Machine Learning Engineer interview question
When do you need a feature store?
Use a feature store when multiple models need governed reusable features with consistent offline and online definitions, point-in-time correctness, lineage and low-latency serving; avoid it when a simple pipeline suffices.
Direct answer
What a strong answer covers
How it works
A feature store solves a specific set of problems: several models need the same features computed the same way, online serving needs low-latency access to values that were computed offline, training needs point-in-time correct joins so features reflect what was known at each label’s timestamp, and the organisation needs to know who owns each feature and where it is used. If one model consumes a handful of features from a batch table, a well-structured pipeline and a table with timestamps deliver the same benefit with far less infrastructure. The store earns its cost when reuse and online-offline consistency become real problems.
Example
A company with eight models sharing customer activity features and two of them served online adopted a feature store: features are defined once, materialised nightly to the offline store and hourly to Redis, and training datasets are built with point-in-time joins against label timestamps. A team with a single nightly churn model kept its dbt tables and skipped the store after estimating the operational cost.
Common mistakes
- Adopting a feature store for one batch model and paying for infrastructure nobody uses.
- Building training sets with a naive join that leaks future feature values.
- Letting online and offline computations diverge because they are implemented twice.
A useful answer structure
Answer from your own work.
Do not memorise a script: use only examples you can explain honestly in follow-up questions.
Open the desktop workspace and rehearse a clear answer in your own words.