Ant Financial's Online Graph Computing: Architecture, Applications, and Core Technologies

Ant Financial has leveraged fifteen years of technical innovation to build a financial‑grade online graph computing platform that powers services for over 1.2 billion users.

The platform enables the "310" model—three‑minute online applications, one‑second approval, and zero manual intervention—by using real‑time graph computation to detect fraud such as cash‑out schemes in products like Huabei.

Key technical requirements include constructing a reliable financial‑level funds network, performing sub‑graph analysis for real‑time decisions, and dynamically building sub‑graphs as user behavior evolves.

Use cases such as the anti‑cashout scenario and the Ant Forest social game illustrate the need for massive concurrency, trillion‑level graph storage, and strong consistency.

The overall architecture provides a unified graph development platform where developers write jobs using a DSL that combines SQL and Gremlin, which the system compiles into a distributed DAG.

Two processing pipelines handle log and event streams: one writes to a high‑performance graph database and builds an in‑memory graph cache for fast sub‑graph extraction; the other decides dynamically whether to traverse and compute sub‑graphs, enabling on‑demand, elastic computation.

Flow‑graph fusion combines stream processing and graph computation in a single system, eliminating the need for multiple heterogeneous engines and reducing latency.

Dynamic DAGs allow the system to adapt computation paths at runtime based on data‑driven decisions, supporting elastic scaling.

Developers can use SQL + Gremlin (GraphView) to create pipelines, lowering the learning curve and simplifying debugging.

High‑performance graph caching uses perfect hash functions and compression to keep the entire graph in memory, achieving about 20 % of the original memory size and delivering sub‑millisecond query latency.

GeaBase, Ant's financial‑grade graph database, implements micro‑sharding, cost‑based data migration, automatic load balancing, and a three‑site, five‑center disaster‑recovery strategy, while providing configurable strong or eventual consistency via the Raft protocol.

The platform now supports over 100 business scenarios across risk control, social interaction, and marketing, running on a 2000‑node cluster 24/7, and is being offered to external financial institutions such as Changshu Rural Commercial Bank.

Future work will continue to enhance online graph computing capabilities and expand them to more partners and use cases.