Cost Optimization¶

Overview¶

Strategies for optimizing ML costs on AWS.

AWS Cost Management Tools¶

Tool	Purpose
Cost Explorer	Visualize and analyze costs
AWS Budgets	Set cost alerts
Savings Plans	Commit to usage for discounts
Spot Instances	Use spare capacity
Compute Optimizer	Right-sizing recommendations

SageMaker Cost Optimization¶

Training Costs¶

Spot Training

Save up to 90% on training costs.

estimator = Estimator(
    ...
    use_spot_instances=True,
    max_wait=3600,
    max_run=1800,
    checkpoint_s3_uri="s3://bucket/checkpoints/"
)

Right-sizing Instances

Instance Type	Use Case
ml.m5.*	General purpose
ml.c5.*	CPU-intensive
ml.p3./ml.g4dn.	Deep learning
ml.r5.*	Memory-intensive

Managed Warm Pools

Reduce training startup time.

estimator = Estimator(
    ...
    keep_alive_period_in_seconds=3600  # 1 hour
)

Inference Costs¶

Serverless Endpoints

Pay per request, no idle costs.

serverless_config = ServerlessInferenceConfig(
    memory_size_in_mb=2048,
    max_concurrency=10
)

Multi-Model Endpoints

Host multiple models on single instance.

mme = MultiDataModel(
    name="mme-endpoint",
    model_data_prefix="s3://bucket/models/",
    image_uri=image_uri,
    role=role
)

Auto Scaling

Scale to zero during low traffic (Serverless) or minimum capacity.

Storage Costs¶

Use S3 lifecycle policies for old data
Clean up unused model artifacts
Use appropriate storage classes

AWS Budgets¶

import boto3

budgets = boto3.client("budgets")

budgets.create_budget(
    AccountId="123456789012",
    Budget={
        "BudgetName": "ML-Monthly-Budget",
        "BudgetLimit": {"Amount": "1000", "Unit": "USD"},
        "TimeUnit": "MONTHLY",
        "BudgetType": "COST"
    },
    NotificationsWithSubscribers=[{
        "Notification": {
            "NotificationType": "ACTUAL",
            "ComparisonOperator": "GREATER_THAN",
            "Threshold": 80,
            "ThresholdType": "PERCENTAGE"
        },
        "Subscribers": [{
            "SubscriptionType": "EMAIL",
            "Address": "team@example.com"
        }]
    }]
)

Cost Allocation Tags¶

Track costs by project, team, or environment.

sagemaker.create_training_job(
    ...
    Tags=[
        {"Key": "Project", "Value": "CustomerChurn"},
        {"Key": "Environment", "Value": "Production"},
        {"Key": "Team", "Value": "DataScience"}
    ]
)

Exam Tips¶

!!! warning "Key Strategies" - Spot Instances for training (up to 90% savings) - Serverless endpoints for variable traffic - Multi-model endpoints for many models - Right-size instances with Compute Optimizer - Use Budgets for cost alerts