`mlflow models build-docker`

One artefact, anywhere

mlflow models build-docker \
  -m 'models:/credit-scoring/Production' \
  -n credit-scoring:latest

What you get is a self-contained image with the model, its Python environment, the MLflow scoring server, and a documented /invocations endpoint. Push it to ECR/GCR/ACR and you can run it on ECS, GKE, AKS, Kubernetes, Cloud Run — anywhere.

Production checklist

• Pin the base image (--base-image python:3.11-slim).
• Run as non-root (USER 1000).
• Add a readiness probe on /ping, liveness on /health.
• Set resource requests/limits — model memory is sticky.
• Log to STDOUT only; the platform aggregates.

Building a model image is canning a stew. The cook is gone but anyone with a can opener can heat it up identically two years later. The label (image tag) tells you which recipe, the can (image hash) makes sure nobody swapped the contents.