Design a Personalized News Feed System with APIs

Context

Design a large-scale personalized news feed for a consumer application. The system must support both push (server-initiated delivery/notifications) and pull (client-initiated fetch) consumption patterns, while ensuring strong ranking, safety, and operational guarantees.

Scale and Latency Targets

• Users: 100 million monthly active users (MAU)
• Publishers: 1 million
• Ingestion Rate: 200,000 writes per second
• Read Rate: 2 million reads per second
• Latency: p99 < 200 ms for feed reads

Functional Requirements

1. Social Graph and Controls:
   • Follow/unfollow functionality.
   • User-level and publisher-level mute and block capabilities.
2. Feed Quality:
   • Deduplication (both exact and near-duplicate content).
   • Diversity across content types, sources, and topics.
   • Freshness and recency constraints for displayed content.
3. Push Channel Safeguards:
   • Daily notification caps per user.
4. Content Lifecycle:
   • Support for publisher-initiated content retractions (takedowns).

API Design Deliverables

Design external APIs, including request/response schemas and idempotency strategies, for the following operations:

• publish (for content creation)
• subscribe (follow) / unsubscribe
• get_feed (must support pagination and consistent cursors)
• ack_consume (to acknowledge content consumption)
• retract (for content takedowns)
• feedback_logging (for user interactions)

For each API, provide:

• Endpoint path and HTTP method.
• Request and response JSON schemas.
• Idempotency strategy.

Architecture Choices and Justification

• Fan-out Strategy: Choose between fan-out-on-write, fan-out-on-read, or a hybrid approach. Justify your decision based on the provided scale and Service Level Objectives (SLOs).
• Storage Layers: Describe the different storage layers, including hot cache/timeline store and cold storage, and their respective roles.
• Event Streaming: Detail the event streaming topology.
• Ranking Pipeline: Outline the components of the ranking pipeline, covering candidate generation, feature engineering, filtering, scoring, re-ranking, and online inference.

Reliability, Safety, and Operations

Address the following non-functional requirements:

• Abuse and spam controls.
• Multi-region replication and read locality.
• Backfill and replay strategies for data consistency and recovery.
• GDPR deletion and data lifecycle management.
• Rate limiting and quotas.
• SLOs, observability (monitoring, logging, alerting), and on-call signals.
• Failure modes and graceful degradation strategies, including handling partial outages.

Experimentation

• Describe how to run online experiments on the ranking algorithm while ensuring user-level traffic consistency across sessions and pagination.
• Propose safe rollout plans and guardrails for new features or model updates.