Day 2 - Amazon SageMaker: A Platform for Directly Training and Deploying ML Models

📌 Key Takeaways

• Amazon SageMaker AI (formerly Amazon SageMaker) is a fully managed ML platform that helps with the entire ML lifecycle, from data preparation through training, deployment, and monitoring.
• The service manages the infrastructure, but you decide the data, the algorithm, and the deployment method. Bedrock is about renting FMs, while SageMaker is about building and controlling your own models.
• JumpStart is a catalog of pre-trained models and solution templates. You pick one and deploy it right away, or fine-tune it on your own data.
• There are four inference methods. Use real-time for immediate responses, serverless for intermittent traffic, asynchronous for large inputs and long processing, and batch transform for processing accumulated data in bulk.
• Costs come from the instance hours while training and the hours an endpoint is up. A real-time endpoint is billed even when there are no requests.

SageMaker Covers the Entire ML Lifecycle

Yesterday's Bedrock was a service for "renting a huge model someone else built." Today's SageMaker is the opposite in character: a platform where you build, train, and deploy your own machine learning models. The AIF-C01 exam does not ask about deep SageMaker technology. Instead it asks, as a big picture, "when is it Bedrock and when is it SageMaker," and how SageMaker helps across the whole ML lifecycle.

• Let's sort out the name first. The current official name is Amazon SageMaker AI, and the exam guide uses this name. Even if a question or document says just "Amazon SageMaker," you can take it to mean the same ML platform.

A traditional ML project goes through a long journey: data preparation → training → evaluation → deployment → monitoring. SageMaker provides tools for each of these stages.

The key point is that SageMaker "manages the infrastructure for you, but I control the model itself." You decide the algorithm, hyperparameters, and data used for training.

There are several more tools that the exam guide names directly. It is enough to pair each name with a one-line role.

💡 Related theory: SageMaker is "managed," not a "serverless abstraction." With Bedrock you only need to call an API without knowing anything about the model's internals, but with SageMaker you design the ML workflow (data, training, evaluation) yourself. In exchange, SageMaker takes over the "tedious infrastructure work" such as provisioning GPU clusters, configuring distributed training, and scaling. This difference is the criterion that separates the two services on the exam.

By analogy, Bedrock is a restaurant where you order a finished dish, and SageMaker is a shared kitchen fully equipped with burners and cooking tools. The kitchen maintains the equipment, but you bring the ingredients (data) and the recipe (algorithm), and in return you can make it to your own taste.

SageMaker JumpStart: Providing a Starting Point

Building a model from scratch is a burden. SageMaker JumpStart offers pre-trained models and solution templates as a catalog so you can get started quickly.

What you can do with JumpStart:

• Deploy publicly available pre-trained models (image classification, text generation, and so on) in a few clicks
• Take a foundation model and fine-tune it on your own data
• Use example templates for common ML solutions

JumpStart may feel like "a bridge between Bedrock and SageMaker." It works with FMs, but the difference is that in the SageMaker environment you host and fine-tune them yourself, with deeper control.

• Whether fine-tuning and evaluation are possible differs by model, and for third-party models the user must check the license terms.
• The exam guide divides the ways of using an FM into a "managed API service" and a "self-hosted API." The former is Bedrock, and the latter is JumpStart and SageMaker endpoints.

As a hypothetical example, a manufacturing company needed a defective-product image classification model. Instead of designing a neural network from scratch, it took a pre-trained image classification model from JumpStart and fine-tuned it on several thousand of its own defect images. A few days were enough, and it could keep its own data and model inside the company under its own control. This is a case where JumpStart answered the requirement "I want to use an FM but keep it under my control."

Training and Deployment: Training Jobs and Endpoints

SageMaker's two pillars are training and inference.

• Training Jobs: Automatically start training instances, train the model, and clean up the instances when finished. You pay only while training.
• Endpoints (inference): Host the trained model and receive prediction requests. You choose the inference method according to the nature of the work.

A real-time endpoint is "always on," so it costs money even with no traffic. That is why serverless inference is cost-efficient when traffic is uneven, and batch transform is cost-efficient when you process data that accumulates at a fixed time each day all at once. When the exam gives clues such as "intermittent traffic" or "nightly bulk processing," it is asking which inference method applies.

The four methods are the items that the latest exam guide names directly as "types of inference" (batch, real-time, asynchronous, serverless). You do not need to memorize the numbers, but looking at the order of the limits' sizes shows why they are divided this way.

Two questions are enough to leave one answer.

① Do you need the answer right now?
   ├─ Yes ─► ② Is the traffic steady?
   │          ├─ Steady / latency is very important ──► Real-time
   │          └─ Occasional, irregular ───────────────► Serverless
   └─ No ──► ② Is all the data already accumulated?
              ├─ Accumulated (a bundle of files) ─────► Batch transform
              └─ Requests arrive one by one, large and slow ─► Asynchronous

• The cue that separates serverless from asynchronous is size and time. If there is only "intermittent," it is serverless; if "large file" or "tens of minutes of processing" is attached, it is asynchronous.
• A serverless endpoint that has been idle may incur a warm-up time (cold start) when it receives its first request. If you are very sensitive to latency, real-time is the right choice.
• The cue that separates asynchronous from batch transform is the shape in which requests arrive. If users upload items one at a time, it is asynchronous; if you run what has piled up in a warehouse all at once, it is batch.

So When Is It Bedrock and When Is It SageMaker?

This distinction is the core of today.

The decision criterion is simple. "Is an existing huge model enough (Bedrock)" vs "Do I need to build and control my own model with my own data (SageMaker)." If you want to quickly add a generative AI feature, use Bedrock. If you need to operate a domain-specific prediction model under fine-grained control, use SageMaker.

💡 Related theory: The two are complementary, not competing. Many companies run generative features on Bedrock and traditional ML such as sales forecasting and anomaly detection on SageMaker side by side. On the exam, splitting "several FMs without management" to Bedrock and "deploy and monitor a custom model trained on your own data" to SageMaker is usually right.

Common misconceptions.

• "SageMaker is only for traditional ML and has nothing to do with generative AI" → No. You deploy and fine-tune foundation models with JumpStart.
• "You cannot change the model at all in Bedrock" → No. Bedrock also has fine-tuning and distillation. The difference is how deeply you control it. If you must decide the training code, the algorithm, and even the instances yourself, it is SageMaker.
• "It is managed, so the cost is minimized automatically" → No. Endpoints and notebooks you leave on are billed even when not in use.
• "Even numeric prediction on tabular data is better with an FM these days" → The exam asks the opposite. Prediction under regulation, explainability, or operational constraints may suit a traditional ML model, and SageMaker is the stage for that.

Where the Costs Come From

Scenario → Answer Decision Table

Starting tomorrow, we look at managed AI services (Rekognition, Textract, Comprehend, Transcribe, Polly, Translate) that process images, documents, text, and speech with a single API call, with no need to build a model.

📖 Glossary

• ML lifecycle: One full turn of a model, running through data preparation, training, evaluation, deployment, and monitoring.
• Training Job: SageMaker's unit of work that starts training instances, trains a model, and cleans up when finished.
• Endpoint: The address where a trained model is deployed and receives prediction requests.
• JumpStart: A catalog for picking, deploying, and fine-tuning pre-trained models and solution templates.
• AutoML: A method that automates algorithm selection and tuning. In SageMaker, Autopilot handles it.
• Drift: The phenomenon where the data coming in during operation or the model quality differs from what it was at training time.
• Batch transform: A method of predicting accumulated data all at once without an always-on endpoint.
• Cold start: The delay that occurs while an idle serverless resource is prepared to receive its first request.
• Payload: The data sent along with a single request.
• MLOps: A way of operating a model's experimentation, deployment, monitoring, and retraining as repeatable procedures.

📝 Practice Questions

Click a choice to reveal the answer and explanation.

Question 1

A team wants to directly train a custom predictive model using their company's unique data and continuously monitor model quality post-deployment. What is the most appropriate service?

• AAmazon Bedrock
• BAmazon SageMaker
• CAmazon Polly
• DAmazon Translate

Question 2

Which most accurately explains the difference between Bedrock and SageMaker?

• ABedrock trains models directly, SageMaker only calls APIs
• BBedrock rents FMs via API; SageMaker builds and operates your own models
• CThey're completely the same service
• DSageMaker doesn't support generative AI at all

Question 3

You want to quickly deploy pre-trained models and solution templates from a catalog or fine-tune them with your company's data. Which SageMaker feature is this?

• AModel Monitor
• BJumpStart
• CData Wrangler
• DBatch Transform

Question 4

You want to process large accumulated data at a set time each night. What's the most cost-effective inference method?

• AReal-time endpoint
• BBatch Transform
• CRetrain the model each time
• DManually operate an always-on GPU instance

Question 5

Traffic is irregular and sporadic; you want costs near zero when there are no requests. What inference method suits this scenario?

• AServerless inference
• BAlways-on real-time endpoint
• CBatch Transform
• DAsynchronous inference is the only option