Video Search System - Design a Scalable Video Search and Ranking Backend
Question Details
Round 1 ML System Design
Problem
Design the backend for a video search system similar to YouTube Search. Given a text query,
return the top 20 most relevant videos ranked by a combination of relevance and engagement.
Requirements
- Corpus: 500M videos, updated continuously.
- Latency: p95 < 200ms.
- Ranking signals: title/tag match, view count, watch time, CTR, freshness.
- Personalization: logged-in users see results biased toward their history.
High-Level Architecture
Query -> Query Understanding (NLU, spelling, intent)
-> Candidate Retrieval (inverted index + ANN vector search)
-> Lightweight Scoring (L1 ranker, ~1000 candidates)
-> Deep Ranking Model (L2 ranker, ~100 candidates, serves top 20)
-> Diversity + Freshness re-rank
-> Response
Discussion Points
- How do you combine sparse (BM25) and dense (embedding) retrieval?
- What training data do you use for the ranking model (clicks, watch time, skip events)?
- How do you avoid position bias in click-through training labels?
Follow-ups
- How do you handle queries in languages the index was not trained on?
- How do you detect and suppress low-quality or spam videos from ranking highly?
- How do you measure search quality in an offline experiment vs. a live A/B test?
- How does the architecture change for real-time trending content that has no historical engagement data?
Full Details
Round 1 ML System Design
Problem
Design the backend for a video search system similar to YouTube Search. Given a text query,
return the top 20 most relevant videos ranked by a combination of relevance and engagement.
Requirements
- Corpus: 500M videos, updated continuously.
- Latency: p95 < 200ms.
- Ranking signals: title/tag match, view count, watch time, CTR, freshness.
- Personalization: logged-in users see results biased toward their history.
High-Level Architecture
Query -> Query Understanding (NLU, spelling, intent)
-> Candidate Retrieval (inverted index + ANN vector search)
-> Lightweight Scoring (L1 ranker, ~1000 candidates)
-> Deep Ranking Model (L2 ranker, ~100 candidates, serves top 20)
-> Diversity + Freshness re-rank
-> Response
Discussion Points
- How do you combine sparse (BM25) and dense (embedding) retrieval?
- What training data do you use for the ranking model (clicks, watch time, skip events)?
- How do you avoid position bias in click-through training labels?
Follow-ups
- How do you handle queries in languages the index was not trained on?
- How do you detect and suppress low-quality or spam videos from ranking highly?
- How do you measure search quality in an offline experiment vs. a live A/B test?
- How does the architecture change for real-time trending content that has no historical engagement data?
About This Question
This is a reported interview question from a way interview during the phone round.
It covers the following topics: Mle, Phone, System Design, System Design, Backtracking .
More Way Interview Questions
About Way Interview Reports
This question was reported by a candidate who interviewed at Way. LeakCode aggregates interview reports from 10+ sources, including 1Point3Acres, Glassdoor, LeetCode Discuss, Blind, Reddit, Indeed, and Nowcoder. Each report is translated where necessary, deduplicated against existing entries, and tagged by company, role, round type, and reporting date.
Use this question as one calibration data point, not a memorization target. Companies typically rotate their question pools every 2-4 months; the exact wording of a 2024 question may differ from what you encounter today. The underlying pattern, difficulty level, and follow-up depth at Way are the higher-signal extractions to take from this report.
For broader preparation context, the Way interview process typically includes a recruiter screen, one or two technical phone screens, and a 4-5 round on-site loop covering coding, system design (at L4+ levels), and behavioral. Reports tagged on LeakCode show the round-by-round distribution and typical difficulty calibration. To browse questions filtered by round type and seniority, use the company hub linked above.
How To Practice This Type of Question
Solve similar problems on LeetCode under timed conditions (25-35 minutes per medium difficulty). The goal is pattern recognition: recognize the underlying technique (sliding window, two-pointer, BFS, memoized recursion, etc.) within 60-90 seconds of reading. Strong candidates verbalize their hypothesis out loud before coding, then iterate based on feedback. Weak candidates dive into implementation immediately, lose time on the wrong approach, and run out of time for follow-ups.
Companies update their question pools every 2-4 months. The exact wording of any given question may have been retired by the time you interview. Focus your prep on the pattern, not the specific problem. The patterns that appear in Way reports consistently are the ones worth investing in; one-off niche problems are not.
During Your Way Round
Apply the standard interview round template: clarify requirements (2-3 minutes), state your approach out loud and confirm direction with the interviewer (3-5 minutes), code with narration (15-25 minutes), test with concrete examples including edge cases (5 minutes), discuss optimization or trade-offs if time permits (5 minutes). This template is universally accepted across FAANG and adjacent companies; deviating from it produces weaker interviewer feedback signal.
The single most predictive failure mode in Way reports tagged "no hire": not asking clarifying questions. Interviewers are explicitly trained to weight this. Strong candidates ask 3-5 clarifying questions even on problems that look obvious; weak candidates dive into code immediately. The clarifying-question check is often the first signal recorded in the interviewer's written notes.