A company runs large batch-processing simulation jobs on data stored in an S3 bucket. The jobs are fault-tolerant (can be interrupted) and outputs are stored in a separate S3 bucket. Each job processes 15–20 GB of input data. Which architecture will meet the requirements most cost-effectively?

The most cost-effective solution is to use AWS Batch with EC2 Spot Instances and the SPOTCAPACITYOPTIMIZED allocation strategy. Batch is designed for large-scale, fault-tolerant batch processing, and Spot Instances offer significant cost savings for interruptible workloads. SPOTCAPACITYOPTIMIZED intelligently selects Spot Instances from the most available pools, minimizing interruptions. Using Lambda with provisioned concurrency is less suitable because the 15-20 GB input size per job exceeds Lambda's ephemeral storage and memory limits, making it inefficient for this type of data processing. A mixed On-Demand and Spot Instance strategy in AWS Batch would be more expensive than an all-Spot strategy for a fault-tolerant workload. While Amazon EKS can run batch jobs, it introduces more operational overhead and complexity compared to AWS Batch for this specific use case, and a mixed instance strategy would still be less cost-effective than an all-Spot approach for fault-tolerant jobs.