How We Scaled WeChat Pay’s Transaction Records to Billions Daily

Problems with the Old System

The original transaction‑record storage used a simple key/value model where each user’s entire transaction history was stored in a single value; as data grew, values exceeded the 20 MB limit, causing write failures. Moreover, writing transaction records was on the payment critical path, the record types were incomplete, and query mechanisms were primitive, forcing users to scroll endlessly.

Technical Solution

To support massive data volume with high performance and reliability, a distributed KV storage platform (tssd) was adopted and enhanced with a multi‑level data‑partitioning scheme. User data is split into several values: a root node storing metadata and multiple data nodes holding actual records. Data is partitioned by time, enabling fast page queries and binary‑search‑style lookup for specific timestamps.

Online Operations

Historical data migration required consolidating incomplete records from multiple sources, taking six months to clean, verify, and import 72.3 billion rows. To handle storage platform timeouts that caused occasional data loss, a full‑link queuing mechanism was introduced, ensuring each user’s requests are processed on the same server and process to avoid overwrites.

During holiday peaks (e.g., Chinese New Year), request volume can be ten times normal. A flexible throttling strategy buffers excess requests to local disks and later drains them to storage, applied only to red‑packet traffic to preserve resources for other payment types.

Data Security

Three layers protect user data: (1) access control with tickets and whitelist for internal services; (2) data masking and encryption of sensitive fields (WeChat ID, UIN, merchant ID) plus ID virtualization; (3) strict personnel policies, minimal permissions, centralized privileged‑access management, and comprehensive operation logging for audit.

Effect

The rebuilt system dramatically improved data completeness and accuracy, eliminating complaints about missing transaction details and reducing overall complaint rate by 67 %. Daily transaction records now approach tens of billions (including red packets), total stored data exceeds one trillion records, and query experience is much smoother.

Conclusion

WeChat Pay’s transaction‑record system redesign showcases how distributed KV storage, time‑based data partitioning, flexible throttling, and rigorous security measures can scale a payment platform to handle billions of daily records while delivering a reliable, secure, and user‑friendly experience.