Scenario & Challenges
As digital transformation in financial services deepens, real-time data analytics has evolved from a decision-support tool into a core foundation for transaction risk control, intelligent marketing, and regulatory compliance. The existing data analytics platform of a large financial institution faced three long-standing pressures.
First, the sharp conflict between data torrents and timeliness. The institution generates billions of market data records, transaction flows, and user behavior logs every day. Traditional batch processing can no longer meet the millisecond-level requirements for risk exposure calculation and abnormal transaction detection. Under extreme market conditions, every millisecond of added analysis latency can create unpredictable exposure risk.
Second, the stringent demands of critical business continuity. The real-time analytics system is directly embedded in the transaction chain and real-time risk control decisions, so any unplanned downtime causes business interruption. Regulators explicitly require core systems to meet 99.999% availability, with recovery time objectives and recovery point objectives approaching zero. The infrastructure must provide strong data consistency and multi-site active-active real-time processing, placing extremely high demands on reliability.
Third, the challenge of balancing agile and stable workloads. Analysis models need rapid iteration as markets change, and peak-to-valley compute demand can differ several-fold. However, traditional infrastructure lacks elasticity, causing resource planning to swing between wasted performance from over-provisioning and degraded service from insufficient elasticity. How to guarantee absolute reliability for stable core business while supporting rapid iteration and burst compute demand for agile analytics became an unavoidable issue in planning.
Implementation Approach
To address these challenges, the financial institution re-examined the computing and data infrastructure supporting real-time analytics. It established an overall planning principle: centered on real-time data flow, anchored by business continuity, and supported by elastic computing. The implementation was carried out across three layers.
Build a distributed low-latency computing plane.
Instead of centralized single-point processing, real-time data ingestion, stream computing, and in-memory analytics are deployed on homogeneous low-latency computing clusters. The clusters use high-bandwidth, low-jitter network interconnects to ensure microsecond-level data transfer between nodes. Compute nodes are selected with a focus on balancing high clock frequency and large cache capacity, with NUMA affinity optimization for all-in-memory computing scenarios. This maximizes single-node processing capability while enabling smooth horizontal scaling to handle traffic growth. The entire chain uses RDMA technology to significantly reduce CPU overhead and latency during data transfer.
Build a geo-redundant active-active continuity system with seamless failover.
The infrastructure is deployed on a three-center architecture combining dual-active sites in the same city and remote disaster recovery. Real-time analytics tasks are deployed in active-active mode based on data sharding, with the same stream computing topology running in parallel with identical state in two data centers. When a cascading failure occurs at any site, the traffic scheduling layer takes over analysis tasks within seconds, with no state loss and no duplicate or missing computation results. The storage layer uses a distributed strong consistency protocol, returning success only after data is synchronously written and confirmed at both active sites. This satisfies the ACID requirements of financial transactions and avoids data inconsistency caused by asynchronous replication. Recovery processes are automated through an orchestration engine, eliminating delays and misoperation risks from manual decisions.
Establish a pooled and elastic compute supply model.
Compute resources are shifted from dedicated physical machines to a unified resource pool. Containerization and real-time orchestration frameworks enable fine-grained task scheduling. In normal conditions, batch analytics, real-time computing, and model training share pooled resources and compete dynamically according to preset weights. When predefined rules are triggered by market events, the elastic controller automatically allocates additional compute from the resource pool to real-time risk control applications, prioritizing tier-one critical tasks. Meanwhile, non-critical reporting tasks are transparently downgraded or suspended to ensure the SLA of the core analytics chain remains unaffected. The platform also incorporates hardware-assisted power management and cooling strategies, allowing compute density to increase significantly without exceeding data center power and cooling limits—balancing performance with energy efficiency.
Results
After the new infrastructure went live, the institution's real-time analytics capability and business continuity assurance were systematically improved and have withstood real-world tests during multiple extreme market fluctuations.
End-to-end latency for real-time risk control was reduced from seconds to a stable sub-millisecond range. Even during peak concurrent trading across the entire market, 99.99% of abnormal transactions could still be identified and blocked within 1 millisecond. Real-time risk exposure calculation now runs on a T+0 uninterrupted basis, shifting decision-making from after-the-fact review to in-process immediate intervention. This helped the institution avoid several potential black swan events.
In terms of business continuity, with the three-center active-active architecture, actual annual availability reached 99.9997%. Neither planned maintenance operations nor unexpected hardware failures caused any service interruption. During two power anomalies at the same-city site level, analysis tasks seamlessly switched to the backup site within 2 seconds, with zero perception by end users—fully meeting and exceeding the recovery targets set by regulators. Multi-site coordinated operation also verified remote failover capability under extreme disaster scenarios, with recovery time 75% shorter than the predetermined target.
From an operational efficiency perspective, the unified resource pool increased average hardware utilization from 18% to 63%. The launch cycle for new real-time analytics services was shortened from weeks to less than three days. The institution successfully passed subsequent annual compliance audits, receiving full marks in data consistency verification and critical-moment recovery drills, and has become a reference example for regional fintech infrastructure planning. This practice shows that infrastructure architecture planning for real-time data analytics must integrate low-latency computing, strong-consistency data synchronization, and elastic scheduling as a whole. Only in this way can the optimal balance be found between the extreme efficiency and absolute stability required by financial business.