Design a recommendation system
Recommend products or videos each user is likely to want, from a catalog of millions, like Amazon’s “customers also bought” or Netflix’s home page.
Functional requirements
- Personalized recommendations on the home page
- Similar items on each item page
- Recommendations react to what the user just did
Non-functional requirements
- Recommendations served in under 100 ms
- Fresh within minutes of new activity
- Works for brand-new users and items
Scale
- 300M users
- 50M items
- 10B interaction events per day
How you’ll be scored
10 checks covering your diagram and your notes, and any valid approach counts. Hints appear for anything you miss. Common mistakes, such as single points of failure or components you never explain, cost points. How grading works
What interviewers look for
- Clarify before you draw. State what is in and out of scope. For “Design a recommendation system”, that starts with: Personalized recommendations on the home page; Similar items on each item page; Recommendations react to what the user just did. Interviewers notice candidates who pin the requirements down first.
- Turn the scale into numbers. The scale (300M users; 50M items; 10B interaction events per day) should become requests per second, storage and bandwidth, and those figures should drive your choices. Estimates with units carry more weight than adjectives.
- Show how each quality is met. Your design should visibly deliver: Recommendations served in under 100 ms; Fresh within minutes of new activity; Works for brand-new users and items. Saying that it does isn’t enough; point to the part of the design that makes it true.
- Expect follow-ups on the hard parts. Interviewers usually probe machine learning, batch + streaming and ranking here. This is a hard problem, so interviewers expect depth: what breaks first at this scale, how each failure is contained, and which trade-offs you accepted on purpose.
- Explain failure and trade-offs. Say what happens when a component fails, where a single point of failure would hurt, and what you chose not to build. Naming a trade-off is a stronger answer than claiming there isn’t one.