Practice/Meta/Design a People You May Know Service

Design a People You May Know Service

ML System DesignMust

Problem Statement

Build a scalable system that generates personalized friend suggestions for users on a social networking platform. The system should analyze various signals including mutual connections, shared interests, geographic proximity, workplace/education affiliations, and user interaction patterns to recommend people a user is likely to know or want to connect with. Your solution must handle hundreds of millions of active users, process billions of connection events daily, and return fresh recommendations with low latency. The system should balance recommendation quality with computational efficiency, and account for privacy considerations when surfacing suggestions.

Key Requirements

Functional

Generate personalized recommendations -- produce a ranked list of suggested connections for each user based on multiple signals
Update recommendations incrementally -- refresh suggestions as users form new connections and interact with the platform
Support ranking customization -- allow different recommendation strategies based on user segments or experimental features
Provide explanation signals -- indicate why a particular person is being recommended (mutual friends, shared group, etc.)
Filter inappropriate suggestions -- exclude blocked users, previously rejected suggestions, and privacy-protected profiles

Non-Functional

Scalability -- support 500M+ active users with 10B+ potential connection pairs to evaluate
Reliability -- maintain 99.9% availability for recommendation serving; tolerate data pipeline failures gracefully
Latency -- return top recommendations in under 200ms at p99; batch generation can tolerate hours of delay
Consistency -- eventual consistency acceptable for recommendations; strong consistency required for blocking/privacy rules

What Interviewers Focus On

Based on real interview experiences, these are the areas interviewers probe most deeply:

1. Graph-Based Recommendation Algorithms

Interviewers want to see you balance algorithmic sophistication with computational feasibility at massive scale. The naive approach of traversing the entire friend graph is computationally infeasible.

Hints to consider:

Discuss collaborative filtering approaches like "friends of friends" and how to limit traversal depth
Consider graph algorithms such as personalized PageRank or node embeddings for similarity computation
Address the tradeoff between exact computations versus approximation techniques like locality-sensitive hashing
Explain how to handle cold start problems for new users with few or no connections

2. Feature Engineering and Signal Processing

The quality of recommendations depends heavily on identifying and combining multiple signals effectively. Interviewers expect you to think beyond just mutual friends.

Hints to consider:

Identify diverse signals: mutual friends, shared groups/pages, geographic proximity, workplace/education history, interaction patterns
Discuss feature normalization and weighting strategies for combining heterogeneous signals
Consider temporal signals like recent profile views, message exchanges, or event co-attendance
Address privacy-preserving feature extraction from sensitive data like location or browsing behavior

3. System Architecture and Data Pipeline

The system involves both offline batch processing for expensive computations and online serving for low-latency retrieval. The architecture must elegantly separate these concerns.

Hints to consider:

Design a Lambda architecture with batch processing for model training and candidate generation plus online serving layer
Discuss data stores for different access patterns: graph databases for traversals, vector databases for similarity search, key-value stores for fast lookup
Consider how to partition users or the social graph to parallelize computation effectively
Address incremental update strategies to avoid full recomputation when the graph changes

4. Ranking and Personalization

Not all recommendations are equally valuable; the system must rank candidates and personalize based on user preferences and context.

Hints to consider:

Discuss machine learning ranking models that combine multiple features into a relevance score
Consider contextual signals like time of day, device type, or user engagement patterns
Address diversity in recommendations to avoid filter bubbles and improve exploration
Explain how to incorporate user feedback (accepted/rejected suggestions) to improve future recommendations

5. Privacy and Abuse Prevention

Friend suggestions can inadvertently leak information or be exploited for harassment, requiring careful consideration of privacy implications.

Hints to consider:

Design privacy controls allowing users to opt out or limit discoverability through certain signals
Implement safeguards against profile scraping or enumeration attacks through the recommendation API
Address asymmetric visibility where one user may see another in suggestions but not vice versa
Consider regulations like GDPR that restrict how personal data can be used for recommendations

Practice/Meta/Design a People You May Know Service

Design a People You May Know Service

ML System DesignMust

Problem Statement

Key Requirements

Functional

Generate personalized recommendations -- produce a ranked list of suggested connections for each user based on multiple signals
Update recommendations incrementally -- refresh suggestions as users form new connections and interact with the platform
Support ranking customization -- allow different recommendation strategies based on user segments or experimental features
Provide explanation signals -- indicate why a particular person is being recommended (mutual friends, shared group, etc.)
Filter inappropriate suggestions -- exclude blocked users, previously rejected suggestions, and privacy-protected profiles

Non-Functional

Scalability -- support 500M+ active users with 10B+ potential connection pairs to evaluate
Reliability -- maintain 99.9% availability for recommendation serving; tolerate data pipeline failures gracefully
Latency -- return top recommendations in under 200ms at p99; batch generation can tolerate hours of delay
Consistency -- eventual consistency acceptable for recommendations; strong consistency required for blocking/privacy rules

What Interviewers Focus On

Based on real interview experiences, these are the areas interviewers probe most deeply:

1. Graph-Based Recommendation Algorithms

Hints to consider:

Discuss collaborative filtering approaches like "friends of friends" and how to limit traversal depth
Consider graph algorithms such as personalized PageRank or node embeddings for similarity computation
Address the tradeoff between exact computations versus approximation techniques like locality-sensitive hashing
Explain how to handle cold start problems for new users with few or no connections

2. Feature Engineering and Signal Processing

The quality of recommendations depends heavily on identifying and combining multiple signals effectively. Interviewers expect you to think beyond just mutual friends.

Hints to consider:

Identify diverse signals: mutual friends, shared groups/pages, geographic proximity, workplace/education history, interaction patterns
Discuss feature normalization and weighting strategies for combining heterogeneous signals
Consider temporal signals like recent profile views, message exchanges, or event co-attendance
Address privacy-preserving feature extraction from sensitive data like location or browsing behavior

3. System Architecture and Data Pipeline

The system involves both offline batch processing for expensive computations and online serving for low-latency retrieval. The architecture must elegantly separate these concerns.

Hints to consider:

Design a Lambda architecture with batch processing for model training and candidate generation plus online serving layer
Discuss data stores for different access patterns: graph databases for traversals, vector databases for similarity search, key-value stores for fast lookup
Consider how to partition users or the social graph to parallelize computation effectively
Address incremental update strategies to avoid full recomputation when the graph changes

4. Ranking and Personalization

Not all recommendations are equally valuable; the system must rank candidates and personalize based on user preferences and context.

Hints to consider:

Discuss machine learning ranking models that combine multiple features into a relevance score
Consider contextual signals like time of day, device type, or user engagement patterns
Address diversity in recommendations to avoid filter bubbles and improve exploration
Explain how to incorporate user feedback (accepted/rejected suggestions) to improve future recommendations

5. Privacy and Abuse Prevention

Friend suggestions can inadvertently leak information or be exploited for harassment, requiring careful consideration of privacy implications.

Hints to consider:

Design privacy controls allowing users to opt out or limit discoverability through certain signals
Implement safeguards against profile scraping or enumeration attacks through the recommendation API
Address asymmetric visibility where one user may see another in suggestions but not vice versa
Consider regulations like GDPR that restrict how personal data can be used for recommendations