Streaming Deep Learning Scenarios with Flink
As a low-latency streaming tool, Flink offers the possibility of using machine learning, even "deep learning" (neural networks), with low latency. The growing FlinkML library provides some of the infrastructure support required for this goal, combined with third-party tools. This talk is a progress report on several scenarios we are developing at Lightbend, which combine Flink, Deeplearning4J, Spark, and Kafka to analyze cluster telemetry for anomaly detection, predictive autoscaling, and other scenarios. I'll focus on the pragmatics of training deep learning models in a streaming context, using batch and mini-batch training, combined with low-latency application of those models. I'll discuss the architecture we're using and highlight trade offs of particular tools for certain design problems in the implementation. I'll discuss the drawbacks and workarounds of our design and finish with a look at how future developments in Flink could improve its support for scenarios like ours.