Seats Are Selling

Design the seat reservation system for a Ticketmaster style flash sale: one stadium show, 50,000 seats, and roughly 2 million people hitting the sale the second it opens.

Requirements

Functional:

• Browse a live seat map and see which seats are already taken.
• Pick specific seats or ask for best available, then receive an exclusive hold.
• Pay inside the hold window and get a confirmed ticket.
• Release the seats automatically when the user does not pay in time.

Non-functional:

• No double booking. A seat is sold at most once, including during a failover.
• An admitted user gets a hold decision in under 300 ms at p99.
• The sale stays fair and legible to the 2 million people who are waiting.
• Payment runs through a third party processor that can take 30 seconds and can fail.

Scale

• 2,000,000 users arrive over about 5 minutes, roughly 60% of them inside the first 60 seconds.
• Peak edge arrivals near 20,000 per second. Unthrottled, that is 60,000 to 100,000 API req/s.
• 50,000 seats with an 8 minute hold window, so useful hold throughput is in the hundreds per second, not the tens of thousands.
• Seat map payload is about 2 MB raw for 50,000 seats, near 300 KB compressed.
• Hold, order, and audit data for one event stays under 1 GB, so storage is not the constraint.
• Targets: p99 hold latency 300 ms, seat map staleness budget 2 seconds.

Out of scope

Ticket transfer and resale, dynamic pricing, fraud scoring, the mobile client itself, and the venue box office integration.