Cloud Computing for Data Science

🎯 Learning Objectives

• Understand cloud service models (IaaS, PaaS, SaaS)
• Provision cloud resources for data science workloads
• Use cloud storage (S3, GCS, Blob) for data lakes
• Deploy ML models on cloud platforms

📖 Core Content

7.1 Cloud Service Models

7.2 Cloud Services for Data Science

7.3 Cloud Storage for Data Science

python
Run Inline
# runnable
# Note: Requires boto3 (AWS) or google-cloud-storage (GCP)
# AWS S3 example:
# import boto3
#
# s3 = boto3.client('s3')
#
# # Upload data
# s3.upload_file('local_data.csv', 'my-bucket', 'data/raw/data.csv')
#
# # Download data
# s3.download_file('my-bucket', 'data/processed/features.parquet', 'features.parquet')
#
# # List files
# response = s3.list_objects_v2(Bucket='my-bucket', Prefix='data/')
# for obj in response['Contents']:
#     print(obj['Key'])

7.4 Cost Management

📝 Practice Questions

Q1: When would you use EC2 vs Lambda vs SageMaker?

EC2: Full control, any workload, persistent. Lambda: Short-running (≤15 min), event-driven, no server management. SageMaker: Managed ML lifecycle (training, tuning, deployment). Use SageMaker for ML, Lambda for simple API triggers, EC2 for general compute. Q2: What is object storage and why is it good for data lakes?

Object storage (S3, GCS) stores data as objects with metadata and unique IDs. It's: (1) infinitely scalable (exabytes), (2) cheap ($0.023/GB/month), (3) accessible via HTTP APIs, (4) durable (99.999999999% durability). Perfect for storing raw data before processing. Q3: How do you avoid cloud cost surprises?

1. Set budget alerts (notify at 50%, 80%, 100% of budget)
2. Use auto-scaling to reduce idle resources
3. Clean up unused resources (forgotten EC2 instances, old snapshots)
4. Use spot instances for non-critical training
5. Monitor costs weekly; spikes usually mean a misconfiguration Join Discord PreviousMLflow & ExperimentsNextRegular Expressions