Of course, exactly once receiving is not same as exactly once. In case of direct kafka stream, the data may actually be pulled multiple time. But even if the data of a batch is pulled twice because of some failure, the final result (that is, transformed data accessed through foreachRDD) will always be the same even if recomputed. In other words, the data in partition x of the RDD of time t, will always be the same even if that partition gets recomputed. Now, to get end-to-end exactly once, you will have also push data out to external data stores in the exactly-once manner - either the updates are idempotent, or you can use the unique id [(batch time, partition ID)] to update the store transactionally (such that each partition is inserted into the data store only once.
This is also explained in my talk. - https://www.youtube.com/watch?v=d5UJonrruHk On Tue, Jul 14, 2015 at 8:18 PM, Chen Song <chen.song...@gmail.com> wrote: > Thanks TD. > > As for 1), if timing is not guaranteed, how does exactly once semantics > supported? It feels like exactly once receiving is not necessarily exactly > once processing. > > Chen > > On Tue, Jul 14, 2015 at 10:16 PM, Tathagata Das <t...@databricks.com> > wrote: > >> >> >> On Tue, Jul 14, 2015 at 6:42 PM, Chen Song <chen.song...@gmail.com> >> wrote: >> >>> Thanks TD and Cody. I saw that. >>> >>> 1. By doing that (foreachRDD), does KafkaDStream checkpoints its offsets >>> on HDFS at the end of each batch interval? >>> >> >> The timing is not guaranteed. >> >> >>> 2. In the code, if I first apply transformations and actions on the >>> directKafkaStream and then use foreachRDD on the original KafkaDStream to >>> commit offsets myself, will offsets commits always happen after >>> transformation and action? >>> >>> What do you mean by "original KafkaDStream"? if you meant the >> directKafkaStream? If yes, then yes, output operations like foreachRDD is >> executed in each batch in the same order as they are defined. >> >> dstream1.foreachRDD { rdd => func1(rdd) } >> dstream2.foreachRDD { rdd => func2(rdd) } >> >> In every batch interval, func1 will be executed before func2. >> >> >> >> >>> Chen >>> >>> On Tue, Jul 14, 2015 at 6:43 PM, Tathagata Das <t...@databricks.com> >>> wrote: >>> >>>> Relevant documentation - >>>> https://spark.apache.org/docs/latest/streaming-kafka-integration.html, >>>> towards the end. >>>> >>>> directKafkaStream.foreachRDD { rdd => >>>> val offsetRanges = rdd.asInstanceOf[HasOffsetRanges] >>>> // offsetRanges.length = # of Kafka partitions being consumed >>>> ... >>>> } >>>> >>>> >>>> On Tue, Jul 14, 2015 at 3:17 PM, Cody Koeninger <c...@koeninger.org> >>>> wrote: >>>> >>>>> You have access to the offset ranges for a given rdd in the stream by >>>>> typecasting to HasOffsetRanges. You can then store the offsets wherever >>>>> you need to. >>>>> >>>>> On Tue, Jul 14, 2015 at 5:00 PM, Chen Song <chen.song...@gmail.com> >>>>> wrote: >>>>> >>>>>> A follow up question. >>>>>> >>>>>> When using createDirectStream approach, the offsets are checkpointed >>>>>> to HDFS and it is understandable by Spark Streaming job. Is there a way >>>>>> to >>>>>> expose the offsets via a REST api to end users. Or alternatively, is >>>>>> there >>>>>> a way to have offsets committed to Kafka Offset Manager so users can >>>>>> query >>>>>> from a consumer programmatically? >>>>>> >>>>>> Essentially, all I need to do is monitor the progress of data >>>>>> consumption of the Kafka topic. >>>>>> >>>>>> >>>>>> On Tue, Jun 30, 2015 at 9:39 AM, Cody Koeninger <c...@koeninger.org> >>>>>> wrote: >>>>>> >>>>>>> You can't use different versions of spark in your application vs >>>>>>> your cluster. >>>>>>> >>>>>>> For the direct stream, it's not 60 partitions per executor, it's 300 >>>>>>> partitions, and executors work on them as they are scheduled. Yes, if >>>>>>> you >>>>>>> have no messages you will get an empty partition. It's up to you >>>>>>> whether >>>>>>> it's worthwhile to call coalesce or not. >>>>>>> >>>>>>> On Tue, Jun 30, 2015 at 2:45 AM, Shushant Arora < >>>>>>> shushantaror...@gmail.com> wrote: >>>>>>> >>>>>>>> Is this 3 is no of parallel consumer threads per receiver , means >>>>>>>> in total we have 2*3=6 consumer in same consumer group consuming from >>>>>>>> all >>>>>>>> 300 partitions. >>>>>>>> 3 is just parallelism on same receiver and recommendation is to use >>>>>>>> 1 per receiver since consuming from kafka is not cpu bound rather >>>>>>>> NIC(network bound) increasing consumer thread on one receiver won't >>>>>>>> make >>>>>>>> it parallel in ideal sense ? >>>>>>>> >>>>>>>> In non receiver based consumer spark 1.3 If I use 5 execuots and >>>>>>>> kafka topic has 300 partions , does kafkaRDD created on 5 executors >>>>>>>> will >>>>>>>> have 60 partitions per executor (total 300 one to one mapping) and if >>>>>>>> some >>>>>>>> of kafka partitions are empty say offset of last checkpoint to current >>>>>>>> is >>>>>>>> same for partitons P123, still it will create empty partition in >>>>>>>> kafkaRDD ? >>>>>>>> So we should call coalesce on kafkaRDD ? >>>>>>>> >>>>>>>> >>>>>>>> And is there any incompatibity issue when I include >>>>>>>> spark-streaming_2.10 (version 1.3) and spark-core_2.10(version 1.3) in >>>>>>>> my >>>>>>>> application but my cluster has spark version 1.2 ? >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> On Mon, Jun 29, 2015 at 7:56 PM, Shushant Arora < >>>>>>>> shushantaror...@gmail.com> wrote: >>>>>>>> >>>>>>>>> 1. Here you are basically creating 2 receivers and asking each of >>>>>>>>> them to consume 3 kafka partitions each. >>>>>>>>> >>>>>>>>> - In 1.2 we have high level consumers so how can we restrict no of >>>>>>>>> kafka partitions to consume from? Say I have 300 kafka partitions in >>>>>>>>> kafka >>>>>>>>> topic and as in above I gave 2 receivers and 3 kafka partitions . >>>>>>>>> Then is >>>>>>>>> it mean I will read from 6 out of 300 partitions only and for rest 294 >>>>>>>>> partitions data is lost? >>>>>>>>> >>>>>>>>> >>>>>>>>> 2.One more doubt in spark streaming how is it decided which part >>>>>>>>> of main function of driver will run at each batch interval ? Since >>>>>>>>> whole >>>>>>>>> code is written in one function(main function in driver) so how it >>>>>>>>> determined kafka streams receivers not to be registered in each >>>>>>>>> batch only >>>>>>>>> processing to be done . >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> On Mon, Jun 29, 2015 at 7:35 PM, ayan guha <guha.a...@gmail.com> >>>>>>>>> wrote: >>>>>>>>> >>>>>>>>>> Hi >>>>>>>>>> >>>>>>>>>> Let me take ashot at your questions. (I am sure people like Cody >>>>>>>>>> and TD will correct if I am wrong) >>>>>>>>>> >>>>>>>>>> 0. This is exact copy from the similar question in mail thread >>>>>>>>>> from Akhil D: >>>>>>>>>> Since you set local[4] you will have 4 threads for your >>>>>>>>>> computation, and since you are having 2 receivers, you are left >>>>>>>>>> with 2 threads to process ((0 + 2) <-- This 2 is your 2 threads.) >>>>>>>>>> And the other /2 means you are having 2 tasks in that stage >>>>>>>>>> (with id 0). >>>>>>>>>> >>>>>>>>>> 1. Here you are basically creating 2 receivers and asking each of >>>>>>>>>> them to consume 3 kafka partitions each. >>>>>>>>>> 2. How does that matter? It depends on how many receivers you >>>>>>>>>> have created to consume that data and if you have repartitioned it. >>>>>>>>>> Remember, spark is lazy and executors are relted to the context >>>>>>>>>> 3. I think in java, factory method is fixed. You just pass around >>>>>>>>>> the contextFactory object. (I love python :) see the signature isso >>>>>>>>>> much >>>>>>>>>> cleaner :) ) >>>>>>>>>> 4. Yes, if you use spark checkpointing. You can use yourcustom >>>>>>>>>> check pointing too. >>>>>>>>>> >>>>>>>>>> Best >>>>>>>>>> Ayan >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> On Mon, Jun 29, 2015 at 4:02 AM, Shushant Arora < >>>>>>>>>> shushantaror...@gmail.com> wrote: >>>>>>>>>> >>>>>>>>>>> Few doubts : >>>>>>>>>>> >>>>>>>>>>> In 1.2 streaming when I use union of streams , my streaming >>>>>>>>>>> application getting hanged sometimes and nothing gets printed on >>>>>>>>>>> driver. >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> [Stage 2:> >>>>>>>>>>> >>>>>>>>>>> (0 + 2) / 2] >>>>>>>>>>> Whats is 0+2/2 here signifies. >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> 1.Does no of streams in topicsMap.put("testSparkPartitioned", 3); >>>>>>>>>>> be same as numstreams=2 ? in unioned stream ? >>>>>>>>>>> >>>>>>>>>>> 2. I launched app on yarnRM with num-executors as 5 . It created >>>>>>>>>>> 2 receivers and 5 execuots . As in stream receivers nodes get fixed >>>>>>>>>>> at >>>>>>>>>>> start of app throughout its lifetime . Does executors gets >>>>>>>>>>> allicated at >>>>>>>>>>> start of each job on 1s batch interval? If yes, how does its fast to >>>>>>>>>>> allocate resources. I mean if i increase num-executors to 50 , it >>>>>>>>>>> will >>>>>>>>>>> negotiate 50 executors from yarnRM at start of each job so does it >>>>>>>>>>> takes >>>>>>>>>>> more time in allocating executors than batch interval(here 1s , say >>>>>>>>>>> if >>>>>>>>>>> 500ms).? Can i fixed processing executors also throughout the app? >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> SparkConf conf = new >>>>>>>>>>> SparkConf().setAppName("SampleSparkStreamingApp"); >>>>>>>>>>> JavaStreamingContext jssc = new >>>>>>>>>>> JavaStreamingContext(conf,Durations.milliseconds(1000)); >>>>>>>>>>> >>>>>>>>>>> Map<String,String> kafkaParams = new HashMap<String, String>(); >>>>>>>>>>> kafkaParams.put("zookeeper.connect","ipadd:2181"); >>>>>>>>>>> kafkaParams.put("group.id", "testgroup"); >>>>>>>>>>> kafkaParams.put("zookeeper.session.timeout.ms", "10000"); >>>>>>>>>>> Map<String,Integer> topicsMap = new HashMap<String,Integer>(); >>>>>>>>>>> topicsMap.put("testSparkPartitioned", 3); >>>>>>>>>>> int numStreams = 2; >>>>>>>>>>> List<JavaPairDStream<byte[],byte[]>> kafkaStreams = new >>>>>>>>>>> ArrayList<JavaPairDStream<byte[], byte[]>>(); >>>>>>>>>>> for(int i=0;i<numStreams;i++){ >>>>>>>>>>> kafkaStreams.add(KafkaUtils.createStream(jssc, byte[].class, >>>>>>>>>>> byte[].class,kafka.serializer.DefaultDecoder.class , >>>>>>>>>>> kafka.serializer.DefaultDecoder.class, >>>>>>>>>>> kafkaParams, topicsMap, StorageLevel.MEMORY_ONLY())); >>>>>>>>>>> } >>>>>>>>>>> JavaPairDStream<byte[], byte[]> directKafkaStream = >>>>>>>>>>> jssc.union(kafkaStreams.get(0),kafkaStreams.subList(1, >>>>>>>>>>> kafkaStreams.size())); >>>>>>>>>>> JavaDStream<String> lines = directKafkaStream.map(new >>>>>>>>>>> Function<Tuple2<byte[],byte[]>, String>() { >>>>>>>>>>> >>>>>>>>>>> public String call(Tuple2<byte[], byte[]> arg0) throws Exception >>>>>>>>>>> { >>>>>>>>>>> ...processing >>>>>>>>>>> ..return msg; >>>>>>>>>>> } >>>>>>>>>>> }); >>>>>>>>>>> lines.print(); >>>>>>>>>>> jssc.start(); >>>>>>>>>>> jssc.awaitTermination(); >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> ------------------------------------------------------------------------------------------------------------------------------------------------------- >>>>>>>>>>> 3.For avoiding dataloss when we use checkpointing, and factory >>>>>>>>>>> method to create sparkConytext, is method name fixed >>>>>>>>>>> or we can use any name and how to set in app the method name to >>>>>>>>>>> be used ? >>>>>>>>>>> >>>>>>>>>>> 4.In 1.3 non receiver based streaming, kafka offset is not >>>>>>>>>>> stored in zookeeper, is it because of zookeeper is not efficient >>>>>>>>>>> for high >>>>>>>>>>> writes and read is not strictly consistent? So >>>>>>>>>>> >>>>>>>>>>> we use simple Kafka API that does not use Zookeeper and >>>>>>>>>>> offsets tracked only by Spark Streaming within its checkpoints. >>>>>>>>>>> This eliminates inconsistencies between Spark Streaming and >>>>>>>>>>> Zookeeper/Kafka, and so each record is received by Spark Streaming >>>>>>>>>>> effectively exactly once despite failures. >>>>>>>>>>> >>>>>>>>>>> So we have to call context.checkpoint(hdfsdir)? Or is it >>>>>>>>>>> implicit checkoint location ? Means does hdfs be used for small >>>>>>>>>>> data(just >>>>>>>>>>> offset?) >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> On Sat, Jun 27, 2015 at 7:37 PM, Dibyendu Bhattacharya < >>>>>>>>>>> dibyendu.bhattach...@gmail.com> wrote: >>>>>>>>>>> >>>>>>>>>>>> Hi, >>>>>>>>>>>> >>>>>>>>>>>> There is another option to try for Receiver Based Low Level >>>>>>>>>>>> Kafka Consumer which is part of Spark-Packages ( >>>>>>>>>>>> http://spark-packages.org/package/dibbhatt/kafka-spark-consumer) >>>>>>>>>>>> . This can be used with WAL as well for end to end zero data loss. >>>>>>>>>>>> >>>>>>>>>>>> This is also Reliable Receiver and Commit offset to ZK. Given >>>>>>>>>>>> the number of Kafka Partitions you have ( > 100) , using High >>>>>>>>>>>> Level Kafka >>>>>>>>>>>> API for Receiver based approach may leads to issues related >>>>>>>>>>>> Consumer >>>>>>>>>>>> Re-balancing which is a major issue of Kafka High Level API. >>>>>>>>>>>> >>>>>>>>>>>> Regards, >>>>>>>>>>>> Dibyendu >>>>>>>>>>>> >>>>>>>>>>>> >>>>>>>>>>>> >>>>>>>>>>>> On Sat, Jun 27, 2015 at 3:04 PM, Tathagata Das < >>>>>>>>>>>> t...@databricks.com> wrote: >>>>>>>>>>>> >>>>>>>>>>>>> In the receiver based approach, If the receiver crashes for >>>>>>>>>>>>> any reason (receiver crashed or executor crashed) the receiver >>>>>>>>>>>>> should get >>>>>>>>>>>>> restarted on another executor and should start reading data from >>>>>>>>>>>>> the offset >>>>>>>>>>>>> present in the zookeeper. There is some chance of data loss which >>>>>>>>>>>>> can >>>>>>>>>>>>> alleviated using Write Ahead Logs (see streaming programming >>>>>>>>>>>>> guide for more >>>>>>>>>>>>> details, or see my talk [Slides PDF >>>>>>>>>>>>> <http://www.slideshare.net/SparkSummit/recipes-for-running-spark-streaming-apploications-in-production-tathagata-daspptx> >>>>>>>>>>>>> , Video >>>>>>>>>>>>> <https://www.youtube.com/watch?v=d5UJonrruHk&list=PL-x35fyliRwgfhffEpywn4q23ykotgQJ6&index=4> >>>>>>>>>>>>> ] from last Spark Summit 2015). But that approach can give >>>>>>>>>>>>> duplicate records. The direct approach gives exactly-once >>>>>>>>>>>>> guarantees, so >>>>>>>>>>>>> you should try it out. >>>>>>>>>>>>> >>>>>>>>>>>>> TD >>>>>>>>>>>>> >>>>>>>>>>>>> On Fri, Jun 26, 2015 at 5:46 PM, Cody Koeninger < >>>>>>>>>>>>> c...@koeninger.org> wrote: >>>>>>>>>>>>> >>>>>>>>>>>>>> Read the spark streaming guide ad the kafka integration guide >>>>>>>>>>>>>> for a better understanding of how the receiver based stream >>>>>>>>>>>>>> works. >>>>>>>>>>>>>> >>>>>>>>>>>>>> Capacity planning is specific to your environment and what >>>>>>>>>>>>>> the job is actually doing, youll need to determine it >>>>>>>>>>>>>> empirically. >>>>>>>>>>>>>> >>>>>>>>>>>>>> >>>>>>>>>>>>>> On Friday, June 26, 2015, Shushant Arora < >>>>>>>>>>>>>> shushantaror...@gmail.com> wrote: >>>>>>>>>>>>>> >>>>>>>>>>>>>>> In 1.2 how to handle offset management after stream >>>>>>>>>>>>>>> application starts in each job . I should commit offset after >>>>>>>>>>>>>>> job >>>>>>>>>>>>>>> completion manually? >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> And what is recommended no of consumer threads. Say I have >>>>>>>>>>>>>>> 300 partitions in kafka cluster . Load is ~ 1 million events per >>>>>>>>>>>>>>> second.Each event is of ~500bytes. Having 5 receivers with 60 >>>>>>>>>>>>>>> partitions >>>>>>>>>>>>>>> each receiver is sufficient for spark streaming to consume ? >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> On Fri, Jun 26, 2015 at 8:40 PM, Cody Koeninger < >>>>>>>>>>>>>>> c...@koeninger.org> wrote: >>>>>>>>>>>>>>> >>>>>>>>>>>>>>>> The receiver-based kafka createStream in spark 1.2 uses >>>>>>>>>>>>>>>> zookeeper to store offsets. If you want finer-grained control >>>>>>>>>>>>>>>> over >>>>>>>>>>>>>>>> offsets, you can update the values in zookeeper yourself >>>>>>>>>>>>>>>> before starting >>>>>>>>>>>>>>>> the job. >>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>> createDirectStream in spark 1.3 is still marked as >>>>>>>>>>>>>>>> experimental, and subject to change. That being said, it >>>>>>>>>>>>>>>> works better for >>>>>>>>>>>>>>>> me in production than the receiver based api. >>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>> On Fri, Jun 26, 2015 at 6:43 AM, Shushant Arora < >>>>>>>>>>>>>>>> shushantaror...@gmail.com> wrote: >>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>> I am using spark streaming 1.2. >>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>> If processing executors get crashed will receiver rest the >>>>>>>>>>>>>>>>> offset back to last processed offset? >>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>> If receiver itself got crashed is there a way to reset the >>>>>>>>>>>>>>>>> offset without restarting streaming application other than >>>>>>>>>>>>>>>>> smallest or >>>>>>>>>>>>>>>>> largest. >>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>> Is spark streaming 1.3 which uses low level consumer api, >>>>>>>>>>>>>>>>> stabe? And which is recommended for handling data loss 1.2 >>>>>>>>>>>>>>>>> or 1.3 . >>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>> >>>>>>>>>>>>>>> >>>>>>>>>>>>> >>>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> -- >>>>>>>>>> Best Regards, >>>>>>>>>> Ayan Guha >>>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>> >>>>>>> >>>>>> >>>>>> >>>>>> -- >>>>>> Chen Song >>>>>> >>>>>> >>>>> >>>> >>> >>> >>> -- >>> Chen Song >>> >>> >> > > > -- > Chen Song > >
