Building Feature Pipelines
When to use
- Engineering features for ML models from warehouse/stream data.
- Preventing label leakage and train/serve skew.
- Setting up a feature store, online serving, or historical backfills.
- Do NOT use for general modeling/aggregation (use dbt/Spark skills) unless it feeds ML features.
Workflow
- [ ] Define each feature with an entity key and an event timestamp
- [ ] Build training sets with point-in-time-correct joins (as-of the label time)
- [ ] Share ONE definition for offline (training) and online (serving)
- [ ] Set freshness/materialization for online features
- [ ] Backfill historical features idempotently for training
- Point-in-time correctness. Join features as of each label's timestamp — use only data that was known before the prediction time. This prevents label leakage, the most damaging feature bug.
- Offline/online parity. Compute a feature the same way for training (offline, batch) and serving (online, low-latency). Divergent logic causes train/serve skew and silent production degradation.
- Freshness. Online features must be materialized on a schedule that meets the model's staleness tolerance.
- Idempotent backfills. Recomputing historical features must be repeatable
(see
designing-backfills-and-replays).
Patterns
Point-in-time (as-of) join — pick the latest feature value strictly before each label event:
SELECT l.entity_id, l.label_ts, f.value AS feature
FROM labels l
LEFT JOIN features f
ON f.entity_id = l.entity_id
AND f.event_ts <= l.label_ts -- only past data; no leakage
QUALIFY ROW_NUMBER() OVER (
PARTITION BY l.entity_id, l.label_ts ORDER BY f.event_ts DESC) = 1;
Single definition, two paths — define the feature once (e.g. Feast
FeatureView); materialize to an offline store for training and an online store
for serving so both use identical logic.
Freshness + materialization — schedule online materialization; monitor feature
freshness like any dataset SLA (implementing-pipeline-observability).
Common pitfalls
- Label leakage — joining features computed after the label time inflates offline metrics and collapses in production; always as-of join.
- Train/serve skew — separate offline and online implementations drift; share one definition.
- Stale online features — model serves on old values; set and monitor freshness.
- Non-reproducible backfills — inconsistent training history; make feature recomputation idempotent.
- No entity/timestamp keys — features can't be joined correctly across time; require both.
- Unversioned features — silent redefinition breaks model comparability; version feature definitions.