Case Study: How Atome Cut Real-Time Feature Engineering from Weeks to Days with RisingWave

For Atome, a leading financial services platform in Southeast Asia, the ability to approve credit and monitor payments in real time is the backbone of its business. This requires a robust system capable of low-latency, high-throughput feature computation to accurately assess risk.

However, their previous Flink-based pipeline was hindering their ability to respond quickly to emerging threats. By adopting RisingWave, a SQL-based streaming database, Atome has dramatically improved its real-time risk management capabilities.

The Challenge: A Slow and Complex Data Pipeline

Atome's previous Flink-based system created significant bottlenecks that slowed down innovation and response times.

• Slow iteration: The specialized skills required for Flink meant that developing and deploying new risk features and rules was a complex process that often took weeks from start to finish.

• Database bottlenecks: Feature lookups were frequently executed against OLTP databases, leading to high query costs and low throughput, especially during peak transaction periods.

• Operational friction: The process of packaging, deploying, and backfilling new Flink jobs was cumbersome, delaying the risk team's ability to react to new fraud patterns.

The Solution: Real-Time SQL and Materialized Views

Atome chose RisingWave to build a more agile and efficient stream processing foundation for its risk engine. The key advantages were clear and impactful.

• Postgres-style SQL for streaming: RisingWave allowed the team to express complex joins, windows, and aggregations using familiar SQL. This eliminated the need for custom operators and specialized expertise, democratizing pipeline development.

• Cascading materialized views: Instead of repeatedly querying OLTP systems, features and aggregates are maintained as continuously fresh materialized views. This drastically reduces database load while providing auditable and debuggable data points for better observability.

The production implementation at Atome ingests data from MySQL CDC and Kafka and processes it entirely with SQL. This includes normalization, generating features like velocity checks and cross-stream joins, and sinking the final results to MySQL tables for consumption by the online risk engine.

“With RisingWave’s Postgres-style SQL pipelines, we cut feature delivery from weeks to days. No job packaging, no bespoke operators, just DDL, materialized views, and scale. RisingWave has been the real-time data backbone for our core products including CRM and risk features.”

— Qian Chao, Senior Manager, Data Platform, Atome

Key Results: A Leap in Performance and Velocity

The migration to RisingWave delivered immediate and quantifiable improvements across the board.

• Drastic reduction in development time: The cycle to add a new stream feature was reduced from weeks to ~1 day.

• High-throughput processing: The system handles a source throughput of 10k rows/s and sinks data at 2k rows/s.

• Ultra-Low Latency: Average end-to-end latency—from data ingestion to the final feature write—is consistently sub-second, with no backpressure observed during peak loads.

Business and Operational Impact

These technical achievements translated directly into significant business value and operational stability.

• Operational Wins:

  • Reduced database load: Peak query pressure on transactional databases was significantly lowered.

  • Latency stability: Sub-second latency enabled reliable in-flow approvals and real-time risk flagging.

  • Rapid change management: The risk team can now deploy new rules and experiments in just 1–3 days.

• Business Advantages:

  • Improved approval confidence: More features are available at decision time with lower latency.

  • Cost control: The need for large OLTP clusters and emergency read replicas was reduced.

  • Extensibility: The same real-time pipeline pattern is now being used for other use cases like CRM personalization, accelerating time-to-value across the organization.

"We're seeing RisingWave become the critical substrate for real-time feature engineering across various industries, especially FSI, including risk control, fraud detection, and recommendation systems. Teams typically adopt it for one use case and then expand because a reliable stream processing platform enables better real-time decisions as their feature sets grow."

— Sam Hu, GM APJ, RisingWave