AI Practices AI实践 3d ago Updated 3d ago 更新于 3天前 43

How Jumio built a real-time feature store on AWS Jumio如何在AWS上构建实时特征存储

Jumio built a streaming-first real-time feature store on AWS to solve data duplication, manual deployment, latency, and delayed event handling challenges in their fraud detection pipeline The architecture uses Amazon Kinesis Data Streams for ingestion, Amazon Managed Service for Apache Flink for in-flight feature engineering, and Amazon SageMaker Feature Store for low-latency feature serving A dual hot/cold storage strategy leverages Amazon ElastiCache for Valkey for hot in-memory data and SageM Jumio在AWS上构建了流式优先的实时特征存储架构,支持欺诈检测场景下亚100毫秒的低延迟特征服务 采用Amazon Kinesis Data Streams + Managed Service for Apache Flink + SageMaker Feature Store的核心技术栈,实现特征工程与部署的自动化 架构同时支持实时与离线特征存储,热数据使用ElastiCache for Valkey内存存储,冷数据存入标准存储,离线数据以Iceberg格式存储于S3 解决了传统特征管理中的数据重复、手动部署错误、延迟挑战和延迟事件处理等核心痛点

55
Hot 热度
70
Quality 质量
60
Impact 影响力

Analysis 深度分析

TL;DR

  • Jumio built a streaming-first real-time feature store on AWS to solve data duplication, manual deployment, latency, and delayed event handling challenges in their fraud detection pipeline
  • The architecture uses Amazon Kinesis Data Streams for ingestion, Amazon Managed Service for Apache Flink for in-flight feature engineering, and Amazon SageMaker Feature Store for low-latency feature serving
  • A dual hot/cold storage strategy leverages Amazon ElastiCache for Valkey for hot in-memory data and SageMaker Feature Store's standard tier for cold data, achieving sub-100ms latency for fraud detection
  • A parallel offline pipeline routes Flink output through Amazon Data Firehose to Amazon S3 (Iceberg format), enabling near real-time model retraining, debugging, and evaluation via Amazon EMR Serverless and Athena
  • The system is deployed across three AWS Regions (us-east-1, eu-central-1, ap-southeast-1) and supports schema evolution, conditional feature creation, and agile cross-functional feature development

Why It Matters

This case study provides a practical blueprint for any organization building real-time ML systems that require sub-100ms feature serving, particularly in fraud detection, identity verification, or other latency-sensitive domains. It demonstrates how to unify fragmented offline and online feature engineering workflows into a single centralized platform, eliminating data duplication and manual re-implementation risks that commonly plague ML production pipelines.

Technical Details

  • Real-time ingestion and processing: Events enter via Amazon Kinesis Data Streams, where Apache Flink applications perform in-flight feature engineering (conditional feature creation, event-time-based selection) and write directly to Amazon SageMaker Feature Store
  • Hot/cold data tiering: Hot features are served from Amazon ElastiCache for Valkey (in-memory, low-latency reads), while cold/less-frequently-accessed features reside in SageMaker Feature Store's standard tier for scalability and durability
  • Offline feature store pipeline: Flink sinks write to Amazon Data Firehose → Amazon S3 (Iceberg format); S3 events trigger AWS Lambda → Amazon EMR Serverless, which updates Iceberg tables for model training, debugging, and evaluation
  • Multi-region deployment: The architecture is deployed across us-east-1, eu-central-1, and ap-southeast-1 to support global identity verification workloads with regional data residency
  • Key AWS services: Amazon SageMaker Feature Store, Amazon Managed Service for Apache Flink, Amazon Kinesis Data Streams, Amazon ElastiCache for Valkey, Amazon Data Firehose, Amazon EMR Serverless, Amazon Athena, and Apache Iceberg

Industry Insight

  • Organizations with fragmented feature engineering workflows should prioritize a centralized feature store to eliminate data duplication and reduce the risk of training-serving skew caused by manual re-implementation in production code
  • The hot/cold data tiering pattern (ElastiCache + SageMaker Feature Store) offers a cost-effective approach to meeting strict latency SLAs without over-provisioning expensive in-memory infrastructure for all features
  • The near real-time offline pipeline (Flink → Firehose → S3/Iceberg → EMR Serverless) demonstrates how streaming and batch can be unified, enabling faster model retraining cycles and improving ML operational maturity

TL;DR

  • Jumio在AWS上构建了流式优先的实时特征存储架构,支持欺诈检测场景下亚100毫秒的低延迟特征服务
  • 采用Amazon Kinesis Data Streams + Managed Service for Apache Flink + SageMaker Feature Store的核心技术栈,实现特征工程与部署的自动化
  • 架构同时支持实时与离线特征存储,热数据使用ElastiCache for Valkey内存存储,冷数据存入标准存储,离线数据以Iceberg格式存储于S3
  • 解决了传统特征管理中的数据重复、手动部署错误、延迟挑战和延迟事件处理等核心痛点

为什么值得看

本文提供了企业级实时特征存储的完整架构实践,对需要构建低延迟ML推理系统的团队具有重要参考价值。Jumio的案例展示了如何在AWS上实现可扩展、高可靠、低延迟的特征工程平台,适用于金融风控、欺诈检测等对实时性要求极高的场景。

技术解析

  • 实时数据流架构:事件通过Amazon Kinesis Data Streams进入系统,Apache Flink应用实时处理并丰富数据,然后将特征直接写入Amazon SageMaker Feature Store,支持亚100毫秒的特征服务延迟。
  • 冷热数据分层存储:实时特征存储采用双层架构,热数据由Amazon ElastiCache for Valkey提供内存级存储以支持低延迟读取,冷数据保留在标准存储中以平衡可扩展性与成本。
  • 离线特征存储链路:Flink输出通过Amazon Data Firehose写入S3,S3事件触发AWS Lambda调用Amazon EMR Serverless更新Apache Iceberg表,实现近实时的历史数据回填,支持模型训练与调试。
  • 多区域部署策略:架构部署在us-east-1、eu-central-1和ap-southeast-1三个AWS区域,确保全球范围内的低延迟访问和高可用性。
  • 特征工程能力:系统支持基于事件时间的条件特征创建和选择,满足复杂ML用例的需求,同时允许跨职能团队独立引入新特征,减少协调开销。

行业启示

  • 实时特征存储已成为支持低延迟ML推理的关键基础设施,企业应优先考虑流式架构而非传统的批处理方案,特别是在欺诈检测、实时推荐等场景。
  • 冷热数据分层存储策略能够有效平衡性能与成本,建议在设计特征平台时采用类似ElastiCache+对象存储的分层架构。
  • 特征工程与部署的自动化是解决数据重复和人工错误的关键,企业应建立统一的特征管理平台,实现从特征开发到生产部署的端到端自动化。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Deployment 部署 Training 训练 Programming 编程