ephizen / wiki

MLflow

Open-source platform for the ML lifecycle — experiment tracking, model registry, packaging, and deployment.

Category
MLOps
Difficulty
Intermediate
When to use
You need experiment tracking and a model registry without buying a full MLops platform, and you want something self-hostable.
When not to use
You're a solo researcher running ten experiments a month — a spreadsheet or Weights & Biases free tier is easier.
Alternatives
Weights & Biases Neptune Comet ClearML
https://mlflow.org →

At a glance

FieldValue
CategoryML lifecycle platform
DifficultyIntermediate
When to useExperiment tracking + registry, self-hosted
When not to useTiny projects; heavy-collaboration teams that want polished UX
AlternativesWeights & Biases, Neptune, Comet, ClearML

What it is

MLflow has four components: Tracking (log params, metrics, artifacts per run), Projects (packaging training code), Models (a standard format with flavors for sklearn, PyTorch, XGBoost, etc.), and Model Registry (versioned, stage-tagged model store). It runs against a backend store (Postgres is standard) and an artifact store (S3/GCS).

When we reach for it at Ephizen

  • Comparing training runs side by side with clear lineage from code commit to metrics.
  • The registry as the source of truth for “which model is in production”.
  • Packaging classical ML and small deep learning models into a standard serve-able format.

Getting started

import mlflow
mlflow.set_experiment("churn")
with mlflow.start_run():
    mlflow.log_params({"lr": 0.05, "max_depth": 6})
    mlflow.log_metric("val_auc", 0.87)
    mlflow.sklearn.log_model(clf, "model", registered_model_name="churn-xgb")

Gotchas

  • The tracking UI is functional but dated. Weights & Biases is much nicer if budget allows.
  • Self-hosting needs real backing storage — SQLite + local files is fine for one user, not a team.
  • log_artifact on large files is slow over the network; prefer logging URIs for huge artifacts.

Related tools