Katkat
Back

Design an ads audience targeting system

NetflixSoftware Engineer · Technical Screen
HardSoftware EngineerTechnical ScreenSystem Design

Last updated

This problem evaluates your ability to design a high-scale, low-latency system for filtering user profiles against complex, dynamic advertising audience segments.

Problem Statement

Design a service that determines in real-time whether a specific user belongs to one or more advertising audiences. Advertisers define audiences based on user attributes (e.g., location, watch history, device type, subscription tier) and behavioral events.

Requirements

  • Functional: Provide an API endpoint is_in_audience(user_id, audience_id) that returns a boolean.
  • Functional: Support asynchronous ingestion of user attribute updates and audience definition changes.
  • Non-Functional: The system must handle millions of concurrent users with sub-50ms latency for membership checks.
  • Non-Functional: Support complex boolean logic in audience definitions (e.g., "Users in 'US' AND ('Action' genre fans OR 'New Subscribers')").

Scale

  • 300 million active users.
  • 50,000 active audience segments.
  • Peak load: 100,000 requests per second for membership checks.

Your Answer Should Include:

  1. High-level architecture diagram and service components.
  2. Data model strategy (storage for user profiles vs. audience definitions).
  3. Mechanism for keeping the audience membership state up-to-date.
  4. Discussion on handling "cold-start" scenarios and query optimization techniques.

Hints

Hint 1

Consider pre-computing membership caches or using Bitsets/Roaring Bitmaps to represent audience sets.

Hint 2

Think about whether to push user-attribute updates to audience segments or pull audience definitions into a local service cache.

Hint 3

Use a publish-subscribe model to propagate changes in audience criteria across your distributed service nodes.

Follow-up questions

  • How would you handle a scenario where an advertiser updates an audience definition that currently contains millions of users?
  • How would you ensure data privacy and PII compliance when storing and processing user attribute data?
  • If the latency budget decreases to 5ms, what changes would you make to the architecture?