Mlflow Mlops Migration

MLflow MLOps Migration

A phased, gated workflow that turns an arbitrary ML codebase — however unstructured — into a production-grade open-source MLflow 3 setup covering the full MLOps cycle: tracked experiments, a domain-modelled registry, dev/staging/prod separation, evaluation-gated promotion, and served models. It is written to be driven with a developer who has no MLflow 3 experience: every phase produces a reviewable artifact before anything is changed, and every API decision defers to the sibling mlflow-3 rule pack (which is pinned to mlflow 3.15.1 and names the MLflow 2-era idioms this migration exists to remove).

When to Apply

Use this skill when:

• A team wants MLflow (or has a messy/partial MLflow 2 setup) and needs the path to a production-grade MLflow 3 deployment — not just API fixes.
• Training code exists but experiments are untracked, models are shipped by copying files, or "deployment" means a pickle in a bucket.
• You are asked to design or review a dev/staging/prod model-promotion story.
• An MLflow 2 → 3 migration touches infrastructure (stages, ./mlruns file stores, MLServer), not only client code.

Don't use it for a single API question — read the relevant mlflow-3 rule directly.

Workflow Overview

0 assess ─▶ 1 domain-model ─▶ 2 environments ─▶ 3 instrument ─▶ 4 promote ─▶ 5 serve ─▶ 6 operate
  audit        registry           tracking per      training code    eval-gated    validate,     retrain loop,
  report       naming, alias      env (dev local,   → MLflow 3       copy_model_   serve,        challenger,
  (script,     + gate design      stg/prod DB+S3    idioms (rule     version +     smoke-test    maintenance
  read-only)   (interview)        + auth)           pack)            alias flip    /invocations  (gated)

Phases run in order — each has entry/exit criteria in references/workflow.md, and scripts/verify.sh is the exit gate for the infrastructure phases. Re-running any phase is safe: 00-assess.sh regenerates only its own report (and refuses to clobber anything else), scaffold-dev-tracking.sh refuses to overwrite (exit code 2 = already done), and verify.sh only reads. The one non-idempotent step is promotion's copy_model_version — see references/promotion.md for how to resume instead of re-copying.

Risk Level: Write

This workflow edits training code, writes infrastructure files, and stands up services. Guardrails:

• Nothing in phase 0–1 modifies anything — always complete both before touching code or infra.
• Confirm with the user before: starting/replacing any tracking server, rewriting a training entrypoint, flipping a prod @champion alias (dev/staging flips may be automated by the phase-4 pipeline), and exposing a serving endpoint beyond localhost.
• Two maintenance commands are destructive and must be run only with explicit user confirmation and a stated reason: mlflow gc (permanently deletes soft-deleted runs and experiments — registry entities are untouched) and mlflow db upgrade (irreversible schema migration — snapshot the database first). A PreToolUse hook in hooks/hooks.json blocks both unless MLFLOW_MAINTENANCE_ACK=yes is set for that command, so they cannot run un-confirmed by accident.

Requirements

• Python ≥ 3.10 with mlflow==3.15.1 installed in the project environment
• bash, curl, jq — the scripts use them
• uv — the serving phase uses --env-manager uv for fast isolated environment rebuilds (substitute virtualenv everywhere if uv is unavailable)
• Docker + docker-compose — for the dev tracking stack and build-docker serving images
• A database + object store per shared environment (staging/prod) — PostgreSQL/MySQL and S3/GCS/Azure; dev runs on the scaffolded local stack
• The sibling mlflow-3 skill — phase 3 cites its rules; if it is not installed, read the MLflow 3 migration guide instead (the workflow still works, with more manual verification)

Setup

config.json starts empty. Phase 2 fills it (tracking URIs per environment, registry namespace, model name, serving URL). If fields are empty when a script needs them, the script says which ones — fill them via the _setup_instructions in the file.

Quick Reference

Gotchas

See gotchas.md — failure points discovered while running this workflow, including the migrate-filestore SQLite-only target and the basic-auth bootstrap credentials.

Related Skills

• mlflow-3 — the sibling library-reference rule pack this workflow cites at every API decision