Building a Scalable Coupon Push System with Redis for Real‑Time Notifications

Background

The coupon‑center must push a reminder to a user exactly one minute before a coupon becomes claimable. With >60 million registered users and an estimated 200 000 concurrent pushes, the system must guarantee sub‑second timeliness and high throughput.

Key Challenges

Push timeliness – delayed notifications lead to user complaints.

Massive volume – a popular coupon can generate hundreds of thousands of push messages simultaneously.

Rejected Approaches

MQ delayed delivery – Message‑queue delay granularity is coarse (seconds to minutes) and cannot guarantee precise timing. Handling cancellations and deduplication would add considerable complexity.

Single‑machine scheduled task – A cron job that scans the database for due reminders becomes a performance bottleneck, suffers from poor timeliness, and introduces a single point of failure.

Proposed Architecture: Redis‑Backed Scheduled‑Task Cluster

Each subscription reminder is stored in a Redis Sorted Set where the reminder timestamp is used as the score. Multiple backend servers run a 1‑second timer; each server pulls a configurable batch of due records from a specific queue determined by userId % queueCount. Queue count matches the number of processing servers and can be changed at runtime.

Why Redis?

High‑performance in‑memory storage, faster than MySQL, with persistence and stability.

Sorted Sets naturally support time‑based ordering, making it trivial to select records ready for push.

Implementation Details

• Key design : Use userId as the base key and hash it to one of several Sorted Set queues to spread load and allow merging of near‑identical push times.

• Dynamic configuration : Queue count and batch size are stored in Alibaba’s Diamond configuration service, allowing on‑the‑fly adjustments without redeployment.

• Batch size is also configurable at runtime, enabling the cluster to scale its throughput according to traffic.

Load Balancing

A simple algorithm uses a Redis auto‑increment key combined with the queue count to distribute work:

nextId = INCR global_counter;
queue = nextId % queueCount;

This reduces contention so that multiple servers rarely compete for the same queue.

Performance Evaluation

With 10 processing servers each pulling 2 000 records per second, the cluster can push 20 000 messages per second. Adding an asynchronous MQ step to deliver messages to the message centre adds ~0.5 s overhead, so 200 000 pushes complete in a few seconds.

Future Improvements

Add monitoring to detect task backlogs.

Provide a visual management console.

Implement priority‑aware scheduling.

Introduce resource‑aware dispatch to guarantee critical tasks when capacity is limited.