Scalable, real-time Machine Learning using Apache Kafka 

Agenda ● Traditional model deployment process ● 90 seconds to WoW ● Let’s process the incoming stream ● Demo ● What’s more? 2 

$ whoami ● Personalisation lead at Hotstar ● Led Data Infrastructure team at Grofers and TinyOwl ● Kafka fanboy ● Usually rant on twitter @jayeshsidhwani 3 

Machine Learning @ Hotstar ● ~150 mn users ● 4.8 mn peak concurrency ● 120K peak recommendation requests per second ● Diverse content in diverse languages 4 

Traditional model deployment process 5 Model Training Data Lake Serialized Model Batch Predictions Recommendation APIs Offline Online ● One-day / few-hours batch pre-compute ● Slow time to react 

Sense of urgency? 6 ● 90 seconds to convert a new user ● To power his experience, we need to know user’s gender, interests and more ● Need an always-thinking machine 

Thinking streams 7 Data at Rest Data in motion ● Slow ● Batch-y ● Fast ● Sub-second 

Enter Apache Kafka 8 ● Kafka is a scalable, fault-tolerant, distributed message queue ● Producers and Consumers ● Uses ○ Real-time applications such as: intelligent notifications, anomaly etc. ○ Asynchronous communication in event-driven architectures Diagram credits: http://kafka.apache.org 

Real-time infrastructure at Hotstar 9 ● All clickstream data pushed into Apache Kafka ● Apache Kafka Streams to process events as they happen ● Incoming data available for everyone Intelligence Apple TV iOS ANDROID Roku STREAM PROCESSING FRAMEWORK Filter Window Join Anomaly Machine Learning 

Demo Predict whether a flight is delayed in real-time 10 

How to process a stream? 11 ML 

Advanced use-cases 12 page-clicks Processor nodes Source / Sink nodes video-plays predict-gender predict-interest 5-min trending videos Recommended for You Hotstar Streaming Platform 

Questions? 13 tech.hotstar.com