MLflow is organized into three components: Tracking, Projects, and Models. You can use each of these components on their own—for example, maybe you want to export models in MLflow’s model format without using Tracking or Projects—but they are also designed to work well together.
MLflow’s core philosophy is to put as few constraints as possible on your workflow: it is designed to work with any machine learning library, determine most things about your code by convention, and require minimal changes to integrate into an existing codebase. At the same time, MLflow aims to take any codebase written in its format and make it reproducible and reusable by multiple data scientists. On this page, we describe a typical ML workflow and where MLflow fits in.
Machine learning requires experimenting with a wide range of datasets, data preparation steps, and algorithms to build a model that maximizes some target metric. Once you have built a model, you also need to deploy it to a production system, monitor its performance, and continuously retrain it on new data and compare with alternative models.
Being productive with machine learning can therefore be challenging for several reasons:
It’s difficult to keep track of experiments. When you are just working with files on your laptop, or with an interactive notebook, how do you tell which data, code and parameters went into getting a particular result?
It’s difficult to reproduce code. Even if you have meticulously tracked the code versions and parameters, you need to capture the whole environment (for example, library dependencies) to get the same result again. This is especially challenging if you want another data scientist to use your code, or if you want to run the same code at scale on another platform (for example, in the cloud).
There’s no standard way to package and deploy models. Every data science team comes up with its own approach for each ML library that it uses, and the link between a model and the code and parameters that produced it is often lost.
Moreover, although individual ML libraries provide solutions to some of these problems (for example, model serving), to get the best result you usually want to try multiple ML libraries. MLflow lets you train, reuse, and deploy models with any library and package them into reproducible steps that other data scientists can use as a “black box,” without even having to know which library you are using.
MLflow provides three components to help manage the ML workflow:
MLflow Tracking is an API and UI for logging parameters, code versions, metrics, and artifacts when running your machine learning code and for later visualizing the results. You can use MLflow Tracking in any environment (for example, a standalone script or a notebook) to log results to local files or to a server, then compare multiple runs. Teams can also use it to compare results from different users.
MLflow Projects are a standard format for packaging reusable data science code. Each project
is simply a directory with code or a Git repository, and uses a descriptor file or simply
convention to specify its dependencies and how to run the code. For example, projects can contain
conda.yaml file for specifying a Python Conda environment.
When you use the MLflow Tracking API in a Project, MLflow automatically remembers the project
version (for example, Git commit) and any parameters. You can easily run existing MLflow
Projects from GitHub or your own Git repository, and chain them into multi-step workflows.
MLflow Models offer a convention for packaging machine learning models in multiple flavors, and a variety of tools to help you deploy them. Each Model is saved as a directory containing arbitrary files and a descriptor file that lists several “flavors” the model can be used in. For example, a TensorFlow model can be loaded as a TensorFlow DAG, or as a Python function to apply to input data. MLflow provides tools to deploy many common model types to diverse platforms: for example, any model supporting the “Python function” flavor can be deployed to a Docker-based REST server, to cloud platforms such as Azure ML and AWS SageMaker, and as a user-defined function in Apache Spark for batch and streaming inference. If you output MLflow Models using the Tracking API, MLflow also automatically remembers which Project and run they came from.
When you specify the location of an artifact in MLflow APIs, the syntax depends on whether you are invoking the Tracking, Models, or Projects API. For the Tracking API, you specify the artifact location using a (run ID, relative path) tuple. For the Models and Projects APIs, you specify the artifact location in the following ways:
<scheme>/<scheme-dependent-path>. For example:
Data is the key to obtaining good results in machine learning, so MLflow is designed to scale to large data sets, large output files (for example, models), and large numbers of experiments. Specifically, MLflow supports scaling in three dimensions:
An individual MLflow run can execute on a distributed cluster, for example, using Apache Spark. You can launch runs on the distributed infrastructure of your choice and report results to a Tracking Server to compare them. MLflow includes a built-in API to launch runs on Databricks.
MLflow supports launching multiple runs in parallel with different parameters, for example, for hyperparameter tuning. You can simply use the Projects API to start multiple runs and the Tracking API to track them.
MLflow Projects can take input from, and write output to, distributed storage systems such as AWS S3 and DBFS. MLflow can automatically download such files locally for projects that can only run on local files, or give the project a distributed storage URI if it supports that. This means that you can write projects that build large datasets, such as featurizing a 100 TB file.
There are multiple ways you can use MLflow, whether you are a data scientist working alone or part of a large organization:
Individual Data Scientists can use MLflow Tracking to track experiments locally on their machine, organize code in projects for future reuse, and output models that production engineers can then deploy using MLflow’s deployment tools. MLflow Tracking just reads and writes files to the local file system by default, so there is no need to deploy a server.
Data Science Teams can deploy an MLflow Tracking server to log and compare results across multiple users working on the same problem. By setting up a convention for naming their parameters and metrics, they can try different algorithms to tackle the same problem and then run the same algorithms again on new data to compare models in the future. Moreover, anyone can download and run another model.
Large Organizations can share projects, models, and results using MLflow. Any team can run another team’s code using MLflow Projects, so organizations can package useful training and data preparation steps that other teams can use, or compare results from many teams on the same task. Moreover, engineering teams can easily move workflows from R&D to staging to production.
Production Engineers can deploy models from diverse ML libraries in the same way, store the models as files in a management system of their choice, and track which run a model came from.
Researchers and Open Source Developers can publish code to GitHub in the MLflow Project format,
making it easy for anyone to run their code using the
mlflow run github.com/... command.
ML Library Developers can output models in the MLflow Model format to have them automatically support deployment using MLflow’s built-in tools. In addition, deployment tool developers (for example, a cloud vendor building a serving platform) can automatically support a large variety of models.