How Ant’s Logic Data Center (LDC) Scales to 580,000 TPS on Double‑11

Background

During the 2022 Double‑11 peak, Ant Group’s payment platform processed 583,000 transactions per second (TPS), a record that required a redesign of the underlying data‑center architecture.

Logic Data Center (LDC)

LDC (Logic Data Center) is a logical abstraction that unifies multiple physical IDC sites into a single coordinated entity. It allows traffic to be isolated and routed by user‑ID, enabling independent scaling of each logical unit.

Architectural Evolution

Monolithic application – a single gateway calls a monolithic service which directly accesses a single database instance. This creates a single‑point bottleneck for both performance and availability.

Distributed micro‑services – services are vertically split into domains (e.g., merchant, product, order) and horizontally scaled across many servers. Service‑level single points are removed, but the storage layer remains a single bottleneck because every server still connects to the same database.

Read‑write separation – writes go to a primary instance, reads are served from replicas. Replication latency is typically 10–20 ms, so about 90 % of writes are not read immediately.

Sharding (database/table partitioning) – data is partitioned by a dimension such as userId (e.g., last two digits) into many databases and tables. Both vertical (different business domains) and horizontal (splitting large tables) sharding are applied. However, each application server must maintain connections to all shards, limiting further scaling.

Unit‑based distributed architecture (LDC) – traffic is routed at the gateway to a specific logical unit, called a Zone , based on the user‑ID shard. A Zone only connects to the subset of databases that belong to its user‑ID range, halving the total connection count and allowing dozens or hundreds of Zones to operate independently.

Zone Types

RZone (Resource Zone) – the smallest self‑contained logical unit. It hosts all services and databases required for its assigned userId range, enabling a complete business flow without cross‑Zone calls.

GZone (Global Zone) – a single copy of services and data deployed for disaster‑recovery. It resides in one data center and is not used for normal traffic.

CZone (City Zone) – synchronizes GZone data between cities, acting as a bridge to avoid costly cross‑city RPC calls and to mitigate latency caused by geographic distance.

Traffic Routing Logic

When a request arrives at the gateway, the user’s userId is hashed (e.g., userId % 100) to determine the target Zone range. For example, userId = 37487834 yields 34, which falls into the 00‑49 range; the request is routed to the corresponding Zone. Inside that Zone, the application only opens connections to the databases that store the 00‑49 shard, reducing the number of active connections by roughly 50 % compared with a full‑shard approach. The model can be extended to up to 100 Zones, each handling a distinct userId segment.

High Availability

If an RZone fails, the routing layer can redirect its traffic to other healthy RZones. GZone provides a backup copy of data, while CZone ensures that the backup is kept consistent across geographic locations, allowing rapid failover without incurring cross‑city latency.

Illustrative Diagrams

Key architecture diagrams: