Barclays delivers real-time fraud decisions with predictable sub-100ms performance

Barclays is one of the world’s leading universal banks, serving more than 20 million customers and processing tens of millions of payment transactions every day. Fraud detection sits directly in the transaction authorization path, where every decision must be fast, accurate, and consistent.

Barclays needed to modernize a fragmented fraud platform while maintaining strict authorization latency budgets and improving fraud accuracy.

Fragmented fraud systems

Multiple bespoke platforms across business units maintained separate rules, models, and customer profiles. This limited visibility across the organization and contributed to false positives and false negatives.

Strict authorization time budgets

Fraud decisions occur at the end of the card payment chain, where the customer is waiting for approval. Latency variability risked declined legitimate transactions, customer frustration, and increased stand-in processing (STIP).

Rapid data and transaction growth

Barclays processed more than 30 million transactions per day while data volumes expanded from 3TB to more than 30TB in three years. The system had to scale without sacrificing performance predictability.

Balancing accuracy and performance

The platform needed to reduce fraud losses and minimize false declines, without adding latency or operational complexity. Performance at the 99.99th percentile mattered as much as average latency.

Barclays consolidated its fraud systems into a unified machine learning platform powered by Aerospike.

Single shared fraud platform

Aerospike enabled Barclays to centralize fraud rules, models, and customer profiles into one consistent, shared system. This eliminated data silos and reduced inconsistencies across business units.

Predictable single-hop data access

With partition-aware clients and a primary index in memory, Aerospike delivers direct, single-hop access to records. Fraud decisions no longer depended on cache warmth or complex multi-layer architectures.

Hybrid Memory Architecture for scale

By storing indexes in RAM and data on SSD, Aerospike provided near in-memory performance without requiring the full dataset in memory. This enabled predictable performance as the dataset expanded beyond 30TB.

Tight tail latency under peak load

Aerospike reduced latency by 80% compared to the previous system and consistently delivered sub-100ms response times at the 99.99th percentile, even at peak transaction volumes.

With Aerospike as the real-time data foundation for fraud detection, Barclays achieved measurable performance, scalability, and accuracy improvements.

80% lower latency

Barclays reduced latency by 80% and achieved response times under 100 milliseconds at the 99.99th percentile, keeping fraud decisions within strict authorization windows.

4× throughput at scale

The platform now supports 10 million or more transactions per day, with four times the throughput of the prior system.

10× data growth without replatforming

The fraud dataset expanded from 3TB to more than 30TB in three years without requiring architectural redesign or cache reengineering.

Reduced stand-in processing and improved accuracy

By consolidating systems and ensuring consistent access to fraud signals, Barclays reduced stand-in processing (STIP) and improved fraud detection outcomes by lowering false positives and false negatives.