Reference ArchitectureAWSMLOps and CI/CDData engineering
AWS MLOps Workflow with GitLab CI/CD
A reference pattern showing how the pieces fit together — not a specific deployed system.
End-to-end MLOps on AWS: governed data, SageMaker training and evaluation, GitLab CI/CD with security checks, staged rollout and monitored retraining — with no automatic promotion to production.
Overview
- Data & model development: ingestion (AWS DMS, Kinesis) into a versioned S3 data lake governed with Glue Data Catalog and Lake Formation; validation with quarantine for failed quality checks; SageMaker Feature Store and time-aware train/validation/test splits; SageMaker Pipelines for training, tuning, evaluation on holdout data, a threshold gate and Model Registry.
- CI/CD & release control (GitLab): linting, unit, data-contract and integration tests; dependency, secrets, container and IaC scanning; signed images in Amazon ECR; pipeline triggers, evaluation results, staging tests, human approval and an immutable release manifest.
- Staging, deployment & monitoring: separate staging and production accounts, blue/green or canary rollout with rollback to the previous approved version, CloudWatch and SageMaker Model Monitor for drift and quality, EventBridge alerts, and a feedback loop that triggers candidate retraining — always with human approval.
Business use case
A reference pattern for teams that need reproducible models, auditable releases and safe rollouts on AWS.
My role
Architecture design and documentation.