Start Here
Prep for senior and staff loops where the job is the data behind machine learning — training data, features, serving, and the plumbing that keeps a model honest in production. What the interviews test, how to structure any ML data design question, and five drills you can run cold.
What the loop actually tests
"ML data architecture" is the seam between data engineering and machine learning, and interview loops for it — ML platform engineer, data engineer for ML, applied-ML infrastructure, staff DE on a ranking or fraud team — are graded on a specific, narrow set of things. Not on whether you can train a model. Not on whether you can recite the modern data stack. On whether you can design and defend the data path that makes a model correct, fresh, reproducible and cheap — offline for training and online for serving — and whether you understand where those two paths diverge and how to keep them from drifting apart.
Every question in these loops is a variation on four themes:
- Correctness across time. Point-in-time joins, label leakage, feature backfills, training/serving skew. Can you build a training set that reflects what the model would actually have known at prediction time?
- Two speeds, one truth. Batch features for training and low-latency features for serving computed from the same definition. Can you keep offline and online consistent, and can you say what "consistent" means with numbers?
- Reproducibility and lineage. Which data, which code, which version produced this model? Can you rebuild it, and can you prove which features a prediction used?
- Operating it. Drift, quality gates, cost, PII in training sets, retention. Can you run this for three years without a quiet catastrophe?
An ML system is a data pipeline with a model bolted on the end. Interviewers for these roles know that; they are hiring the person who owns the pipeline. Treat every question as a data-architecture question that happens to have a model as its most demanding consumer — one that cares about time correctness more than any dashboard ever will.
How this pairs with the deep dive
This guide is deliberately thin on explanation and thick on execution. The mechanisms — how a point-in-time join actually works, what an online store is, how dataset versioning is done, what drift detection measures — live in the companion reference, ML Data Architecture (fifteen chapters, company-agnostic). Every drill here cites the chapter that carries the depth, so you can go as far as a given interviewer will push. Read that guide when you want to understand; read this one when you want to perform.
| Interview theme | Drill in this guide | Depth in the deep dive |
|---|---|---|
| Feature definitions, offline/online stores, point-in-time retrieval | 03 · Design a Feature Store | 04 · Feature Engineering & the Feature Store |
| Training sets, leakage, splits, dataset versioning | 04 · Training Data with Point-in-Time Correctness | 02, 03, 05 |
| Streaming features, freshness SLAs, skew | 05 · Real-Time Features for Fraud & Ranking | 06, 08 |
| Embeddings, vector stores, RAG corpora | 06 · Embedding & RAG Data Infrastructure | 07 |
| Validation, drift, monitoring, lineage | 07 · Quality & Drift Monitoring | 09, 10 |
| Batch vs streaming, warehouse vs lakehouse, build vs buy | 08 · The Big Trade-offs | 13 · Reference Architectures |
Reading order
Eleven chapters in three sections. The first three set the frame; the five drills are the core; the last three are what you take into the room.
- Orient (00–02). This page, then 01 for what these roles own and how the loop is structured, then 02 for the five-move method that this whole guide hangs off. Do not skip 02 — every drill is the method applied.
- Drills (03–07). Five design questions, each worked end to end: the prompt as an interviewer would give it, the clarifying questions, the estimate, the board, the trade-offs, the probes, and the answer at four levels. Read them; then close the page and re-derive the board from the prompt alone. If you can't, read it again.
- Execution (08–10). 08 is the decision matrices — the forks every interviewer pulls on. 09 is a bank of prompts with graded answers for drill mode. 10 is the morning-of read, plus a 30-60-90 for the offer stage.
These are discussion interviews. Reading a drill is a third of the work; the rest is saying it, with a marker in your hand, under a clock. For each drill, set a 35-minute timer and narrate to a blank board. The method in 02 tells you what to say in each minute.
If you only have an hour
- 15 minutes: 02 · The Method. Memorise the five moves and the decision sentence.
- 25 minutes: 03 · Design a Feature Store — it is the question you will almost certainly get in some form, and it exercises every move.
- 15 minutes: the "big four" in 08 — batch vs streaming features, warehouse vs lakehouse for ML, build vs buy the feature store, and how to version a training set.
- 5 minutes: the checklist at the top of 10 · Day-Of Tactics.
If you have an evening, add 04 and 05, then drill 09 with drill mode on.
How you'll be graded
Loops at this level don't grade on right or wrong; they grade on how far you can go. Every drill in this guide ends with the answer at four levels, and the distinctions are consistent across questions:
- L1 · Foundational. Knows the parts exist — "we'd put features in a feature store" — but can't explain what the store does that a table doesn't.
- L2 · Practitioner. Has built the happy path: batch features, a training job, a serving endpoint. Names offline vs online. Misses time correctness and skew until prompted.
- L3 · Senior. Leads with the failure modes — leakage, skew, backfill semantics, freshness — and designs around them with numbers. Makes defended trade-off calls with a flip condition.
- L4 · Staff. Frames the platform for many models and teams: contracts between producers and consumers, cost per prediction, governance of training data, migration paths, and what to not build. Knows when the feature store is the wrong answer.
The gap between L2 and L3 is where most offers are decided. It is almost entirely about leading with the time-correctness story instead of waiting to be asked. Next: what these roles actually own, and how the loop is put together — 01 · The Role Decoded & the Interview Loop.