A complete MLOps architecture for training and serving a sentiment-analysis model on AWS, built around the Amazon Reviews dataset.
Problem
The model is the easy part of an MLOps system. The harder problem is everything around it (data movement, encryption, orchestration and observability), designed as a single coherent pipeline rather than a collection of scripts.
Architecture
- Ingestion: raw data lands in S3, encrypted with KMS.
- ETL: AWS Glue transforms the raw reviews into a training-ready format, stored in a second bucket.
- Training: SageMaker trains the sentiment model from the transformed data and stores the artifacts in S3.
- Batch inference: SageMaker Batch Transform scores new data from an inference-input bucket into an output bucket.
- Analytics: Amazon Athena queries the results directly on S3 with SQL.
- Orchestration & monitoring: Step Functions coordinates the stages; CloudWatch watches them run.