Build a reusable preprocessing artifact

Training code and serving code drift when they reimplement imputation, scaling, category order and model scoring separately. A fitted pipeline packages the transformations and estimator into one versioned object.

Test roundtrip equivalence

Our small pipeline sends numeric amount through median imputation and standardization, channel through one-hot encoding with unknown-ignore, and the combined features into logistic regression.

The probe has two rows: one with unseen phone channel and one with missing amount. We serialize the trusted in-memory object and restore it locally. Maximum probability difference before and after the roundtrip is exactly zero. Both probe probability vectors remain finite, and restored classes are [0,1].

The ColumnTransformer documentation defines column-specific transformations, while Pipeline chains preprocessing and prediction.

from feature_cases import artifact_case

r = artifact_case()
assert r['probe_rows'] == 2
assert r['maximum_roundtrip_probability_difference'] == 0
assert r['classes'] == [0, 1]
assert r['unknown_category_probe_finite']
print(r)

Run the roundtrip in the feature-engineering lab. It uses a trusted local pickle byte stream. Loading untrusted pickle files can execute code and is outside this evidence.

Package more than model weights

Store input schema, feature order, fitted categories, software versions, training-data manifest, model version and expected outputs. Add smoke cases for missing, unknown, boundary and malformed inputs. Hash artifacts and verify them before loading through a trusted release path.

Cross-version compatibility is not guaranteed for every serialization format. Rebuild in a locked clean environment or choose an appropriate portable representation, then test numerical equivalence. Document platform constraints.

Serving should call the same predict_proba path used in offline evaluation. Log model version and safe validation outcomes. Avoid logging sensitive raw features by default.

A successful roundtrip is necessary but limited. It does not verify the source query, point-in-time features, latency, concurrency or monitoring. Add integration tests for those layers.

Exercise: write the artifact to a temporary trusted path, record its hash and load it in a clean environment with pinned dependencies. Compare transformed features and probabilities, then deliberately reorder input columns and confirm name-based selection preserves or rejects the request as designed.

NeuraPath's Data Science course connects notebook results to reproducible serving artifacts. Prediction equivalence is a release check, not an assumption.

Continue learning

This article is part of the Feature engineering and data quality sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in Text features: control vocabulary and fit boundaries.
• Continue with Audit proxy variables in a predictive model.