Cloud ML: AWS SageMaker and GCP Vertex AI

Cloud ML platforms provide managed infrastructure for training, tuning, deploying, and monitoring models at scale.

Cloud ML Landscape

1. AWS SageMaker Workflow

SageMaker SDK Example

import sagemaker
from sagemaker.sklearn.estimator import SKLearn

estimator = SKLearn(
    entry_point="train.py",
    role=sagemaker.get_execution_role(),
    instance_count=1,
    instance_type="ml.m5.xlarge",
    framework_version="1.0-1",
    hyperparameters={"n_estimators": 100, "max_depth": 8},
    output_path="s3://bucket/output"
)

estimator.fit({"train": "s3://bucket/train/", "test": "s3://bucket/test/"})

# Deploy to endpoint
predictor = estimator.deploy(
    initial_instance_count=1,
    instance_type="ml.t2.medium"
)

2. GCP Vertex AI

Vertex AI Pipeline (Kubeflow)

from google_cloud_pipeline.pipeline import pipeline
from kfp import dsl

@dsl.pipeline(name="training-pipeline")
def training_pipeline(
    data_path: str = "gs://bucket/data/",
    n_estimators: int = 200
):
    prepare = dsl.ContainerOp(
        name="prepare",
        image="gcr.io/project/prepare:latest",
        arguments=["--input", data_path]
    )

    train = dsl.ContainerOp(
        name="train",
        image="gcr.io/project/train:latest",
        arguments=["--n-estimators", n_estimators]
    ).after(prepare)

    evaluate = dsl.ContainerOp(
        name="evaluate",
        image="gcr.io/project/evaluate:latest"
    ).after(train)

3. Feature Store Comparison

4. Managed Inference

# SageMaker Real-time
from sagemaker.model import Model

model = Model(
    image_uri=inference_image,
    model_data="s3://bucket/model.tar.gz",
    role=role
)
predictor = model.deploy(instance_type="ml.g4dn.xlarge", initial_instance_count=1)

# Serverless inference
from sagemaker.serverless import ServerlessInferenceConfig

predictor = estimator.deploy(
    serverless_inference_config=ServerlessInferenceConfig(
        memory_size_in_mb=2048,
        max_concurrency=50
    )
)

5. Cost Optimization

6. Multi-Cloud Strategy

• Avoid lock-in: Use abstracted interfaces (Kubeflow, MLflow)
• Data gravity: Keep data where compute lives
• Cost comparison: Profile workloads across providers
• Compliance: Consider data residency requirements

Key Takeaways

• SageMaker: Most comprehensive; best for AWS-native shops
• Vertex AI: Strong AutoML and BigQuery integration; Google Research models
• Azure ML: Enterprise integration; OpenAI access; responsible AI tools
• Cost: Spot training + right-sized inference = 50-70% savings