You don't need to. Just sample 1% of data, let the model do the feature engineering, ask model model to replicate everything in Pandas (or R, or whatever else), so you can run it for a full dataset.