ML Data Engineering & Feature Pipelines

Question 1

What is the main role of data engineering in machine learning systems?

Answer 1

To provide reliable, well-documented pipelines that deliver feature and label data in the right shape, quality, and cadence for training and serving models.

Question 2

What is a feature in ML?

Answer 2

An input variable used by a model to make predictions, typically derived from raw data through transformations and aggregations.

Question 3

What is a label in supervised learning?

Answer 3

The target variable that the model is trained to predict, representing known outcomes for historical examples.

Question 4

Why is feature engineering often more impactful than model tweaking?

Answer 4

Better features can reveal relevant structure in data and improve signal-to-noise, while model changes may offer smaller gains on poor features.

Question 5

What is the difference between a feature and a raw field from a source system?

Answer 5

Features are cleaned, standardized, and engineered for modeling, while raw fields are direct outputs from source systems that may include noise and idiosyncrasies.

Question 6

What is a feature pipeline?

Answer 6

A data pipeline that computes, stores, and serves features for training and inference, often from raw logs or transactional data.

Question 7

Why is it useful to separate feature pipelines from model code?

Answer 7

It decouples data preparation from modeling logic, enabling reuse across models and clearer ownership for data vs algorithms.

Question 8

What is an offline feature store conceptually?

Answer 8

A storage layer that holds historical feature values for training and batch scoring, usually on a warehouse or data lake.

Question 9

What is an online feature store conceptually?

Answer 9

A low-latency store that serves the latest feature values for real-time inference, often based on key-value or cache technology.

Question 10

Why is consistency between offline and online features critical?

Answer 10

If features differ between training and serving, models will see a different input distribution in production, causing performance degradation (train–serve skew).

Question 11

What is train–serve skew?

Answer 11

A mismatch between the data and features used during training and those seen at inference time in production.

Question 12

What are common causes of train–serve skew?

Answer 12

Different feature code paths offline vs online, time-dependent features not computed the same way, or using future information in training that is unavailable at inference.

Question 13

How can you reduce train–serve skew?

Answer 13

Reuse the same feature definitions in offline and online pipelines, avoid using future data, and thoroughly test end-to-end feature flows.

Question 14

What is a snapshot training dataset?

Answer 14

A static extract of features and labels at a given time, used to train a model on a specific snapshot of history.

Question 15

What is a point-in-time correct training dataset?

Answer 15

A dataset that ensures features for each training example are computed only from information available up to the event time, avoiding leakage from the future.

Question 16

What is data leakage in ML datasets?

Answer 16

Using information in training that would not be available at prediction time, leading to overly optimistic metrics and poor performance in production.

Question 17

What are examples of leakage from feature engineering?

Answer 17

Using post-outcome flags as features, aggregating over future periods, or joining labels back into features inadvertently.

Question 18

Why is point-in-time correctness challenging?

Answer 18

It requires storing historical states of features and carefully designing joins and windows so future updates do not contaminate past examples.

Question 19

What is a common pattern for building point-in-time features?

Answer 19

Store time-stamped events and compute features using window functions or aggregations restricted to times before the label event.

Question 20

What is target leakage via look-ahead windows?

Answer 20

When feature windows accidentally extend beyond the label event time, effectively including information about the outcome in inputs.

Question 21

What is a label generation pipeline?

Answer 21

A pipeline that derives target values from raw events or transactional data, applying definitions and time windows consistently.

Question 22

Why should label definitions be documented and versioned?

Answer 22

Changing label logic over time can alter what the model is learning; versioning makes experiments reproducible and auditable.

Question 23

What is a positive-unlabeled (PU) situation in labels?

Answer 23

A setting where positive cases are labeled but negatives are mixed with unlabeled and potentially positive cases, common in fraud or anomaly detection.

Question 24

Why is sampling important for ML training datasets?

Answer 24

Full historical data may be too large or imbalanced; sampling can improve training efficiency and class balance while preserving key patterns.

Question 25

What is stratified sampling?

Answer 25

Sampling that preserves the proportion of classes or important subgroups, improving representativeness of the training set.

Question 26

Why must sampling decisions be recorded?

Answer 26

They affect metrics and reproducibility; future comparisons and retraining must know how data was sampled.

Question 27

What is a feature drift in ML systems?

Answer 27

A change over time in the distribution of input features compared to the training data.

Question 28

What is concept drift in ML systems?

Answer 28

A change over time in the relationship between features and labels, such that the original model becomes less appropriate.

Question 29

How can feature drift be detected?

Answer 29

By monitoring feature distributions over time and comparing them to historical baselines, using statistical tests or divergence measures.

Question 30

How can label distribution drift be detected?

Answer 30

By monitoring label proportions and outcome rates over time, when labels are available, and comparing to training distributions.

Question 31

What is a retraining pipeline?

Answer 31

An orchestrated process that periodically or conditionally rebuilds a model using recent data, evaluates it, and, if acceptable, promotes it.

Question 32

Why must retraining pipelines be automated?

Answer 32

Manual retraining is error-prone and slow; automation ensures models stay current and reduces operational overhead.

Question 33

What is a model registry conceptually?

Answer 33

A system that stores model artifacts, metadata, and versions, supporting promotion from staging to production.

Question 34

What metadata should be tracked per model version?

Answer 34

Training data snapshot, feature and label definitions, hyperparameters, metrics, code version, and deployment history.

Question 35

Why is reproducibility important in ML data pipelines?

Answer 35

To debug issues, audit model decisions, compare versions, and comply with regulations or internal governance.

Question 36

What is a feature importance analysis used for in data engineering?

Answer 36

To understand which features are most predictive, guiding feature pipeline investment and data quality prioritization.

Question 37

Why is feature availability and latency as important as predictive power?

Answer 37

A highly predictive feature that is slow or unavailable at inference time may not be usable in production settings.

Question 38

What is feature freshness in an ML context?

Answer 38

How up-to-date feature values are relative to when predictions are made, critical for time-sensitive models.

Question 39

How can batch feature pipelines support low-latency serving?

Answer 39

By precomputing features on a schedule and storing them in fast key-value stores or caches for quick lookup at inference time.

Question 40

What is online feature computation?

Answer 40

Computing features on the fly at request time from recent events or streaming data, balancing freshness with latency constraints.

Question 41

What is the tradeoff between precomputed and on-demand features?

Answer 41

Precomputed features reduce inference latency but may be less fresh; on-demand features are more current but can increase latency and complexity.

Question 42

What is a feature view or feature definition object?

Answer 42

A declarative specification of how to compute a feature set from underlying tables or streams, used by feature stores.

Question 43

Why is decoupling feature definitions from model code beneficial?

Answer 43

It enables reuse across models, standardizes feature logic, and reduces the risk of inconsistent implementations.

Question 44

What is A/B testing in the context of ML deployment?

Answer 44

Running two or more model versions in parallel on different subsets of traffic to compare performance and business impact.

Question 45

Why do A/B tests depend on solid data engineering?

Answer 45

Because reliable tracking of exposures, predictions, and outcomes is required to attribute effects correctly and avoid bias.

Question 46

What is logging in an ML pipeline context?

Answer 46

Recording inputs, predictions, model version, and possibly outcomes for each inference to support monitoring, debugging, and retraining.

Question 47

Why must logging avoid leaking sensitive data unnecessarily?

Answer 47

Logs often have broad access and long retention; logging too much sensitive data increases risk and compliance burden.

Question 48

What is a good one-sentence mental model for ML data engineering?

Answer 48

Design feature and label pipelines that are point-in-time correct, reusable across training and serving, and continuously monitored for drift and quality.