Перейти к содержанию

Ранжирование ленты новостей -- прохождение интервью

~4 минуты чтения

Предварительно: Компоненты системы | Определение задачи

Интервью по News Feed Ranking -- один из самых частых кейсов на ML System Design собеседованиях в FAANG (задаётся в ~40% всех MLSD раундов). Кандидат должен за 45--60 минут пройти путь от clarifying questions до trade-offs. Ключевой дифференциатор -- умение показать multi-objective thinking (engagement + quality + diversity) и знание реальных ограничений: 500K QPS, <200 мс total latency, position bias в training data. Кандидаты, которые фокусируются только на модели без системных аспектов, проваливают 70%+ таких интервью.

Interview Framework (45-60 min)

Фаза Время Фокус
Clarifying questions 0--5 мин Платформа, масштаб, цели
High-level design 5--15 мин 3-stage pipeline
Deep dive 15--30 мин Ranking model, features
Trade-offs & challenges 30--45 мин Multi-objective, bias
Extensions 45--60 мин Cold start, ads, safety

Step 1: Clarifying Questions (5 min)

**Platform:**
- What type? (social, news, professional)
- What content formats? (text, images, video)
- Social graph or interest-based?

**Scale:**
- DAU?
- Posts per day?
- Average friend/follow count?

**Objectives:**
- Primary metric? (time spent, engagement)
- Quality constraints? (misinformation, harassment)
- Ad integration?

Step 2: High-Level Design (10 min)

Architecture

graph TD
    REQ["Feed Request"] --> PIPE["Feed Ranking Pipeline"]

    subgraph PIPE_INNER["Pipeline"]
        CG["Candidate<br/>Generation"] --> RM["Ranking<br/>Models"]
        RM --> BP["Blending<br/>& Policy"]
    end

    PIPE --> PIPE_INNER
    PIPE_INNER --> CS["Content Store"]
    PIPE_INNER --> FS["Feature Store"]
    PIPE_INNER --> UP["User Profile"]

    style CG fill:#e8eaf6,stroke:#3f51b5
    style RM fill:#e8eaf6,stroke:#3f51b5
    style BP fill:#fff3e0,stroke:#ef6c00
    style CS fill:#e8f5e9,stroke:#4caf50
    style FS fill:#e8f5e9,stroke:#4caf50
    style UP fill:#e8f5e9,stroke:#4caf50

Pipeline Stages

"Three main stages:

1. **Candidate Generation** (10K → 1000)
   - Friends' posts (social graph)
   - Followed pages/accounts
   - Groups
   - Recommendations (explore)

2. **Ranking** (1000 → 50)
   - Multi-task model
   - Predict: P(like), P(comment), P(share), watch_time
   - Combine into single score

3. **Blending & Policy** (50 → final feed)
   - Diversity enforcement
   - Ad insertion
   - Policy filtering (safety)
   - Slot-based composition"

Step 3: Deep Dive (15 min)

Ranking Model

Multi-task learning for engagement prediction. Почему multi-task: разные типы engagement имеют разные сигналы, shared representation эффективнее, лучшая калибровка per objective.

graph TD
    UF["User Features"] --> UE["Embedding"]
    PF["Post Features"] --> PE["Embedding"]
    CF["Context Features"] --> CE["Embedding"]
    UE --> SL["Shared Layers"]
    PE --> SL
    CE --> SL
    SL --> T1["P(like)"]
    SL --> T2["P(comment)"]
    SL --> T3["P(share)"]

    style SL fill:#e8eaf6,stroke:#3f51b5
    style T1 fill:#e8f5e9,stroke:#4caf50
    style T2 fill:#e8f5e9,stroke:#4caf50
    style T3 fill:#e8f5e9,stroke:#4caf50

Final Score = w1 x P(like) + w2 x P(comment) + w3 x P(share) + ...

Key Features

"Critical features for feed ranking..."

"1. **User-Creator Affinity**
   - Interaction history (likes, comments)
   - Social connection strength
   - Time since last interaction

2. **Content Features**
   - Post type (text, image, video)
   - Text embeddings
   - Image/video quality
   - Content freshness

3. **Engagement Signals**
   - Post's current engagement rate
   - Viral velocity (growing fast?)
   - Creator's avg engagement

4. **Context**
   - Time of day
   - User's session depth
   - Device type

Cross features are key:
- User × Creator type affinity
- User × Content topic affinity
- User × Post format preference"

Diversity & Exploration

"Avoiding filter bubbles..."

"Problem:
- Only show content user engaged with before
- Creates echo chambers
- Reduces discovery

Solutions:

1. **MMR Re-ranking**
   - Score = relevance - λ × max_similarity_to_selected
   - Penalize similar items

2. **Slot-based Diversity**
   - Reserve slots for different content types
   - Slot 1-3: High relevance
   - Slot 4: Different topic
   - Slot 5: From less-engaged creator

3. **Exploration Budget**
   - 5% of feed for exploration
   - Random content from new creators
   - Track engagement for learning"

Step 4: Trade-offs & Challenges (15 min)

Multi-Objective Optimization

"Balancing competing objectives..."

"Objectives:
1. Engagement (maximize)
2. Quality (maximize)
3. Freshness (balance)
4. Creator fairness (constraint)
5. User safety (hard constraint)

Approach:
1. Scalarization:
   Score = Σ wi × objective_i

2. Constraints:
   - Min quality threshold
   - Max clickbait score
   - Safety must pass

3. Pareto optimization:
   - Find Pareto-optimal solutions
   - Business chooses point on frontier"

Handling Viral Content

"Content can go viral in minutes..."

"Challenge:
- Model was trained on historical data
- Viral post has no history
- Must react quickly

Solutions:

1. **Real-time Signals**
   - Stream engagement events
   - Compute trending score
   - Boost trending content

2. **Velocity Features**
   - Engagement rate in last 1h
   - Acceleration (rate of rate)

3. **Fast Retraining**
   - Online learning for trending
   - Hourly model updates"

Position Bias

"Users scroll from top..."

"Bias:
- Top positions get more engagement
- Training data reflects this bias
- Model overfits to position

Debiasing:
1. **Inverse Propensity Weighting**
   - Weight = 1 / P(view | position)

2. **Position as Feature**
   - Train with position
   - Predict at position=1

3. **Randomization**
   - Small % of random ranking
   - Use for unbiased evaluation"

Step 5: Extensions (10 min)

Common Questions

Q: How to handle new users (cold start)? 

"New user has no history:

1. Demographic-based recommendations
2. Popular content in their region
3. Onboarding interests selection
4. Rapid learning from first interactions
5. Social graph initialization (import contacts)"

Q: How to insert ads? 

"Ads are part of the feed:

1. Separate ad ranking system
2. Fixed ad slots (every 5th position)
3. Relevance matching to organic content
4. Frequency caps per user
5. Quality constraints (no jarring experience)"

Q: Content safety? 

"Filter harmful content:

1. Pre-screening before indexing
2. Real-time moderation signals
3. Hard filter in ranking (never show flagged)
4. Soft filter (demote borderline content)
5. User controls (hide, report)"

Interview Checklist

Must Cover:

  • Three-stage pipeline
  • Multi-task ranking model
  • User-creator affinity features
  • Diversity mechanisms
  • Real-time updates for viral

Good to Cover:

  • Position bias handling
  • Multi-objective optimization
  • Cold start
  • Ad integration
  • Content safety

Red Flags:

  • Single objective (engagement only)
  • No diversity consideration
  • Ignoring content freshness
  • Not mentioning viral handling
  • Forgetting creator fairness

Sample Script

Interviewer: "Design Instagram feed ranking"

You: "Let me clarify - are we including Reels
and Stories, or just main feed?"

Interviewer: "Main feed with photos and videos"

You: "And the scale?"

Interviewer: "1 billion users, 200 million posts/day"

You: "Here's my approach:
[Draw architecture]

Three-stage pipeline:
1. Candidate generation from social graph
2. Multi-task ranking (like, comment, watch_time)
3. Blending with diversity and ads

For ranking, I'd use a multi-task neural network
predicting multiple engagement types.

Key features:
- User-creator affinity (interaction history)
- Content quality signals
- Real-time engagement velocity

For diversity, I'd use MMR re-ranking to avoid
showing all similar content.

For viral content, stream engagement events
and compute trending scores in real-time.

Shall I dive into any component?"

Типичные заблуждения

Заблуждение: на интервью нужно сразу рисовать архитектуру

Первые 5 минут -- clarifying questions. Кандидаты, которые пропускают этот этап, в 60%+ случаев проектируют систему под неправильные constraints. "Social network" и "news aggregator" -- совершенно разные ленты: social граф vs interest graph, 500 друзей vs 10M+ publisher posts. Без уточнения масштаба (100K DAU vs 2B DAU) pipeline может быть over-engineered или under-engineered.

Заблуждение: достаточно описать модель, чтобы получить strong hire

ML System Design -- это не ML моделирование. Интервьюеры оценивают 5 аспектов: (1) problem formulation, (2) data/features, (3) model architecture, (4) serving infrastructure, (5) trade-offs. Кандидат, который идеально описал модель, но не упомянул latency budget, position bias debiasing или diversity mechanisms, получает в лучшем случае lean hire. Системное мышление важнее quality модели.

Заблуждение: position bias -- это нерешаемая проблема

3 проверенных подхода: (1) Inverse Propensity Weighting -- вес sample = 1/P(view|position), убирает bias в training data; (2) Position as Feature -- тренируем с position feature, на inference подставляем position=1; (3) Randomization -- 1--5% трафика получают случайный порядок для сбора unbiased data. Facebook и YouTube используют комбинацию этих подходов и получают 2--5% improvement в offline метриках после debiasing.

Интервью

Как бы вы разделили время на 45-минутном MLSD интервью?

❌ "Потрачу 30 минут на модель -- это самая важная часть"

✅ "5 минут на clarifying questions (тип платформы, масштаб, objectives), 10 минут на high-level architecture (3-stage pipeline с конкретными latency бюджетами), 15 минут на deep dive (multi-task model, ключевые features, training pipeline), 10 минут на trade-offs (diversity vs engagement, position bias, viral content), 5 минут на extensions (cold start, ads, safety). Такое распределение покрывает все dimensions, которые оценивают интервьюеры."

Как вы обработаете viral content, который набирает 100K лайков за 10 минут?

❌ "Перетренируем модель на свежих данных"

✅ "Ретрейн занимает часы -- не подходит. Три подхода: (1) real-time streaming engagement events через Kafka, вычисляем velocity features (engagement rate за последний час, acceleration -- рост скорости engagement); (2) boosting score для контента с высокой velocity -- множитель 2--5x в post-ranking; (3) online learning для trending signal -- обновление lightweight модели каждые 15--30 минут. Ключевое: velocity features должны быть в Feature Store с <5 мс latency, обновляемые near-real-time."