Feature Versioning and Backfill Safety
Feature versioning tracks exactly which definition of a feature produced which historical values, so that fixing or improving a feature's logic later doesn't silently corrupt past backtests or retrain models on data that never really existed.
Prerequisites: Point-in-Time Feature Computation Engines
A quant team fixes a bug in how a feature computes trailing volatility, recomputes it for all history, and overwrites the old values. Every backtest ever run on that feature is now silently invalid, because it was tested against numbers that no longer exist. Feature versioning prevents this by tagging every feature definition with a version identifier and never overwriting old values in place — a changed calculation becomes a new version, stored alongside the old one.
Versioning a feature means treating any change to its calculation as a brand-new feature with its own history, rather than an in-place edit — so a backtest can always be reproduced against exactly the definition it was originally run with.
Backfill safety is the related discipline of making sure that when a new feature version is computed retroactively over historical dates, it only uses information that would genuinely have been available on each historical date — not, for instance, an updated data vendor field that wasn't actually published until later.
Worked example. A momentum feature, version 1, used a 20-day lookback. An analyst changes it to a 60-day lookback for better performance. Rather than overwriting the stored values, the pipeline creates "momentum_v2" with its own backfilled history, computed strictly from data available as of each historical date, while "momentum_v1" remains untouched so any model or backtest still referencing it keeps producing the same results it always did.
Without this discipline, a research team can spend months debugging performance regressions that turn out to be nothing more than a feature quietly changing definition underneath an unchanged model.
Further reading
- Zaharia et al., 'Feature Stores for ML' (various vendor engineering blogs)