Design Real-Time ML Feature Pipeline — System Design Practice | AmanAI Lab

Sheet›Design Real-Time ML Feature Pipeline

GoogleMetaAmazonNetflix

45:00

Hard

The Problem

Problem

Design a real-time ML feature pipeline that computes features used by multiple ML models (fraud detection, recommendation, ads CTR) and serves them with:

•Online serving latency ≤ 5ms
•Feature freshness ≤ 1 minute for real-time features
•Point-in-time correct joins for training (no data leakage)
•Support for 10,000 feature definitions shared across 50 ML teams

What you'll be assessed on

This tests your understanding of the full feature store architecture — both online and offline paths, point-in-time correctness, and operational complexity.

Scale & Constraints

▸10K feature definitions used by 50 ML teams
▸Online feature reads ≤ 5ms P99
▸Real-time features must be ≤ 1 minute stale
▸Historical features needed for training with point-in-time correct joins
▸100K feature read QPS at peak serving
▸Feature backfill: compute historical features for any 2-year window

Must Cover0/9

Hints (if stuck)

💡 Point-in-time join: for each training sample (user, item, timestamp), join the feature value that existed at that timestamp — not the latest.

💡 Training-serving skew is the #1 cause of model degradation in production.

💡 Keep stream and batch feature definitions in the same DSL — compute once, serve everywhere.

💡 Redis key format: feature_group:entity_id → JSON or msgpack blob.

💡 Flink watermarks handle late-arriving events — critical for fraud velocity features.

0 words · auto-saved

Back to sheet

Mock interviewTake quizFlashcards

Problem

Design a real-time ML feature pipeline that computes features used by multiple ML models (fraud detection, recommendation, ads CTR) and serves them with:

•Online serving latency ≤ 5ms

•Feature freshness ≤ 1 minute for real-time features

•Point-in-time correct joins for training (no data leakage)

•Support for 10,000 feature definitions shared across 50 ML teams

What you'll be assessed on