Spark Summit 2014: Spark Streaming for Realtime Auctions

  • Published on

  • View

  • Download

Embed Size (px)


Spark Summit (2014-06-30) What do you do when you need to update your models sooner than your existing batch workflows provide? At Sharethrough we faced the same question. Although we use Hadoop extensively for batch processing we needed a system to process click stream data as soon as possible for our real time ad auction platform. We found Spark Streaming to be a perfect fit for us because of its easy integration into the Hadoop ecosystem, powerful functional programming API, and low friction interoperability with our existing batch workflows. In this talk, well present about some of the use cases that led us to choose a stream processing system and why we use Spark Streaming in particular. Well also discuss how we organized our jobs to promote reusability between batch and streaming workflows as well as improve testability


<ul><li> Spark Streaming for Realtime Auctions @russellcardullo Sharethrough </li> <li> Agenda Sharethrough? Streaming use cases How we use Spark Next steps </li> <li> Sharethrough vs </li> <li> The Sharethrough Native Exchange </li> <li> How can we use streaming data? </li> <li> Use Cases Creative Optimization Spend Tracking Operational Monitoring $* = max { k } k </li> <li> Creative Optimization Choose best performing variant Short feedback cycle required impression: {! device: iOS,! geocode: C967,! pkey: p193,! ! } Variant 1 Variant 2 Variant 3 Click! Content Here Hello Friend ? ? ? </li> <li> Spend Tracking Spend on visible impressions and clicks Actual spend happens asynchronously Want to correct prediction for optimal serving Predicted Spend Actual Spend </li> <li> Operational Monitoring Detect issues with content served on third party sites Use same logs as reporting Placement Click Rates P1 P5 P2 P3 P4 P6 P7 P8 </li> <li> We can directly measure business impact of using this data sooner </li> <li> Why use Spark to build these features? </li> <li> Why Spark? Scala API Supports batch and streaming Active community support Easily integrates into existing Hadoop ecosystem But it doesnt require Hadoop in order to run </li> <li> How weve integrated Spark </li> <li> Existing Data Pipeline web! server log le ume HDFS web! server log le ume web! server log le ume analytics! web! service database </li> <li> Pipeline with Streaming web! server log le ume HDFS web! server log le ume web! server log le ume analytics! web! service database </li> <li> Batch Streaming Real-Time reporting Low latency to use data Only reliable as source Low latency &gt; correctness Daily reporting Billing / earnings Anything with strict SLA Correctness &gt; low latency </li> <li> Spark Job Abstractions </li> <li> Job Organization nterSource Transform Sink Job </li> <li> Sources ! case class BeaconLogLine(! timestamp: String,! uri: String,! beaconType: String,! pkey: String,! ckey: String! )! ! object BeaconLogLine {! ! def newDStream(ssc: StreamingContext, inputPath: String): DStream[BeaconLogLine] = {! ssc.textFileStream(inputPath).map { parseRawBeacon(_) }! }! ! def parseRawBeacon(b: String): BeaconLogLine = {! ...! }! }! case class! for pattern! matching generate! DStream encapsulate ! common! operations </li> <li> Transformations ! ! def visibleByPlacement(source: DStream[BeaconLogLine]): DStream[(String, Long)] = {! source.! filter(data =&gt; {! data.uri == "/strbeacon" &amp;&amp; data.beaconType == "visible"! }).! map(data =&gt; (data.pkey, 1L)).! reduceByKey(_ + _)! }! ! type safety! from! case class </li> <li> Sinks ! ! class RedisSink @Inject()(store: RedisStore) {! ! def sink(result: DStream[(String, Long)]) = {! result.foreachRDD { rdd =&gt;! rdd.foreach { element =&gt;! val (key, value) = element! store.merge(key, value)! }! }! }! }! ! custom! sinks for! new stores </li> <li> Jobs ! object ImpressionsForPlacements {! ! def run(config: Config, inputPath: String) {! val conf = new SparkConf().! setMaster(config.getString("master")).! setAppName("Impressions for Placement")! ! val sc = new SparkContext(conf)! val ssc = new StreamingContext(sc, Seconds(5))! ! val source = BeaconLogLine.newDStream(ssc, inputPath)! val visible = visibleByPlacement(source)! sink(visible)! ! ssc.start! ssc.awaitTermination! }! }! source transform sink </li> <li> Advantages? </li> <li> Code Reuse ! object PlacementVisibles {! ! val source = BeaconLogLine.newDStream(ssc, inputPath)! val visible = visibleByPlacement(source)! sink(visible)! ! }! ! ! ! object PlacementEngagements {! ! val source = BeaconLogLine.newDStream(ssc, inputPath)! val engagements = engagementsByPlacement(source)! sink(engagements)! ! } composable! jobs </li> <li> Readability ! ssc.textFileStream(inputPath).! map { parseRawBeacon(_) }.! filter(data =&gt; {! data._2 == "/strbeacon" &amp;&amp; data._3 == "visible"! }).! map(data =&gt; (data._4, 1L)).! reduceByKey(_ + _).! foreachRDD { rdd =&gt;! rdd.foreach { element =&gt;! store.merge(element._1, element._2)! }! }! ? </li> <li> Readability ! ! val source = BeaconLogLine.newDStream(ssc, inputPath)! val visible = visibleByPlacement(source)! redis.sink(visible)! ! </li> <li> Testing def assertTransformation[T: Manifest, U: Manifest](! transformation: T =&gt; U,! input: Seq[T],! expectedOutput: Seq[U]! ): Unit = {! val ssc = new StreamingContext("local[1]", "Testing", Seconds(1))! val source = ssc.queueStream(new SynchronizedQueue[RDD[T]]())! val results = transformation(source)! ! var output = Array[U]()! results.foreachRDD { rdd =&gt; output = output ++ rdd.collect() }! ssc.start! rddQueue += ssc.sparkContext.makeRDD(input, 2)! Thread.sleep(jobCompletionWaitTimeMillis)! ssc.stop(true)! ! assert(output.toSet === expectedOutput.toSet)! }! function,! input,! expectation test </li> <li> Testing ! ! test("#visibleByPlacement") {! ! val input = Seq(! "pkey=abcd, ",! "pkey=abcd, ",! "pkey=wxyz, ",! )! ! val expectedOutput = Seq( ("abcd",2),("wxyz", 1) )! ! assertTransformation(visibleByPlacement, input, expectedOutput)! }! ! use our! test helper </li> <li> Other Learnings </li> <li> Other Learnings Keeping your driver program healthy is crucial 24/7 operation and monitoring Spark on Mesos? Use Marathon. Pay attention to settings for spark.cores.max Monitor data rate and increase as needed Serialization on classes Java Kryo </li> <li> Whats next? </li> <li> Twitter Summingbird Write-once, run anywhere Supports: Hadoop MapReduce Storm Spark (maybe?) </li> <li> Amazon Kinesis web! server Kinesis web! server web! server other! applications mobile! device app! logs </li> <li> Thanks! </li> </ul>


View more >