Problem Statement: Extreme Concurrency Under Time Pressure
Frames the shopping festival as a bounded-duration, extreme-throughput distributed systems challenge distinct from steady-state e-commerce.
Problem Statement
Design a shopping festival platform that handles a 48-hour event with 50x normal traffic, time-limited flash sales, layered promotions, personalized coupons, real-time inventory, and instant analytics — without losing a single order, overselling inventory, or degrading below acceptable latency.
A shopping festival is not merely a busy day. It is a bounded-duration adversarial workload. Amazon Prime Day 2024 generated over 400 million items sold across 4 days. Alibaba's Singles' Day 2023 processed orders at a peak rate exceeding 583,000 per second in 2019 and sustained 459 billion RMB GMV. Shopify's BFCM 2023 processed $9.3 billion in sales with 4.2 million orders on the peak day. These are not organic traffic patterns — they are engineered demand explosions with known start times, unknown peak shapes, and zero tolerance for failure.
Why This Problem Is Different
A normal e-commerce system optimizes for steady-state throughput. A festival system optimizes for three orthogonal pressures simultaneously:
- Throughput explosion: 50x baseline QPS for 48 hours, with 5x spikes within the festival window during flash sales.
- Correctness under contention: Inventory decrements must be atomic across millions of concurrent checkouts. A single oversell on a limited-item deal creates customer trust damage worth more than the item.
- Latency preservation: The 99th percentile checkout must stay under 2 seconds even at 50x load. Degradation must be invisible to the buyer.
The problem requires time-based deals, lightning deals, multi-day promotions, marketing banners, real-time inventory per deal item, personalized offers, autoscaling, queueing, distributed caching across regions, and resilience to partial outages.
The Four Architectural Planes
- Demand plane: Traffic ingestion, rate limiting, load shedding, CDN offload, and static asset delivery.
- Commerce plane: Catalog, cart, inventory, promotions, checkout, payment, and order creation.
- Intelligence plane: Personalization, real-time analytics, deal effectiveness, and dynamic pricing signals.
- Operations plane: Autoscaling, monitoring, incident response, configuration management, and rollback.
A strong answer separates these planes. The demand plane sheds load before the commerce plane sees it. The intelligence plane reads projections and never blocks the commerce plane. The operations plane acts on leading indicators, not trailing alerts.
Public Operating Baselines
Amazon reports that Prime Day infrastructure handles millions of concurrent shoppers with DynamoDB processing over 100 million reads per second during peak. Alibaba's OceanBase processed 61 million transactions per minute at peak during Singles' Day 2021. Shopify handled 4.2 million orders in a single day during BFCM 2023 using their pod-based sharding architecture. These are published figures that establish the category is operationally real at extreme scale.
For capacity planning in this answer, we explicitly assume a platform with 200 million daily active users normally, scaling to 500 million during the festival, with 50x checkout QPS increase and 200x flash-sale item QPS increase during lightning deals.
Key Highlights
- •A shopping festival is a bounded-duration adversarial workload, not steady-state traffic.
- •Amazon Prime Day 2024 sold 400M+ items; Alibaba 11.11 peak hit 583K orders/sec in 2019.
- •Three orthogonal pressures: throughput explosion, correctness under contention, latency preservation.
- •Four architectural planes: demand, commerce, intelligence, operations.
- •The demand plane must shed load before the commerce plane sees it.
- •All uncited scale numbers are explicit design assumptions for this interview.
Section Rescue Kit
Buzzwords to use:
Safe statements:
- "I will separate the demand plane from the commerce plane so load shedding happens before core services are impacted."
- "Before drawing boxes, let me define what correctness means for inventory, what latency means for checkout, and what availability means for the platform."