Feature Store

Scattered CSV files with undocumented columns create silent schema drift - yesterday's momentum column used a 60-day window, today's uses 20 days, and nothing recorded the change. A structured feature store prevents this.

The Problem

Without versioned storage, feature definitions drift silently. A researcher recomputes features with different parameters, overwrites the file, and downstream models train on inconsistent data. Point-in-time correctness is also lost: loading features "as of January 2024" returns data that was computed using information from March 2024.

The Pattern

WRONG

import polars as pl

# Unversioned, unstructured, no metadata - silent drift guaranteed
features.write_csv("features.csv")  # What version? What parameters? When computed?
# Later: someone overwrites with different parameters
features_v2.write_csv("features.csv")  # Old version gone forever