Serverless Architecture in Decisioning using AWS

Elastic, low latency, and cost-effective systems are becoming increasingly important as both business decisioning systems (real and batch time) as well as processing analytic models and policies, and business rules across volatile workloads. This paper describes this type of serverless architecture for decisioning implemented on Amazon Web Services (AWS) and evaluates the performance of the solution on throughput, latency, cost, maintainability and operational resiliency, by comparison with traditional deployments on containers or virtual machines. We have proposed an architecture which is a combination of AWS Lambda for stateless policy evaluation, Step Functions for orchestration of multi-step decision flows, API Gateway for secure ingestion, DynamoDB for low-latency storage of state and features and Event Bridge/SQS for asynchronous eventing. We bring together model inference with Amazon sage Maker endpoints/lambda hosted lightweight models, and share best practices around cache/ cold- start mitigation, concurrency control, and transactional consistency. An end-to-end decision latency metric, scalability under burst traffic, cost/decision metric, and operational complexity metric (deployment and monitoring) are also used to evaluate the performance of the decision support system for synthetic and natural workloads. Results show that a properly architected serverless decisioning platform is capable of providing sub-100ms median latency for typical decision profiles, near-linear cost reduction at low to moderate utilization levels and simplified operations through managed services, but with trade-offs relating to cold starts, stateful workflows and deterministic performance at very high sustained throughput. As a conclusion, we provide best practice suggestions, design patterns for hybrid serverless/stateful components, as well as suggestions for future work (adaptive provisioning, distributed feature stores, and privacy preserving decisioning).