Awesome

Open Source MLOps

This is the Fuzzy Labs guide to the universe of free and open source MLOps tools.

Contents

• What is MLOps, anyway?
• What counts as open source?
• Data annotation
• Data validation
• Data version control
• Experiment tracking
• Model training
• Model validation
• Model registries
• Feature stores
• Feature engineering
• Model deployment and serving
• Model monitoring
• Full stacks
• More resources

What is MLOps anyway?

MLOps (machine learning operations) is a discipline that helps people to train, deploy and run machine learning models successfully in production environments. Because this is a new and rapidly-evolving field, there are a lot of tools out there, and new ones appear all the time. If we've missed any, then please do raise a pull request! (but please ensure it meets the definitions of open source below).

What counts as open source?

In deciding what to include in this list we apply 3 criteria.

👮‍ Fits the definition

Rather than invent our own definition, we defer to the Open Source Initiative's definition of open source. The OSI have laid out a set of clear and unambiguous requirements for something to be considered free and open source.

⚖ Open source license

So it follows that any OSI-approved license is acceptable, with examples including the various Apache, GPL, and BSD licenses. But this also means there are licenses that we don't include, such as the increasingly-popular Server Side Public License; to understand why, see the OSI's writeup on the subject.

🔋 Batteries included

We only list tools that are fully-functional and able to stand on their own. Some vendors offer limited open source versions of their tools, but ultimately those vendors want to guide you towards a SaaS offering. Sometimes there's a genuinely open-source client, but a proprietary server, meaning that the tool is not truly stand-alone. Everything in this collection comes complete with "batteries included".

Data annotation

Data annotation (or labelling) tools help us create labelled training data for use in supervised learning. This is one of the most labour-intensive processes in machine learning, and one where good tooling can save a great deal of effort. Because data comes in a huge variety of types and formats, there are also a huge variety of different tools which specialise in certain types of data.

For a more in-depth list of data annotation tools, check out Open Source Data Annotation & Labeling Tools by ZenML.

Data validation

The golden rule of machine learning is that a model is only as good as the data used to train it. Data validation is the process of verifying that data is accurate and consistent. As errors in data can severely impact the performance of your model, it is crucial to spot data errors early.

Sometimes your model might achieve good accuracy during training , but performs poorly on real-world data. This could be an indication that there are errors in your data which were not spotted during validation. Even a tiny mismatch in the format of the data provided to your machine learning model can significantly impact its performance.

Data version control

Just like code, data grows and evolves over time. Data versioning tools help you to keep track of these changes.

You might wonder why you can't just store data in Git (or equivalent). There are a few reasons this doesn't work, but the main one is size: Git is designed for small text files, and typical datasets used in machine learning are just too big. Some tools, like DVC, store the data externally, but also integrate with Git so that data versions can be linked to code versions.

Experiment tracking

Machine learning involves a lot of experimentation. We end up training a lot of models, most of which are never intended to go into production, but represent progressive steps towards having something production-worthy. Experiment tracking tools are there to help us keep track of each experiment. What exactly do we need to track? typically this includes the code version, data version, input parameters, training performance metrics, as well as the final model assets.

Model training

Model validation

Data validation has been around for some time, but what about validating ML? Although ML models are very dependant on the data that you feed them, models must also be validated before being released into the real world.

Model registries

Model registries are used to track the lifecycle of trained machine learning models.

Workflows

Feature stores

Feature Engineering

Model deployment and serving

Model serving is the process of taking a trained model and presenting it behind a REST API, and this enables other software components to interact with a model. To make deployment of these model servers as simple as possible, it's commonplace to run them inside Docker containers and deploy them to a container orchestration system such as Kubernetes.

Model monitoring

Monitoring means making sure that each deployed model is both functioning, and producing sensible results. We don't just want to check for errors in the traditional sense, but also for things like drift and signs of bias in the predictions and decisions that come from a model.

Data Catalogues

Tools for data discovery and data access control. Can be very important to workout and manage what data can be trained on or tested on, and how these rules vary from client to client. Great summary at https://atlan.com/open-source-data-governance-tools/. This space is pretty young with no real convergence on an expected set of features. They vary in focus from data discovery to data accesss management.

Full stacks

Governance

More resources

Here are some more resources for MLOps, both open-source and proprietary

• Top 10 Open Source MLOps Tools
• Awesome MLOps - a mixture of open source and proprietory tools and platforms.
• Best open source MLOps tools