How to create a KinesisSink for Spark Structural Streaming

Question

I am using Spark 2.2 on Databricks and trying to implement a Kinesis sink to write from Spark to a Kinesis stream.

I am using the following provide sample from here https://docs.databricks.com/_static/notebooks/structured-streaming-kinesis-sink.html

/**
* A simple Sink that writes to the given Amazon Kinesis `stream` in the given    `region`. For authentication, users may provide
* `awsAccessKey` and `awsSecretKey`, or use IAM Roles when launching their cluster.
*
* This Sink takes a two column Dataset, with the columns being the `partitionKey`, and the `data` respectively.
* We will buffer data up to `maxBufferSize` before flushing to Kinesis in order to reduce cost.
*/
class KinesisSink(
    stream: String,
    region: String,
    awsAccessKey: Option[String] = None,
    awsSecretKey: Option[String] = None) extends ForeachWriter[(String, Array[Byte])] { 

 // Configurations
 private val maxBufferSize = 500 * 1024 // 500 KB

 private var client: AmazonKinesis = _
 private val buffer = new ArrayBuffer[PutRecordsRequestEntry]()
 private var bufferSize: Long = 0L

 override def open(partitionId: Long, version: Long): Boolean = {
   client = createClient
   true
 }

 override def process(value: (String, Array[Byte])): Unit = {
   val (partitionKey, data) = value
   // Maximum of 500 records can be sent with a single `putRecords` request
   if ((data.length + bufferSize > maxBufferSize && buffer.nonEmpty) ||    buffer.length == 500) {
  flush()
}
buffer += new PutRecordsRequestEntry().withPartitionKey(partitionKey).withData(ByteBuffer.wrap(data))
bufferSize += data.length
 }

 override def close(errorOrNull: Throwable): Unit = {
   if (buffer.nonEmpty) {
     flush()
   }
   client.shutdown()
 }

 /** Flush the buffer to Kinesis */
 private def flush(): Unit = {
   val recordRequest = new PutRecordsRequest()
     .withStreamName(stream)
     .withRecords(buffer: _*)

   client.putRecords(recordRequest)
   buffer.clear()
   bufferSize = 0
 }

 /** Create a Kinesis client. */
 private def createClient: AmazonKinesis = {
   val cli = if (awsAccessKey.isEmpty || awsSecretKey.isEmpty) {
     AmazonKinesisClientBuilder.standard()
       .withRegion(region)
       .build()
   } else {
     AmazonKinesisClientBuilder.standard()
       .withRegion(region)
       .withCredentials(new AWSStaticCredentialsProvider(new BasicAWSCredentials(awsAccessKey.get, awsSecretKey.get)))
       .build()
   }
   cli
 }
}

I then implement the KinesisSink class using

val kinesisSink = new KinesisSink("us-east-1", "MyStream", Option("xxx..."), Option("xxx..."))

Finally i create a stream using this sink. This KinesisSink takes a two column Dataset, with the columns being the partitionKey , and the data respectively.

case class MyData(partitionKey: String, data: Array[Byte])

val newsDataDF = kinesisDF
   .selectExpr("apinewsseqid", "fullcontent").as[MyData]
   .writeStream
   .outputMode("append")
   .foreach(kinesisSink)
   .start

but i am still getting the following error

error: type mismatch;
found   : KinesisSink
required: org.apache.spark.sql.ForeachWriter[MyData]
   .foreach(kinesisSink)

Answer 1

You need to change the signature of KinesisSink.process method, it should take your custom MyData object and then extract partitionKey and data from there.

Answer 2

I used the exact same KinesisSink provided by databricks and got it working by creating dataset using

val dataset = df.selectExpr("CAST(rand() AS STRING) as partitionKey","message_bytes").as[(String, Array[Byte])]

And used the dataset to write to kinesis stream

val query = dataset
.writeStream
.foreach(kinesisSink)
.start()
.awaitTermination()

While I used dataset.selectExpr("partitionKey","message_bytes") I got the same type mismatch error:

error: type mismatch;
found: KinesisSink
required: org.apache.spark.sql.ForeachWriter[(String, Array[Byte])]
.foreach(kinesisSink)

selectExpr is not needed in this case since the dataset drives the data type of the ForeachWriter.