7. Extending Open Source feature Store, Joost & Thijs, Adyen

https://www.youtube.com/watch?v=Yidl2VmrgVs&ab_channel=Tecton (opens in a new tab)

• Use cases

  • Fraud classification
  • Predict best time to retry a failed payment
  • Detect illegal activity on the platform

• Open source feast-spark-offline-store and move it into the feast project

  

  • Very custom solution

• In house build package, in early versions Spark was supported but it became incompatible, hence the library.

  • Offline features are mainly aggregations over day ranges
  • A lot of internal ETL pipelines use PySpark
  • SparkOfflineStore as the interface to query existing feature tables, following BigQuery and Redshift template