Feature stores and data pipelines quiz

Question 1

What is the primary purpose of a feature store in an ML platform?

Accepted Answer

A centralized system to define, store, and serve features consistently for training and inference

Answer

A replacement for the data warehouse that stores all raw event data

Answer

A model registry that versions trained model artifacts

Answer

A dashboarding tool for monitoring model accuracy

Question 2

What is 'training-serving skew' and how do feature stores mitigate it?

Accepted Answer

A mismatch between training and serving feature values due to divergent computation paths

Answer

The latency gap between training time and inference time

Answer

Overfitting caused by too many features

Answer

Data drift in the target label over time

Question 3

What is a point-in-time correct join and why is it essential for generating training data?

Accepted Answer

An as-of join that uses only feature values known at or before each event timestamp

Answer

An inner join that matches on the primary key only

Answer

A join that always uses the latest available feature value

Answer

A join that deduplicates feature rows by entity

Question 4

Why do feature stores typically maintain both an offline and an online store?

Accepted Answer

Offline serves high-throughput training; online serves low-latency real-time lookups

Answer

Offline is a backup of the online store

Answer

Online stores raw data while offline stores aggregates

Answer

They are redundant copies for high availability only

Question 5

What does 'materialization' mean in a feature store like Feast?

Accepted Answer

Loading computed feature values from the offline store into the online store

Answer

Compiling feature definitions into SQL

Answer

Persisting model weights to disk

Answer

Converting features into one-hot encodings

Question 6

In feature store terminology, what is an 'entity'?

Accepted Answer

The join/lookup key identifying the object that features describe (e.g., user_id)

Answer

A single column of a feature table

Answer

The transformation function applied to raw data

Answer

The timestamp of a feature record

Question 7

What is a 'feature view' (or feature group) and what does it encapsulate?

Accepted Answer

A group of features sharing an entity and source, with serving metadata

Answer

A SQL view in the data warehouse only

Answer

A trained model's input layer

Answer

A monitoring dashboard for a feature

Question 8

What is the role of a TTL (time-to-live) on features in a feature store?

Accepted Answer

Maximum age a feature value can have to be considered valid at lookup time

Answer

How long the online store keeps connection pools open

Answer

The retention period for raw event logs

Answer

The cron interval for materialization jobs

Question 9

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

Accepted Answer

Precomputed scheduled vs. continuous stream-updated vs. computed at request time

Answer

Cheap vs. expensive vs. free features

Answer

Numeric vs. categorical vs. text features

Answer

Offline-only vs. online-only vs. cached features

Question 10

Why are on-demand / request-time transformations needed even with a feature store?

Accepted Answer

Because some feature inputs exist only in the request and cannot be precomputed

Answer

Because the online store is too slow for all features

Answer

Because batch jobs are unreliable

Answer

Because they reduce storage costs

Question 11

What latency characteristics typically distinguish an online feature store backend choice?

Accepted Answer

Millisecond key-value lookups (Redis/DynamoDB), not columnar scans

Answer

Highest possible scan throughput like a data warehouse

Answer

Full SQL analytics capability

Answer

Strong multi-table join performance

Question 12

What is feature lineage and why does it matter operationally?

Accepted Answer

Tracking sources/transforms producing a feature and its downstream consumers

Answer

The order features appear in the model input vector

Answer

The git history of the model code

Answer

The feature's statistical distribution over time

Question 13

How does data drift differ from concept drift in a feature pipeline context?

Accepted Answer

Data drift changes P(X); concept drift changes P(y|X)

Answer

Data drift is faster than concept drift

Answer

Data drift affects labels; concept drift affects features

Answer

They are synonyms for the same phenomenon

Question 14

What is idempotency in a data pipeline and why is it important?

Accepted Answer

Re-running a step yields the same result, avoiding duplicates on retry

Answer

Running steps in parallel without locks

Answer

Guaranteeing exactly-once network delivery

Answer

Encrypting data at rest in the pipeline

Question 15

What does 'backfill' mean for feature pipelines and what is a common pitfall?

Accepted Answer

Recomputing historical features; pitfall is leaking future data into the past

Answer

Deleting old features to save space

Answer

Copying online store data back to offline

Answer

Retraining the model on new data only

Question 16

What delivery guarantee do most production streaming feature pipelines target, and why not exactly-once everywhere?

Accepted Answer

At-least-once with idempotent sinks to approximate exactly-once

Answer

At-most-once to maximize throughput

Answer

Exactly-once is trivial so it's always used

Answer

Best-effort with no retries

Question 17

In stream processing, what is the difference between event time and processing time?

Accepted Answer

Event time = when event occurred; processing time = when system handles it

Answer

They are identical in distributed systems

Answer

Event time is always later than processing time

Answer

Processing time is embedded in the event payload

Question 18

What is a watermark in a streaming pipeline?

Accepted Answer

A threshold marking event-time progress so windows can finalize

Answer

A checksum verifying message integrity

Answer

A rate limit on the stream

Answer

The offset of the last consumed message

Question 19

What is the Lambda architecture and what problem motivated the Kappa alternative?

Accepted Answer

Lambda has dual batch+speed layers; Kappa unifies on streaming to avoid duplicate logic

Answer

Kappa adds a third caching layer to Lambda

Answer

Lambda is streaming-only; Kappa is batch-only

Answer

They are storage formats, not architectures

Question 20

Why is schema enforcement / a schema registry important in feature pipelines?

Accepted Answer

It enforces agreed field types/evolution, preventing silent data corruption

Answer

It compresses data to reduce storage

Answer

It assigns primary keys automatically

Answer

It schedules pipeline jobs

Question 21

What is a data contract and how does it relate to feature reliability?

Accepted Answer

An enforced producer-consumer agreement on schema, semantics, and quality

Answer

A legal NDA covering training data

Answer

A license for using a feature store

Answer

A SQL constraint on a single table

Question 22

What is the medallion (bronze/silver/gold) architecture in a lakehouse pipeline?

Accepted Answer

Raw (bronze) -> cleaned (silver) -> curated/aggregated (gold) layering

Answer

Three model accuracy tiers

Answer

Three replication zones for HA

Answer

Three pricing tiers of storage

Question 23

What advantage do table formats like Apache Iceberg/Delta Lake bring to feature pipelines?

Accepted Answer

ACID transactions, schema evolution, and time travel over object storage

Answer

Faster GPU training

Answer

Built-in feature transformations

Answer

Automatic model deployment

Question 24

Why is feature reuse across teams a key value proposition of a feature store, and what governance is needed?

Accepted Answer

Reuse cuts duplication/skew but needs catalog, ownership, versioning, and access control

Answer

Reuse is purely a storage cost optimization

Answer

Reuse requires no metadata management

Answer

Reuse only benefits a single team

Question 25

How should you version features when their transformation logic changes in a breaking way?

Accepted Answer

Add a new immutable version and migrate consumers, never mutate in place

Answer

Overwrite the existing feature definition immediately

Answer

Delete the feature and recompute from scratch

Answer

Version only the model, not the feature

Question 26

What is the 'cold start' problem for online feature serving after deploying a new feature?

Accepted Answer

New features lack online values until backfilled, causing nulls at serving

Answer

The model is too slow on first request

Answer

The entity key is unknown at training time

Answer

The offline store is empty after deployment

Question 27

What metrics best characterize the freshness and health of a feature pipeline?

Accepted Answer

Freshness lag, completeness, volume anomalies, distribution shift, and SLA timeliness

Answer

Only model accuracy on holdout data

Answer

GPU utilization and memory

Answer

Number of git commits to the repo

Question 28

What is a DAG in the context of workflow orchestrators like Airflow, and why use one?

Accepted Answer

A directed acyclic dependency graph defining task order and enabling scheduling/retries

Answer

A database normalization technique

Answer

A streaming windowing function

Answer

A model architecture diagram

Question 29

What is the key difference between Apache Airflow and dbt in a data stack?

Accepted Answer

Airflow orchestrates arbitrary workflows; dbt does SQL transformations in the warehouse

Answer

They are competing schedulers with identical scope

Answer

dbt orchestrates Airflow tasks

Answer

Airflow only does SQL transformations

Question 30

Why is ELT often preferred over ETL in modern cloud data pipelines?

Accepted Answer

ELT loads raw first and transforms in the warehouse, enabling scale and reprocessing

Answer

ELT skips transformation entirely

Answer

ETL is always faster and cheaper

Answer

ELT cannot keep raw data

Test yourself