MLflow On-Premise Deployment using Docker Compose

Easily deploy an MLflow tracking server with 1 command.

MinIO S3 is used as the artifact store and MySQL server is used as the backend store.

How to run

1. Clone (download) this repository

   git clone https://github.com/sachua/mlflow-docker-compose.git

2. cd into the mlflow-docker-compose directory

3. Build and run the containers with docker-compose

   docker-compose up -d --build

4. Access MLflow UI with http://localhost:5000

5. Access MinIO UI with http://localhost:9000

Containerization

The MLflow tracking server is composed of 4 docker containers:

• MLflow server
• MinIO object storage server
• MySQL database server

Example

1. Install conda

2. Install MLflow with extra dependencies, including scikit-learn

   pip install mlflow[extras]

3. Set environmental variables

   export MLFLOW_TRACKING_URI=http://localhost:5000
   export MLFLOW_S3_ENDPOINT_URL=http://localhost:9000

4. Set MinIO credentials

   cat <<EOF > ~/.aws/credentials
   [default]
   aws_access_key_id=minio
   aws_secret_access_key=minio123
   EOF

5. Train a sample MLflow model

   mlflow run https://github.com/mlflow/mlflow-example.git -P alpha=0.42

   • Note: To fix ModuleNotFoundError: No module named 'boto3'

     #Switch to the conda env
     conda env list
     conda activate mlflow-3eee9bd7a0713cf80a17bc0a4d659bc9c549efac #replace with your own generated mlflow-environment
     pip install boto3

6. Serve the model (replace with your model's actual path)

   mlflow models serve -m S3://mlflow/0/98bdf6ec158145908af39f86156c347f/artifacts/model -p 1234

7. You can check the input with this command

   curl -X POST -H "Content-Type:application/json; format=pandas-split" --data '{"columns":["alcohol", "chlorides", "citric acid", "density", "fixed acidity", "free sulfur dioxide", "pH", "residual sugar", "sulphates", "total sulfur dioxide", "volatile acidity"],"data":[[12.8, 0.029, 0.48, 0.98, 6.2, 29, 3.33, 1.2, 0.39, 75, 0.66]]}' http://127.0.0.1:1234/invocations

Personal Note

This note is based on changes to the .env and docker-compose.yml files. The changes to the Minio Access Key must first be made in the Minio Console.

MLFlow

Jupyter Notebook Usage

1. Make Minio Access Keys on Minio, then save access key id and secret access key.

2. Import environment on notebook

%env MLFLOW_TRACKING_URI=http://localhost:5000
%env MLFLOW_S3_ENDPOINT_URL=http://localhost:9000
%env AWS_ACCESS_KEY_ID=2vSrPs21nZYaUQvovgRL
%env AWS_SECRET_ACCESS_KEY=yCqF29KU1qbykEnsceWMDRNvPelgGAVBmyD6PeU5

3. Check it again and setup experiment name

import os
import mlflow

assert "MLFLOW_TRACKING_URI" in os.environ
assert "MLFLOW_S3_ENDPOINT_URL" in os.environ
assert "AWS_ACCESS_KEY_ID" in os.environ
assert "AWS_SECRET_ACCESS_KEY" in os.environ

# you can also use this method for set tracking uri, instead using environment
mlflow.set_tracking_uri("http://localhost:5000/")
mlflow.set_experiment("nyc-taxi")

4. Use with statement in trainer code

with mlflow.start_run():
    # your trainer code

Python Code Usage

1. Make Minio Access Keys on Minio, then save access key id and secret access key.

2. Put environment in .env file

MLFLOW_TRACKING_URI=http://localhost:5000
MLFLOW_S3_ENDPOINT_URL=http://localhost:9000
AWS_ACCESS_KEY_ID=2vSrPs21nZYaUQvovgRL
AWS_SECRET_ACCESS_KEY=yCqF29KU1qbykEnsceWMDRNvPelgGAVBmyD6PeU5

3. Import .env file in python code and setup experiment name

from dotenv import load_dotenv

load_dotenv()

# you can also use this method for set tracking uri, instead using environment
mlflow.set_tracking_uri("http://localhost:5000/")
mlflow.set_experiment("nyc-taxi")

4. Use with statement in trainer code

with mlflow.start_run():
    # your trainer code

Changelog

• 2023-12-01 Add Adminer and Grafana service; Remove Prefect agent service (outdated)
• 2023-11-28 Migrate to Postgres DB and add Prefetch server and agent service