Design Case Study: A Flash-Sale Ticketing Platform
Module 6 · Putting It All Together

Design Case Study: A Flash-Sale Ticketing Platform

Apply all four domains to one design: gather requirements for a concert ticketing site with extreme on-sale spikes, build the architecture step by step, record the trade-offs, and practice the whiteboard design round used in Solutions Architect interviews.

The Brief

A regional promoter sells concert tickets online. Today it runs on two servers in a colocation facility, and the site crashed during the last big on-sale. They want to move to AWS before the next season.

From your requirements interview:

RequirementDetail
TrafficAbout 200 requests/second normally; up to 50,000 requests/second for the first 10 minutes of a big on-sale
CorrectnessA seat must never be sold twice
HoldsA buyer gets 5 minutes to pay for held seats; unpaid seats return to inventory
PaymentsThrough a third-party payment provider; the company wants its card-data (PCI) scope as small as possible
Availability99.9%; an outage during an on-sale is a headline
RecoveryRPO 1 minute; RTO 4 hours for a Regional disaster
UsersMostly North America
TeamSix developers, no dedicated operations staff
CostBusy for a few hours a week; idle most of the time

Before designing, notice what these requirements imply. Extreme, short spikes and a mostly idle week point to serverless. “Never sold twice” points to conditional writes or transactions. A small team points to managed services. A 4-hour RTO for a Regional disaster does not require active/active.


The Architecture

UsersCloudFront+ AWS WAFS3 static siteWaiting roomadmits N users/secAPI GatewayCognito auth, throttlingLambdaAPI handlersDynamoDB seatsconditional write = holdStep Functions checkoutpay → confirm, or release seatPayment provider (external)EventBridgeSQS → SESticket emailsFirehose → S3sales analyticsCross-cuttingSecrets Manager: payment API keyKMS encryption · CloudTrail · GuardDutyCloudWatch alarms · X-Ray tracingDynamoDB PITR + AWS Backup to 2nd RegionInfrastructure as code to rebuild anywhere

Walking through it, domain by domain

Performance (Domain 3). The static site comes from S3 through CloudFront, so 50,000 people loading the page never touch the backend. A virtual waiting room (AWS publishes a reference solution) gives each visitor a place in line and admits them to the purchase API at a rate the backend can handle. API Gateway throttling is a second limit. Lambda scales out quickly, and DynamoDB on-demand absorbs bursts of writes.

Correctness. Holding a seat is one conditional write to the seat’s item in DynamoDB: set holder = me and holdExpires = now + 5 min only if the seat is available or its hold has expired. If two buyers try at once, exactly one write succeeds. DynamoDB TTL isn’t used to release holds, because TTL deletes can lag; the expiry timestamp in the condition is what makes a seat available again.

What happens when two buyers click the same seat at the same moment:

Buyer ADynamoDB: seat 14CBuyer Bhold 14C if available or expiredsame write, same momentEach conditional writeis atomic: one wins✓ held by A until 20:05✕ ConditionalCheckFailedApp shows “seat just taken”Checkout startsIf A doesn’t pay by 20:05, the next conditional write succeeds because the hold has expired.

Resilience (Domain 2). Every service here is Multi-AZ by default. Checkout is a Step Functions workflow: call the payment provider with retries, confirm the order, and on failure release the seat (a compensating step). Order events go through EventBridge to an SQS queue for email, so a slow email service never blocks a sale. For Regional disaster, DynamoDB point-in-time recovery meets the 1-minute RPO within the Region, AWS Backup copies to a second Region, and infrastructure as code rebuilds the stack there within the 4-hour RTO. That’s backup and restore, the cheapest strategy that meets the requirement.

Security (Domain 1). AWS WAF rate-limits by IP and filters known bots. Cognito handles sign-in. The browser sends card details directly to the payment provider (hosted fields or a redirect), so card data never touches the company’s systems and PCI scope stays minimal. The provider’s API key lives in Secrets Manager. Each Lambda function has its own least-privilege role.

Cost (Domain 4). Between on-sales, almost nothing runs: Lambda, API Gateway, DynamoDB on-demand, and Step Functions all charge for use. CloudFront reduces egress and origin load. No Savings Plan is needed for a workload that is idle most of the week.


Trade-Offs to Write Down

DecisionChosenAlternativeWhy
Seat inventoryDynamoDB conditional writesAurora with row locksHandles extreme write bursts with no connection limits or capacity planning; Aurora would work at lower peaks and suits complex reporting
API computeLambdaECS on FargateScales to zero and absorbs bursts; Fargate is better for steady load or long-running connections
Surge controlWaiting room + throttlingPre-scaling onlyProtects the payment provider and DynamoDB from a thundering herd, and is fairer to buyers
DRBackup and restore to a second RegionWarm standbyA 4-hour RTO allows it, at a fraction of the cost
⚠️ Common Pitfall: Load-Testing Only Your Own Side

Common mistake: the team load-tests Lambda and DynamoDB at 50,000 requests per second and declares victory.

The payment provider has its own rate limits, and service quotas (Lambda concurrency, API Gateway throttles) have default values that may be too low. Find each downstream limit, set the waiting room’s admission rate below the tightest one, and request quota increases weeks before the on-sale.


The Design Round in SA Interviews

Many Solutions Architect interviews include a design round very like this lesson. A structure that works:

  1. 1 Clarify

    5 min of questions: scale, RTO/RPO, constraints

  2. 2 Assume

    Say your numbers out loud

  3. 3 Draw

    Simple boxes and arrows first

  4. 4 Deep dive

    The hardest part (here: no double sells)

  5. 5 Break it

    What fails? What does it cost?

  6. 6 Summarize

    Decisions and trade-offs

Interviewers care more about how you reason than about naming every service. Say why you chose each component and what you’d change if a requirement changed.


Key Takeaway

Let the requirements drive the design: spikes and idle time point to serverless, correctness points to conditional writes, a small team points to managed services, and the RTO decides the DR strategy. Protect the backend with caching, a waiting room, and throttling, isolate slow work with events and queues, and write down every trade-off.

🧪 Try It: Change One Requirement

Redesign the ticketing platform for each change below, one at a time. For each, list what changes and what stays the same.

  1. The promoter expands to Europe and Asia, and users must get local latency for browsing.
  2. The RTO for a Regional disaster drops to 5 minutes.
  3. Traffic becomes steady at 5,000 requests per second all day, every day.

Hints: (1) CloudFront already helps browsing; consider where the inventory lives and whether seats are Regional. (2) Warm standby with DynamoDB global tables and Route 53 failover. (3) Revisit Lambda vs Fargate and DynamoDB on-demand vs provisioned capacity with Savings Plans and reserved capacity.

🧠 Knowledge Check

1. Two buyers try to hold the same seat at the same moment. Which design guarantees only one succeeds?

2. The workload is idle most of the week and spikes to 50,000 requests per second during on-sales. Which compute choice best balances cost and scaling?

3. The RTO for a Regional disaster is 4 hours and the company wants to minimize cost. Which DR strategy is appropriate?

4. How does the design keep the company's card-data (PCI) scope small?