Skip to main content
Instruction and application
In Progress

Feature engineering in production

Good features are only useful if they remain stable, reproducible and available when a model is live. Production feature engineering is about operational discipline as much as modelling skill.

Microscope illustration

Building scalable feature pipelines​

Production pipelines automate extraction, transformation and delivery of features for both training and inference.

  • ETL workflows for batch processing
  • Streaming features for low-latency use cases
  • Modular transformations that can be reused and updated safely

Feature stores and management​

Feature stores centralise feature definitions and make them reusable across multiple models.

  • Online stores support real-time inference
  • Offline stores support historical analysis and retraining
  • Governance controls track ownership, access and dependencies

Monitoring feature drift​

Feature drift occurs when the distribution of production data changes over time. Without monitoring, models can silently degrade.

  • Use statistical checks such as KL divergence, KS tests or PSI
  • Set thresholds and alerts for unusual distribution shifts
  • Pair drift detection with retraining and feature review workflows

Handling feature versioning​

Versioning lets teams update features without breaking downstream models.

  • Tag feature definitions clearly
  • Keep rollback paths when new versions underperform
  • Maintain audit logs so experiments and production states can be reproduced
Lightbulb icon

Production mindset​

A feature is not finished when it improves offline accuracy. It is finished when it can be computed reliably, served consistently and monitored safely over time.

How can a feature store help ensure consistent customer recommendations across multiple models?
0 words
Would real-time computation improve personalisation, or would batch processing be more effective? How might a hybrid approach enhance accuracy?
0 words
What strategies would you use to detect feature drift in customer shopping behaviour, and how would you adapt your feature transformations to maintain relevance?
0 words
In progress