Historically, machine learning training pipelines have been heavily utilizing batch training models, i.e., getting retrained every few hours. However, industrial practitioners have proved that real-time training can yield a more adaptive and personalized user experience. The transition from batch to real-time is full of tradeoffs to get the benefits of accuracy and freshness while keeping the costs low and having a predictable, maintainable system.

This session will delve deeper into our journey of migrating to a streaming pipeline for our ML models using Kafka and Flink. You will learn how to transition from Pub/Sub to Kafka for incoming real-time events and leverage Flink for streaming joins using RocksDB and checkpointing. We will also discuss navigating nuances like causal dependency between events, event-time versus processing time, and exactly-once vs atleast-once delivery, among others.

Furthermore, you will see how we utilized topic partitioning in Kafka to improve scalability, reduced the throughput of events by 85% using Avro schema and compression, decreased cost by 40%, and set up a separate pipeline to ensure correctness.

Attending this session, you'll gain a deeper understanding of the tradeoffs and nuances in real-time systems, allowing you to make well-informed decisions that suit your specific requirements.