Interview Prep Hub

ML Data Architecture — Interview Prep

Company-agnostic prep for senior and staff loops where the job is the data behind machine learning — ML platform, data engineering for ML, applied-ML infrastructure. A five-move method that structures any ML data design question, five whiteboard drills, the decision matrices that win the trade-off conversation, graded practice questions, and day-of tactics.

How to read this guide — the method

These rounds are graded on whether you can hold the offline and online paths in your head at once and say exactly where they diverge — that is where real systems break. The five moves in 02 force that structure; the drills in Section B are the questions you are most likely to be asked; 08 gives you the matrices to defend every fork. The underlying material lives in the ML Data Architecture deep dive.

Section A · Orient

00 Start Here What these loops actually test, how this guide pairs with the deep dive, the reading order, and what to do if you only have an hour. 01 The Role Decoded & the Interview Loop What ML platform and ML-data roles own, how the loop is assembled, what each round scores, and the four-level rubric every drill here is calibrated against. 02 The Method: Five Moves for Any ML Data Design Question Requirements and the signal → sources and labels → the offline path → the online path → operating it. With the estimates, the board layout and the narration that make it land in 45 minutes. The method

Section B · Drills

03 Drill: Design a Feature Store The question you will almost certainly get in some form, worked end to end — prompt, clarifying questions, estimate, board, point-in-time retrieval, trade-offs and the follow-up probes. Classic 04 Drill: Training Data with Point-in-Time Correctness "It looks great offline and falls over in production." Designing the pipeline that builds an honest training set — spine, as-of join, availability time, versioning, splits. 05 Drill: Real-Time Features for Fraud & Ranking Fifty milliseconds to decide, and the feature is "what happened in the last two minutes" — streaming computation, the online store, event time, and the consistency contract with batch. 06 Drill: Embedding & RAG Data Infrastructure Design the data plane that turns a corpus into embeddings, keeps them fresh and serves retrieval — with the freshness, cost and evaluation questions that separate a demo from a platform. 07 Drill: Data Quality & Drift Monitoring A model that was fine last quarter is quietly wrong today. Design the system that would have caught it — checks at every hop, skew detection, drift with a decision, and the incident path.

Section C · Trade-offs

08 The Big Trade-offs Six forks that appear in every ML data conversation, each as a decision matrix with the condition that flips it — and the one-line answer that defends the fork instead of "it depends". Decision matrices

Section D · Execution

09 Practice Questions, Graded Fourteen questions across design, deep-dive, coding-adjacent and behavioural rounds — each answer graded L1 to L4, each naming the trap that pulls strong candidates down a level. 10 Day-Of Tactics & Your 30-60-90 How to spend the minutes, lay out the board, sound like an owner and recover from a wrong turn — plus the plan you'll be asked for once the offer lands.

Reading paths

If you only have an hour

Read 02 · The Method, run 03 · Design a Feature Store out loud against a timer, then drill 09 · Practice Questions with the answers collapsed.

Full prep (about a week)

00 → 10 in order, one drill a day through Section B, reading the matching deep-dive chapter first. Reread 10 · Day-Of Tactics the morning of.

Coming from data engineering

You already know pipelines. The new material is point-in-time correctness (04), the online path and skew (05), and drift (07). Spend your time there.

Coming from ML / data science

You know models and metrics; the gap is usually systems — the online store and latency budget in 05, the trade-off matrices in 08, and the operating story in 07.