Kinesis Application State in Dynamo DB

[VikramBPurohit]

Hi @itsvikramagr,

Thanks again for writing this library and providing constant support to community.

Our spark structure streaming code doesn't need to be stateful. We are ok with cluster or executors going down as long as Kinesis is managing consumer's application state.

https://docs.aws.amazon.com/streams/latest/dev/kinesis-record-processor-ddb.html

Could you shed some light on -

1. Why Kinesis isn't able to manage offset/application state in dynamo db although kinesis sql is calling describe stream API?
2. I don't see any place in kinesis-sql code which is setting application name for Kinesis.

Thanks in advance for your response!

[itsvikramagr]

Kinesis SQL library does not use KCL for most of its working. KCL and structured streaming APIs does not go along so well. Also, KCL comes with an extra cost of DynamoDB which can be avoided for structured streaming

But at the same time, while designing kinesis-sql, we knew that users would want more reliable shard progress committer. Hence the module provides an option of pluggable committer. We can do the following to use another committer

• Implement https://github.com/qubole/kinesis-sql/blob/master/src/main/scala/org/apache/spark/sql/kinesis/MetadataCommitter.scala
• Following code only consider HDFS as valid committer type. we can change it to use new DynamoDB based committer.

private def metadataCommitter: MetadataCommitter[ShardInfo] = {
    metaDataCommitterType.toLowerCase(Locale.ROOT) match {
      case "hdfs" =>
        new HDFSMetadataCommitter[ ShardInfo ](metaDataCommitterPath, hadoopConf(sqlContext))
      case _ => throw new IllegalArgumentException("only HDFS is supported")
    }
  } 

private def metaDataCommitterType: String = {
    sourceOptions.getOrElse("executor.metadata.committer", "hdfs").toString
  } 

I am relying on the community to implement dynamodb committer. cc @VikramBPurohit