Tracking at Scale

Two upgrades that matter

1. Remote tracking server — a single MLflow / W&B instance shared by all teams, backed by a real database (Postgres) and object store (S3).
2. Naming + tagging conventions — agreed at team level, not invented per-experiment. Without conventions, search becomes archaeology.

A minimal convention

experiment: <business_goal>-<quarter>
tags: owner=<team>, model_family=<xgb|nn|linear>, dataset_version=<hash>

Six months later, anyone can find 'the best fraud model from Q2 by the risk team' in seconds.

A shared tracking server is a library catalog. Without the catalog, every researcher has their own pile of books in their own office. The library is only useful when everyone files books with the same Dewey decimal scheme.