3 tips for starting your MLOps journey

Avoid common obstacles by applying these 3 tips to your MLOps strategy.

3 tips for starting your MLOps journey

As firms begin to think about incorporating MLOps practices into their analytics initiative, they must be aware of one of the biggest obstacles: change management. Having a plan for how to implement and manage the changes is paramount to the success of an MLOps implementation.

Organizations will find that there are gaps in some combination of their existing processes, role palettes, and technology stack. Depending on the maturity level in each of these areas, some firms will experience change management from more manageable tweaks here and there to larger overhauls for how things are done. Below are some tips for managing change as you begin on your MLOps journey.

3 tips for starting:

  1.  Choose 1-3 focus areas to kickstart your MLOps implementation
  2. Leverage the right mix of skillsets
  3. Provide adequate resources

1. Choose 1-3 focus areas to kickstart your MLOps implementation

software engineer focusing on computer screen

If your organization is just beginning its MLOps journey, remember one thing: don’t try to do everything at once. Trying to implement too many changes at once often results in an unorganized rollout with little evidence of progress at the end of the day.

When getting started, consider targeting a select few focus areas for implementing MLOps concepts. For example, you may want to focus on model reproducibility. Achieving true reproducibility requires the thoughtful implementation of source control management processes, environment tracking, model portability, as well as model registries. Rather than try to achieve all of these at once, pick one. For example, a good place to start would be to implement source control processes for the data science team.

One common observation is that data scientists tend to keep their work on a laptop (perhaps using something like file naming conventions to keep track of their code versions), rather than version their code and artifacts using a source code management system like GIT. When implementing MLOps into your analytics organization, data scientists need to be encouraged to become more like their software engineering counterparts. This change will need to be planned, tracked, and managed to ensure new processes are adequately supported. A plan which provides training materials and review sessions will help ensure that changes are being adhered to. For more information on how to develop your MLOps maturity, check out this article from GigaOm.

2. Leverage the right mix of skillsets

handshake between data science and machine learning


Many organizations are eager to implement MLOps but do not stop to think if there is a role/skill gap in their existing structure. One common misconception is that data scientists should be able to manage the models throughout the entire model lifecycle. This is not the way things should be approached, as it requires a highly specialized workforce that is difficult to identify and hire.

Instead, data scientists should be left to do the work they do best (i.e., building accurate machine learning models) while leaving the production machinery to another role: ML Engineers. ML Engineers should own the responsibility of taking the models built by a data scientist and building the machinery around them to be production ready. These ML Engineers will handle the packaging, deployment, and monitoring mechanisms for getting the model into a production state and ensuring it continues to be performant.

Working together with data scientists, agreements will be made about how often to retrain a model, defining adequate model performance metrics, as well as when to require further investigations as conditions evolve over time. For additional insights on the key roles to consider for MLOps, please reference this article from Gartner.

3. Provide adequate resources

woman holding icon of an employee


As new processes and tools are rolled out to the business, remember to make adequate training and resources available. Whether incorporating a new tool to handle part of the MLOps model lifecycle, or introducing a new business process, you must give the data scientist and ML engineers the appropriate resources.

We recommend an appointed team of practitioners who can help answer any questions about new processes and how to best adhere to them. On the technology side, many technology vendors can provide training on the various platforms and can provide advisory services on how to use the technology to achieve desired outcomes.

Ready to implement MLOps? Neal Analytics can help you with maturity assessments, roadmap planning, and more with our flexible engagement models. Start your MLOps journey today!

This article was originally published on 7/28/2020.