System Design: Low-Latency Targeted Ad Serving (≤100 ms E2E)

You are designing an online advertising serving system for web/app inventory with a strict end-to-end latency SLA of 100 ms (client request to ad response). Assume multi-region deployment and peak global load.

Functional Requirements

• Real-time ad selection and auction within 100 ms end-to-end.
• Targeting: keywords/context, user segments, and geography.
• Campaign management: budgets (daily/total), pacing, schedules, frequency capping, creatives.
• Ranking: combine advertiser bid with predicted CTR/CVR and quality signals.
• Real-time auctions and pricing.
• Near-real-time reporting for impressions, clicks, spend, basic conversions.

What to Deliver

1. APIs
   • Define an ad request API (input fields) and ad response API (output fields).
2. Core Data Models
   • Specify entities for campaign/line item, creative, targeting, budget/pacing, user segments, and event/log schemas.
3. Architecture
   • Outline components and data flow: edge gateways, ad selector, feature store, model service, auctioneer, throttling/pacing, caching layers, logging/streaming pipeline, and offline analytics/warehouse.
4. Auction Mechanics and Strategy
   • Choose an auction (e.g., first-price, second-price/GSP, VCG), explain incentives and trade-offs.
5. Additional Topics
   • Handling cold start for new ads and new users.
   • Deduplication (impressions/clicks/events; creative dedupe within response).
   • A/B testing strategy and experiment bucketing.
   • Privacy and compliance (GDPR/CCPA, consent, data retention, DSAR).
6. Non-Functional Requirements
   • Scaling approach, storage and indexing choices, consistency guarantees.
   • Fault tolerance, backpressure strategies.
   • Capacity planning with back-of-the-envelope estimates (traffic, CPU for scoring/auction, storage for logs/features, network bandwidth).

Constraints and Assumptions (you may refine)

• 100 ms total E2E budget (assume ~50–70 ms server-side budget to account for network/client).
• Peak traffic: choose and justify a realistic peak QPS (e.g., 20k–100k QPS globally); design should scale.
• Near-real-time reporting: ≤5 minutes latency.
• Availability target: ≥99.9% per region with graceful degradation.