MLOps Tools Universe
I have been trying to grasp all aspects of MLOps for quite a while. There are “all-in-one” platforms from 3 major clouds:
Then there are a number of companies and ecosystems either offering somewhat all-in-one platforms or piecemeal tools: MLFlow, Metaflow, Dataiku, H2O.ai, Valohai, Weights & Biases, Neptune, Comet ML, ClearML, OctoML, BentoML, ZenML, and many many more.
The number of tools and companies is overwhelming. Probably there are more companies than customers! Here is the relevant part of Matt Turck’s 2021 Machine Learning, AI, and Data (MAD) Landscape:
MLOps — The Dust is Settling
I think the dust is beginning to settle. I seek simplicity, and this landscape is anything but simple. I will default to choosing the all-in-one offering from the cloud providers for my project or company. Seeing the push from the big 3 to enhance and simplify their MLOps offerings, this decision will become more common.
Apart from the big 3, there is space for 2–3 more all-in-one-offerings. Consolidation of these tools and companies is inevitable. There will be a bunch of open-source tools that will become the foundation of all this, much like how it happened in Big Data storage and processing.
Key Piecemeal Functionalities
MLOps itself looks pretty complex, but there are 7 key functionalities:
- Data Versioning
- Feature Store
- Pipeline Orchestration
- Experiment Tracking
- Model Versioning
- Model Serving
- Model Monitoring
If you do not want to go with the big 3, pick the least number of the most popular open-source source offerings that cover the maximum of these 7 functionalities. Look for a coalition of 2–3 tools that work well together out of the box and give you almost all of the 7 functionalities.
May the force be with you!
