You already know: transactions and row-oriented databases (D1 Databases); partitioned logs and change data capture (D14 Messaging); object storage and partitioning (D14 Distributed storage); partitioned, idempotent pipelines and validation (D11 ML pipelines); content-addressed data versioning (D11 Platform tools).
Why this matters for Atlas. Every model is downstream of data, and AI data comes from everywhere: production databases, event streams, logs, files from partners, scraped web pages, human annotations. Ingestion brings it into one place reliably, with provenance; storage keeps it in formats that make the dominant access pattern (scanning columns of billions of rows) fast and cheap; table formats add the database guarantees (atomic writes, versioning, time travel) that make datasets reproducible. This lesson builds each piece: OLTP vs OLAP, the lakehouse architecture, why columnar files like Parquet are 10–100× better for analytics and training data, how a table format turns a pile of files into a versioned table, and how to ingest without duplicates or loss.
1. OLTP versus OLAP: two workloads, two designs
The question. Why can't Atlas's analysts just query the production Postgres database?
Start from nothing. Two workloads look nothing alike: