Hybrid batch/real-time architectures (sometimes called “lambda architectures”) are a powerful pattern for building robust, production-quality, up-to-the-minute data analytics systems.

We’ll discuss why you may want to go hybrid, the sorts of challenges that can arise when building production data systems, and effective techniques for making them easier to deploy and manage. We’ll take the data system at Metamarkets as an example, which uses Hadoop, Storm, Kafka, and Druid to ingest over 10TB of new data every day and to offer the ability to query trillions of aggregated events.

We’ll talk about our experience running this system in production for the past year, with particular focus on data pipeline development and operations. We took two pronged approach involving both software and operations practices. The most helpful pieces of software so far have been a Scala library we developed to express common data processing needs, paired with an execution engine that can run those programs on both Hadoop and Storm. We'll cover its design and implementation and what kinds of patterns we've seen emerge in its usage. 

We'll also discuss operations practices, which we have found to be just as important as software. Ours center around a robust metrics collection system that feeds into alerting and metrics visualization. We'll talk about what kinds of metrics we've found most valuable, what kinds of things we alert on, and show some of the most useful visualizations.