AI and machine learning (ML) are everywhere now, but the tools we have for putting these models into use aren’t keeping up. While AI/ML technology has been quickly adopted and applied, the infrastructure to support it (like MLOps tools) isn’t as advanced. This has made wonder whether we should continue developing MLOps as a separate field or integrate it more closely with the already established DevOps practices?
The main challenge in AI/ML right now is moving models from dev to prod. This process, is complicated and (potentially) risky due to the lack of standard procedures. To address this, most of us have begun using Docker for packaging applications for AI/ML models too. This method simplifies deployment and management, making life easier for everyone involved, from developers to operations teams.
Unfortunately, Docker wasn’t designed for ML models, and though it works there are a few areas where the Docker files fall short. One being that Docker files aren’t modular, so you can’t pull them apart. Another being that you can’t see the model’s history from a Docker file. And, just to be clear, nothing wrong with Docker, it just wasn’t built for packaging up something like a Jupyter Notebook.
To solve this, we built an open-source MLOps project named Kit. Kit allows data scientists and developers to package their models, data, code, and settings in a way that’s easy to manage and deploy. This standardization aims to make AI/ML models as manageable as application code, reducing the friction in deploying these models to production environments.
Kit uses the industry-standard container format (OCI, just like Docker) to package models, ensuring that they can be easily shared, managed, and deployed across different environments without needing to understand complex containerization technologies like Docker or Kubernetes. This approach not only makes the model handoff process smoother but also enhances collaboration across teams by allowing them to use the tools they prefer (Eg, data scientist can use a Jupyter Notebook and everyone else can use what they prefer.)
Moreover, Kit emphasizes model traceability and reproducibility, offering a comprehensive package that includes everything needed to reproduce or test a model. This approach is more practical than traditional methods, which often overlook the inclusion of essential elements like hyperparameters and datasets. Kit also supports secure model deployment and compliance, addressing the growing concerns about security in AI/ML model deployment.
You can learn more about Kit here: https://kitops.ml, and support us by [giving Kit a star on GitHub]!(https://github.com/jozu-ai/kitops)
