Let’s talk about MLOps. But, to do that, we need to know what it is.

MLOps stands for Machine Learning Operations. Some experts say it is a subset of ModelOps.

ModelOps is a practice of collaboration and communication between data scientists and operations professionals to help manage the production of machine learning (ML) lifecycles.

But let’s not get too technical.

The diagram below shows a great depiction MLOps:

Why is this diagram important?

Scaling has become a difficult process for data scientists, especially when it comes to transferring ML models from a development environment to a production environment.

Additionally, stakeholders in an organization are usually siloed across different teams with varying responsibilities. This causes an extended delivery time frame to implement a ML model to production.

MLOps was developed to optimize the ML lifecycle. The communication and implementation process between those stakeholders is streamlined and made more effective with MLOps.

With so many stakeholders in the ML lifecycle, it can be hard to keep track of who does what. Keep in mind, these responsibilities can vary at different organizations, but it is still the same across the entire board.

ML Lifecycle Stakeholders

Subject Matter Expert (SME)

• Asks essential business questions
• Ensures model performance meet business needs/goals

Data Analyst

• Handles data analysis and exploratory data analysis
• Assists with developing features for ML model consumption
• Optimizes and builds data extractions for use in ML processes (ETL)

Data Engineer

Data Scientist

• Develops models to answer business questions brought up by SME’s
• Responsible for testing models and delivering them into production to produce business value
• Reviews model results, accuracy, and retrains models

Software Engineer

• Develops API’s or applications that work with ML models
• Verifies ML models work correctly with other software platforms

Machine Learning Architect

• Enables scaling for ML models in production
• Improves and optimizes the architecture for ML models in production

DevOps Engineer

• Manages security and performance of architecture that supports ML models
• Handles Continuous Integration/Continuous Deployment (CI/CD) pipelines for ML models in all environments

As mentioned before, the stakeholders are usually segregated across different groups and teams. But, with MLOps in practice, the stakeholders can assess and mitigate risks for the organization as a collaborative group. Especially because there are different risk levels for implementing and managing ML models.

ML Lifecycle Break Down

How do these stakeholders work collaboratively to implement a healthy and cohesive ML lifecycle with checks and balances in place? Let’s break down the ML lifecycle.

• Business Questions
• Who does this concern?
• Subject Matter Experts (SMEs)
• Data Transformation
• Cycle of Data Acquisition --> Data Transformation
• Who does this concern?
• Data Analysts
• Model Development
• Cycle of Feature Engineering --> Model Training/Experimentation --> Model Evaluation and Comparison --> Data Transformation --> back to Feature Engineering
• Who does this concern?
• Data Scientists
• Data Engineers
• Subject Matter Experts
• Model Packaging
• Runtime Environment
• Risk Evaluation/QA
• Who does this concern?
• Software Engineers
• ML Architects
• Development to Production
• Scaling
• Continuous Integration/Continuous Deployment (CI/CD)
• Who does this concern?
• DevOps
• Monitoring
• Logging
• Alerts
• Performance Drift
• Who does this concern?
• Data Scientists
• DevOps

Conclusion

There are a lot of moving pieces in the ML lifecycle. From asking the right business questions to data transformation, model development, monitoring, development to production, and packaging the models together.

Are you ready to begin the journey to ML? Are you going to undertake this extensive process alone? Or would you like to partner with an experienced firm?

Contact us today for a custom demo!