Inside Tencent’s Early Mobile Backend: LBS, Microservices, and Snowball Crises

In August 2012 the author joined Tencent as a backend developer for the mobile version of Tencent Weibo, at a time when mobile internet was just taking off and all mobile apps were grouped in a single business unit.

The first major feature was a location‑based service (LBS) that displayed nearby posts and people, implementing the SoLoMo (Social, Location, Mobile) model. Because the built‑in geographic functions of PostgreSQL were limited, the team created a simple grid‑based index: latitude and longitude were divided into cells, each cell receiving an ID that linked to post IDs. Variants such as geohash were explored, but the lightweight grid approach proved sufficient.

Beyond grid indexing, the system also used administrative region codes derived from latitude/longitude road‑network data to map posts to provinces, cities, and districts, enabling region‑based features like virtual “Weibo walls” that aggregated posts around points of interest.

A particularly interesting feature was location‑based friend recommendation: when two friends who usually posted from fixed locations (home, office) suddenly appeared near each other in an unfamiliar area, the system would push a notification saying “a familiar face nearby,” creating a surprise effect. To identify a user’s frequent locations, the team combined the grid method with a stay‑point algorithm inspired by a Microsoft paper, defining concepts such as StayPoint and StayArea based on trajectory and time.

At that time the team used Tencent’s internal RPC framework TAF (Total Application Frame). They began refactoring the monolithic application into microservices, separating concerns by functionality, traffic volume, latency, and importance. TAF’s flexible reporting mechanism allowed detailed metrics (count, sum, avg) and per‑call success/failure rates.

Through code inspection the author discovered that TAF’s implementation resembled the Half‑Sync/Half‑Async pattern with an epoll‑based thread pool, and that its design borrowed concepts from ICE and Thrift. Later, the team examined Google’s Protocol Buffers and realized that many binary protocols (including TAF’s JCE) share a TLV (type‑length‑value) foundation.

The author also experienced his first large‑scale production avalanche when the synchronous access layer, which performed QQ‑based authentication, became blocked after the QQ auth service crashed on a holiday. The inability to scale or restart servers highlighted the dangers of single‑point‑of‑failure designs.

Beyond LBS, the team built a fast image upload/download service that later impressed early WeChat developers, though WeChat’s subsequent focus on voice (PTT) and relentless product iteration eventually outpaced Weibo’s image‑centric approach.

The narrative concludes with reflections on the team’s culture—recruited from multiple cities, cohesive, and relatively free from office politics—providing a personal glimpse into early mobile internet development at Tencent.