2.1. Data Ingestion and Storage

💡 First Principle: Your model can only learn from data it can access efficiently. Choosing the wrong storage format or ingestion path doesn't just slow things down—it creates bottlenecks that cascade through the entire ML lifecycle. A model trained on a fraction of available data because ingestion was too slow is a model making decisions with incomplete information.

What happens when a team stores their training data as uncompressed CSV files in a single S3 bucket with no partitioning? Every training job reads every row of every column, even when the model only needs three features from last month. Training that should take 20 minutes takes 4 hours, costs 12x more, and the team stops iterating because each experiment feels too expensive. The storage decision made weeks ago now dictates the pace of model improvement.

Think of data ingestion like a restaurant's supply chain. You wouldn't have a Michelin-star kitchen receive all ingredients in unlabeled, mixed-together boxes. You'd want ingredients sorted by type, labeled, delivered on schedule, and stored at the right temperature. In ML, "sorted by type" means columnar formats with proper schemas, "delivered on schedule" means streaming pipelines for real-time data, and "stored at the right temperature" means choosing the right AWS storage tier for access patterns.

⚠️ Common Misconception: S3 is "just storage" — candidates underestimate its role in ML performance. S3 partitioning, file format, and file size directly affect training speed. SageMaker's Pipe mode streams data directly from S3 into training, but only works with RecordIO/Protobuf or specific file formats. Choosing CSV when Parquet would work means paying more for slower training — and the exam expects you to make this connection.