Core concept · Foundation

Features & Feature Engineering

Turn raw signals into available, stable, and useful decision-time features.

Content reviewed September 2026Read to the end

What you will learn

Turn raw signals into available, stable, and useful decision-time features.

Decision trigger

What incremental quality gain would justify this feature's latency and ownership cost?

Interview impact

Feature availability and cost shape both the training graph and hot serving path.

Prediction-time feature budget

1Static profile
2Recent aggregates
3Request context
4Freshness gate
5Model input

Mental model

A feature is a timed dependency with semantics, not just a column.

An ETA model uses traffic aggregates refreshed every hour for a decision that must react within minutes.

How it works

Classify features as static, dynamic, or request context and set windows, time semantics, and owners.
Define missing, stale, and fallback behavior before adding them to the serving contract.

Product consequences

Fresh context can improve decisions but stale context can be worse than a simpler baseline.

Operational consequences

Every online feature adds compute, storage, serving, validation, and incident surface.

Alternatives

Use raw or model-learned representations when manual features add brittle maintenance.
Drop a feature whose incremental quality does not justify latency, privacy, or ownership cost.

Failure modes

Training sees a feature that online serving cannot reproduce.
Default values hide an upstream outage.

Worked example

ETA starts with route distance and historical duration, then adds traffic and location context only within a latency budget.
Each feature has an event-time window, freshness threshold, and fallback.

Interviewer probes

What exists at prediction time?
What happens when this feature is stale or missing?

Expected depth by level

Entry: Name plausible available features.
Mid-level: Add freshness, cost, and skew controls.
Senior+: Define feature ownership, deprecation, privacy, and incident policy.
Role emphasis: SWE-ML emphasizes serving contracts; Applied Scientist emphasizes signal validity.

System Design prerequisites

Caching Stream processing

Sources and review notes

These sources support the technical claims on this page. Product-specific examples remain Engineering Foundry teaching scenarios.

Point-in-time joinsFeast · official documentation · reviewed 2026-09-04
Historical feature retrieval and point-in-time correctness.
Hidden Technical Debt in Machine Learning SystemsGoogle Research · original paper · reviewed 2026-09-04
Feedback loops, data dependencies, configuration debt, and system-level ML risks.

Reach this point to mark the lesson as read.

Finish this lesson

Record completion when you have worked through the material. Completion is learning progress, not evidence of mastery.

Loading completion…

Previous: Dataset Construction, Splits & Leakage

Loading Engineering Foundry…

Sources and review notes

These sources support the technical claims on this page. Product-specific examples remain Engineering Foundry teaching scenarios.

Point-in-time joinsFeast · official documentation · reviewed 2026-09-04

Historical feature retrieval and point-in-time correctness.

Hidden Technical Debt in Machine Learning SystemsGoogle Research · original paper · reviewed 2026-09-04

Feedback loops, data dependencies, configuration debt, and system-level ML risks.

Decision trigger

Interview impact

Mental model

How it works

Product consequences

Operational consequences

Alternatives

Failure modes

Worked example

Interviewer probes

Expected depth by level

System Design prerequisites

Practice this concept

Sources and review notes

Finish this lesson

Decision trigger

Interview impact

Mental model

How it works

Product consequences

Operational consequences

Alternatives

Failure modes

Worked example

Interviewer probes

Expected depth by level

System Design prerequisites

Practice this concept

Sources and review notes

Finish this lesson