What problems do MLOps solve?
MLOps promises to bring manageability to the entire process of machine learning for production use cases through good software engineering practices.
The challenges in MLOps are many:
- Variety of roles collaborate for building an ML use case. E.g., Domain SME, Data Engineer, Data Scientists, Operations team, Security and so on.
- Understanding and then collaborating on a large number of data sources, often through a complex process of preparing a large volume of data sets.
- Keeping track of source data, features, ML experiments, model versions, hyper-parameters, model characteristics
- Taking models to production also has many challenges such as model performance, monitoring, scaling, deploying A/B versions, drift management etc.
- Complexity of the environment e.g., libraries, access control and security are also not trivial.
Below is a schematic view of elements of MLOps.
Why is MLOps Difficult to Implement?
MLOps is difficult to implement with a definitive view because the current ML platforms and tooling options are still evolving at a rapid pace. No one platform or tool is uniformly capable of addressing all the challenges through the whole ML lifecycle and for variety of ML use cases that exists today. ML use case development process & data engineering processes are extremely varied today. The cost and performance needs of the production environment often needs yet another approach to deploy and operate the models at scale. Adopting an appropriate MLOps approach for an organization has to take into consideration various elements of people, processes, tools, cost and business alignment.
