Arango logo

Arango MLflow Service

The Arango MLflow Service integrates the MLflow platform for managing the full machine learning lifecycle into the Arango Contextual Data Platform

Overview

The ArangoDB MLflow service is a service that hosts the official MLflow application in your Kubernetes cluster and connects automatically to the ArangoDB environment, e.g. for registering the LLM to be self-hosted and used by services requiring LLMs (such as the Importer and AutoRAG).

MLflow is an open-source platform, purpose-built to assist machine learning practitioners and teams in handling the complexities of the machine learning process. It focuses on the full lifecycle for machine learning projects, ensuring that each phase is manageable, traceable, and reproducible.

The main purpose of the ArangoDB’s MLflow integration is to provide a seamless experience to manage your machine learning models and experiments within the ArangoDB environment. For example, any spawned LLM host service is automatically linked to the MLflow service and is able to fetch any registered model from the MLflow model registry.

ArangoDB’s MLflow integration is only required when working with private LLMs. For more information, see the Triton LLM host documentation.
You can find detailed instructions about how to organize the format of a model for a dedicated LLM host service in the official MLflow  documentation.

Core components

MLflow consists of the following core components:

  • Model Registry: A centralized model store, set of APIs, and UI to collaboratively manage the full lifecycle of an MLflow Model, including model lineage, versioning, stage transitions, and annotations.
  • Experiment Tracking: Provides an API and UI for logging parameters, code versions, metrics, and artifacts during the ML process, allowing for comparison of multiple runs across different users.
  • Model Packaging: Offers a standard format for packaging models from any framework.
  • Serving: Facilitates the deployment of models to various platforms. Within the ArangoDB environment, this enables the integration with services that utilize self-hosted LLMs.
  • Evaluation: Provides tools for in-depth model analysis, facilitating objective model comparison.
  • Observability: Ensures that the ML lifecycle is traceable and reproducible through various metrics and logs.

Quickstart

The ArangoDB MLflow service is started by default and is automatically spawned. You can interact with it in two ways:

  • Web interface: Through the Arango Contextual Data Platform web interface
  • Programmatically: Using the official MLflow client

Web interface

The MLflow web interface is only available through the Arango Contextual Data Platform web interface:

  1. Sign in to the Arango Contextual Data Platform web interface with your platform credentials. MLflow does not have a separate login.
  2. Expand AI Tools in the main navigation and click MLFlow.
The MLflow route requires authentication. Opening https://<EXTERNAL_ENDPOINT>:8529/mlflow/ directly in your browser is no longer possible. Use the AI Tools section of the Arango Contextual Data Platform web interface instead.

Programmatic access

For the official MLflow client, as well as for scripts and other services that call MLflow over HTTP, use the following endpoint:

https://<EXTERNAL_ENDPOINT>:8529/mlflow/

Every request to this endpoint needs to include a valid JWT, either as an Authorization: Bearer <token> header or via the MLFLOW_TRACKING_TOKEN environment variable. Requests without a valid token are rejected with an HTTP 401 or 403 status.

Obtaining a Bearer Token

Before you can authenticate with the MLflow service, you need to obtain a Bearer token. You can generate this token using the ArangoDB authentication API:

curl -X POST https://<EXTERNAL_ENDPOINT>:8529/_open/auth \
  -d '{"username": "your-username", "password": "your-password"}'

This returns a JWT token that you can use as your Bearer token. For more details about ArangoDB authentication and JWT tokens, see the ArangoDB Authentication documentation.

Installation

First, install the MLflow client:

pip install mlflow==3.10.1
The ArangoDB MLflow service runs MLflow 3.x. When registering Triton model bundles, each bundle must include an MLmodel file at its root; see Triton LLM Host for details.

There are two approaches for programmatic access to your ArangoDB MLflow service:

  • Configuration in Python
  • Using environment variables

Configuration in Python 

import mlflow
import os

# Set authentication and tracking URI
os.environ['MLFLOW_TRACKING_TOKEN'] = 'your-bearer-token-here'
mlflow.set_tracking_uri("https://<EXTERNAL_ENDPOINT>:8529/mlflow/")

# Start logging your experiments
with mlflow.start_run():
    mlflow.log_artifact("local_file.txt")

Using environment variables

Set the environment variables in your shell:

export MLFLOW_TRACKING_URI="https://<EXTERNAL_ENDPOINT>:8529/mlflow/"
export MLFLOW_TRACKING_TOKEN="your-bearer-token-here"

Then use MLflow normally in your Python code:

import mlflow

# MLflow automatically uses the environment variables
with mlflow.start_run():
    mlflow.log_artifact("local_file.txt")

Health check

To test whether the service is running, you can use the following snippet:

curl -H "Authorization: Bearer your-bearer-token-here" https://<EXTERNAL_ENDPOINT>:8529/mlflow/health

Expected output on success: HTTP 200 status with response body OK.

API Reference

For detailed API documentation, refer to the official MLflow REST API Reference .

On this page