Introduction: Classic ML Design Questions

1. The Questions Repeat

ML system design interviews feel unpredictable. They are not.

Under all the product names and company logos, the same small set of designs keeps coming back. Recommendations. Feed ranking. Search ranking. Ad click prediction. Fraud detection. These are the classics. Interviewers know them, expect you to know them, and ask them first.

Learn these few well and you can handle most of what you will be asked, including the rarer questions like "design job matching" or "design duplicate detection." Those are the same classic designs applied to a different product.

This module teaches them, one lesson at a time.

2. They All Share One Shape

Here is the part that makes this module easy: these are not a set of unrelated systems. They are all the same idea, with different settings.

That idea is a funnel. You start with millions of items and you need to show the user a handful. You cannot score millions of items in a few milliseconds, so you narrow the list in stages:

1. Retrieve. A cheap, fast stage selects a few hundred promising items out of millions.
2. Rank. A slower, smarter stage scores those few hundred and sorts them.
3. Apply the rules. A final layer adjusts the top handful for business reasons: freshness, variety, "don't show two ads in a row."

Every lesson in this module is this funnel with different settings. Once you see the shape, each new design becomes one simple question: what is different here?

And each one changes one thing:

• Recommendations is the plain funnel. Start here.
• Feed ranking adds many goals at once. A good post is one you will like, comment on, and not report. Balancing those goals is the main problem.
• Search ranking adds a query, and a trap: people click whatever sits at the top, even when it is not the best result. Your own click logs are biased.
• Ad click prediction adds money. The predicted number is not just for sorting; it sets the price in an auction. So it has to be an accurate probability, not just "higher than the others."
• Fraud detection adds an opponent. The fraudster studies your model and changes tactics to beat it. And missing one fraud costs far more than a false alarm.
• Harmful content detection adds scale and high stakes. You are filtering the whole platform, and a wrong decision in either direction either silences a real user or lets abuse through.

If some of those terms (position bias, calibration, an adaptive adversary) are new to you right now, do not worry. Each lesson defines its own new idea in plain language before using it.

3. What You Should Already Know

This module builds on Module 2. We assume you already have the infrastructure from there: a feature store, online inference, and a monitoring stack. Here we build real product designs around that infrastructure.

Every design lesson also walks the seven-step framework in the same order. The first seven sections of each lesson are the seven steps, so the order you study is the order you will use in the interview.

Each lesson then ends with an interview kit: a 30-second spoken answer with the follow-ups you are likely to get, the common mistakes that weaken an answer, and two variation questions that apply the same design to a similar product.

4. The Lessons

5. Two Words You Will See Everywhere

Two terms show up in every lesson. Learn them now and the rest reads easily.

1. Candidate generation (also called retrieval) is the first stage of the funnel. It cheaply pulls a few hundred items that might be a good fit out of the whole catalog, usually with a precomputed index or by finding items whose embeddings sit close together. (An embedding is just a list of numbers that captures what an item is about; similar items get similar numbers.) Its job is recall: of all the items the user would have wanted, how many made it through this stage? If the right item is not pulled here, nothing later can recover it. Ranking can only sort what retrieval hands it.
2. Ranking is the second stage. It scores those few hundred items with a richer model and sorts them by predicted value, whatever value means for the product: a click, a watch, a new connection, a purchase. Its job is precision at the top: of the few items you show first, how many are actually good? The user sees only that top handful, so this is where ranking quality matters most.

6. Before You Start

Pick a product you use that has a feed or recommendations. Then answer three quick questions:

1. Roughly how many items are in its catalog, and how many do you actually see in one session? That ratio tells you how hard the funnel is working.
2. Name three things the app probably logged about you today that could help it rank tomorrow's session.
3. Find one item it showed you that you did not care about. Was that a retrieval miss (nothing better existed to pull) or a ranking miss (something better was pulled, but shown below it)? Telling those two apart is this whole module in one question.