Amazon SageMaker AI is a fully managed service for building, training and deploying machine learning models.
Build, train, and deploy models
Amazon SageMaker AI is the AWS platform for custom machine learning. Data scientists prepare data, train models and deploy them to production in one managed service, without running their own servers.
How it works
- Prepare and store data.
- Train a model with built-in algorithms or your own framework, distributed across managed instances.
- Deploy the model to a secure, scalable endpoint.
- Lower ongoing costs with SageMaker AI Savings Plans.
| Need | Service type |
|---|---|
| Common AI task through an API (text, images, speech) | Pre-built AI services such as Comprehend, Rekognition, Transcribe |
| Your own model trained on your data | Amazon SageMaker AI |
Exam cues
- If the question says build, train and deploy custom ML models, think Amazon SageMaker AI.
- If the question says a common task with no ML expertise, think a pre-built AI service instead.
Remember
- Fully managed: build, train, deploy.
- Supports your own algorithms and frameworks.
Trap: SageMaker AI is for teams that build models. For ready-made capabilities like translation or face detection, pick the dedicated AI service.
From the AWS docs (5)
- Amazon SageMaker AI is a fully managed machine learning service where data scientists and developers build, train and deploy ML models into a production-ready hosted environment.
- With SageMaker AI you store and share data without building and managing your own servers, and it provides managed ML algorithms that run efficiently on very large data in a distributed environment.
- SageMaker AI supports bring-your-own algorithms and frameworks with flexible distributed training, and deploys models into a secure, scalable environment.
- SageMaker AI Savings Plans lower prices for SageMaker AI instance usage regardless of instance family, size, component or Region.
- With EnableNetworkIsolation set, SageMaker training and inference containers cannot make any outbound network calls, not even to Amazon S3, and get no AWS credentials.
Inference options
SageMaker AI offers four ways to serve a model, chosen by latency need, payload size and traffic pattern.
How it works
| Option | Latency | Best for |
|---|---|---|
| Real-time | Milliseconds | Interactive, steady traffic |
| Serverless | Milliseconds plus cold starts | Intermittent traffic, CPU models |
| Asynchronous | Near real-time, queued | Payloads up to 1 GB, up to 1 hour of processing |
| Batch transform | Minutes to hours | Whole datasets, no endpoint |
Exam cues
- If the question says immediate, low-latency predictions, think real-time inference.
- If the question says large payloads or long processing, think asynchronous inference.
- If the question says score a whole dataset on a schedule, think batch transform.
- If the question says sporadic traffic and tolerates cold starts, think Serverless Inference.
Remember
- Asynchronous: up to 1 GB payload, up to one hour.
- Serverless: no GPUs.
Trap: Batch transform doesn’t keep an endpoint running, so it can’t answer individual live requests.
From the AWS docs (5)
- SageMaker AI real-time inference suits interactive workloads with low-latency requirements.
- SageMaker AI Serverless Inference deploys models without managing infrastructure and suits workloads with idle periods between traffic spurts that can tolerate cold starts.
- SageMaker AI Serverless Inference doesn't support GPUs, so it suits smaller models that run on CPU; large foundation models typically need GPU-based real-time endpoints.
- Asynchronous inference queues requests and processes them asynchronously, for payloads up to 1 GB, processing times up to one hour and near real-time latency needs.
- Batch transform gets inferences for large datasets when you don't need a persistent endpoint, and can also preprocess datasets.
JumpStart, Canvas and Unified Studio
SageMaker AI covers the range from no-code to full code. Canvas builds predictions without code, JumpStart deploys and tunes pre-trained models, and Unified Studio puts data, analytics and AI work in one place.
How it works
| Tool | For | Gives you |
|---|---|---|
| SageMaker Canvas | Business analysts | No-code ML predictions |
| SageMaker JumpStart | Builders | Pre-trained FMs and models to deploy and fine-tune |
| SageMaker Unified Studio | Data and AI teams | One interface for data, analytics, ML and Bedrock |
Exam cues
- If the question says no code, business analyst, predictions, think SageMaker Canvas.
- If the question says pre-trained or open-source models to deploy or fine-tune on SageMaker AI, think SageMaker JumpStart.
- If the question says one environment for data, analytics and AI, think SageMaker Unified Studio.
Remember
- JumpStart supports PEFT and LoRA fine-tuning.
- Least effort: pre-trained model; most effort: custom container.
Trap: JumpStart models run on SageMaker AI endpoints that you pay for by the hour, not serverlessly like Bedrock’s catalog.
From the AWS docs (6)
- Amazon SageMaker JumpStart is an ML hub with hundreds of open-weight and proprietary foundation models that you can deploy in a few steps and customize with techniques such as Parameter-Efficient Fine-Tuning (PEFT) and LoRA.
- Pre-trained models from SageMaker JumpStart need the least effort, built-in algorithms need more, and custom code or a custom Docker image needs the most.
- Amazon SageMaker Canvas generates predictions with machine learning without writing any code, for use cases such as predicting customer churn or planning inventory.
- Amazon SageMaker Unified Studio is a single development experience that brings together AWS data, analytics, AI and ML services to build, deploy and monitor workflows from one interface.
- Amazon Bedrock in SageMaker Unified Studio (previously Amazon Bedrock IDE) lets you discover models, try prompts in a playground and build generative AI apps with knowledge bases and guardrails, with little or no code.
- SageMaker Canvas Ready-to-use models make predictions on your data without building a model, using Amazon Textract, Amazon Rekognition and Amazon Comprehend behind the scenes.
Data preparation and human review
Good models need prepared data and, often, people in the loop. SageMaker AI has tools to prepare and share data, to label it, and to send uncertain predictions to human reviewers.
How it works
| Tool | Job |
|---|---|
| Data Wrangler (in Canvas) | Prepare, transform and analyze data |
| Feature Store | Store and share features for training and inference |
| Ground Truth | Label training data with people plus automated labeling |
| Augmented AI (A2I) | Human review of low-confidence predictions |
Exam cues
- If the question says human review of low-confidence predictions, think Amazon A2I (human-in-the-loop).
- If the question says label training data with a workforce, think SageMaker Ground Truth.
- If the question says reuse features across teams, think SageMaker Feature Store.
- If the question says visual data preparation, think Data Wrangler in SageMaker Canvas.
Remember
- Ground Truth labels training data before training; A2I reviews predictions after deployment.
- Ground Truth and A2I are closed to new customers.
Trap: Ground Truth and A2I both involve people, but at different stages: labeling training data versus checking live predictions.
From the AWS docs (6)
- SageMaker Data Wrangler imports, prepares, transforms, featurizes and analyzes data for ML; it is now integrated into SageMaker Canvas, with a natural language interface as well as a visual one.
- Amazon SageMaker Feature Store creates, stores and shares features, the transformed inputs for ML models, for training and inference across teams and accounts.
- Amazon SageMaker Ground Truth builds labeled training datasets with a vendor or private workforce, and automated data labeling uses ML to decide which data needs human labels.
- Amazon Augmented AI (A2I) adds human review of low-confidence or randomly sampled predictions from Amazon Rekognition, Amazon Textract or custom ML models.
- SageMaker Ground Truth and Amazon A2I are no longer open to new customers; existing customers can keep using them, but no new features are planned.
- The Data Wrangler Data Quality and Insights Report analyzes an imported dataset and shows issues such as missing values and outliers to guide cleaning.
MLOps, monitoring and governance
MLOps applies DevOps habits to ML: automated, repeatable pipelines, versioned models, monitoring in production and retraining when the world changes. SageMaker AI has a tool for each step and for documenting models.
How it works
| Need | SageMaker AI feature |
|---|---|
| Repeatable training workflows | Pipelines |
| Versioned, approved models | Model Registry |
| Drift and quality in production | Model Monitor |
| Model documentation and risk | Model Cards |
| One view of all models | Model Dashboard |
- A pipeline prepares data, trains, evaluates and registers a model version.
- An approved version is deployed.
- Monitoring compares live data and predictions with the training baseline.
- Drift triggers retraining through the same pipeline.
Exam cues
- If the question says repeatable, automated ML workflow, think SageMaker Pipelines.
- If the question says document intended use and risk for governance, think SageMaker Model Cards.
- If the question says performance degrades as data changes, think drift; monitor and retrain.
Remember
- Model Monitor watches data quality, model quality, bias drift and feature attribution drift.
- Model Monitor is closed to new customers.
Trap: Model Cards document a model; they don’t monitor it. Monitoring production behavior is Model Monitor’s job.
From the AWS docs (10)
- A SageMaker AI pipeline is a series of connected steps in a directed acyclic graph (DAG), built with a drag-and-drop UI or the Pipelines SDK, that automates ML workflows.
- SageMaker Model Registry catalogs models for production and manages model versions.
- SageMaker Model Monitor watches models in production for drift in data quality, model quality (such as accuracy), bias and feature attribution, on real-time endpoints or batch transform jobs.
- SageMaker Model Monitor is no longer open to new customers; existing customers can keep using it, but no new features are planned.
- SageMaker Model Cards document a model's intended use and risk rating, training details and metrics, evaluation results and other considerations in one place, for governance and responsible AI.
- SageMaker Model Dashboard brings together model monitoring, model cards, workflow lineage and endpoint performance in one view.
- Data drift is a significant change in input data compared with the training data; concept drift is a change in the properties of the target; both degrade a model and trigger retraining.
- In SageMaker Model Registry, setting a model version's approval status to Approved can trigger CI/CD deployment; the status can be set manually or by a condition step in a pipeline.
- SageMaker Role Manager builds least-privilege roles for ML personas such as data scientists and MLOps engineers, but it is no longer open to new customers.
- Fully managed MLflow on SageMaker AI tracks experiments and monitors the performance of models and AI applications in one tool.
Flashcards
Nothing to review here right now. Pick other topics or another filter.
·
All 27 cards as a list
Data scientists want to build, train and deploy their own custom machine learning model without managing servers. Which AWS service fits?
Amazon SageMaker AI, a fully managed ML service.
When should a company use Amazon SageMaker AI instead of a service like Amazon Rekognition?
When it needs its own custom model trained on its own data. Rekognition and similar services offer pre-built AI for common tasks.
Can Amazon SageMaker AI train models with the team's own algorithms and frameworks?
Yes. It supports bring-your-own algorithms and frameworks with distributed training.
Is Amazon SageMaker AI the simplest choice to add sentiment analysis to an app with no ML team?
No. A pre-built service such as Amazon Comprehend does it through an API; SageMaker AI is for building custom models.
Which AWS service helps developers build, train and deploy machine learning models without managing their own servers?
Amazon SageMaker AI provides managed tools for building, training and deploying machine learning models.
A fraud check must score each card payment in milliseconds while the customer waits. Which SageMaker AI inference option fits?
Real-time inference, a persistent endpoint for interactive, low-latency requests.
A company must score a 50 GB file of customer records once a week and doesn't need an endpoint the rest of the time. Which SageMaker AI option fits?
Batch transform, which runs inference over a whole dataset without a persistent endpoint.
Requests carry 500 MB images and take about 20 minutes each to process. Which SageMaker AI inference option fits?
Asynchronous inference, built for payloads up to 1 GB and processing up to one hour.
An internal tool gets a few requests per hour with long idle gaps, and occasional cold starts are fine. Which SageMaker AI option minimizes cost?
Serverless Inference, which needs no instances and suits intermittent traffic.
Which SageMaker AI inference option suits an interactive application that needs low-latency predictions?
Real-time inference suits interactive workloads with low-latency requirements.
A team wants to deploy and fine-tune an open-weight foundation model from a catalog of pre-trained models, with its own instances. Which SageMaker AI feature fits?
Amazon SageMaker JumpStart.
A business analyst with no coding skills wants to build a churn prediction model. Which SageMaker AI tool fits?
Amazon SageMaker Canvas, which builds ML predictions without code.
What is the difference between Amazon Bedrock and SageMaker JumpStart for foundation models?
Bedrock serves FMs serverlessly through an API.
JumpStart deploys pre-trained models to SageMaker AI endpoints you manage and can fine-tune.
Which approach needs the least effort to get an ML model into use on SageMaker AI?
A pre-trained model from SageMaker JumpStart, ahead of built-in algorithms and custom code.
A data team wants one interface for data, analytics and AI work, including Amazon Bedrock prototyping. Which tool fits?
Amazon SageMaker Unified Studio.
A loan app built on Amazon Textract must send low-confidence extractions to a person to check. Which AWS service was built for this?
Amazon Augmented AI (A2I), which routes low-confidence predictions to human reviewers. It's no longer open to new customers.
Data scientists across teams keep rebuilding the same customer features. Which SageMaker AI feature lets them store and share features?
Amazon SageMaker Feature Store.
A team must label 100,000 images for training, using its own staff plus ML to label the easy ones automatically. Which SageMaker AI feature was built for this?
SageMaker Ground Truth with automated data labeling. It's no longer open to new customers.
Which SageMaker AI tool prepares and transforms data visually, now inside SageMaker Canvas?
SageMaker Data Wrangler.
How can Amazon SageMaker Ground Truth help prepare data for machine learning?
It helps build labeled training datasets using human workers and automated data labeling.
A team must automate data preparation, training, evaluation and registration of models as repeatable steps. Which SageMaker AI feature fits?
SageMaker Pipelines (a DAG of ML workflow steps).
An auditor asks for one document per model with its intended use, risk rating, training data and evaluation results. Which SageMaker AI feature fits?
SageMaker Model Cards.
A model's accuracy drops months after launch because customer behavior changed. What is this, and what fixes it?
Drift (data or concept drift). Detect it with monitoring and retrain the model on recent, representative data.
Which SageMaker AI feature stores approved model versions ready for deployment?
SageMaker Model Registry.
Which SageMaker AI feature was built to detect data quality, model quality, bias and feature attribution drift in production?
SageMaker Model Monitor. It's no longer open to new customers.
Is training a model once enough for MLOps?
No. MLOps means repeatable pipelines, versioned models, monitoring in production and retraining when data drifts.
How does SageMaker Model Registry help teams manage models for production?
It catalogs models for production and manages their versions.