Skip to content

Commit

Permalink
Merge remote-tracking branch 'upstream/master' into mllib-stats-api-c…
Browse files Browse the repository at this point in the history 
…heck
  • Loading branch information
@jkbradley
jkbradley committed Aug 18, 2014
2 parents 4e5d15e + c77f406 commit b20d90a
Show file tree
Hide file tree
Showing 10 changed files with 166 additions and 105 deletions.
97 changes: 49 additions & 48 deletions docs/streaming-kinesis.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -3,56 +3,57 @@ layout: global
title: Spark Streaming Kinesis Receiver
---

### Kinesis
Build notes:
<li>Spark supports a Kinesis Streaming Receiver which is not included in the default build due to licensing restrictions.</li>
<li>_**Note that by embedding this library you will include [ASL](https://aws.amazon.com/asl/)-licensed code in your Spark package**_.</li>
<li>The Spark Kinesis Streaming Receiver source code, examples, tests, and artifacts live in $SPARK_HOME/extras/kinesis-asl.</li>
<li>To build with Kinesis, you must run the maven or sbt builds with -Pkinesis-asl`.</li>
<li>Applications will need to link to the 'spark-streaming-kinesis-asl` artifact.</li>
## Kinesis
###Design
<li>The KinesisReceiver uses the Kinesis Client Library (KCL) provided by Amazon under the Amazon Software License.</li>
<li>The KCL builds on top of the Apache 2.0 licensed AWS Java SDK and provides load-balancing, fault-tolerance, checkpointing through the concept of Workers, Checkpoints, and Shard Leases.</li>
<li>The KCL uses DynamoDB to maintain all state. A DynamoDB table is created in the us-east-1 region (regardless of Kinesis stream region) during KCL initialization for each Kinesis application name.</li>
<li>A single KinesisReceiver can process many shards of a stream by spinning up multiple KinesisRecordProcessor threads.</li>
<li>You never need more KinesisReceivers than the number of shards in your stream as each will spin up at least one KinesisRecordProcessor thread.</li>
<li>Horizontal scaling is achieved by autoscaling additional KinesisReceiver (separate processes) or spinning up new KinesisRecordProcessor threads within each KinesisReceiver - up to the number of current shards for a given stream, of course. Don't forget to autoscale back down!</li>

Kinesis examples notes:
<li>To build the Kinesis examples, you must run the maven or sbt builds with -Pkinesis-asl`.</li>
<li>These examples automatically determine the number of local threads and KinesisReceivers to spin up based on the number of shards for the stream.</li>
<li>KinesisWordCountProducerASL will generate random data to put onto the Kinesis stream for testing.</li>
<li>Checkpointing is disabled (no checkpoint dir is set). The examples as written will not recover from a driver failure.</li>
### Build
<li>Spark supports a Streaming KinesisReceiver, but it is not included in the default build due to Amazon Software Licensing (ASL) restrictions.</li>
<li>To build with the Kinesis Streaming Receiver and supporting ASL-licensed code, you must run the maven or sbt builds with the **-Pkinesis-asl** profile.</li>
<li>All KinesisReceiver-related code, examples, tests, and artifacts live in **$SPARK_HOME/extras/kinesis-asl/**.</li>
<li>Kinesis-based Spark Applications will need to link to the **spark-streaming-kinesis-asl** artifact that is built when **-Pkinesis-asl** is specified.</li>
<li>_**Note that by linking to this library, you will include [ASL](https://aws.amazon.com/asl/)-licensed code in your Spark package**_.</li>

Deployment and runtime notes:
<li>A single KinesisReceiver can process many shards of a stream.</li>
<li>Each shard of a stream is processed by one or more KinesisReceiver's managed by the Kinesis Client Library (KCL) Worker.</li>
<li>You never need more KinesisReceivers than the number of shards in your stream.</li>
<li>You can horizontally scale the receiving by creating more KinesisReceiver/DStreams (up to the number of shards for a given stream)</li>
<li>The Kinesis libraries must be present on all worker nodes, as they will need access to the Kinesis Client Library.</li>
<li>This code uses the DefaultAWSCredentialsProviderChain and searches for credentials in the following order of precedence:<br/>
1) Environment Variables - AWS_ACCESS_KEY_ID and AWS_SECRET_KEY<br/>
2) Java System Properties - aws.accessKeyId and aws.secretKey<br/>
3) Credential profiles file - default location (~/.aws/credentials) shared by all AWS SDKs<br/>
4) Instance profile credentials - delivered through the Amazon EC2 metadata service<br/>
</li>
<li>You need to setup a Kinesis stream with 1 or more shards per the following:<br/>
http://docs.aws.amazon.com/kinesis/latest/dev/step-one-create-stream.html</li>
<li>Valid Kinesis endpoint urls can be found here: Valid endpoint urls: http://docs.aws.amazon.com/general/latest/gr/rande.html#ak_region</li>
<li>When you first start up the KinesisReceiver, the Kinesis Client Library (KCL) needs ~30s to establish connectivity with the AWS Kinesis service,
retrieve any checkpoint data, and negotiate with other KCL's reading from the same stream.</li>
<li>Be careful when changing the app name. Kinesis maintains a mapping table in DynamoDB based on this app name (http://docs.aws.amazon.com/kinesis/latest/dev/kinesis-record-processor-implementation-app.html#kinesis-record-processor-initialization).
Changing the app name could lead to Kinesis errors as only 1 logical application can process a stream. In order to start fresh,
it's always best to delete the DynamoDB table that matches your app name. This DynamoDB table lives in us-east-1 regardless of the Kinesis endpoint URL.</li>
###Example
<li>To build the Kinesis example, you must run the maven or sbt builds with the **-Pkinesis-asl** profile.</li>
<li>You need to setup a Kinesis stream at one of the valid Kinesis endpoints with 1 or more shards per the following: http://docs.aws.amazon.com/kinesis/latest/dev/step-one-create-stream.html</li>
<li>Valid Kinesis endpoints can be found here: http://docs.aws.amazon.com/general/latest/gr/rande.html#ak_region</li>
<li>When running **locally**, the example automatically determines the number of threads and KinesisReceivers to spin up based on the number of shards configured for the stream. Therefore, **local[n]** is not needed when starting the example as with other streaming examples.</li>
<li>While this example could use a single KinesisReceiver which spins up multiple KinesisRecordProcessor threads to process multiple shards, I wanted to demonstrate unioning multiple KinesisReceivers as a single DStream. (It's a bit confusing in local mode.)</li>
<li>**KinesisWordCountProducerASL** is provided to generate random records into the Kinesis stream for testing.</li>
<li>The example has been configured to immediately replicate incoming stream data to another node by using (StorageLevel.MEMORY_AND_DISK_2)
<li>Spark checkpointing is disabled because the example does not use any stateful or window-based DStream operations such as updateStateByKey and reduceByWindow. If those operations are introduced, you would need to enable checkpointing or risk losing data in the case of a failure.</li>
<li>Kinesis checkpointing is enabled. This means that the example will recover from a Kinesis failure.</li>
<li>The example uses InitialPositionInStream.LATEST strategy to pull from the latest tip of the stream if no Kinesis checkpoint info exists.</li>
<li>In our example, **KinesisWordCount** is the Kinesis application name for both the Scala and Java versions. The use of this application name is described next.</li>

Failure recovery notes:
<li>The combination of Spark Streaming and Kinesis creates 3 different checkpoints as follows:<br/>
1) RDD data checkpoint (Spark Streaming) - frequency is configurable with DStream.checkpoint(Duration)<br/>
2) RDD metadata checkpoint (Spark Streaming) - frequency is every DStream batch<br/>
3) Kinesis checkpointing (Kinesis) - frequency is controlled by the developer calling ICheckpointer.checkpoint() directly<br/>
###Deployment and Runtime
<li>A Kinesis application name must be unique for a given account and region.</li>
<li>A DynamoDB table and CloudWatch namespace are created during KCL initialization using this Kinesis application name. http://docs.aws.amazon.com/kinesis/latest/dev/kinesis-record-processor-implementation-app.html#kinesis-record-processor-initialization</li>
<li>This DynamoDB table lives in the us-east-1 region regardless of the Kinesis endpoint URL.</li>
<li>Changing the app name or stream name could lead to Kinesis errors as only a single logical application can process a single stream.</li>
<li>If you are seeing errors after changing the app name or stream name, it may be necessary to manually delete the DynamoDB table and start from scratch.</li>
<li>The Kinesis libraries must be present on all worker nodes, as they will need access to the KCL.</li>
<li>The KinesisReceiver uses the DefaultAWSCredentialsProviderChain for AWS credentials which searches for credentials in the following order of precedence:</br>
1) Environment Variables - AWS_ACCESS_KEY_ID and AWS_SECRET_KEY<br/>
2) Java System Properties - aws.accessKeyId and aws.secretKey<br/>
3) Credential profiles file - default location (~/.aws/credentials) shared by all AWS SDKs<br/>
4) Instance profile credentials - delivered through the Amazon EC2 metadata service
</li>
<li>Checkpointing too frequently will cause excess load on the AWS checkpoint storage layer and may lead to AWS throttling</li>
<li>Upon startup, a KinesisReceiver will begin processing records with sequence numbers greater than the last checkpoint sequence number recorded per shard.</li>
<li>If no checkpoint info exists, the worker will start either from the oldest record available (InitialPositionInStream.TRIM_HORIZON)
or from the tip/latest (InitialPostitionInStream.LATEST). This is configurable.</li>
<li>When pulling from the stream tip (InitialPositionInStream.LATEST), only new stream data will be picked up after the KinesisReceiver starts.</li>
<li>InitialPositionInStream.LATEST could lead to missed records if data is added to the stream while no KinesisReceivers are running.</li>
<li>In production, you'll want to switch to InitialPositionInStream.TRIM_HORIZON which will read up to 24 hours (Kinesis limit) of previous stream data
depending on the checkpoint frequency.</li>
<li>InitialPositionInStream.TRIM_HORIZON may lead to duplicate processing of records depending on the checkpoint frequency.</li>

###Fault-Tolerance
<li>The combination of Spark Streaming and Kinesis creates 2 different checkpoints that may occur at different intervals.</li>
<li>Checkpointing too frequently against Kinesis will cause excess load on the AWS checkpoint storage layer and may lead to AWS throttling. The provided example handles this throttling with a random backoff retry strategy.</li>
<li>Upon startup, a KinesisReceiver will begin processing records with sequence numbers greater than the last Kinesis checkpoint sequence number recorded per shard (stored in the DynamoDB table).</li>
<li>If no Kinesis checkpoint info exists, the KinesisReceiver will start either from the oldest record available (InitialPositionInStream.TRIM_HORIZON) or from the latest tip (InitialPostitionInStream.LATEST). This is configurable.</li>
<li>InitialPositionInStream.LATEST could lead to missed records if data is added to the stream while no KinesisReceivers are running (and no checkpoint info is being stored.)</li>
<li>In production, you'll want to switch to InitialPositionInStream.TRIM_HORIZON which will read up to 24 hours (Kinesis limit) of previous stream data.</li>
<li>InitialPositionInStream.TRIM_HORIZON may lead to duplicate processing of records where the impact is dependent on checkpoint frequency.</li>
<li>Record processing should be idempotent when possible.</li>
<li>Failed or latent KinesisReceivers will be detected and automatically shutdown/load-balanced by the KCL.</li>
<li>If possible, explicitly shutdown the worker if a failure occurs in order to trigger the final checkpoint.</li>
<li>A failed or latent KinesisRecordProcessor within the KinesisReceiver will be detected and automatically restarted by the KCL.</li>
<li>If possible, the KinesisReceiver should be shutdown cleanly in order to trigger a final checkpoint of all KinesisRecordProcessors to avoid duplicate record processing.</li>
8 changes: 8 additions & 0 deletions external/flume-sink/src/main/scala/org/apache/spark/streaming/flume/sink/SparkSink.scala
View file
Original file line number Diff line number Diff line change
Expand Up @@ -131,6 +131,14 @@ class SparkSink extends AbstractSink with Logging with Configurable {
blockingLatch.await()
Status.BACKOFF
}

private[flume] def getPort(): Int = {
serverOpt
.map(_.getPort)
.getOrElse(
throw new RuntimeException("Server was not started!")
)
}
}

/**
Expand Down
55 changes: 26 additions & 29 deletions external/flume/src/test/scala/org/apache/spark/streaming/flume/FlumePollingStreamSuite.scala
View file
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,8 @@ import java.net.InetSocketAddress
import java.util.concurrent.{Callable, ExecutorCompletionService, Executors}
import java.util.Random

import org.apache.spark.TestUtils

import scala.collection.JavaConversions._
import scala.collection.mutable.{SynchronizedBuffer, ArrayBuffer}

Expand All @@ -39,9 +41,6 @@ import org.apache.spark.util.Utils

class FlumePollingStreamSuite extends TestSuiteBase {

val random = new Random()
/** Return a port in the ephemeral range. */
def getTestPort = random.nextInt(16382) + 49152
val batchCount = 5
val eventsPerBatch = 100
val totalEventsPerChannel = batchCount * eventsPerBatch
Expand Down Expand Up @@ -77,17 +76,6 @@ class FlumePollingStreamSuite extends TestSuiteBase {
}

private def testFlumePolling(): Unit = {
val testPort = getTestPort
// Set up the streaming context and input streams
val ssc = new StreamingContext(conf, batchDuration)
val flumeStream: ReceiverInputDStream[SparkFlumeEvent] =
FlumeUtils.createPollingStream(ssc, Seq(new InetSocketAddress("localhost", testPort)),
StorageLevel.MEMORY_AND_DISK, eventsPerBatch, 1)
val outputBuffer = new ArrayBuffer[Seq[SparkFlumeEvent]]
with SynchronizedBuffer[Seq[SparkFlumeEvent]]
val outputStream = new TestOutputStream(flumeStream, outputBuffer)
outputStream.register()

// Start the channel and sink.
val context = new Context()
context.put("capacity", channelCapacity.toString)
Expand All @@ -98,10 +86,19 @@ class FlumePollingStreamSuite extends TestSuiteBase {

val sink = new SparkSink()
context.put(SparkSinkConfig.CONF_HOSTNAME, "localhost")
context.put(SparkSinkConfig.CONF_PORT, String.valueOf(testPort))
context.put(SparkSinkConfig.CONF_PORT, String.valueOf(0))
Configurables.configure(sink, context)
sink.setChannel(channel)
sink.start()
// Set up the streaming context and input streams
val ssc = new StreamingContext(conf, batchDuration)
val flumeStream: ReceiverInputDStream[SparkFlumeEvent] =
FlumeUtils.createPollingStream(ssc, Seq(new InetSocketAddress("localhost", sink.getPort())),
StorageLevel.MEMORY_AND_DISK, eventsPerBatch, 1)
val outputBuffer = new ArrayBuffer[Seq[SparkFlumeEvent]]
with SynchronizedBuffer[Seq[SparkFlumeEvent]]
val outputStream = new TestOutputStream(flumeStream, outputBuffer)
outputStream.register()
ssc.start()

writeAndVerify(Seq(channel), ssc, outputBuffer)
Expand All @@ -111,18 +108,6 @@ class FlumePollingStreamSuite extends TestSuiteBase {
}

private def testFlumePollingMultipleHost(): Unit = {
val testPort = getTestPort
// Set up the streaming context and input streams
val ssc = new StreamingContext(conf, batchDuration)
val addresses = Seq(testPort, testPort + 1).map(new InetSocketAddress("localhost", _))
val flumeStream: ReceiverInputDStream[SparkFlumeEvent] =
FlumeUtils.createPollingStream(ssc, addresses, StorageLevel.MEMORY_AND_DISK,
eventsPerBatch, 5)
val outputBuffer = new ArrayBuffer[Seq[SparkFlumeEvent]]
with SynchronizedBuffer[Seq[SparkFlumeEvent]]
val outputStream = new TestOutputStream(flumeStream, outputBuffer)
outputStream.register()

// Start the channel and sink.
val context = new Context()
context.put("capacity", channelCapacity.toString)
Expand All @@ -136,17 +121,29 @@ class FlumePollingStreamSuite extends TestSuiteBase {

val sink = new SparkSink()
context.put(SparkSinkConfig.CONF_HOSTNAME, "localhost")
context.put(SparkSinkConfig.CONF_PORT, String.valueOf(testPort))
context.put(SparkSinkConfig.CONF_PORT, String.valueOf(0))
Configurables.configure(sink, context)
sink.setChannel(channel)
sink.start()

val sink2 = new SparkSink()
context.put(SparkSinkConfig.CONF_HOSTNAME, "localhost")
context.put(SparkSinkConfig.CONF_PORT, String.valueOf(testPort + 1))
context.put(SparkSinkConfig.CONF_PORT, String.valueOf(0))
Configurables.configure(sink2, context)
sink2.setChannel(channel2)
sink2.start()

// Set up the streaming context and input streams
val ssc = new StreamingContext(conf, batchDuration)
val addresses = Seq(sink.getPort(), sink2.getPort()).map(new InetSocketAddress("localhost", _))
val flumeStream: ReceiverInputDStream[SparkFlumeEvent] =
FlumeUtils.createPollingStream(ssc, addresses, StorageLevel.MEMORY_AND_DISK,
eventsPerBatch, 5)
val outputBuffer = new ArrayBuffer[Seq[SparkFlumeEvent]]
with SynchronizedBuffer[Seq[SparkFlumeEvent]]
val outputStream = new TestOutputStream(flumeStream, outputBuffer)
outputStream.register()

ssc.start()
writeAndVerify(Seq(channel, channel2), ssc, outputBuffer)
assertChannelIsEmpty(channel)
Expand Down
2 changes: 1 addition & 1 deletion mllib/src/main/scala/org/apache/spark/mllib/stat/Statistics.scala
View file
Original file line number Diff line number Diff line change
Expand Up @@ -155,7 +155,7 @@ object Statistics {
* :: Experimental ::
* Conduct Pearson's independence test for every feature against the label across the input RDD.
* For each feature, the (feature, label) pairs are converted into a contingency matrix for which
* the chi-squared statistic is computed.
* the chi-squared statistic is computed. All label and feature values must be categorical.
*
* @param data an `RDD[LabeledPoint]` containing the labeled dataset with categorical features.
* Real-valued features will be treated as categorical for each distinct value.
Expand Down
Loading

0 comments on commit b20d90a

Please sign in to comment.