DevOps Habits ML Apps Still Need

An ML endpoint without versioned config is a science experiment. Pin packages, models, and feature flags like you pin code.

Bake models into artifacts when they fit. Ambiguous downloads at deploy time make incidents non-reproducible.

Health checks should exercise the model path, not only return 200 OK from a static route.

Separate secrets from training notebooks. Leaked keys in .ipynb files are still common and still bad.

Automate the path from merge to canary. Manual SSH releases do not get safer as models get larger.