Feature engineering at scale quiz

Question 1

What is the primary purpose of a feature store in a large-scale ML system?

Accepted Answer

To centralize feature definitions and serve consistent values for training and inference, avoiding training-serving skew

Answer

To store raw training datasets as immutable Parquet files

Answer

To replace the model registry and track experiment metrics

Answer

To automatically select the best features via wrapper methods

Question 2

What is a point-in-time (as-of) join and why is it essential when building training datasets at scale?

Accepted Answer

A join that attaches only feature values known at or before each event's timestamp to prevent label leakage

Answer

A join that always uses the latest feature value for each entity

Answer

A join that randomly samples timestamps to balance classes

Answer

A spatial join based on geographic proximity

Question 3

What is training-serving skew and what is a common cause of it?

Accepted Answer

Discrepancy between feature values at training vs serving time, often from divergent code paths

Answer

Overfitting caused by too many features

Answer

The latency gap between batch and streaming jobs

Answer

Class imbalance between training and test sets

Question 4

When serving features online with strict latency budgets, why are online and offline stores typically split into different backends?

Accepted Answer

Because batch analytical reads and millisecond single-key lookups have different optimal storage backends

Answer

Because regulations require physical data separation

Answer

Because online stores cannot hold numeric data

Answer

To reduce cloud egress costs only

Question 5

What is the difference between batch, streaming, and on-demand (request-time) features?

Accepted Answer

Batch is precomputed on schedule, streaming updates near real-time from events, on-demand is computed at request time from the payload

Answer

They are three names for the same scheduled computation

Answer

Streaming features must always be categorical

Answer

On-demand features are deprecated in modern feature stores

Question 6

Why is feature backfilling computationally challenging, and what technique makes it tractable?

Accepted Answer

It requires point-in-time correct historical recomputation, made tractable via incremental windowed aggregations and partition pruning

Answer

It is trivial because you only need the latest value

Answer

It is solved by simply increasing online store TTL

Answer

It only matters for streaming features

Question 7

What is the main risk of target (mean) encoding a high-cardinality categorical feature, and how do you mitigate it?

Accepted Answer

It leaks the target and overfits rare categories; mitigate with out-of-fold encoding and smoothing

Answer

It increases dimensionality uncontrollably; mitigate with PCA

Answer

It cannot represent unseen categories; mitigate by dropping the column

Answer

It only works for regression; mitigate by binning

Question 8

For very high-cardinality categoricals (e.g., millions of IDs) in deep models, why is the hashing trick or embedding preferred over one-hot encoding?

Accepted Answer

Hashing/embeddings give bounded memory and handle unseen IDs, while one-hot explodes dimensionality

Answer

One-hot is preferred because it never collides

Answer

Embeddings require labeled data while one-hot does not

Answer

They produce identical memory footprints

Question 9

What is the key consideration when fitting feature scalers/encoders inside a cross-validation loop at scale?

Accepted Answer

Fit transforms only on the training fold and apply to validation to avoid leakage

Answer

Fit on the entire dataset once for efficiency, then split

Answer

Always refit on the validation fold for better calibration

Answer

Scalers should be fit on the test set to match production

Question 10

In Spark, why can a naive groupBy aggregation for feature computation cause data skew, and how do you address it?

Accepted Answer

Hot keys overload a few partitions; fix with salting, partial aggregation, and adaptive query execution

Answer

Skew is impossible in Spark because partitions are equal-sized

Answer

Increasing executor count alone always removes skew

Answer

Caching the DataFrame eliminates skew

Question 11

What is feature drift versus concept drift, and how do they differ in monitoring implications?

Accepted Answer

Feature drift is a change in P(X) detectable without labels; concept drift is a change in P(Y|X) needing labels/performance

Answer

They are synonyms for model overfitting

Answer

Concept drift only affects images; feature drift only affects tabular data

Answer

Both require retraining on every batch

Question 12

What does the Population Stability Index (PSI) measure and how is it commonly interpreted?

Accepted Answer

The distributional shift of a feature between baseline and current, with >0.25 indicating significant drift

Answer

The correlation between two features

Answer

The model's AUC over time

Answer

The fraction of missing values in a column

Question 13

Why are tree-based gradient boosting models relatively insensitive to monotonic feature scaling, while still benefiting from other feature engineering?

Accepted Answer

Tree splits depend only on value ordering, so monotonic scaling is irrelevant, but interactions/encodings still help

Answer

Trees require scaling more than linear models do

Answer

Trees ignore all engineered features

Answer

Scaling changes tree splits but improves accuracy

Question 14

What is the advantage of computing time-windowed aggregation features using sliding windows over event streams (e.g., Flink) rather than recomputing in batch?

Accepted Answer

Streaming windows update incrementally for fresh low-latency features and handle late events via watermarks

Answer

Streaming is always less accurate so it is avoided

Answer

Batch recomputation is always faster than streaming

Answer

Streaming cannot compute aggregations

Question 15

In a fraud detection pipeline, why might you avoid using a feature like 'account_is_now_blocked' even if it is highly predictive?

Accepted Answer

Because the block is a downstream consequence of fraud detection, making it leakage unavailable at true scoring time

Answer

Because blocked accounts are rare

Answer

Because boolean features hurt tree models

Answer

Because it increases inference latency

Question 16

What is the role of a feature's TTL (time-to-live) in an online feature store?

Accepted Answer

It bounds how long a materialized feature value is served before being treated as stale/null

Answer

It sets the model's prediction timeout

Answer

It determines the number of CV folds

Answer

It encrypts feature values at rest

Question 17

When normalizing/standardizing numeric features for a model deployed online, where should the fitted statistics (mean, std) be stored and why?

Accepted Answer

Persist training-fit statistics as versioned artifacts and apply identically at serving to avoid skew

Answer

Recompute them live from each request batch

Answer

Hardcode them as 0 and 1 always

Answer

Store only in the offline store, never used at serving

Question 18

What is the difference between filter, wrapper, and embedded feature selection methods?

Accepted Answer

Filter uses model-agnostic stats, wrapper retrains on subsets, embedded selects during training (e.g., Lasso)

Answer

They are three names for L1 regularization

Answer

Filter is the most expensive and wrapper the cheapest

Answer

Embedded methods never use the model

Question 19

Why is mutual information often preferred over Pearson correlation for filter-based feature selection?

Accepted Answer

MI captures nonlinear dependencies that Pearson correlation misses

Answer

MI is faster to compute than correlation

Answer

Correlation works for categorical targets while MI does not

Answer

They always give identical rankings

Question 20

What problem does the 'cold start' for new entities cause in feature serving, and how is it commonly handled?

Accepted Answer

Handle missing history via sensible defaults, segment averages, content/contextual features, or a cold-start model

Answer

It is solved by increasing the TTL

Answer

New entities should always be rejected

Answer

Cold start only affects training not serving

Question 21

Why might you store precomputed embeddings as features in a feature store rather than recompute them at inference?

Accepted Answer

To avoid recomputing expensive embeddings per request and keep inference fast, at the cost of some staleness

Answer

Because embeddings cannot be computed online at all

Answer

Because stored embeddings are always more accurate

Answer

To reduce model size on disk

Question 22

What is the danger of computing global statistics (e.g., overall mean for imputation) before splitting into train/test, and how does it manifest at scale?

Accepted Answer

It leaks test/future information into training, biasing evaluation; compute stats only on the training partition

Answer

No danger; it is more efficient

Answer

It only affects deep learning models

Answer

It causes runtime errors but not bias

Question 23

How does feature lineage/versioning help in a large ML platform?

Accepted Answer

It enables reproducibility, drift debugging, impact analysis, and safe rollouts of transformation changes

Answer

It only matters for compliance audits

Answer

It replaces the need for monitoring

Answer

It compresses feature storage

Question 24

What is the benefit of computing features in a declarative transformation framework (e.g., dbt, Spark SQL, or a feature DSL) over imperative per-service code?

Accepted Answer

A single declarative definition compiles to batch and serving, reducing skew and enabling lineage and reuse

Answer

It runs slower but is easier to read

Answer

It removes the need for an online store

Answer

It only works for SQL databases

Question 25

Why is naive count/frequency encoding sometimes safer than target encoding for high-cardinality features?

Accepted Answer

Because it does not use the label, so it cannot leak the target, while still encoding popularity

Answer

Because it produces dense embeddings

Answer

Because it is always more predictive than target encoding

Answer

Because it reduces cardinality to two values

Question 26

In streaming feature computation, what is a watermark and why is it important?

Accepted Answer

An event-time progress estimate that lets windows finalize results and bound state while tolerating late data

Answer

A security token for the stream

Answer

A checksum for message integrity

Answer

The maximum throughput of the stream

Question 27

What is the purpose of feature value monitoring (e.g., null rate, range, cardinality) in production beyond drift detection?

Accepted Answer

It catches data-quality breakages (null spikes, range/cardinality anomalies) that silently degrade models

Answer

It is redundant with model accuracy metrics

Answer

It only matters during training

Answer

It replaces unit tests

Question 28

When would you choose log or Box-Cox transformation of a numeric feature, and does it help tree models?

Accepted Answer

For skewed, multiplicative-scale features to help linear/distance models, but it does not change tree splits

Answer

Always; it improves all models equally

Answer

Only for categorical features

Answer

Only when the target is binary

Question 29

What is the main scalability advantage of the hashing trick, and what is its principal downside?

Accepted Answer

Advantage: O(1) fixed-memory encoding with no vocabulary; downside: collisions conflate distinct categories

Answer

Advantage: zero collisions; downside: high memory

Answer

Advantage: it learns embeddings; downside: it needs labels

Answer

Advantage: perfect interpretability; downside: slow lookup

Question 30

Why are interaction and ratio features (e.g., spend/income, clicks/impressions) often valuable even with powerful models?

Accepted Answer

They encode domain relationships compactly, improving sample efficiency and normalizing across scales

Answer

They are unnecessary because models learn everything from raw inputs

Answer

They only help linear models

Answer

They always cause leakage

Test yourself