What is the best way to handle point-in-time joins using Feast for training data?

All answers to this question.

The point-in-time join is essentially the "killer feature" of this platform. When you call the historical feature method, you provide an entity dataframe containing the IDs and timestamps of your events. The system then looks back into your offline store—like Parquet files or Snowflake—to find the most recent feature values for each entity that existed before that specific timestamp. It’s all handled through a join logic that is notoriously difficult to write manually in SQL. It saved my team weeks of custom engineering and made our backtesting far more accurate.

   Answered 2025-09-23 by Sandra Hayes

How does the performance hold up when joining an entity dataframe with millions of rows against a very large offline feature table?

   Answered 2025-09-24 by Gregory Ward

• Gregory, it depends on your offline provider. If you use Spark or BigQuery as your provider, it pushes the join logic to the compute engine, which handles millions of rows easily. If you are running it locally on a small machine, you might hit memory limits, so definitely use a distributed backend for scale.

     Commented 2025-09-25 by Ronald Simmons

This functionality is critical for finance models where a few seconds of data leakage can lead to completely unrealistic backtest results.

   Answered 2025-09-26 by Deborah Ortiz

• You hit the nail on the head, Deborah. In fintech, temporal integrity is everything, and this tool makes it much easier to maintain.

     Commented 2025-09-27 by Sandra Hayes