RETURN TO SYSTEM DESIGN VAULT (55)BLUEPRINT #55 // Very Hard
#55Very HardFinTech & Transactional APIs
Develop an Ads Management and Display System for a Social Feed
ASKED IN INTERVIEWS AT:MetaGoogleAmazonPinterest
EXECUTIVE SUMMARY & PROBLEM STATEMENT
Design an ad-serving platform executing real-time bidding (RTB) auctions in under 50ms, budget pacing, user targeting, and conversion attribution.
SCALE ESTIMATION & TRAFFIC ENVELOPE
DAILY ACTIVE USERS / TPS1 Billion ad impressions per day, 50,000 auctions/sec
READ / WRITE TRAFFIC RATIOReal-time read auctions with high volume telemetry click writes
5-YEAR STORAGE CAPACITYDistributed in-memory candidate indices in Redis + ClickHouse for telemetry
FUNCTIONAL REQUIREMENTS
- [1]Advertisers create campaigns, target demographics, and set daily budgets
- [2]Run second-price Vickrey auction on every feed ad slot
- [3]Track impressions, clicks, and conversions
NON-FUNCTIONAL REQUIREMENTS
- •Auction completion latency < 40ms
- •Zero budget overspend (smooth pacing throughout 24 hours)
- •High availability under global feed traffic
CORE ARCHITECTURAL COMPONENTS
Auction Engine (C++ / Go)
Targeting Candidate Index
Probabilistic Pacing Service (Redis)
Ad Serving Mixer
Billing & Attribution Ledger
DATA SCHEMA & PERSISTENCE
Campaigns (id, advertiser_id, daily_budget_cents, bid_cpm, targeting_json), UserProfiles (id, age, gender, interest_tags)
DISTRIBUTED TRADE-OFFS & FAILURE MODES
- ⟡Probabilistic pacing artificially throttles bids during early hours to avoid burning an entire daily budget before peak evening shopping.