Image by author
Following the popularity of our “7 End-to-End MLOps Platforms to Try in 2024” blog, we are creating another list of open source end-to-end MLOP tools.
Open source tools protect privacy and give you more control over your data and models, but they also require you to manage and deploy these tools yourself, hire additional staff for maintenance, and are responsible for security and outages.
In short, both paid MLOps platforms and open source tools have their pros and cons, and you need to choose the one that suits you best.
In this blog, you will learn about five end-to-end open source MLOps tools for training, tracking, deploying, and monitoring models in production.
1. Cube Flow
kubeflow/kubeflow makes all your machine learning operations simple, portable, and scalable on Kubernetes. It is a cloud-native framework that allows you to create machine learning pipelines, train and deploy models in production.
Image from Kubeflow
Kubeflow is compatible with cloud services (AWS, GCP, Azure) and self-hosted services. It allows machine learning engineers to integrate any kind of AI framework for training, fine-tuning, scheduling, and deploying models. It also provides a centralized dashboard for monitoring and managing pipelines, editing code using Jupyter Notebooks, tracking experiments, a model registry, and storing artifacts.
2. ML Flow
mlflow/mlflow is commonly used for tracking and logging experiments, but over time it has become an end-to-end MLOps tool that works with all kinds of machine learning models, including LLMs (large-scale language models).
Image from MLflow
MLFlow has six core components:
- TrackingVersion and store parameters, code, metrics and output files, and also features interactive metric and parametric visualization.
- project: Package data science source code for reusability and reproducibility.
- modelIt stores machine learning models and metadata in a standard format for later use by downstream tools, and also provides options for serving and deploying the models.
- Model Registry: A centralized model store for managing the life cycle of MLflow models, providing features such as versioning, model lineage, model aliases, model tagging, and annotations.
- Recipes (pipelines): A machine learning pipeline that can quickly train high-quality models and deploy them in production.
- LLM (Master of Laws): Supports evaluation, rapid engineering, tracking, and deployment of LLMs.
Manage your entire machine learning ecosystem with CLI, Python, R, Java, and REST APIs.
3. Metaflow
Netflix/metaflow enables data scientists and machine learning engineers to rapidly build and manage machine learning/AI projects.
Metaflow was originally developed at Netflix to improve data scientists' productivity, and is now open source so anyone can benefit from it.
Image from Metaflow Docs
Metaflow provides unified APIs for data management, versioning, orchestration, training and deployment modes, and computing. It is compatible with major cloud providers and machine learning frameworks.
4. Seldon Core V2
SeldonIO/seldon-core is another popular end-to-end MLOps tool that lets you package, train, deploy, and monitor thousands of machine learning models in production.
Image from seldon-core
Key features of Seldon Core:
- Deploy the model locally using Docker or to a Kubernetes cluster.
- Track model and system metrics.
- Deploy drift and outlier detectors in parallel with the model.
- It supports most machine learning frameworks, including TensorFlow, PyTorch, Scikit-Learn, and ONNX.
- A data-centric MLOP approach.
- The CLI is used to manage workflows, inference, and debugging.
- Save costs by transparently deploying multiple models.
Seldon Core transforms machine learning models into REST/gRPC microservices that allow you to easily scale and manage thousands of machine learning models, while also providing additional features such as metric tracking, request logging, explanations, outlier detection, A/B testing, canaries, and more.
5. ML Run
The mlrun/mlrun framework makes it easy to build and manage machine learning applications in production by streamlining production data ingestion, machine learning pipelines, and online applications, dramatically reducing engineering effort, time to production, and computational resources.
Image from MLRun
Core components of MLRun:
- project management: A centralized hub for managing various project assets such as data, functions, jobs, workflows, and secrets.
- Data and deliverables: Connect disparate data sources, manage metadata, catalogue artifacts, and version them.
- Feature Store: Store, prepare, catalog, and serve model features for training and deployment.
- Batch Execution and Workflow: Execute one or more functions and collect, track, and compare all of their results and deliverables.
- Real-time Delivery Pipeline: Rapid deployment of scalable data and machine learning pipelines.
- Real-time monitoring: Monitor data, models, resources, and production components.
Conclusion
Instead of using one tool for each step in your MLOps pipeline, you can use just one tool to do it all. Train, track, store, version, deploy, and monitor your machine learning models with just one end-to-end MLOPs tool. Simply deploy locally with Docker or to the cloud.
Open source tools provide more control and privacy, but they also come with challenges like managing the tools, updating them, dealing with security issues and downtime, etc. If you are starting out as an MLOps engineer, we recommend focusing on open source tools and then moving to managed services like Databricks, AWS, Iguazio, etc.
I hope you liked my content on MLOps, if you want to read more, mention it in the comments or connect with me on LinkedIn.
Abid Ali Awan (@1abidaliawan) is a Certified Data Scientist professional who loves building machine learning models. Currently, he focuses on content creation and writing technical blogs on Machine Learning and Data Science techniques. Abid holds a Masters in Technology Management and a Bachelors in Communication Engineering. His vision is to build AI products using Graph Neural Networks for students suffering from mental illness.
