[DEPRECATED] MLflow Pipelines Regression Example

NOTE: This repository is deprecated as of 2022/11/07, and will be removed soon. If you are using MLflow 2.0, please refer to MLflow Recipes Regression Example instead.

This repository serves as an example project for the MLflow Regression Pipeline. Follow the instructions below to download the latest MLflow and this repository to create a linear regressor and evaluate its performance, all out of box!

Note This example repo is intended for first-time MLflow Pipelines users to learn its fundamental concepts and workflows. For users already familiar with MLflow Pipelines, seeking a template repository to solve a specific regression ML problem, consider using mlp-regression-template instead.

Note: MLflow Pipelines is an experimental feature in MLflow. If you observe any issues, please report them here. For suggestions on improvements, please file a discussion topic here. Your contribution to MLflow Pipelines is greatly appreciated by the community!

The ML problem: NYC taxi fare prediction

In this example, we demonstrate how to use MLflow Pipelines to predict the fare amount for a taxi ride in New York City, given the pickup and dropoff locations, trip duration and distance etc. The original data was published by the NYC gov. We show how to build and evaluate a very simple linear regressor step by step, following the best practices of machine learning engineering. By the end of this example, you will learn how to use MLflow Pipelines to

• Ingest the raw source data.
• Splits the dataset into training/validation/test.
• Create a feature transformer and transform the dataset.
• Train a linear model (regressor) to predict the taxi fare.
• Evaluate the trained model, and improve it by iterating through the transform and train steps.
• Register the model for production inference.

All of these can be done with Jupyter notebook or on Databricks environment. Finally, challenge yourself to build a better model. Try the following:

• Find a better data source with more training data and more raw feature columns.
• Clean the dataset to make it less noisy.
• Find better feature transformations.
• Fine tune the hyperparameters of the model.

Installation instructions

Follow the MLflow Pipelines installation guide. You may need to install additional libraries for extra features:

• Hyperopt is required for hyperparameter tuning.
• PySpark is required for distributed training or to ingest Spark tables.
• Delta is required to ingest Delta tables. These libraries are available natively in the Databricks Runtime for Machine Learning.

Log to the designated MLflow Experiment

To log pipeline runs to a particular MLflow experiment:

1. Open profiles/databricks.yaml or profiles/local.yaml, depending on your environment.
2. Edit (and uncomment, if necessary) the experiment section, specifying the name of the desired experiment for logging.

Development Environment -- Databricks

Sync this repository with Databricks Repos and run the notebooks/databricks notebook on a Databricks Cluster running version 11.0 or greater of the Databricks Runtime or the Databricks Runtime for Machine Learning with workspace files support enabled.

Note: When making changes to pipelines on Databricks, it is recommended that you edit files on your local machine and use dbx to sync them to Databricks Repos, as demonstrated here

Note: data profiles display in step cards are not visually compatible with dark theme. Please avoid using the dark theme if possible.

Accessing MLflow Pipeline Runs

You can find MLflow Experiments and MLflow Runs created by the pipeline on the Databricks ML Experiments page.

Development Environment -- Local machine

Jupyter

1. Launch the Jupyter Notebook environment via the jupyter notebook command.
2. Open and run the notebooks/jupyter.ipynb notebook in the Jupyter environment.

Note: data profiles display in step cards are not visually compatible with dark theme. Please avoid using the dark theme if possible.

Command-Line Interface (CLI)

First, enter the example root directory and set the profile via environment variable

cd mlr-regression-example

export MLFLOW_PIPELINES_PROFILE=local

Then, try running the following MLflow Pipelines CLI commands to get started. Note that the --step argument is optional. Pipeline commands without a --step specified act on the entire pipeline instead.

Available step names are: ingest, split, transform, train, evaluate and register.

• Display the help message:

mlflow pipelines --help

• Run a pipeline step or the entire pipeline:

mlflow pipelines run --step step_name

• Inspect a step card or the pipeline dependency graph:

mlflow pipelines inspect --step step_name

• Clean a step cache or all step caches:

mlflow pipelines clean --step step_name

Note: a short cut to mlflow pipelines is installed as mlp. For example, to run the ingest step, instead of issuing mlflow pipelines run --step ingest, you may type

mlp -s ingest

Accessing MLflow Pipeline Runs

To view MLflow Experiments and MLflow Runs created by the pipeline:

1. Enter the example root directory: cd mlr-regression-example

2. Start the MLflow UI

mlflow ui \
   --backend-store-uri sqlite:///metadata/mlflow/mlruns.db \
   --default-artifact-root ./metadata/mlflow/mlartifacts \
   --host localhost

3. Open a browser tab pointing to http://127.0.0.1:5000