A data pipeline moves and prepares data, and ends with tables that people or dashboards use. An ML pipeline does that too, and then continues: it trains a model on the data, evaluates it, stores it, deploys it, and monitors it after release. So an ML pipeline contains a data pipeline as its first part, and then adds model-specific steps.
Steps that are unique to ML
- Feature engineering: turning raw data into model inputs (counts over windows, encodings, aggregates per customer). These must be computed the same way for training and for live predictions.
- Training: running an algorithm on a dataset, often with hyperparameter search, and tracking each run's parameters, data version and results.
- Evaluation: measuring the model on held-out data against metrics and against the current model, before any release.
- Model registry: storing versioned models with their metadata and approval state.
- Deployment: serving the model as a batch job or an online API, often with gradual rollouts.
- Monitoring: watching the quality of predictions and the data feeding them for drift, since the world changes and models decay.
- Retraining: triggered by schedule or by drift signals.
Key differences from a typical data pipeline
- Reproducibility is stricter: you must be able to recreate the exact training data and code that produced a model, which needs data versioning.
- Point-in-time correctness: a training row must use only information available at that moment, otherwise the model "sees the future" (leakage).
- The output is a model and not a table, and its quality is statistical, so a pipeline can run without errors and still produce a worse model. Testing includes data validation, performance thresholds and bias checks.
- Training-serving skew is a special failure mode with no equivalent in classic ETL.
Where data engineers fit
Usually they own the data side: reliable ingestion, clean datasets, feature pipelines and stores, data quality and lineage, and the infrastructure and orchestration. ML engineers and data scientists own the modelling, though the roles overlap in small teams.
Say it briefly in an interview
An ML pipeline is a data pipeline plus feature engineering, training, evaluation, deployment and drift monitoring, with higher demands on reproducibility and point-in-time data. Give leakage as an example, since it shows you understand the real problem.