Deep Dive with Spark Streaming - Tathagata Das - Spark Meetup 2013-06-17

download Deep Dive with Spark Streaming - Tathagata  Das - Spark Meetup 2013-06-17

of 49

  • date post

    26-Jan-2015
  • Category

    Technology

  • view

    108
  • download

    0

Embed Size (px)

description

Slides from Tathagata Das's talk at the Spark Meetup entitled "Deep Dive with Spark Streaming" on June 17, 2013 in Sunnyvale California at Plug and Play. Tathagata Das is the lead developer on Spark Streaming and a PhD student in computer science in the UC Berkeley AMPLab.

Transcript of Deep Dive with Spark Streaming - Tathagata Das - Spark Meetup 2013-06-17

  • 1. Deep dive into Spark StreamingTathagata Das (TD)Matei Zaharia, Haoyuan Li, Timothy Hunter,Patrick Wendell and many othersUC BERKELEY

2. What is Spark Streaming? Extends Spark for doing large scale stream processing Scales to 100s of nodes and achieves second scale latencies Efficient and fault-tolerant stateful stream processing Integrates with Sparks batch and interactive processing Provides a simple batch-like API for implementing complexalgorithms 3. Discretized Stream ProcessingRun a streaming computation as a series of very small,deterministic batch jobs3SparkSparkStreamingbatches of Xsecondslive data streamprocessedresults Chop up the live stream into batches of Xseconds Spark treats each batch of data as RDDsand processes them using RDD operations Finally, the processed results of the RDDoperations are returned in batches 4. Discretized Stream ProcessingRun a streaming computation as a series of very small,deterministic batch jobs4 Batch sizes as low as second, latency ~ 1second Potential for combining batch processingand streaming processing in the samesystemSparkSparkStreamingbatches of Xsecondslive data streamprocessedresults 5. Example 1 Get hashtags from Twitterval tweets = ssc.twitterStream(, )DStream: a sequence of distributed datasets (RDDs)representing a distributed stream of databatch @ t+1batch @ t batch @ t+2tweets DStreamstored in memory as an RDD(immutable, distributed dataset)Twitter Streaming API 6. Example 1 Get hashtags from Twitterval tweets = ssc.twitterStream(, )val hashTags = tweets.flatMap (status => getTags(status))flatMap flatMap flatMaptransformation: modify data in one DStream to createanother DStreamnew DStreamnew RDDs createdfor every batchbatch @ t+1batch @ t batch @ t+2tweets DStreamhashTags Dstream*#cat, #dog, + 7. Example 1 Get hashtags from Twitterval tweets = ssc.twitterStream(, )val hashTags = tweets.flatMap (status => getTags(status))hashTags.saveAsHadoopFiles("hdfs://...")output operation: to push data to external storageflatMap flatMap flatMapsave save savebatch @ t+1batch @ t batch @ t+2tweets DStreamhashTags DStreamevery batchsaved to HDFS 8. DStream of dataExample 2 Count the hashtags over last 1 minval tweets = ssc.twitterStream(, )val hashTags = tweets.flatMap (status => getTags(status))val tagCounts = hashTags.window(Minutes(1), Seconds(1)).countByValue()sliding windowoperationwindow length sliding intervalwindow lengthsliding interval 9. tagCountsExample 2 Count the hashtags over last 1 minval tagCounts = hashTags.window(Minutes(1), Seconds(1)).countByValue()hashTagst-1 t t+1 t+2 t+3sliding windowcountByValuecount over allthe data in thewindow 10. Key concepts DStream sequence of RDDs representing a stream of data- Twitter, HDFS, Kafka, Flume, ZeroMQ, Akka Actor, TCP sockets Transformations modify data from one DStream to another- Standard RDD operations map, countByValue, reduceByKey, join, - Stateful operations window, countByValueAndWindow, Output Operations send data to external entity- saveAsHadoopFiles saves to HDFS- foreach do anything with each batch of results 11. Arbitrary Stateful Computations Maintain arbitrary state, track sessions- Maintain per-user mood as state, and update it with his/her tweetsmoods = tweets.updateStateByKey(tweet => updateMood(tweet))updateMood(newTweets, lastMood) => newMoodtweetst-1 t t+1 t+2 t+3moods 12. Combine Batch and Stream Processing Do arbitrary Spark RDD computation within DStream- Join incoming tweets with a spam file to filter out bad tweetstweets.transform(tweetsRDD => {tweetsRDD.join(spamHDFSFile).filter(...)}) 13. Fault-tolerance RDDs remember the operationsthat created them Batches of input data arereplicated in memory for fault-tolerance Data lost due to worker failure,can be recomputed fromreplicated input data Therefore, all transformed data isfault-tolerantinput datareplicatedin memoryflatMaplost partitionsrecomputed onother workerstweetsRDDhashTagsRDD 14. Agenda Overview DStream Abstraction System Model Persistence / Caching RDD Checkpointing Performance Tuning 15. Discretized Stream (DStream)A sequence of RDDs representinga stream of dataWhat does it take to define a DStream? 16. DStream InterfaceThe DStream interface primarily defines how to generate anRDD in each batch interval List of dependent (parent) DStreams Slide Interval, the interval at which it will compute RDDs Function to compute RDD at a time t 17. Example: Mapped DStream Dependencies: Single parent DStream Slide Interval: Same as the parent DStream Compute function for time t: Create new RDD by applying mapfunction on parent DStreams RDD of time t 18. Window operation gather together data over a sliding window Dependencies: Single parent DStream Slide Interval: Window sliding interval Compute function for time t: Apply union over all the RDDs ofparent DStream between times t and (t window length)Example: Windowed DStreamParent DStreamwindow lengthsliding interval 19. Example: Network Input DStreamBase class of all input DStreams that receive data from the network Dependencies: None Slide Interval: Batch duration in streaming context Compute function for time t: Create a BlockRDD with all theblocks of data received in the last batch interval Associated with a Network Receiver object 20. Network ReceiverResponsible for receiving data and pushing it into Sparks datamanagement layer (Block Manager)Base class for all receivers - Kafka, Flume, etc.Simple Interface: What to do on starting the receiver- Helper object blockGenerator to push data into Spark What to do on stopping the receiver 21. Example: Socket Receiver On start:Connect to remote TCP serverWhile socket is connected,Receiving bytes and deserializeDeserialize them into Java objectsAdd the objects to blockGenerator On stop:Disconnect socket 22. DStream Grapht = ssc.twitterStream().map()t.foreach()t1 = ssc.twitterStream()t2 = ssc.twitterStream()t = t1.union(t2).map()t.saveAsHadoopFiles()t.map().foreach()t.filter().foreach()TMFETwitter Input DStreamMapped DStreamForeach DStreamTUMTM FFEFEFEDStream GraphSpark Streaming programDummy DStream signifyingan output operation 23. DStream Graph RDD Graphs Spark jobs Every interval, RDD graph is computed from DStream graph For each output operation, a Spark action is created For each action, a Spark job is created to compute itTUMTM FFEFEFEDStream GraphBUMBM FAA ARDD GraphSpark actionsBlock RDDs withdata received inlast batch interval3 Spark jobs 24. Agenda Overview DStream Abstraction System Model Persistence / Caching RDD Checkpointing Performance Tuning 25. Components Network Input Tracker Keeps track of the data received by each networkreceiver and maps them to the corresponding input DStreams Job Scheduler Periodically queries the DStream graph to generate Sparkjobs from received data, and hands them to Job Manager for execution Job Manager Maintains a job queue and executes the jobs in Sparkssc = new StreamingContextt =ssc.twitterStream()t.filter().foreach()Your programDStream graphSpark ContextNetwork Input TrackerJob ManagerJob SchedulerSpark ClientRDD graph SchedulerBlock managerShuffletrackerSpark WorkerBlockmanagerTaskthreadsClusterManager 26. Execution Model Receiving DataSpark Streaming + Spark Driver Spark WorkersStreamingContext.start()NetworkInputTrackerReceiverData recvdBlockManagerBlocks replicatedBlockManagerMasterBlockManagerBlocks pushed 27. Spark WorkersExecution Model Job SchedulingNetworkInputTrackerJobScheduler SparksSchedulersReceiverBlockManagerBlockManagerJobs executed onworker nodesDStreamGraphJobManagerJobQueueSpark Streaming + Spark DriverJobsBlock IDsRDDs 28. Job Scheduling Each output operation used generates a job- More jobs more time taken to process batches higher batchduration Job Manager decides how many concurrent Spark jobs to run- Default is 1, can be set using Java propertyspark.streaming.concurrentJobs- If you have multiple output operations, you can try increasing thisproperty to reduce batch processing times and so reduce batch duration 29. Agenda Overview DStream Abstraction System Model Persistence / Caching RDD Checkpointing Performance Tuning 30. DStream Persistence If a DStream is set to persist at a storage level, then all RDDsgenerated by it set to the same storage level When to persist?- If there are multiple transformations / actions on a DStream- If RDDs in a DStream is going to be used multiple times Window-based DStreams are automatically persisted in memory 31. DStream Persistence Default storage level of DStreams is StorageLevel.MEMORY_ONLY_SER(i.e. in memory as serialized bytes)- Except for input DStreams which have StorageLevel.MEMORY_AND_DISK_SER_2- Note the difference from RDDs default level (no serialization)- Serialization reduces random pauses due to GC providing more consistentjob processing times 32. Agenda Overview DStream Abstraction System Model Persistence / Caching RDD Checkpointing Performance Tuning 33. What is RDD checkpointing?Saving RDD to HDFS to prevent RDD graph from growing too large Done internally in Spark transparent to the user program Done lazily, saved to HDFS the first time it is computedred_rdd.checkpoint()HDFS fileContents of red_rdd savedto a HDFS file transparent toall child RDDs 34. Why is RDD checkpointing necessary?Stateful DStream operators can have infinite lineagesLarge lineages lead to Large closure of the RDD object large task sizes high task launch times High recovery times under failuredatat-1 t t+1 t+2 t+3states 35. Why is RDD checkpointing necessary?Stateful DStream operators can have infinite lineagesPeriodic RDD checkpointing solves thisUseful for iterative Spark programs as welldatat-1 t t+1 t+2 t+3statesHDFSHDFS 36. RDD Checkpointing Periodicity of checkpoint determines a tradeoff- Checkpoint too frequent: HDFS writing will slow things down- Checkpoint too infrequent: Task launch times may increa