Design and Implementation of a High‑Concurrency Flash Sale System for Automotive E‑Commerce

Introduction

Flash‑sale (seckill) is a marketing tool that uses limited‑time, limited‑quantity offers to stimulate purchases; in the automotive e‑commerce platform it involves high‑value discount coupons (often >30,000 CNY) that attract both genuine users and malicious actors.

The system must handle massive concurrent requests while preventing overselling and ensuring accurate stock deduction.

Overall Architecture

The architecture follows the principles of suitability, simplicity, and evolution, and has supported more than 70 seckill events since April 2020, with up to 12 simultaneous events.

Key components include a gateway layer, Redis cache, MySQL database, and a micro‑service order center.

Seckill Model Selection

Two models are discussed: synchronous (immediate result) and asynchronous (MQ‑based peak‑shaving). The synchronous model is chosen for high‑value, low‑stock coupons because it provides better user experience and integrates with reservation and risk‑control checks.

Stock Deduction Design

Three common approaches—order‑time deduction, payment‑time deduction, and pre‑reservation—are evaluated. The system adopts order‑time deduction for its simplicity and precise control, using database transactions to guarantee that inventory never becomes negative.

High‑Performance Architecture

Gateway Layer Technology

OpenResty (Nginx + Lua) is used for its event‑driven, non‑blocking I/O model and built‑in support for MySQL, Redis, and Memcached.

Cache Warm‑up

Scheduled tasks preload reservation lists, product info, and activity data into Redis, enabling the API to handle tens of thousands of QPS.

Layered Rate Limiting

Because traffic spikes form a sharp vertical curve, a multi‑layer rate‑limiting strategy intercepts requests before they reach the application layer.

Stock Deduction Implementation

Two transaction patterns are provided. Pattern 1 inserts a change‑log before updating stock; Pattern 2 updates stock first then logs. The SQL for conditional stock update is:

UPDATE item SET inventory = CASE WHEN inventory >= xxx THEN inventory-xxx ELSE inventory END

Pattern 1 is recommended to avoid InnoDB row‑lock contention.

Rate Limiting Code Example

RateLimiter rateLimiter = RateLimiter.create(20);
System.out.println(rateLimiter.acquire(5));

High‑Availability Architecture

Cache HA

Redis clusters provide high availability; hotspot keys are mitigated by sharding inventory across multiple keys.

Circuit Breaker

Hystrix is used on critical paths (stock query, risk check, promotion query, order creation) to prevent cascade failures.

Overload Protection

When traffic exceeds predefined thresholds, the system quickly returns a static page to shed excess load, while preserving normal traffic.

Security Architecture

Reservation Risk Control

After a user makes a reservation, an asynchronous call to a risk‑assessment service evaluates behavior, profile, and third‑party credit scores; users above a risk threshold are blacklisted.

User‑Behavior Risk Control

Order tokens generated on the checkout page must be present and older than 1 second; missing or stale tokens indicate bot activity and are rejected.

Future Outlook

Static/Dynamic Separation

Product detail pages are static‑served via CDN before the seckill starts, with the countdown fetched from an API, reducing bandwidth pressure.

Deployment Optimization

Order‑processing paths will be isolated and deployed in dual‑data‑center active‑active mode to ensure resilience.

Monitoring Enhancements

Existing infrastructure monitoring covers service latency, JVM GC, and network health; future work will add business‑level metrics for full visibility.

Conclusion

The design demonstrates how a flash‑sale system can evolve from simple to complex architectures based on traffic levels, balancing performance, consistency, and user experience.