Conclusion : Streaming model workflows

Streaming model pipelines are useful for systems that need to apply ML models in real-time. To build these types of pipelines, we explored two message brokers that can scale to large volumes of events and provide data sources and data sinks for these pipelines. Streaming pipelines often constrain the types of operations you can perform due to latency requirements. For example, it would be challenging to build a streaming pipeline that performs feature generation on user data because historic data would need to be retrieved and combined with the streaming data while maintaining low latency. There are patterns for achieving this type of result, such as precomputing aggregates for a user and storing the data in an application database. However, it can be significantly more work getting this type of pipeline to work in a streaming mode versus a batch mode.