Skip to content

80% of Feature Store Tables Used for Single Projects: A Data Scientist’s Perspective

A data scientist at a company using DBT extensively for data transformation shares insights on the utility of feature stores. While feature stores are designed to promote reusability of complex logic across projects, the reality often differs. In this company, 80% of the tables in the feature store are used for just one project. This leads to a complex hierarchy of tables, causing issues like data leakage and ambiguity. The onboarding process is challenging due to this complexity. Although pre-computing calculations can be beneficial for real-time inference, this need is exceptional rather than the norm. Most models operate on scheduled runs, suggesting that the extensive infrastructure might not justify its complexity, as it primarily prevents the need to copy and paste a few dozen lines of SQL.

Source: www.reddit.com

Related Videos