MLflow Registry

One registry the whole company trusts

MLflow is the single supported place every data scientist and AI team tracks experiments, registers models, and manages aliases, with 120+ active projects in flight. I own it as production infrastructure: Cloud SQL for metadata, GCS for artifacts, and a tracking server that proxies all artifact access so no client ever needs bucket credentials.

Run boring on purpose

• Autoscales 2 to 10 pods on CPU, with health probes, Datadog APM, and Prometheus metrics, so it stays observable and heals itself.
• Artifact proxying means teams log and load models without ever touching GCS credentials.

Knowing what not to ship

• Tested MLflow 3.x’s new GenAI job queue and caught it filling a tiny in-pod disk and pinning CPU at 600 percent in a retry loop. Disabled it until the platform can run it safely across pods.
• MLflow does not auto-migrate its Postgres schema, so I made version bumps require an explicit upgrade plan before they ship. No silent schema drift.
• Dropped an IAM-proxy auth experiment back to managed Postgres when the extra complexity was not earning its keep.

Why it matters

One trusted registry, kept boring on purpose. Teams get a reliable place to track and ship models, and the platform gets clear operational rules and no surprises.