Big Data on AWS Deep Dive (Part 9): SageMaker and the ML Platform — From Training to Production

Data is ready, features are computed, the online serving layer is built — now it is time to train and deploy models. This is where Amazon SageMaker takes the stage.

This chapter explains SageMaker as an umbrella product: SageMaker is not a single service — it is a collection of sub-services covering the entire ML lifecycle.

---

What Is SageMaker

One-Sentence Definition

Amazon SageMaker = AWS’s fully managed ML platform for the entire workflow. Everything you need for ML — data exploration, feature engineering, training, hyperparameter tuning, deployment, and monitoring — can be done within a single AWS service.

Naming Changes (Easy to Confuse — Post-2024 re:Invent Structure Through 2026)

Name	Meaning	Status (2026-05)
SageMaker AI	The original SageMaker. All ML/FM training and inference capabilities discussed in this chapter belong here	GA
SageMaker Lakehouse	Unified access to S3 + Redshift + federated data sources	GA
SageMaker Unified Studio	Unified IDE combining AI + Lakehouse + Glue Studio + EMR Notebook + Bedrock IDE	2025 GA
SageMaker Catalog	Governance layer based on DataZone (data + models + projects)	GA
Amazon SageMaker (new brand)	The umbrella covering all four pillars above	—

When we say “SageMaker” we typically mean SageMaker AI (the ML platform). The Lakehouse piece was already covered in Parts 2-4 (S3 + Iceberg + Glue Catalog).

2025-2026 additions: Unified Studio GA, Q Developer embedded in the IDE, Bedrock Foundation Models callable directly from SageMaker Studio, HyperPod task governance and flexible training plans.

---

The 5 Core Capabilities of SageMaker

Data Preparation: Studio and Processing Jobs

SageMaker Studio is a browser-based JupyterLab environment integrated with AWS services.

• Write Python notebooks, run wr.athena.read_sql_query(...) to query data directly
• Integrated Spark / SKLearn containers
• JupyterLab Spaces: isolated environments per user

Processing Job runs data preprocessing scripts:

• Spins up EC2, runs your Python script, tears down when finished
• Ideal for: feature generation, ETL, data cleansing
• Similar to EMR but more lightweight (no Spark cluster needed)

from sagemaker.sklearn.processing import SKLearnProcessor

processor = SKLearnProcessor(
    role=role, instance_type='ml.m5.xlarge', instance_count=2,
)
processor.run(
    code='preprocess.py',
    inputs=[ProcessingInput(source='s3://.../sample/', destination='/opt/ml/processing/input/')],
    outputs=[ProcessingOutput(source='/opt/ml/processing/output/', destination='s3://.../processed/')],
)

Feature Management: SageMaker Feature Store

This is especially important for recommendation scenarios.

The Problem: Training and inference must maintain feature consistency.

• During training you use feature X read from ads_user_features
• At serving time inference also needs the same feature X, but via a DynamoDB point lookup
• Field names, encoding, and default values on both sides must be identical — manual synchronization is extremely error-prone

The Feature Store Solution:

Define a Feature Group:
  user_features:
    - user_id (record id)
    - event_time (event time)
    - age, city, ctr_7d, last_5_clicks, ...
        |
        | ingest()
        v
  +-----------------------------+
  | Offline Store (S3+Iceberg)  |  <-- Read during training, auto PIT queries
  | Online Store (DynamoDB)     |  <-- Read at inference, millisecond latency
  +-----------------------------+
        ^
        | auto sync

Key characteristics:

• Offline Store: S3 + Iceberg, automatically partitioned by event_time
• Online Store: Built-in KV (DynamoDB-like, fully managed by AWS)
• Auto Sync: Write to Offline -> immediately synced to Online
• PIT Queries: get_features_at(event_time) automatically retrieves the snapshot at that point in time

Feature Store vs. Self-Built (DynamoDB + Iceberg):

	Feature Store	Self-Built
Online/Offline consistency	Automatic	Maintain sync jobs yourself
PIT queries	One API call	Write Iceberg time travel SQL yourself
Flexibility	Medium (schema constrained by framework)	High
Cost	Slightly more expensive	Cheaper

Recommendation: In the POC phase, self-build (DynamoDB + Iceberg); switch to Feature Store when productionizing. For complex scenarios, start with Feature Store from day one.

Official docs: https://docs.aws.amazon.com/sagemaker/latest/dg/feature-store.html

Training: Training Jobs

Built-in Algorithms

AWS provides 17+ built-in algorithms (XGBoost / LinearLearner / FactorizationMachine / KMeans / common deep learning):

from sagemaker.xgboost.estimator import XGBoost

xgb = XGBoost(
    entry_point='train.py',
    instance_type='ml.m5.2xlarge',
    instance_count=1,
    role=role,
    framework_version='1.7-1',
)
xgb.fit({'train': 's3://.../train/', 'val': 's3://.../val/'})

Custom Containers

PyTorch / TensorFlow / JAX all have official containers; for custom code and libraries:

from sagemaker.pytorch import PyTorch

estimator = PyTorch(
    entry_point='train_two_tower.py',
    framework_version='2.1.0',
    instance_type='ml.g5.xlarge',  # GPU
    role=role,
    hyperparameters={'lr': 0.001, 'embedding_dim': 64},
)
estimator.fit({'train': 's3://.../sample/'})

Hyperparameter Tuning

Automatic tuning (Bayesian optimization / grid / random):

from sagemaker.tuner import HyperparameterTuner, ContinuousParameter

tuner = HyperparameterTuner(
    estimator=xgb,
    objective_metric_name='validation:auc',
    hyperparameter_ranges={'eta': ContinuousParameter(0.01, 0.5)},
    max_jobs=20, max_parallel_jobs=4,
)
tuner.fit(...)

Distributed Training

Data parallelism / model parallelism / pipeline parallelism across multiple machines and GPUs.

Pricing

Billed per training instance type + duration, per-second billing. Stop training and billing stops immediately.

• ml.m5.xlarge: ~$0.23/hour
• ml.g5.xlarge (GPU): ~$1.4/hour
• ml.p4d.24xlarge (8x A100): ~$32/hour

Official docs: https://docs.aws.amazon.com/sagemaker/latest/dg/train-model.html

Deployment: Endpoints

After training, the model is saved to S3 and deployed to an Endpoint for online inference.

Real-Time Endpoint

Launches EC2 + loads model + exposes HTTP API:

predictor = xgb.deploy(
    initial_instance_count=2,
    instance_type='ml.m5.xlarge',
    endpoint_name='rank-model-v1',
)

# Invoke
result = predictor.predict({'features': [...]})

Characteristics:

• P99 ~ 30-50ms (including network)
• 24x7 online, billed per instance-hour
• Supports Auto Scaling (automatically adds nodes as QPS increases)

Serverless Inference

Pay-per-invocation, no cost when idle:

from sagemaker.serverless import ServerlessInferenceConfig

predictor = model.deploy(
    serverless_inference_config=ServerlessInferenceConfig(
        memory_size_in_mb=4096, max_concurrency=20,
    )
)

Cold start has a few seconds of latency — not suitable for the recommendation hot path (which requires stable millisecond latency).

Multi-Model Endpoint (MME)

A single Endpoint loads N models (lazy-loaded on demand), ideal for long-tail models (each with low individual QPS).

Async Endpoint / Batch Transform

Batch inference (e.g., scoring all users site-wide once per day).

A/B Testing in Endpoints

An Endpoint can host multiple Production Variants with traffic split by weight:

predictor.update_endpoint(
    initial_instance_count=2, instance_type='ml.m5.xlarge',
    new_endpoint_config_name='ab-test-config',
    # variant A 90% + variant B 10%
)

This makes gradual (Canary) model rollouts very friendly.

Official docs: https://docs.aws.amazon.com/sagemaker/latest/dg/deploy-model.html

Monitoring: Model Monitor

After a model goes online, the most important thing is not speed but sustained quality. Model Monitor tracks:

Monitor Type	What It Detects
Data Quality	Whether input feature distributions have drifted (concept drift)
Model Quality	Prediction vs. actual label AUC / error
Bias Drift	Prediction fairness across different user groups
Feature Attribution	Whether feature importance remains stable

Outputs alarms -> CloudWatch -> Slack / email. When drift is detected, it can automatically trigger retraining.

Official docs: https://docs.aws.amazon.com/sagemaker/latest/dg/model-monitor.html

---

Customer Scenario: SageMaker Usage Patterns

Two-Tower Recall Model

# Training (SageMaker Training Job)
PyTorch(
    entry_point='train_two_tower.py',
    instance_type='ml.g5.xlarge',
    role=role,
).fit({'train': 's3://.../ads_sample_follow/'})

# Output:
#   user_tower.pt      (user tower weights)
#   item_tower.pt      (item tower weights)

# Item embedding offline pre-computation (Processing Job)
SKLearnProcessor(...).run(
    code='compute_item_embedding.py',
    # Load item_tower.pt, compute vectors for all items on the platform
    # Output: item_id, embedding[64]
    outputs=[ProcessingOutput(destination='s3://.../item_embeddings/')]
)

# Write to OpenSearch k-NN (via Glue Job or Lambda)

# Deploy user_tower as an Endpoint
PyTorchModel(model_data='s3://.../user_tower.tar.gz').deploy(
    instance_type='ml.m5.xlarge', initial_instance_count=2,
    endpoint_name='user-tower-encoder',
)
# Recommendation service calls -> compute user vector in real-time -> query OpenSearch

LightGBM Ranking Model

# Training
SKLearn(
    entry_point='train_lgb.py',
    instance_type='ml.m5.4xlarge',
).fit({'train': 's3://.../ads_sample_ctr/'})

# Deployment
predictor = SKLearnModel(...).deploy(
    instance_type='ml.c5.xlarge', initial_instance_count=4,
    endpoint_name='rank-lgb-v1',
)

# Recommendation service scores each candidate
scores = predictor.predict(features_for_50_candidates)

End-to-End Orchestration

MWAA DAG:
  04:00 SageMaker Training: train_recall_two_tower
  04:00 SageMaker Training: train_rank_lgb
  05:00 SageMaker Processing: compute_item_embeddings
  05:30 Glue Job: write item_embeddings -> OpenSearch
  06:00 Lambda: deploy new endpoints (Canary 10%)
  10:00 Lambda: ramp up to 50%
  14:00 Lambda: full traffic

---

SageMaker and Bedrock in the GenAI Era (2026 Status)

Although the customer scenario is classic recommendation rather than LLM, the AWS GenAI stack is mature by 2025-2026. Here is the full landscape for future reference when you need RAG / Agents / content generation:

SageMaker AI Side (Self-Train and Self-Deploy)

• JumpStart: Open-source model marketplace (Llama 3 / Mistral / Stable Diffusion / DeepSeek etc.), one-click deployment
• HyperPod: Large model training clusters (thousand-GPU scale); 2025 adds task governance + flexible training plans (reserve GPU capacity by time slot)
• Inference Components: Multiple models share the same GPU instance, dramatically reducing long-tail LLM deployment costs
• MLflow on SageMaker: Managed MLflow for experiment tracking

Amazon Bedrock Side (API-Based Model Access)

Capability	Description
Foundation Models	Claude 4 series / Llama 3.x / Mistral / Cohere / Amazon Nova (Micro/Lite/Pro/Premier text + Canvas image + Reel video)
Bedrock Knowledge Bases	Managed RAG supporting S3 Vectors / OpenSearch / Aurora pgvector backends, structured data retrieval
Bedrock AgentCore + Agent Registry	New in 2025, managed Agent runtime + central registry for multi-agent / tool calling
Bedrock Guardrails	Content safety + PII filtering
Bedrock Prompt Management / Flows	Prompt versioning + visual orchestration

Amazon Q

• Q Developer: AI coding assistant embedded in the IDE (integrated into SageMaker Unified Studio)
• Q in QuickSight: BI natural language queries, Scenarios (What-if analysis), auto-generated Topics
• Q Business: Enterprise knowledge base Q&A

For the customer’s social app, if they later need “AI comment generation / content moderation / intelligent customer service,” start from Bedrock; if they need private LLM fine-tuning, use SageMaker HyperPod.

---

Feature Store: To Use or Not to Use

This decision often causes debate in customer scenarios. Here is a decision framework:

Situation	Choice
Team < 5 people, only 1-2 models	Do not use — self-built DynamoDB + Iceberg is more lightweight
Models > 5, strong feature reuse	Use it — avoid duplicating feature definitions
Multi-team collaboration	Use it — provides a unified registry
Strict online/offline consistency requirements	Use it
Extreme performance / cost sensitivity	Do not use — self-built allows finer tuning

For the POC phase with 1-2 models, self-build first to prove business value, then switch to Feature Store.

---

Chapter Summary

Sub-Service	Role
Studio / Notebook	Data exploration + model experimentation
Processing Job	Data preprocessing (lightweight ETL)
Feature Store	Online/offline consistent feature platform
Training Job	Managed model training (per-second billing)
Hyperparameter Tuning	Automated hyperparameter optimization
Endpoint	Online inference (Real-time / Serverless / Async / MME)
Model Monitor	Data / quality drift monitoring
JumpStart	One-click open-source model deployment (for GenAI)

The core value of SageMaker: let ML engineers focus on models, not infrastructure.

Next chapter assembles all components into the final architecture with cost estimates.