AWS has a service or feature for almost every stage of the AI and machine learning (ML) pipeline. The exam often describes one stage in a scenario and asks which service fits it. This lesson maps the main services to the stages where they help, from preparing data to watching a deployed model.
Preparing Data
- AWS Glue is a serverless data integration service. It finds, combines, cleans and moves data from many sources, and keeps a catalogue of what data exists. It's the general-purpose choice for getting data ready at scale.
- Amazon SageMaker Data Wrangler is a visual tool inside SageMaker AI for preparing data specifically for ML. You can explore data, clean it and build features with little or no code, then pass the result into training.
Rule of thumb: Glue for broad data pipelines across the organisation, Data Wrangler for an ML practitioner shaping a training dataset.
Storing and Sharing Features
Amazon SageMaker Feature Store is a central place to store, find and reuse features: the prepared inputs a model learns from. Teams can share features instead of rebuilding them. The same feature values can be served both for training and for real-time predictions, which keeps the two consistent.
Training, Tuning and Deploying
Amazon SageMaker AI covers the core of building traditional ML models. Its training jobs run training on managed computing resources. Automatic model tuning tries many hyperparameter combinations to find the best-performing model. Deployment options then serve the model to applications.
For generative AI, Amazon Bedrock gives managed access to foundation models through an API, with tools to customise them and build applications on them, so many teams never train a model at all.
Monitoring in Production
Monitoring compares live data and predictions against a baseline from training time, and raises alerts when data quality or model quality drifts. Those alerts are what prompt retraining. On AWS this was the job of Amazon SageMaker Model Monitor, which is now closed to new customers. AWS points new customers to its open-source SageMaker AI monitoring solutions, used with Amazon QuickSight dashboards. Amazon CloudWatch covers the operational side: every endpoint reports metrics such as latency and errors, and CloudWatch alarms fire when they cross a threshold.
Building and Using AI Applications
Two newer services sit at the edges of the pipeline:
- Kiro is an agentic coding service built on Amazon Bedrock. It helps developers turn a request into specifications, implementation steps, code and tests. It speeds up building the software around a model, not the model itself.
- Amazon Quick is an AI assistant for work. Business users can ask questions in plain language, research topics, build dashboards, and automate repetitive tasks across their company's data and apps. It's how many people consume insights rather than build models.
On the Exam
| Need | Service |
|---|---|
| Large-scale data integration and cataloguing | AWS Glue |
| Visual data preparation for ML | SageMaker Data Wrangler |
| Share and reuse features for training and inference | SageMaker Feature Store |
| Train and tune custom models | SageMaker AI (training jobs, automatic model tuning) |
| Use foundation models without hosting them | Amazon Bedrock |
| Detect drift in a deployed model | Model monitoring (SageMaker Model Monitor for existing customers; CloudWatch with AWS's open-source monitoring solutions for new ones) |
| Alarm on an endpoint's latency or errors | Amazon CloudWatch |
| AI help writing application code | Kiro |
| Business users asking questions and automating work | Amazon Quick |
Common traps: picking Model Monitor for data preparation, or Feature Store for monitoring. Their names hint at the stage, so read them literally.