Comparing with MLflow
From MLflow runs to approved releases.
EUVYRA tracks experiments and versions models as MLflow users expect, and continues where a registry stops: releases that need an evaluation and named approvals, a gateway that serves only those, and the evidence an auditor asks for.
MLflow today
- MLflow is open source under the Apache 2.0 licence and is a project of LF Projects, LLC, part of the Linux Foundation.
- MLflow 3 was released on 11 June 2025. Its current documentation describes OpenTelemetry-compatible tracing and evaluation with built-in judges for generative AI applications.
- Since MLflow 3.7.0 of 5 December 2025, SQLite is the default backend of the MLflow tracking server.
Many teams run MLflow next to other tools for jobs, serving and compliance. EUVYRA is for teams that want those in the same installation, with one data model from run to release to audit.
Your MLflow work in EUVYRA
| What you do | In EUVYRA |
|---|---|
| Log parameters, metrics, tags and artifacts | Runs with configuration, metrics per step, tags, artifacts, git state and system load |
| Autolog from frameworks | Callbacks for Hugging Face, Lightning, Keras, XGBoost, LightGBM and scikit-learn |
| Register models with versions and aliases | Model versions with model cards and artifact digests, and artifact aliases such as best or production |
| Promote a model to production | A release with pre-flight checks, an evaluation gate and approvals bound to people's functions, then direct, canary or shadow rollout |
| Run training somewhere | Jobs on your Kubernetes cluster or GPU machines, with queues, quotas and cost per run |
| Serve a model | Endpoints with vLLM, llama.cpp or any container, behind a gateway that answers only for approved releases |
| Trace and evaluate GenAI apps | Traces from the SDK or OpenTelemetry, versioned test suites and gates for CI |
How to move
The importer reads MLflow's file store, the mlruns/ folder, directly. MLflow itself does not need to be installed.
Sign in from the CLI
pip install euvyra euvyra login --host https://euvyra.example.eu --api-key euv_...Import your runs
Parameters become the run configuration, metrics keep their steps and timestamps, the
mlflow.source.gittags become the run's git record, and start and end times and status are kept. Limit the import to one experiment with--experiment.euvyra import mlflow ./mlruns --project credit-scoring --experiment baselinePort your logging calls
# before: mlflow.log_params(params); mlflow.log_metric("auc", auc, step=epoch) import euvyra with euvyra.init(project="credit-scoring", config=params) as run: for epoch in range(epochs): run.log({"auc": evaluate()}, step=epoch) run.log_model("model.pkl", name="credit-gbm")run.log_modelregisters a model version whose source is this run, which is what a release approval later reads.
What does not carry over
- The importer reads the file store format only. Runs kept by a tracking server with a database backend, which is the default since MLflow 3.7.0, are not read directly.
- Artifacts, registered models, traces and evaluation datasets are not imported. Log the models you still need with
run.log_model, or reference files that stay in your own bucket.
Questions about moving from MLflow
Do we need MLflow installed to import?
No. euvyra import mlflow reads the files in the mlruns/ folder directly.
Can EUVYRA replace the MLflow model registry?
It covers the same job in its own way: model versions carry their model card, artifact digest and the run they came from, and releases add approvals, evaluation gates and rollback on top.
Does EUVYRA run on Databricks?
EUVYRA is self-hosted with Docker Compose or Helm on your own servers or any Kubernetes cluster. It does not depend on a data platform.