Hello, I am building a Spark streaming application that ingests data from an Amazon Kinesis stream. My application keeps track of the minimum price over a window for groups of similar tickets. When I deploy the application, I would like it to start processing at the start of the previous hours data. This will warm up the state of the application and allow us to deploy our application faster. For example, if I start the application at 3 PM, I would like to process the data retained by Kinesis from 2PM to 3PM, and then continue receiving data going forward. Spark Streaming’s Kinesis receiver, which relies on the Amazon Kinesis Client Library, seems to give me three options for choosing where to read from the stream: read from the latest checkpointed sequence number in Dynamo start from the oldest record in the stream (TRIM_HORIZON shard iterator type) start from the most recent record in the stream (LATEST shard iterator type)
Do you have any suggestions on how we could start our application at a specific timestamp or sequence number in the Kinesis stream? Some ideas I had were: Create a KCL application that fetches the previous hour data and writes it to HDFS. We can create an RDD from that dataset and initialize our Spark Streaming job with it. The spark streaming job’s Kinesis receiver can have the same name as the initial KCL application, and use that applications checkpoint as the starting point. We’re writing our spark jobs in Python, so this would require launching the java MultiLang daemon, or writing that portion of the application in Java/Scala. Before the Spark streaming application starts, we could fetch a shard iterator using the AT_TIMESTAMP shard iterator type. We could record the sequence number of the first record returned by this iterator, and create an entry in Dynamo for our application for that sequence number. Our Kinesis receiver would pick up from this checkpoint. It makes me a little nervous that we would be faking Kinesis Client Library's protocol by writing a checkpoint into Dynamo Thanks in advance! Neil
