Model predictions are as good as the features you provide as input. Feature quality, feature freshness and latency SLA are critical to successfully deploy real-time machine learning systems. Hopsworks provides all the building blocks to effectively and safely deploy real-time machine learning systems. It provides an open API, based on DataFrames, for low latency ingestion of features that can be pre-computed using frameworks such as Spark Streaming, Apache Flink and Apache Beam. It provides the tools to build feature functions that can be executed on-demand using input from the request data and in batch to backfill feature data for training.

‍
Hopsworks also provides the necessary monitoring infrastructure to make sure the system is operating correctly and the model is emitting relevant predictions. In this live coding session we are going to explore:

• How to build and deploy feature pipelines on Hopsworks that provide sub-second feature freshness.
• How to retrieve features from the Hopsworks online feature store with single digit latency.
• How to compute features on-demand by combining input data from the request and data from the online feature store.
• How to leverage feature store change data capture to preemptively make predictions based on new data.
• How to monitor the machine learning system and receive alerts when the system encounters errors.