Presentation: Custom, Complex Windows @Scale Using Apache Flink

Recently, I’ve primarily been building data platforms. That is, platforms to enable Data and Software Engineers to collect and process data.

That's been the focus of my work since about 2013, which recently led me to working with Apache Flink.

We have a Spark installation here, and naturally Spark Streaming. A team also built a cloud native stream processing system from the ground up called Mantis. The platform that our data pipeline was built on used Kafka and Samza as a pair. Samza was primarily used just for very light use cases - very simple transformations. It was also used to route messages from Kafka to different sinks. It was a natural next step to introduce richer stream processing and offer that to our consumers. That's when we started to look at different technologies. Given all of the use cases that we wanted to support, Apache Flink was the best fit.

Flink won out on several fronts. Key amongst them were: native streaming (event by event processing), an amazing asynchronous distributed checkpointing mechanism which enables robust recovery, and a really great network stack with elegant back-pressure. Another feature, and the subject of my talk, is the window API that allows you to implement very nuanced, featureful windowing capabilities that aren't there out of the box.

The talk targets Data and Software Engineers who want an advanced talk on windowing. I will, however, offer some introductory concepts so this talk will be accessible to an eager learner.

I'd like them to come away with an understanding of a key advantage of Apache Flink, and when it's appropriate for them. Many people are already aware what the other choices offer (e.g., Spark Streaming, Samza, Trident/Heron, etc.). Apache Flink is seen as relatively new, but it's been around for a long time (since 2008). In terms of getting mindshare though, it's relatively new on the scene. So, I’m also hoping to help awareness and adoption increase.

I'll essentially present Apache Flink windowing capabilities. The subject of the talk is “Bounding a Window by Types of Events.” This requires a system that supports unaligned windows and has an extensible window API. Most of the other stream processing systems offer predefined window types that are aligned - by aligned I mean that the window fires at a regular interval aligned on wall clock time.

With all of the data that we have available flowing from many devices and systems, the big challenge at this point is extracting useful information. So, creating platforms for people in the data engineering and data science field to be able to do that with ease - to really allow them to focus as much on their problem and their domain, rather than on the nuts and bolts of the infrastructure pieces… this is something that is a strong focus for me at this point.