Apache: Big Data North America 2017

We discuss lessons learned from our first production deployment of a Spark Streaming pipeline for processing game data. Deployment is to the AWS Cloud where we use managed services (i.e. EMR, S3 and Redshift). However, having downstream dependencies with outages and unpredictable response latencies can pose significant challenges. To address, we evolved the architecture by separating data processing from post-processing tasks (i.e. copying data into Redshift). Post-processing tasks are sent downstream from Spark to a task executor that was built using Akka Streams and Reactive Kafka. The end result is a loosely coupled architecture where the Spark streaming job is a firehose to S3 and fault-tolerant when Redshift is unavailable.