If a feature is used in an online inference pipeline and it is created using data only available at request-time, then it is an on-demand feature. If you have historical data, you would like to be able to create the same feature values in your feature pipeline using the same code as is used in the online inference pipeline (to avoid online-offline skew).
On-demand features are typically used when real-time data is critical to the machine learning model's performance or when data availability constraints require the feature to be calculated at request-time.
In general, it is preferable to precompute features in feature pipelines and store them in a feature store, retrieving them when needed in your online inference pipeline. Having a single place (the feature pipeline) where features are computed prevents online-offline skew, and enables materialized feature values to be shared across multiple models, promoting collaboration and reducing duplication of effort. By centralizing and precomputing features, teams can maintain a single source of truth for features, which simplifies the development process and fosters consistency across projects. On-demand can lead to increased latency or computational overhead in the online inference pipeline, negatively impacting prediction latency. However, if the input data used to compute a feature is only available at request time, then you will need to implement an on-demand feature. For example, if you are predicting credit-card fraud online, some credit card transaction features (computed from the transaction purchase amount, category of purchase, location of purchase, timestamp for the purchase) will only be available as on-demand features.
.svg)