1699) On Mon, Nov 30, 2015 at 12:23 PM, Cody Koeninger <c...@koeninger.org> wrote: > You'd need to get the earliest or latest available offsets from kafka, > whichever is most appropriate for your situation. > > The KafkaRDD will use the value of refresh.leader.backoff.ms, so you can > try adjusting that to get a longer sleep before retrying the task. > > On Mon, Nov 30, 2015 at 1:50 PM, swetha kasireddy < > swethakasire...@gmail.com> wrote: > >> Hi Cody, >> >> What if the Offsets that are tracked are not present in Kafka. How do I >> skip those offsets and go to the next Offset? Also would specifying >> rebalance.backoff.ms be of any help? >> >> Thanks, >> Swteha >> >> On Thu, Nov 12, 2015 at 9:07 AM, Cody Koeninger <c...@koeninger.org> >> wrote: >> >>> To be blunt, if you care about being able to recover from weird >>> situations, you should be tracking offsets yourself and specifying offsets >>> on job start, not relying on checkpoints. >>> >>> On Tue, Nov 10, 2015 at 3:54 AM, Adrian Tanase <atan...@adobe.com> >>> wrote: >>> >>>> I’ve seen this before during an extreme outage on the cluster, where >>>> the kafka offsets checkpointed by the directstreamRdd were bigger than what >>>> kafka reported. The checkpoint was therefore corrupted. >>>> I don’t know the root cause but since I was stressing the cluster >>>> during a reliability test I can only assume that one of the Kafka >>>> partitions was restored from an out-of-sync replica and did not contain all >>>> the data. Seems extreme but I don’t have another idea. >>>> >>>> @Cody – do you know of a way to recover from a situation like this? Can >>>> someone manually delete folders from the checkpoint folder to help the job >>>> recover? E.g. Go 2 steps back, hoping that kafka has those offsets. >>>> >>>> -adrian >>>> >>>> From: swetha kasireddy >>>> Date: Monday, November 9, 2015 at 10:40 PM >>>> To: Cody Koeninger >>>> Cc: "user@spark.apache.org" >>>> Subject: Re: Kafka Direct does not recover automatically when the >>>> Kafka Stream gets messed up? >>>> >>>> OK. But, one thing that I observed is that when there is a problem with >>>> Kafka Stream, unless I delete the checkpoint directory the Streaming job >>>> does not restart. I guess it tries to retry the failed tasks and if it's >>>> not able to recover, it fails again. Sometimes, it fails with StackOverFlow >>>> Error. >>>> >>>> Why does the Streaming job not restart from checkpoint directory when >>>> the job failed earlier with Kafka Brokers getting messed up? We have the >>>> checkpoint directory in our hdfs. >>>> >>>> On Mon, Nov 9, 2015 at 12:34 PM, Cody Koeninger <c...@koeninger.org> >>>> wrote: >>>> >>>>> I don't think deleting the checkpoint directory is a good way to >>>>> restart the streaming job, you should stop the spark context or at the >>>>> very >>>>> least kill the driver process, then restart. >>>>> >>>>> On Mon, Nov 9, 2015 at 2:03 PM, swetha kasireddy < >>>>> swethakasire...@gmail.com> wrote: >>>>> >>>>>> Hi Cody, >>>>>> >>>>>> Our job is our failsafe as we don't have Control over Kafka Stream as >>>>>> of now. Can setting rebalance max retries help? We do not have any >>>>>> monitors >>>>>> setup as of now. We need to setup the monitors. >>>>>> >>>>>> My idea is to to have some kind of Cron job that queries the >>>>>> Streaming API for monitoring like every 5 minutes and then send an email >>>>>> alert and automatically restart the Streaming job by deleting the >>>>>> Checkpoint directory. Would that help? >>>>>> >>>>>> >>>>>> >>>>>> Thanks! >>>>>> >>>>>> On Mon, Nov 9, 2015 at 11:09 AM, Cody Koeninger <c...@koeninger.org> >>>>>> wrote: >>>>>> >>>>>>> The direct stream will fail the task if there is a problem with the >>>>>>> kafka broker. Spark will retry failed tasks automatically, which should >>>>>>> handle broker rebalances that happen in a timely fashion. >>>>>>> spark.tax.maxFailures controls the maximum number of retries before >>>>>>> failing >>>>>>> the job. Direct stream isn't any different from any other spark task in >>>>>>> that regard. >>>>>>> >>>>>>> The question of what kind of monitoring you need is more a question >>>>>>> for your particular infrastructure and what you're already using for >>>>>>> monitoring. We put all metrics (application level or system level) into >>>>>>> graphite and alert from there. >>>>>>> >>>>>>> I will say that if you've regularly got problems with kafka falling >>>>>>> over for half an hour, I'd look at fixing that before worrying about >>>>>>> spark >>>>>>> monitoring... >>>>>>> >>>>>>> >>>>>>> On Mon, Nov 9, 2015 at 12:26 PM, swetha <swethakasire...@gmail.com> >>>>>>> wrote: >>>>>>> >>>>>>>> Hi, >>>>>>>> >>>>>>>> How to recover Kafka Direct automatically when the there is a >>>>>>>> problem with >>>>>>>> Kafka brokers? Sometimes our Kafka Brokers gets messed up and the >>>>>>>> entire >>>>>>>> Streaming job blows up unlike some other consumers which do recover >>>>>>>> automatically. How can I make sure that Kafka Direct recovers >>>>>>>> automatically >>>>>>>> when the broker fails for sometime say 30 minutes? What kind of >>>>>>>> monitors >>>>>>>> should be in place to recover the job? >>>>>>>> >>>>>>>> Thanks, >>>>>>>> Swetha >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> -- >>>>>>>> View this message in context: >>>>>>>> http://apache-spark-user-list.1001560.n3.nabble.com/Kafka-Direct-does-not-recover-automatically-when-the-Kafka-Stream-gets-messed-up-tp25331.html >>>>>>>> Sent from the Apache Spark User List mailing list archive at >>>>>>>> Nabble.com. >>>>>>>> >>>>>>>> >>>>>>>> --------------------------------------------------------------------- >>>>>>>> To unsubscribe, e-mail: user-unsubscr...@spark.apache.org >>>>>>>> For additional commands, e-mail: user-h...@spark.apache.org >>>>>>>> >>>>>>>> >>>>>>> >>>>>> >>>>> >>>> >>> >> >
