You already know: lineage and reproducibility (ML lifecycle); DAGs, caching and partitions (ML pipelines); processes vs threads and the GIL (D1 Python internals, GIL concurrency); a model registry as a table (D1 Databases).
Why this matters for Atlas. The ML tooling landscape is crowded, and tools overlap: MLflow tracks experiments and serves models; Kubeflow orchestrates and tunes; cloud platforms bundle all of it. The way to choose without being sold to is to know the handful of problems these tools solve and the mechanism each uses: an append-only log of runs, content-addressed storage, layered immutable images, a task scheduler. This lesson builds a tiny working version of each mechanism, so the real tools stop being magic, and ends with how to decide what to build, buy or skip.
1. The map: problems first, tools second
The question. Which tool does what?
Start from nothing. Group tools by the problem they solve: